Adaptive Dynamic Neural Network Intelligent Control of a Nonlinear DC Motor Drive System
Xiaohua WANG
Nonlinear, parameter variations and external load disturbance are widespread in a DC motor drive system. The PID controller has poor self-regulation and adaptivity and robustness. In this paper, A model-following adaptive control approach based on the neural networks is proposed for the speed control of a nonlinear motor drive system with high stick slip and Coulomb friction and the compensation of the nonlinearities. Two neural networks formulate the identification and control approaches. A Wiener-type neural network (WNN) is proposed to identify the motor system, which formulates the Wiener model with a linear dynamic block in cascade with a nonlinear static gain. A feedforward neural network is proposed to formulate the PID controller, termed as PID-type neural network (PIDNN). In this way, the WNN is used as a motor system identifier to recognize the output of motor servo system and provide Jacobian identification information for PID controller in real-time, which realize self-adjusting and self-adaptive control of motor servo system. Based on MATLAB/Simulink simulation, Wiener-type dynamic neural network intelligent control strategy of motor servo system effectiveness is verified. The results obtained confirm the excellent disturbance rejection and tracking performance properties of the system.
Frequency Separation Based Adaptive Feedforward Control for Rejecting Wideband Vibration with Application to Hard Disk Drives
Jinwen PAN, Zhi CHEN, Qing GAO, Yong WANG, and Roberto HOROWITZ
In this paper, a frequency separation based adaptive feedforward control algorithm is developed with the ability to identify the plant and do compensation region by region. In this algorithm, the accelerometer signal is filtered by a series of uniformly distributed bandpass filters to generate a bunch of subband signals which are mutually exclusive in spectrum. In each subband, the corresponding subband signal acts as the feedforward signal and only the frequency response of system in that region needs to be identified, thus a pretty low order model can be expected to have efficient compensation. Starting from the first region, the feedforward control parameters are learned simultaneously with the low order plant model in the same region and then moves to the next region until all the regions are performed.
Mean-Square Stabilization for a Class of Networked Multi-Rate Sampled-Data Systems
Weiwei MA and Yanjun HE
This work investigates the modelling and H∞ control problem for a class of networked control systems with uncertainties and external disturbances, where the system state variables are divided and sampled by multiple groups of sensors with different sampling periods, i.e., multi-rate sampled-data systems. The multi-rate sampled data are transmitted over multiple networked channels with different network constraint parameters, such as different packet dropout rates and network-induced delays. Based on received asynchronous sampled data, a multi-rate state-feedback controller is introduced. To formulate the complex effects of the multi-rate sampling, network-induced delays, and statistic packet dropout, a new integral inequality is developed and utilized to construct a Lyapunov-like functional. Then the mean square stability criterion and control design method are given under a prescribed H∞ performance level. Finally, the validity of the proposal result is testified by a vehicle active suspension system.
UAV Multi-Target Tracking Based on the Principle of Information Concentration
Renshuo CHENG, Hongbin MA, Jingxiang MA, and Ying JIN
At present, unmanned aerial vehicle (UAV) target tracking technology mainly focuses on single target tracking, rather than single UAV tracking multiple targets. Based on the principle of information concentration and numerical calculation, a new trajectory prediction algorithm for UAV tracking is proposed in order to improve the execution efficiency of target tracking. The principle of information concentration which uses the set of tracking objects and numerical calculation method are used to determine the current position information of UAV. Considering the real motion form of UAV, Kalman filter algorithm is used to smooth the tracking position information, which improves the robustness of the algorithm. In addition, since the complexity of the algorithm is not affected by the number of tracking targets, the new multi-target tracking algorithm can be applied to the engineering background with more targets
Research on Semantic Representation Based on Syntactic Dependency Structure and Type
Qiao KANG, Can ZENG, Pengfei HU, Fangyan DONG, and Kewei CHEN
Long short-term memory (LSTM) has become the state-of-the-art for modeling semantic similarity computation. However, it cannot fully encode hierarchical syntactic dependencies and cannot identify differences in meaning caused by changes in the semantic roles of words or phrases. Therefore, a new relative syntactic distance dependency type long short-term memory (RSD-DT-LSTM) neural model is proposed in this paper. This model modifies the gating mechanism in Dependency type long short-term memory (DT-LSTM) with relative syntactic distance and provides a transformed gate to more directly model syntactic structure. Make sentence structure dependencies and dependency types embed sentence meanings into dense vectors. The RSD-DT-LSTM model was used to perform semantic similarity scoring experiments on the SICK dataset and the MRPC dataset. The Pearson correlation coefficient obtained by the model on the SICK dataset is very close to the DT-LSTM model. The mean square error (MSE) between the actual similarity score y and the predicted score y ^ ' is 0.1% better than the DT-LSTM model. In the MRPC dataset, its accuracy is 0.9% better than that of DT-LSTM. And its recall is 0.3% better than that of DT-LSTM model.
Classifying the Machine Components: A Case of Deep Feature Fusion Classification Model
Amina BATOOL, Yaping DAI, Hongbin MA, and Sijie YIN
Automatic components classification is inevitable for industrial and logistics applications in the contemporary digital world. How to classify different machine components (such as bolt, nut, locatingpin and washers) automatically by computer is a challenging task for image-based object recognition and classification. Despite varying shapes and classes, components are difficult to distinguish when they appear identical in many ways – particularly in images. For solving the task, a case of Deep Feature Fusion Classification Model (DFFCM) is proposed in this article, which is designed through a convolutional neural network combined with Adaboost classifier. By DFFCM, the extracted multi-layered features (containing precise information emanated from image edges) are fused through the direct fusion method to differentiate similar images and used as input to the ensemble Adaboost classifier for classification. Resultantly, DFFCM has achieved the highest accuracy with 99.34% compared with CNN and One-Shot Learning models.
A Method for Improving the Diversity of Movie Recommendation with Knowledge Graph
Wen-Kai MA, Yuan WEN, and Xia HOU
This paper proposes a diversity enhancement method with knowledge graph for improving the diversity of recommendation lists in movie recommendation systems. Firstly, this method uses the IMDb movie dataset to build a knowledge graph of movie-character and movie-company relationships, and expands the fine-grained connectivity between movie entities through 20 relationships. Secondly, this paper proposes an aggregation algorithm to improve the diversity feature expression of movies in the knowledge graph module, and the outputs are embedded into the recommendation module as auxiliary information. Finally, the multi-layer perceptron is used to mine the deep expression of user and movie features. Considering the important impact of movie category on the diversity of recommendation list, this paper improved the calculation method of diversity. The experimental results show that on the MovieLens and Netflix dataset, the proposed method greatly improves the diversity while ensuring the accuracy.
Identification and Classification of Wheat Quality Grading Based on Deep Learning
Zhaohui ZHANG, Zengyang ZUO, Zhi LI, Tianyao ZHANG, and Xiaoyan ZHAO
China is one of the world’s largest producers of wheat. Because of its storability, wheat has become one of the country’s three largest grain reserves. However, the detection methods of wheat quality are mainly manual detection. There are many problems with this identification and classification method. Initially, different identification personnel will produce different identification results. Secondly, the identification process is long, the identification personnel are prone to visual fatigue. Finally, the identification takes a lot of time and money, and the classification results are difficult to reproduce. Based on the above background, this paper designs a wheat unsound kernel identification method based on YOLO algorithm. Image acquisition is carried out with sound kernel, injured kernel, broken kernel, sprouted kernel, spotted kernel, and foreign material in wheat kernels. And image data enhancement methods such as image cropping, rotation, scaling, and segmentation are used to determine the number of data sets expanded. The deep learning recognition model is used to identify wheat quality categories. The wheat quality detection experiment of YOLOv5 recognition model is realized, and the performance indicators of the model are evaluated. The recognition results show that the overall recognition rate of YOLOv5 for wheat is about 93%. And the detection can be completed quickly. The average accuracy of YOLOv5 is evaluated. The experimental results show that YOLOv5 is more suitable for the recognition of unsound wheat kernels.
Detection of Public Chair Uncivilized Lying Behavior Based on YOLOv5
Liu-Ying ZHOU, Dong WEI, Yi-Bing RAN, Chen-Xi LIU, Si-Yue FU, and Zhi-Yi REN
To monitor and manage the uncivilized lying behavior of the urban public chair, this paper studies the urban public chair lying behavior image recognition algorithm based on YOLOv5. At present, the object studied in the field of object detection is a single object, but for the uncivilized lying behavior of the urban public chair, the object to be measured is composed of the chair and the human posture combined, so its characteristics will be more complex. To accurately identify the uncivilized lying behavior of urban public chairs, this paper designs various dataset classification schemes according to the feature differences among various combined objects, and tests them using the YOLOv5x object detection algorithm. The experimental results show that the best performance index mAP is 97.38%, and the overall average recall rate (Recall) is 95.51%, which can effectively realize the detection and identification of uncivilized lying public chair behavior.
Human Motion Capture Based on Sparse Inertial Sensor
Huailiang XIA, Xiaoyan ZHAO, Yan CHEN, and Zhaohui ZHANG
In recent years, human motion capture has a broad market prospect, widely used in film and television animation production, human-computer interaction, virtual reality, sports training, medical rehabilitation and other interdisciplinary disciplines. Unlike the regular motion of mechanical objects, the movement of the human body is complex, changeable and unpredictable, and has a flexibility and variety that no mechanical object has. Therefore, at present, the capture, recognition and reproduction of human posture have become a research hotspot. Body motion capture usually requires placing sensor markers above and below the joint to calculate joint angles, but this will lead to a large number of body sensors, affecting the performer’s normal movements. Using the method of deep learning can reduce the number of sensors, but it requires a large number of training datasets and sufficient computing resources. In this paper, we propose a new deep learning-based approach, which learns data-driven time priors, merges past and future inertial sensor information into superimposed input vectors, which are then fed back to shallow neural networks to estimate human movements. Six inertial measurement unit (IMU) sensors were used to estimate human motion.
Copyright Protection Method of Big Data Based on Color Image Watermark
Yabin XU, Minling ZHU, and Bin SHI
Data watermarking technology is an effective means of copyright protection of big data. In order to embed more real, reliable and practical watermarking information, the method of embedding color image watermarking information to protect big data intellectual property is presented. Firstly, based on JPEG image coding standard algorithm, the color image is compressed without affecting the image quality, and the embedded watermark information is reduced. Then, the watermark is embedded in two dimensions: data tuple and attribute. Finally, under the constraint of data usability, the lowest bit of watermark embedding is calculated and the data is changed to complete the watermark bit embedding. Experimental results show that the proposed method can extract color image watermarking information under different attack conditions. The robustness of watermark is high, and the comprehensive effect of data watermark is better than that of existing methods.
A Dangerous Behavior Identification System for Public Monitoring
Renkai HOU, Yaping DAI, and Shuai SHAO
At the present stage, the identification of dangerous behaviors in public places mostly relies on manual work, which is subjective and has low identification efficiency. This thesis proposes an automatic identification method for dangerous behaviors in public places, which analyzes human behavior and speech emotion through deep learning network and then performed multimodal information fusion. Based on the fusion results, people can judge the emotional atmosphere of the crowd and make early warning and alarm for possible dangerous behaviors. Experiments show that the algorithm adopted in this paper can accurately identify dangerous behaviors and has great application value.
Object Recognition and Search Based on Machine Vision
Junjie CHEN and Yuan LI
The loss of personal belongings often occurs in daily life. This paper studies the object recognition algorithm and its implementation on the mobile terminal. Two offline-based mobile terminal object recognition apps are developed, which can realize real-time object recognition without network, and has good privacy security. In this paper, the object recognition function is realized based on SIFT and YOLOX respectively. The transplantation of these two methods on the mobile terminal is realized and two apps are developed based on Android system. The developed android app was tested through some pictures collected in the surrounding living environment or downloaded from the network. The test results realized the function of real-time object detection, and achieved a prediction accuracy of more than 90% based on the deep learning method.
Adaptive RBF Neural Network Sliding Model Control for Path Tracking of DP Vessel
Yixi YANG, Bin XIN, Yang LI, and Jian WEI
With external environmental disturbances and model uncertainties, a path tracking control method for Dynamic Positioning (DP) vessel is proposed. It addressed the problem of the accuracy of path tracking and improved the convergence rate of errors. First, a 3-degree of freedom horizontal motion model is established. Then, considering input saturation, an adaptive radial basis function neural network (RBFNN) sliding mode controller is designed to realize the tracking of desired path, and detailed stability proof illustrates that all errors are bounded. At last, a series of simulations verify the accuracy and validity of the proposed control strategy.
Sampling Transformation Based Path Planning for Messenger UAV in Urban Environment Messenger
Benxiang WANG, Bin XIN, Yulong DING, and Yang LI
In recent years, unmanned platforms are considered to be an important part of the Internet of Things system. However, in the face of communication-limited special tasks such as large-scale logistics and multi-obstacle urban warfare, information isolation will greatly affect the efficiency of the system. In this case, indirect communication between unmanned platforms can be realize by the mobile relay of messenger UAV. This paper mainly introduces a path planning scheme of messenger UAV in urban environment, aiming at minimizing the cost time to access all mobile UGVs (unmanned ground vehicles). The paper models this problem as a special dynamic generalized traveling salesman problem (DGTSP), and proposes a sampling transformation algorithm. The algorithm discretizes the paths of UGVs through a sampling mechanism, and effectively transforms the problem into asymmetric traveling salesman problem (ATSP) by using motion relationship. Finally, the effectiveness of the algorithm is verified by computational experiments with instances of different sizes.
Efficiently Solving Multi-Stage Dynamic Weapon Target Assignment Problem by a Knowledge-Based Heuristic
Weijie MA, Bin XIN, Yujue WANG, Hanzhong ZHANG, and Yang LI
Based on the dynamic weapon target assignment problem (DWTA), this paper is devoted to solving the real-time and dynamically changing requirements of air defense operations. Taking maximizing the elimination of the enemy’s threat as the objective function, a new DWTA model is established, which makes multi-stage decision making for the enemy’s target according to the battlefield situation at different times and simulates the entire dynamic combat process. The decision-making process regarding different enemy targets is processed in parallel, the survival state of defender combat resources is considered, and the DWTA decision is making from the global battlefield situation at different times. A knowledge-based heuristic algorithm is proposed to solve the problem. In order to test the performance of the proposed heuristic algorithm, this paper selects the KM (Kuhn-Munkres) algorithm for finding the optimal matching of weighted bipartite graphs as the comparison algorithm and selects multiple evaluation indicators to evaluate the performance of the two algorithms. The experimental results show that the algorithm proposed in this paper is better than the KM algorithm in the solution quality of most cases.
Obstacle Avoidance and Path Planning of Intelligent Robots Based on Optimized A* and DWA Algorithms
Zhenjiang HU, Ying JIN, Hongbin MA, and Tao FU
Typically, mobile robots work in an environment that is not fully known, such as wilderness, where they need to move from a specified starting point to a desired location. The traditional A* algorithm is based on Dijkstra algorithm to increase the heuristic function (Euclidean distance between the end point), which greatly reduces the number of grids included while obtaining the optimal path. However, in the search process, the number of traversed grids is still large and the planned path is not smooth. In the view of the above problems, the Euclidean distance in the heuristic function is changed to the diagonal distance and the Manhattan distance, and different weights are given. The Bezier curve is used for smoothing. Finally, the optimized A* algorithm is combined with the DWA algorithm to quickly reach the destination while bypassing the obstacles.
Task Allocation of Multi-Robot Cooperative Coverage Based on Circular Coding Genetic Algorithm
Bin XIN, Heng WANG, and Ming LI
Aiming at the problem of multi-robot cooperative area coverage, this paper adopts genetic algorithm based on circular coding to complete the allocation of robot coverage area. In this paper, the given map is divided into rectangular regions and merged into several areas to be allocated. The genetic algorithm with circular coding method is used to complete the coverage task allocation of these areas. Combined with the result of task allocation, this paper designs a navigation point set for each robot. Robots will cover the whole target area by passing through these navigation points in sequence. This paper compares the state-of-the-art algorithm and demonstrates that the algorithm used in this paper has advantages in the balance of task allocation. By running the task allocation algorithm in different situations, it is verified that the task allocation algorithm has situation scalability. This paper also carries out simulation verification on the Gazebo simulation platform which is under the Robot Operating System (ROS).
Cooperative Control of Rotating Inverted Pendulum Based on Fuzzy Control
Yujue WANG, Weining MAO, Bin XIN, Qing WANG, and Jian WEI
The rotating inverted pendulum is a nonlinear, multivariate, strongly coupled unstable system, and its study can effectively reflect many typical problems in control. In this paper, a parameter self-tuning fuzzy controller is proposed to achieve intelligent control of a single rotating inverted pendulum. Particle swarm optimization is used to adjust its structural parameters, and simulation experiments are given to show that the system can achieve stability with the designed parametric self-tuning fuzzy controller, and the control performance is better than that of the conventional fuzzy controller. On the other hand, the leader-follower control strategy is used to realize the cooperative control of multiple rotating inverted pendulums. Two Qube-Servo2 rotating inverted pendulums are used for the cooperative pendulum starting and stabilization experiments, and the effectiveness of the proposed cooperative control strategy is verified.
Decentralized Task Assignment Based on Information Fusion in Communication-Constrained Environments
Zihao CHEN, Juan LI, Jing XIONG, Jie LI, Chang LIU, Jinkun LIN, Ziwei XIN, and Yiliang LI
In communication-constrained environments, the wireless communication quality of unmanned aerial vehicle (UAV) swarms is unstable. This paper proposes an information fusion strategy (IFS) based on information integrity and authenticity for the task assignment problem in communication-constrained environments. The proposed IFS is embedded into the classical sequential and G-Prim decentralized task assignment algorithms, and the two improved variants with the proposed IFS are denoted as Seq-IFS and G-Prim-IFS, respectively. The Gilbert-Elliott (G-E) model, which can model burst errors and bandwidth saturation, is adopted to describe unreliable communication channels. A series of test instances with different swarm sizes and different levels of communication channel reliability are used to compare the performances of Seq-IFS and G-Prim-IFS with their original forms. Numerical experimental results demonstrate that the Seq-IFS and the G-Prim-IFS outperform their original versions significantly in the majority of test instances, especially in cases with low communication channel reliability.
Air Conditioning Load Prediction Based on HRD-TCN Model
Ning HE, Liqiang LIU, Danlei CHU, Cheng QIAN, and Shang LI
Accurate air conditioning load prediction is of great importance for building energy management. This paper presents a hybrid data decomposition method (HRD) combined with temporal convolutional network (TCN) for air conditioning load prediction. First, the original data are decomposed using HRD to extract its linear features and frequency domain features. Then, the extracted significant features are input to TCN for predicting the next day’s air conditioning load. Finally, the real values of the next day’s air conditioning load are added to the training set, which are decomposed and predicted again to form a rolling mechanism. The experimental results show that the proposed method has excellent prediction performance.
Resilient Self-Triggered Model Predictive Control Strategy Against Replay Attack
Ning HE, Shixuan LIN, Kai MA, and Shuoji CHEN
In this paper, contrapose cyber-physical system (CPS), a resilient control strategy based on input sample reconstruction is proposed to solve the security problem of self-triggered model predictive control (STMPC) in the presence of malicious replay attacks. Firstly, a novel replay attack model is established by considering the STMPC non-periodic sampling features. Secondly, by analyzing the characteristics of established replay attack model, an input reconstruction mechanism based on key data is designed, which ensures that even if the optimal control sample is attacked, the controlled system can still generate feasible control samples to preserve the system. In addition, closed-loop stability of the system under replay attacks is strictly demonstrated. Finally, the effectiveness of the developed resilient control mechanism is verified by a robotic system.
Research on Network Public Opinion Situation Awareness
Yabin XU, Minling ZHU, and Ran LIANG
In order to perceive the development of network public opinion effectively and provide technical support for the supervision of network public opinion, the incremental prediction model GG-ICP of information dissemination is constructed by integrating the structural and temporal characteristics of social networks. Firstly, the GCN is combined with the GRU to learn the dynamic network structure information and timing information together. Then, the time decay function is improved to more effectively reflect the value decay of public opinion information with the passage of time. By coupling the spatial and temporal characteristics, the historical and edge information can be captured more reasonably to improve the prediction accuracy of the model, and then make a scientific prediction of the increment of information forwarding. The results of comparison experiments on real datasets show that, the prediction accuracy of the information dissemination model proposed in this paper is significantly improved compared with the existing models.
A Review of Research Methods on PLI Robots
Jingzhe HE, Hao CHEN, and Yifan LU
This paper mainly introduces a common power line inspection (PLI) robot model, and introduces several main solutions to the model, that is, to prove that its motion equation can converge to a given constant.
Synchronization of Positive Networks with Impulsive Control Under False Data Deception Attacks
Kai ZHOU, Yijun ZHANG, and Zhiqiang CHEN
The aim of this paper is to investigate the synchronization of a network, which consists of a set of positive systems. The false data deception attacks are involved in the communication channels which will destroy the data completeness of the considered network. Distinguished from general networks, the condition retaining positivity of the interconnected positive systems is firstly considered. Considering constraints on the positivity of the systems, a distributed impulsive control approach is proposed to solve the synchronization issue of the network. A data detector is designed as well in the sensor. By rigorous mathematical analysis, a sufficient condition for the mean-square bounded synchronization of continuous-time positive systems is provided. Finally, the effectiveness of main results is illustrated by a specific numerical example.
Thickness Determination of Plastic Products Using Continuous Wave Terahertz Frequency-Domain Spectroscopy
Tianyao ZHANG, Boyang LI, Xiaoyan ZHAO, and Zhaohui ZHANG
In this paper, a novel method for thickness determination of plastic products using continuous wave (CW) THz frequency-domain spectroscopy is proposed. This method begins with the refractive index measurement of target material over THz frequencies. The optimal frequency band with high SNR and minimum water vapor affection is then selected for thickness prediction. For samples with unknown thickness, the THz wave’s optical path is initially reconstructed from the phase delay of the recorded data. The calibrated refractive index is then applied to calculate the physical thickness of the sample. The proposed method is exemplified by the widely used plastic material Teflon. The thicknesses predicted by THzFDS measurements are coincided with conventional caliper measurements perfectly.
Label Design and Identification in High Temperature Logistics Based on Concave Coding and DeepLabV3+ Model
Xiaoyan ZHAO, Pengfei ZHAO, Yuguo YIN, Luqi TAO, and Zhaohui ZHANG
With the development of the Internet of Things, the logistics tracking technology at normal temperature has been quite mature, but there are few tracking methods for products in the high-temperature production process. The main difficulties include that the label materials cannot withstand high temperature for a long time, and the detection devices are vulnerable to environmental impact. This paper takes the logistics tracking of carbon anode for aluminum electrolysis as the application background to carry out specific research. This paper designs a high temperature resistant label, that is, printing concave coding on the surfaces of the anode carbon blocks, which ensures the integrity of labels in high temperature environment without affecting product performance. There is no obvious color distinction between the coding area and the background, so the direct recognition is not effective. Therefore, based on DeepLabV3+ model, MobileNetV2 and Xception are selected as the backbone network to semantically segment the concave coding in order to realize the effective segmentation of the coding area and background, and lay the foundation for the subsequent recognition of the concave coding. The experimental results show that the DeepLabV3+ model based on MobileNetV2 is superior, with MIoU of 95.45% and MPA of 97.38%. In summary, the scheme is feasible and provides a new idea for the development of high-temperature logistics tracking technology in smart manufacturing.
A Laser Pointer Trajectory Recognition Method for Human-Computer Interaction
Chong LI, Bin XIN, and Qing WANG
In the mainstream teaching equipment, because the laser pointer is limited by distance and has a relatively single function, it cannot meet the needs of classroom teaching; in order to facilitate teaching, this paper designs a human-computer interaction oriented laser pointer trajectory recognition method, and proposes a hybrid algorithm that combines the traditional light spot recognition algorithm, it can ensure the real-time update of the background image, at the same time accurately identify the tiny moving target; the improved BP neural network algorithm is proposed for trajectory recognition, which gives full play to the nonlinear fitting ability of the neural network. Compared with the previous trajectory recognition methods, the advantages and characteristics of this method are: the characteristics of the light spot and its background can be trained before use, which ensures the adaptability of the whole system; also, it can continuously identify different trajectories, after arranging and combining them, the laser pointer can be used to realize more complex functions, which is convenient for the speaker’s operation. Therefore, it provides efficiency as well as scalability for us. Finally, after experimental verification, the accuracy rate of this method is as high as 95% when using an infrared laser pointer and normal light conditions, which proves the feasibility and robustness of the method.
Analysis of Player Engagement with Eye Tracking in Game-Based Training
Kemeng HAN
Human factor engineering (HFE) concentrates on designs that take human characteristics into account, aiming at providing leisure and reaction, maximizing human capabilities, improving human’s quality of life and so on. In this field, game-based training is an essential method. Combining passive content with interactive games, users can stay engaged with certain level, which is significantly affected by game difficulty level. Hence, it is crucial to establish the relationship between game difficulty level and player engagement level. To this end, in this paper, an example hooking game is developed based on Field Programmable Gate Array (FPGA) and simulated successfully on an online simulation platform, which makes it possible to construct an offline embedded system for the purpose of investigating player engagement. Eye-tracking methodology was adopted as an objective way to track participants’ reaction, which is hard to capture by typical observation method. User performance and eye-tracking heatmaps of the user interface are analyzed, which reveals that game difficulty is a key factor of player engagement level.
Speaker Detection Based on Attention Mechanism
Haozhen SONG, Mintao DONG, Kewei CHEN, and Fangyan DONG
Speaker detection is a necessary pre-processing step for speech recognition and its downstream applications. Unimodal design architectures are currently widely used, but still have limitations in specific environments. Recent research has shown that bimodal fusion learning of audio and visual is a promising solution to the speaker detection task. Thus, in this paper, a bimodal model SdNet is proposed to implement speaker detection by introducing a cross-attention mechanism. Moreover, in order to optimize the different feature sizes caused by different attention levels at different locations of the network, a large number of residual connections are used to map to higher dimensional spaces. After the model is built, the largest AVA-Speech dataset is used for training, the mean average precision (mAP) reaches 80.1%, which is a certain improvement compared with the previously published models, and the experimental results show that the method can improve the accuracy of speaker recognition.
Speaker Tracking Method Based on Audio-Visual Bimodal Fusion and its Application to Mobile Robot
Hao-Yan ZHANG, Qi-Feng SHI, Long-Bo ZHANG, and Zhen-Tao LIU
Initiative service is one of the key research directions of the new generation of service robots. It is important to track human automatically for initiative service in human-robot interaction. To solve the problems of low precision and poor anti-interference capability of only using single-modal (audio or visual information), a speaker localization and tracking method based on audio-visual bimodal fusion is proposed. Firstly, the azimuth of speaker is obtained based on time difference of arrival (TDOA) using microphone array and face detection based on AdaBoost algorithm is carried out using camera. Frame rate tracker and spatial coordinate fusion are used to obtain the location of speaker. Secondly, path planning according to the preset tracking mode is designed, by which the azimuth and distance between the robot and the speaker is maintained in a limited range. To simulate different situations, different angle and distance for speaker tracking were set. Finally, a mobile robot is driven to follow the planned path by the real-time control system using STM32, and information from the microphone array and the camera are collected by Raspberry Pi. The proposed method was tested by setting 16 different sets of azimuth and distance data (the speaker moves slowly in the experimental place). Experimental results demonstrate that the running trajectory of the mobile robot is close to the ideal preset trajectory, which ensures effective tracking of the target speaker.
Adaptive Cruise Control Optimized for Multi-Objective
Salvador N. Obama OYANA and Jun LI
With the great objective of improving the adaptive cruise control (ACC) system, in terms of safety between the vehicles that travel the route, both the ego car and the lead car, energy-saving, and driver comfort, we analyze these terms as restrictions system control for them, we analyze our system from the driving data of a real vehicle. A hierarchical control architecture is adopted, and the top controller is designed based on Model Predictive Control (MPC). The numerical simulation shows that the proposed MPC controller considerably reduces the energy consumption of the ACC system if we compare it with the energy consumption in cruise control (CC) system and the comfort under the premise of guaranteeing the performance and the inter-vehicular distance between the two vehicles is also quite stable.
Traffic Flow Detection Based on Improved YOLOX and DeepSORT
Zewen DU, Ying JIN, Hongbin MA, and Ping LIU
Aiming at the shortcomings of low detection rate, low tracking accuracy and poor robustness in traditional traffic flow detection algorithms in surveillance video, we introduce a traffic flow detection method, which divides the traffic flow detection task into three modules, including vehicle detection, vehicle tracking and vehicle counting. Firstly, the improved YOLOX model is used to detect the vehicles in a single frame image. Secondly, the DeepSORT algorithm is used to track the detected vehicles. Finally, the two-way traffic flow statistics are completed by the improved virtual coil method. In order to make full use of the feature information of different scales, we use the ASFF module to perform adaptive feature fusion on the output feature maps of different scales. In addition, the convolutional attention modules are also used in the detection algorithm to improve the classification and localization performance of the network. Through experimental tests, the traffic flow detection method proposed in this paper has achieved high accuracy and adaptability in a series of traffic surveillance videos under different weather and different perspectives.
A Decision-Making Method of Multimodal Fusion Based on Deep Reinforcement Learning for Highway Autonomous Driving
Ping LIU, Hongbin MA, Ying JIN, Yazhe DING, Bin CAO, and Zewen DU
This paper focuses on deep reinforcement learning-based decision-making and planning methods for autonomous vehicles to drive on dynamic highways in a high-speed, comfortable, and safe manner. Due to the highly dynamic and complex road environment, the ability to make decisions and plan autonomously as humans has always been a challenging topic for autonomous vehicles. Different from the previous only taking dynamic traffic state as input, a multimodal fusion method is designed in this paper to enrich the input feature information, and the longitudinal acceleration (or deceleration) and lateral rotation angle are used for the output of the decision-making model. The total decision reward for the highway autonomous driving problem is designed into three parts, namely, the non-collision reward, the high-speed driving reward, and the successful-to-destination reward. During the experiment, the training and testing parameters of the decision-making model are set according to the driving parameters of the real highway. The experimental results show that the method accelerates the rapid convergence of the decision-making model by fusing the multi-modal data in the high-speed driving environment, and can make reasonable decisions autonomously in the high-speed environment.
A Mixed Encoding Genetic Algorithm for Train Platforming Rescheduling Under Train Delays
Shuxin DING, Tao ZHANG, Rongsheng WANG, Yanhao SUN, Xiaozhao ZHOU, and Chen CHEN
To improve the operational efficiency of the high-speed railway station with emergencies, a train platforming rescheduling problem (TPRP) at a high-speed railway station is proposed in this paper. The TPRP problem addresses the adjustment of train track assignment and train arrival/departure time under train arrival delays. A mixed-integer nonlinear programming model is proposed to minimize the weighted sum of total train delays and rescheduling costs. A mixed encoding genetic algorithm (GA) is proposed. The individual of the GA is represented as the platform track assignment and train departure priority, which is a mixed encoding scheme with integer and permutation. The individual is decoded to a feasible platform track assignment and arrival/departure time of trains using a rule-based method for conflict resolution in platform tracks and arrival/departure route. Experimental results demonstrate the superiority of the mixed encoding GA in solving the proposed problem compared with other algorithms.
A Joint Task Allocation Method for Heterogeneous Multi-Agent Based on Distributed Auction Algorithm
Jia ZHANG, Xin DU, and Bin XIN
The coordination between sensor resources and firepower resources is one of the key factors to improve the operational efficiency of modern war. The paper establishes a mathematical model for the typical heterogeneous multi-agent joint task allocation problem of sensor-weapon-target joint task allocation. Considering the bias of different agents to different tasks, the paper combines the marginal utility theory in economics with the auction algorithm, and proposes a distributed joint auction algorithm that can quickly and effectively generate task schemes. Simulation results show that the method can get 90% performance effect by consuming less than 10% of the calculation time of the centralized algorithm and get a better allocation scheme. It is more suitable for task scenarios with large task scale, high computing time requirements and unstable communication topology.
Mutation-Based Group Teaching Optimization Algorithm for Scheduling Cloud Workflows with Deadline Constraints
Jiwei MENG, Guanghao XU, Bing CHEN, Huifang LI, Lingguo CUI, and Senchun CHAI
Cloud computing shows great potential in dealing with immense amounts of data, which satisfies users’ demand for high performance computing capability. The elasticity of the cloud and the heterogeneity of virtual machines provisioned provide a cost-effective deploying environment for executing user applications. How to manage and schedule these applications in order to guarantee users’ QoS requirements and improve resource utilization, is still a great challenge. In this paper, we design a Group Teaching Optimization Algorithm (GTOA)-based workflow scheduler to minimize the workflow execution cost under deadline constraints. To improve the global search ability and avoid falling into local optima, we introduce a mutation strategy into the original GTOA and propose a Mutation-based GTOA, i.e., M-GTOA. A set of experiments are conducted to verify its effectiveness and experimental results show that M-GTOA outperforms GTOA in success rate and cost optimization. Compared with GA, PSO and OSAM, M-GTOA achieves better optimization performance in most cases, especially for Montage and Sipht workflows.
Identifying Stroke Types Using Machine Learning Method Based on Multi-Channel Frequency Domain Electromagnetic Signals
Shang GAO and Guohun ZHU
Stroke is an emergency and requires treatment within 3-6 hours. Therefore, rapid diagnosis of stroke subtypes is very important for timely clinical intervention. This study assesses four traditional machine learning methods with frequency domain features to classify stroke subtypes based on multichannel electromagnetic signals. Firstly, redundant data are removed from 256 channels scattering signals on each subject. Then 16 frequency domain features are extracted by a principle component analysis (PCA) method. Finally, features are forwarded into four classifiers to be built and used for stroke type prediction. The results indicate that the logistic regression classifier achieves better results than K-nearest neighbour, XGBoost and MLP neural network methods, achieving the highest accuracy (98%) while retaining one feature on each channel.
Monotonic Feature Extraction for Equipment Remaining Life Prediction Using Dynamic Time Warping
Xiaotong WU, Ningyun LU, and Jianhua LYU
This paper proposes a method to screen out monotonic features for equipment remaining life prediction (RUL) using Dynamic Time Warping (DTW). First, Spearman correlation coefficient is used to pre-screen the original features to build a candidate feature set. Second, trend sequences are constructed from the diagonal regions of DTW cost matrix and they have the same height and width as the feature sequences. Then, the similarity index between the candidate features and the trend sequences is calculated, the features with high similarity are screened out since they have strong correlation with the trend sequences, and all screen-out features form a “core” feature set. Finally, these core features are input to a Long Short Term Memory (LSTM) neural network model to predict the RUL. Experimental verification is carried out using a real-world dataset for low-voltage circuit breakers, and the results show that the features screened out by the proposed method are effective for accurate RUL prediction.
A Review of Knowledge Graph Research in Military Domain
Xudong GUO, Linxiu CHEN, Yuan LI, Qinglin WANG, and Weili GUAN
With the development of information technology as well as the continuous improvement of military informatization level, the complex and changing battlefield situation put forward a huge challenge on the military’s massive data processing capability. This paper presents the technical framework related to the construction of knowledge graph in military field, summarizes different aspects of the application of knowledge graph in military field, and reveals the directions for improvement in view of the current situation of the application of knowledge graph in military field
Control Strategy and Optimization of the Indoor Ski Simulator
Hao SHI, Zhen CHEN, Ning DONG, Qing FEI, Xiaolan YAO, and Wanshuai HU
This paper is based on a self-developed indoor ski simulator for ski training and entertainment, which can simulate slalom and combine it with virtual scenes to allow skiers to have a more realistic skiing experience. Firstly, the model of the skiing movement is analyzed, and the nonlinear control strategy of the motor velocity of the ski simulator is proposed. For the relative lag between the virtual scenes and the current skier’s position, we use a neural network to make predictions and use a sparrow search algorithm (SSA) to optimize neural network weights and thresholds. The simulation and experimental results indicate that this method can effectively predict.
Assessment of Cognitive Functioning System Based on Computer Graphics Technology
Yu Qi ZHANG and Naoyuki KUBOTA
Early detection of cognitive impairment is crucial for dementia prevention. We aim to assess those at risk of dementia’s cognitive functioning system. We developed a computer graphics-based short-term memory and attention test system that utilizes a spot-the-difference task. The system can recognize the position and information of particular things in photographs and then substitutes the original object with a photo of a comparable entity. The system may effectively produce several spot-the-difference task pictures, decreasing the repetitiveness and tediousness of studies. Furthermore, it allows for remote, real-time judgment and helps healthcare professionals provide better telemedicine services. The findings revealed that the proposed strategies were effective.
Comparative Study on Estimating the External Torque of Flexible Robot Driven by Series Elastic Actuators
Shenghong LIU, Gao WANG, Ning LIU, and Feiyan MIN
Accurate and fast estimation of external torque applied to the robot body directly determines the final performance of force control for flexible robot driven by series elastic actuators. The external torque estimation based on dynamics model requires the state information of robot link side, including angle, angular velocity and angular acceleration. Due to the noise and uncertainty of angular acceleration signal obtained by discrete differentiation of the angle, most methods of external torque estimation take the method of avoiding acceleration calculation. In this paper, an extended Kalman filter (EKF) algorithm is used to fuse the information of the photoelectric encoder and accelerometer to obtain the accurate link side angular acceleration signal, which is used to directly estimate (DE) the external torque based on the dynamic model. The results show that the momentum observer (MOB) has the same effect as reaction torque observer (RTOB), and the direct estimation method of precise acceleration signal has high accuracy and wide band. Also, the experiments in turn demonstrate the accuracy of the angular acceleration signal obtained by the EKF fusion.
Experimental Comparison of SNR and RSSI for LoRa-ESL Based on Machine Clustering and Arithmetic Distribution
Malak Abid Ali KHAN, Hongbin MA, Syed Muhammad AAMIR, and Cekderi Anil BARIS
LoRa lacks the sensing capabilities of channel status. Received signal strength indicator (RSSI) decreases due to collision, interference, and near-far effect while for signal-to-noise ratio (SNR), the packets are rejected by decreasing the transmission power (TP) at a higher spreading factor (SF). To overcome these challenges in the case of electric shelf label (ESL) to minimize the dependency on retransmission and acknowledgment, the end devices (EDs) are allocated around gateways (GWs) based on machine clustering with dynamic SF for SNR while dynamic TP for RSSI. The experimental results determined that the RSSI approach is more dominant than SNR because of determining the exact locality of the ED that diminished the capture effect. Arithmetic distribution of EDs for various GWs in different clusters helps to minify the near-far effect. The resultant received power (RP) at each cluster is higher for most of the connected EDs than the threshold RP.
An Arithmetic Sequence Method for Counting Ordered Pillboxes
Maofan CHENG, Hongbin MA, and Ying JIN
It is often necessary to count pillboxes in the medical and health care industry. However, there are many challenges in the counting process, such as complex patterns on the surface of the pillboxes, surface reflection, etc. It is difficult to achieve satisfactory results using existing methods, such as object detection and image segmentation. In this paper, a simple and novel method for arithmetic sequence algorithm based on deep learning and traditional image processing technology fusion is proposed, which can effectively and efficiently count ordered pillboxes. Firstly, we use one deep-learning-based real-time instance segmentation model to segment the pillbox image, and secondly use the arithmetic sequence algorithm we proposed to count the pillboxes. The method requires little computation overhead to obtain an accurate result. Our experiment shows that the algorithm we proposed is effective, efficient, with counting accuracy up to 99.89 % in our dataset.
Streamflow Prediction Using Cascaded-ANFIS Algorithm in Kelani River, Sri Lanka
Namal RATHNAYAKE, Upaka RATHNAYAKE, Tuan Linh DANG, and Yukinobu HOSHINO
Sri Lanka is rich in rainfall throughout the entire year, and river overflow incidents occur continuously. Heavy rainfall and poor infrastructure management cause the runoff of rivers. Therefore, this study proposes an accurate rainfall-runoff prediction model development using the Cascaded-ANFIS algorithm. The performance of the proposed method is compared with five algorithms, namely, Linear Regression, Lasso Regression, Ridge Regression, Gated Recurrent Unit, and Long Short-Term Memory. Moreover, the evaluation was done using six parameters: correlation coefficient, root mean squared error, percent bias, Nash-Sutcliffe Efficiency, and Kling-Gupta Efficiency. The proposed algorithm shows 0.9139, 1.15×10-10, 0.2100, 0.8351, and 0.8776 results, respectively. The results indicate that the Cascaded-ANFIS algorithm outperforms other algorithms in every aspect.
A Comparative Study of Statistical Models and Univariate and Multivariate RNN for Forecasting the Number of COVID-19 Cases
Naoki DOHI, Namal RATHNAYAKE, and Yukinobu HOSHINO
This study forecast COVID-19 cases in Japan to assist in decisions. There are statistical and machine learning models for forecasting time series. Traditionally, Statistical models are used in this field, and many machine learning models are recently adopted. Therefore, both models are needed to be compared for forecasting COVID-19 cases. Auto-Regressive Integrated Moving Average (ARIMA) and Recurrent neural network (RNN) are compared by Akaike information criterion (AIC) and Root Mean Square Error (RMSE). Results show that RNN (with vector inputs and outputs) was better than statistical models. Specifically, the RMSE of RNN is 47 and that of ARIMA is 103.
Proposal of Personal Value-Based User Modeling Without Attribute Evaluation
Kaichi NIHIRA, Hiroki SHIBATA, and Yasufumi TAKAMA
This paper proposes a personal values modeling method that does not require attribute ratings. The proposed method is applied to memory-based collaborative filtering (CF) to demonstrate its effectiveness. Personal values modeling using rate matching rate (RMRate) has been proposed, which needs an evaluation of each item attributes in addition to the rating value to items. Therefore, it cannot be used when the evaluation of item attributes is not available. To obtain personal values model of users only from a rating matrix, this paper defines user’s personal values as their tendency to select popular/unpopular items. Five attributes are proposed for modeling user’s personal values, of which effectiveness is evaluated in combination with user-based CF. Experimental results with three datasets show the proposed method can achieve higher precision and recall than conventional user-based CF.
Proposal of Generation of Rating Matrix Based on Rational Behaviors of Users
Kenshin MORIYOSHI, Hiroki SHIBATA, and Yasufumi TAKAMA
This paper proposes a method to generate an artificial rating matrix based on user’s rational behavior, with the aim of generating a large-scale rating matrix at low cost. Collaborative filtering is one of the major techniques for recommender systems, which is widely used because it can recommend items using only a history of ratings given to the items by users [1]. However, collaborative filtering has some problems such as the cold-start problem [2] and the sparsity problem [3], both of which are caused by the shortage of ratings in a database (rating matrix). This problem is particularly serious for services that have just started operation or do not have a large number of users. The proposed method generates a rating matrix with no missing values using users’ rating probabilities, which are obtained from the distribution of their actual ratings. The final synthetic rating matrix is generated after adjusting its sparsity by introducing missing values. The validity of the proposed method is evaluated by comparing the synthetic data in terms of the similarity of the distribution of several statistics with that of the real data. Synthetic data are generated by the proposed method and generated based on regular but meaningless behavior.
Vulnerability in Impression Based Strategy Against Bribery Behavior
Takuya OTOGURO, Kazushi OKAMOTO, and Atsushi SHIBATA
Electronic marketplaces have adopted centralized reputation systems that aggregate the ratings given by reviewers. This type of systems is used to express the overall evaluation of the products or sellers. Malicious sellers or users may manipulate the ratings to build false evaluations using these systems. Therefore, approaches to maintain high robustness against such attacks and to correctly evaluate products or sellers are being devised. Simultaneously, new types of reputation attacks are increasing. This study shows that an impression-based defensive strategy is vulnerable against a bribery behavior in which a seller offers bribes to and demands excellent evaluations from reviewers who have made transactions. The results of experiment suggest that this strategy fails to identify approximately half or more of the dishonest sellers under the ratio of bribe-giving manipulators and bribed ratings exceed by 80% and 30% respectively.
Müller-Lyer Illusion Is Replicated by Higher Layer of Pre-Trained Deep Neural Network for Object Recognition
Hongtao ZHANG, Zhen LI, and Shinichi YOSHIDA
Deep neural networks (DNNs) are commonly used in neuroscience research, however comparing neural networks to human perception remains challenging. With the development of the disciplinary intersection of deep learning and neuroscience, researchers have begun to use deep neural networks (DNNs) and human behavioral perception to assess the relevance of DNNs to brain neural networks. Here, we explore the similarity between DNNs and human visual perception through the well-known Müller-Lyer illusion. Specifically, by presenting the classical visual illusory stimuli, we compared representation by VGG-19 and ResNet-101 with human perception. Results indicated that similar to human visual perception, Müller-Lyer illusion is also shown in DNNs. Also, we further find that DNNs exhibit low physicalperceptual bias but present similar judgments in analogy with the layers of human perceptual hierarchy.
Image Classification of MRI for Alzheimer’s Disease Using Two Deep Neural Networks and Their Explainability
Boyue SONG and Shinichi YOSHIDA
Currently, many studies have realized the automatic classification of various neurological diseases with a synergy of raw fMRI data and convolutional neural networks (CNNs). However, CNN model has been criticized for its uninterpretability as a black box, which can greatly affect the credibility of the model, especially in the field of clinical diagnosis. In this paper, we used two modified CNNs to classify different stages of Alzheimer’s disease (AD) and applied Gradientsweighted Class Activation Mapping (Grad-CAM) on these models to explain them. However, the images of Grad-CAM didn’t highlight different regions of the brains according to the stages of AD as expected. Instead, the highlighted regions appeared randomly, which indicated the models didn’t effectively extract useful higher-level features for the different stages of AD. Instead, specific lower-level features of the patients’ brain might become the basis for classification.
Trust Towards the Robot with the Presence and Absence of Confidence
Youdi LI, Yasufumi TAKAMA, Eri SATO-SHIMOKAWARA, and Toru YAMAGUCHI
With the increasing presence of social robots in various circumstances, there is a great need to adapt robots’ behaviors to the needs of the individuals they interact with. Expression of confidence is one important capability in either interpersonal or human-robot interaction. In this study, volunteer subjects engaged in 2 phases involving a social robot Kebbi. In the Perceptive Phase, participants watched several videos showing the robot’s confident or unconfident expressions with different settings to pinpoint the factors of confidence expression including voice, motion, and response time. In the Game Phase, participants finished four sections of a treasure hunt game with support from the robot. In this phase, the robot gave hints while showing different combinations of confident/unconfident utterances and non-verbal information to explore how the robot’s expressions consisting of verbal and non-verbal factors affect people’s trust. Results have shown that both verbal and non-verbal factors are essential for people’s estimation of the degree of the robot’s confidence and participants tended to evaluate the robot with non-verbal information as more trusting and are more likely to trust the robot again after mistakes. This study provides insights into how advanced social features, like confidence, can be used to improve human-robot interaction.
Toward Construction of Handwritten Indus Signs Dataset
Sujata SAINI, Hiroki SHIBATA, and Yasufumi TAKAMA
Indus script is a writing system developed in the Indus civilization. Despite so many attempts over the years, it is still undeciphered. The decipherment of the Indus sign has yet to be resolved due to a lack of knowledge and limited data available. The demand for the large volume of datasets has increased recently with the hope of applying the state-of-the-art machine learning algorithms. The aim of the paper is to construct the dataset of the top 10 frequently appeared Indus signs found in multiple sites in the Indus civilization. As the number of images of actual Indus signs available is limited, a dataset is constructed by asking people to draw the Indus signs referring to the actual images. The handwritten images of Indus signs are collected from 36 participants with a web application. To examine the quality and characteristics of the constructed dataset from the viewpoint of machine learning, the dataset is used to train Convolutional Neural Networks (CNN) with the Tensorflow. The result shows the obtained model achieved 93% classification accuracy.
Proposal of Dynamic Generation of Subordinate Clusters Based on Bayesian Information Criterion for Must-Link Constrained K-Means
Shota SHIMIZU, Shun SAKAYAUCHI, Hiroki SHIBATA, and Yasufumi TAKAMA
This paper proposes a constrained K-means clustering method that dynamically generates subordinate clusters based on Bayesian Information Criterion. COP K-means, which considers a pair-wise constraints in partition-based clustering, have difficulty handling the case that a Must-link is given to instances located far away from each other. To address this problem, the proposed method generates subordinate clusters that have a Must-link to a master cluster during a clustering process. The final clustering result is obtained by merging the subordinate clusters. In addition, this paper proposes to determine whether to generate subordinate clusters or not based on the Bayesian Information Criterion. This paper evaluates the effectiveness of the proposed method with synthetic datasets.
Design of a Demonstration Space, the “Living Lab,” for Monitoring Daily Activities in the Older Adult and the Assessment of Assistive Technologies
Kenji KATO, Tatsuya YOSHIMI, Daiki SHIMOTORI, Keita AIMOTO, Naoki ITOH, Kohei OKABE, Naoyuki KUBOTA, Yasuhisa HIRATA, and Izumi KONDO
In this study, we designed and constructed a demonstration space, Living Lab, with specifications that can simulate daily life in a typical house or small nursing home. The basic concept is to provide a living space equipped with an entrance, living room, dining room, kitchen, bathroom and restroom within which the typical daily activities of an elderly individual can be evaluated. To safeguard against falls, the living space contains a suspension system. The size of the restroom and bathroom can be changed as required to simulate either the home or the nursing home environment. The Living Lab can also be used for training as part of a rehabilitation program prior to discharge from hospital, and also serves as a safe place to demonstrate new nursing-care devices and robots as they are developed.
Consideration of Strategy for Selecting Probing Items in Interactive Recommendation
Yasufumi TAKAMA, Mingman XU, and Hiroki SHIBATA
This paper discusses a strategy to determine probing item. The definition of probing item is provided in the paper: the probing item is an item by showing it to a user the system can obtain effective feedback for further recommendation. This paper compares several strategies for determining probing items. Probing items are presented to users especially at their Signup process to obtain their feedback, from which initial user profiles are constructed. Therefore, how to determine probing items are critical for designing recommender systems. This paper reports the results of comparing several strategies by simulating user behaviors using real dataset.
Joint Design of Alarm Deadbands and Delay Timers to Reduce the Number of Nuisance Alarms
Jiandong WANG and Poku GYASI
Alarm deadbands and delay timers are widely used in practice to reduce the number of nuisance alarms. This paper proposes a method to jointly design alarm deadbands and delay timers that are connected in a serial or parallel manner. Serial and parallel alarm systems are effective for removing nuisance alarms with different types of nuisance alarms, while alarm deadbands or delay timers alone are suitable only for one type of false alarms. The principle idea is that nuisance alarms with alarm deviations or alarm durations less than a certain threshold can be effectively removed by an alarm deadband or a delay timer. A minimization problem is formulated to design optimal values of the alarm deadband width and delay timer factor. Bayesian estimation approach is exploited to achieve reliable estimates of joint probabilities of alarm deviations and alarm durations. Existing methods are limited to the design of either alarm deadbands or delay timers based on probability density functions of process variables, and to a restrictive assumption that process variables are independent and identically distributed. The proposed method is for the joint design of serially/parallelly-connected alarm deadbands and delay timers; it is based on joint probabilities of alarm deviations and alarm durations, and does not require the restrictive assumption. Numerical and industrial examples are presented to illustrate the proposed method and compare with existing ones.
Alarm Floods Management by Extracting Structure Features
Chang TIAN and Chunhui ZHAO
Alarm floods indicate specific abnormal episodes, each of which has uniquely causal mechanisms between sources of industrial alarms, ultimately leading to causally dependent alarms. In that regard, online alarm flood classification based on causality differences could provide support for alarm handling, root cause analysis, and predictive maintenance. In this paper, the structure feature of alarm floods is explored for alarm flood classification. The structure feature reflects causal dependencies among alarm variables in alarm flood, achieving desirable discriminate properties for classification due to its interpretable physical meaning. Technically, we develop a probabilistic function that characterizes the alarm data distribution and parametrizes causal dependency probabilities. Instead of solving the parameters by traditional optimization methods, a deep learning model is exploited to generate the parameters and thereby gains the ability to transform an alarm flood to structure features in representation space. Besides, a general classifier is established based on structure features, enabling the online application. Experimental results illustrate the interpretability of structure features and show their effectiveness for alarm floods classification.
Pattern Mining for Prediction and Suppression of Industrial Alarm Floods
Wenkai HU
Alarm floods are among the most difficult issues in industrial alarm monitoring and have been identified as causes of many industrial accidents. During an alarm flood situation, the operator is overwhelmed by a large number of alarms and thus may fail to reorganize critical alarms. As a result, it could be almost impossible to make corrective actions, leading to deterioration of the situation and even causing serious consequences. To handle alarm floods and mitigate their negative influences to process operation, pattern mining has been exploited to extract interesting patterns that are then used for decision supports in alarm monitoring. This abstract mainly introduces the advanced methods and results on pattern mining for prediction and suppression of industrial alarm floods, including similarity analysis, pattern extraction, and real-time matching. Case studies are presented to demonstrate the effectiveness and practicality of these techniques.
Research on Correlation of Alarm Data of Refining and Chemical Plant Based on Optimized CBOW Method
Jinqiu HU, Qin CHEN, and Zhenqiang WEI
As the refining equipment gradually becomes more complex and intelligent, more and more alarm points need to be set in the industrial process. However, the functions of some alarm points are repetitive and unnecessary, and some alarm thresholds are set unreasonably, which will lead to a large number of alarm prompts in a short period of time and interfere with judgment of operator, and brings challenges to the alarm management. Therefore, it is necessary to analyze the alarm log and optimize the alarm management system. In order to process the alarm information more conveniently and quickly, and improve the efficiency of alarm log information analysis. This paper introduces the CBOW model and negative sampling optimization method commonly used in natural language processing to embed the alarm log, embed the alarm information, retain the correlation of the alarm information in time and semantics, calculate the Euclidean distance after the dimensionality reduction of high-dimensional vectors, and obtain the clustering results, which provides help for optimizing the process alarm point setting and alarm threshold setting. Experimental results show that the proposed method can retain the correlation of alarm data in time and semantics, and effectively identify alarms with high correlation.
Multiple Root-Cause Alarm Flood Classification with Weighted Dissimilarity Indices
Jing ZHOU, Jun SHANG, and Tongwen CHEN
Modern industrial facilities and processes have been becoming increasingly complex since a few decades ago. Massive numbers of sensors, actuators, and controllers can inevitably experience failures and induce alarm floods. Efficient fault detection and isolation (FDI) for these components based on analytical models and historical process/alarm data have attracted extensive research interests. Existing studies on root cause analysis are mostly restricted to the scenario that an alarm flood is caused by a single fault, which can be identified by machine learning and pattern matching algorithms. In real industrial processes, it is not rare that some units fail simultaneously, or more than one fault occurs within a short period of time. Owing to physical inter-connections and coupled material/information flows, a faulty component can sometimes cause its neighbor components to fail. Early isolation of all root causes that induce an alarm flood is an important topic. It not only allows field operators to respond quickly to tackle all malfunctioning units, but also is a core requirement for advanced alarm management systems. Motivated by the deficiency identified above, this paper proposes a pattern matching algorithm to achieve multiple root-cause alarm flood classification with weighted dissimilarity indices. For each labeled subset of historical data, a binary signature vector is firstly extracted as the alarm flood template; then a weighting factor is assigned to each unique tag to enhance the importance of relevant alarms and deemphasize irreverent ones. For alarm floods with multiple root causes, a predict-update strategy is adopted to fuse the information in historical data. Alarm flood classification is then achieved by comparing the weighted dissimilarity scores of an ongoing alarm flood with each template vector. The effectiveness of the proposed method is verified with simulated data from a Vinyl Acetate Monomer (VAM) plant model.
Assessment of Alarm Systems for Mixture Processes and Intermittent Faults
Mohsen ASAADI, Iman IZADI, Amin HASSANZADEH, and Fan YANG
Alarm systems are utilized in the process industries to notify operators and other plant staff of abnormal process conditions or equipment faults. Alarm system, as an essential part of distributed control systems (DCS), assists operators in safely operating the process under normal and abnormal situations. The alarm system must be appropriately constructed to maximize the possibility of safe and efficient operation. Nowadays, advanced industrial monitoring systems such as Supervisory Control and Data Acquisition (SCADA) and Centralized Monitoring System (CMS) are equipped with an enormous number of sensors, thanks to the rapid growth in sensing technologies. As an advantage, the number of alarm tags in a plant grows noticeably and makes it possible to monitor the whole industrial plant with more details. As a drawback, the number of alarms may confuse the operator to originate the faults. Chattering alarms are also another disruptive phenomenon in modern alarm systems. Poor alarm system rationalization is the primary source of the aforementioned issues. As a reason for poor alarm rationalization, the behavior of the faulty process variable is not taken into consideration in alarm system design, and again this will bring many meaningless alarms and no-alarms states to the operator. As an attempt to solve this problem, we tried to propose a model to reflect the behavior of an intermittent fault and utilize it to design alarm systems more specifically to the fault type. The Time-variant Finite Mixture Model is utilized to model the behavior of an Intermittent Fault (IF), which is a fault that alternates between faulty and non-faulty behavior at discrete random intervals. Time-variant finite mixture models could be a helpful tool for modelling complex data collections whose statistical properties change over time. A finite mixture model is a convex combination of two or more probability density (or mass) functions (PDF(PMF)). By combining the properties of the individual PDFs, mixture models are capable of approximating any arbitrary distribution. Our work's contribution is twofold. [For the rest of contents, please refer to the PDF.]
Identification of Autonomous Underwater Vehicles Model Based on Recursive Least Square with Bounded-Gain Forgetting
Yingying SHE, Gao WANG, and Feiyan MIN
The hydrodynamic characteristics of autonomous underwater vehicle (AUV) are complex and vary with the working environment, so it is difficult to establish an accurate mathematical model. To solve this problem, an online estimation method of AUV dynamical model parameters is proposed in this paper, including the establishment of AUV dynamical model based on recursive least square; A bounded gain forgetting estimator is used fast tracking of parameter variation. The proposed algorithm and result are verified by simulation experimental data. The results show that this method has good convergence and adaptability. In particular, the identification results of turning force and torque parameters have high consistency in different identification experiments, indicating that the method can extract the maneuvering characteristics of AUV well, which is helpful to the controller design of AUV. The research in this paper has potential application value to the modeling and control of underwater robot.
Collision Monitoring of Manipulators Based on Variational Mode Decomposition
Yingying SHE, Gao WANG, and Feiyan MIN
Collision detection and identification is an important subject for industrial robot research, such as contact location, touching a person or metal, etc. In this paper, a model independent collision detection method is presented, based on the variational mode decomposition of vibration signals generated by collisions. Firstly, the vibration modal features of the manipulator related to collision shock, collision direction and contact position are analysed with dynamic models. Then, the non-stationary vibration symptoms of collisions are extracted with a variational mode decomposition algorithm. The vibration modal symptoms are utilised for the construction and training of the artificial neural network for the collision detection, locating and classification. Finally, the proposed method is validated with both a two-link planar robot. The experimental results show the validity of the collision detection and identification scheme, which can achieve considerable performance.
Task-Driven Rule-Based Algorithm with Mode-Switching for UAV Swarms in Bounded Areas
Ziwei XIN, Juan LI, Chang LIU, Jie LI, Zihao CHEN, and Yiliang LI
The decision-making of UAV swarms under restricted environments is very important to exert the swarm effectiveness. On the basis of bio-inspired rule-based (Bio-RB) decision-making algorithms, this paper proposes a task-driven rule-based (Task-RB) decision-making algorithm. A mapping mechanism from behavioral rules to task modes and a probability-based mode-switching strategy are established for search and target attracting tasks. In order to improve the efficiency of performing tasks at regional boundaries, a boundary handling strategy based on the combination of path prediction and virtual agents is proposed. Taking search and target attracting tasks as examples, the Task-RB and Bio-RB algorithms are compared and tested. Simulation results demonstrate the effectiveness and adaptability of the proposed decision-making method to perform the above tasks in bounded areas.
Embedded Design of Automatic Pesticide Spraying Robot Control System
Ahamed MUSTAK, Hongbin MA, Lepeng SONG, and Ying JIN
In agriculture, crops need to apply pesticide spraying flow control precisely to reduce costs, protect the environment, and increase yield production. Although there have several variable control methods for spraying flow control because indirect control flow techniques and having a slow response could cause inaccuracy and mismanagement, also noted that those systems also suffer from complicated design and debugging, etc. In this paper, an embedded design of the fuzzy PID variable spraying control method is adopted. The experimental results show that the overshoot of Proportional–Integral–Derivative (PID) control is 10.76%, and the overshoot of fuzzy PID control is 7.17% which can meet the requirements of an advanced spray control flow system. For further investigation, a novel spray flow control method based on Programmable Logic Control (PLC) is proposed in this paper.
Fault-Tolerant Control for Manipulator Under Actuator Fault with Output Constraint
Yong HU, Fengjuan XU, Yun LI, Yaobing LI, Jianlei GAO, and Tianyu GONG
This paper investigates motion control problems of robotic manipulators under the actuator fault. The uncertainty induced by actuator faults brings up great challenges to control method design. Besides, with the consideration of output constraint and disturbance simultaneously, a fault-tolerant control (FTC) scheme, oriented to solve motion control problems is proposed in this paper. The barrier Lyapunov function is employed to deal with output constraint, an extended state observer (ESO) is used to compensate the system perturbation such as parameter uncertainties and external disturbance, and a fault estimation observer (FEO) is designed to compensate the actuator fault. Theoretical results are shown to confirm the output constraints are not violated and all the desired signals in the system are uniformly bounded. The simulation results verify the effectiveness and robustness of the proposed control scheme.