Abstract:To address the problems of inaccurate interaction information with the environment, sparse feedback, and unstable convergence of deep reinforcement learning algorithms in path planning, a dueling deep Q-network algorithm based on adaptive ε-greedy strategy and reward design is proposed. When exploring the environment, the agent uses the ε-greedy strategy with self-adjusting greedy factor, while the exploration rate ε is determined by the convergence degree of the learning algorithm, so that the probability of exploration and exploitation can be reasonably assigned. According to the physical theory of artificial potential field method, a potential field reward function is created which contains a larger gravitational potential field in the target, a repulsive potential field reward near the obstacle, making the agent to reach the end point faster. Simulation experiments are conducted in a 2D grid environment, the results show that the algorithm achieves higher average reward and more stable convergence under different scale maps, with an improvement of 48.04% in path planning success rate, which verifies the effectiveness and robustness of the algorithm in path planning. The method proposed in this paper is compared with the Q-learning, which has a 28.14% improvement in path planning success rate with better environment exploration and path planning capabilities.

Abstract:Aiming at the problems of low tracking accuracy, slow response speed and easy jitter in the track tracking and control method of construction vehicles, a robust nonsingular fast terminal sliding mode control method is proposed. First, the kinematic model and posture error model of the construction vehicle are established. Considering the singularity problem of the traditional terminal sliding mode, an integral non-singular fast terminal sliding mode controller is designed, so that the system error can be quickly converged while suppressing the jitter of the controller. Secondly, considering the problems such as the slow approaching speed of the traditional approach law, the linear velocity and angular velocity control laws are designed based on the reverse footwork, and the combination of the fal function and the inverse hyperbolic sine function is selected to design the approach law, which improves the stability and approaching speed of the system and weakens the jitter vibration of the sliding mode control. Finally, the control effect of the traditional non-singular terminal sliding mode controller is compared and simulated in Matlab software. Experimental results show that compared with the traditional nonsingular terminal sliding mode controller, the nonsingular fast terminal sliding mode control strategy proposed in this paper has obvious advantages in tracking accuracy and robustness of different movements.

Abstract:Aiming at the problems that the existing gesture recognition methods have too few datasets, less feature information and insufficient information extracted by the neural network, a gesture recognition method based on RAI image of millimeter-wave radar sensor is proposed.First, use TI′s IWR1443 millimeter-wave radar sensor to collect 10 types of gesture data to build a dataset, and then perform time-frequency analysis on the radar signal reflected by the hand to obtain a fixed frame number of RDI and RAI. In order to fully extract gesture features and classify them accurately, residual blocks and channel attention blocks are introduced on the basis of convolutional neural network. Experimental results show that compared with other features such as RDI, RAI can more accurately characterize gestures, the accuracy of the proposed network is increased by 11.78% compared with the CNN method, the accuracy rate of VGG16-Net and single-parameter VGG16-Net is increased by 7.98% and 11.78%, the parameter volume is reduced by 90.68%, and the time complexity is reduced by 17.2%.

Abstract:This paper proposes a Main-Partial Transfer Region-CNN method for insulator defect detection to address the problem that the Faster R-CNN algorithm is not accurate in detecting insulator defects with small samples in complex environments. The whole method uses a convolutional neural network with a fused residual module and feature pyramid structure as the backbone network for feature extraction, which is used to adapt to different scales of defect targets and retain more valid information. Then, the detected insulator body is automatically cropped to improve the effectiveness of the complex background in the detection of the defective area, so that the model can be more effective in mining the defective features. The insulator defect images collected by UAV aerial photography are detected. The results show that the average accuracy of the method in this paper is improved by 37.5% compared with the Faster R-CNN baseline model, reaching 89.6%. The accuracy is improved by 34.9% and 60.2% on the detection of missing and damaged, respectively.

Abstract:Aiming at the problem of co-channel interference when users reuse spectrum resources in D2D communication, an improved Hunter prey optimization algorithm is proposed to control the power of D2D users. It adjusts the D2D transmission power according to the rules that the hunter moves to the prey and the prey moves to the safest position under the constraints of satisfying the communication quality of the system users. The Sobol sequence initializes the population, and introduces the water wave dynamic adaptive factor into the predator position update formula, so as to determine the optimal transmit power of D2D users. The simulation results show that the algorithm can not only improve the total throughput of the system and reduce the interference to cellular users, but also improve the convergence speed and optimization accuracy.

Abstract:Aiming at the problems of high chattering, slow convergence speed and poor observation accuracy of sliding mode observer in permanent magnet synchronous motor position sensorless control system, a new sliding mode observer based on adaptive sliding mode gain variable power reaching law is proposed. Sliding mode observer, this reaching law adds an adaptive sliding mode gain variable exponential term consisting of the observed rotational speed, flux linkage and external input expected target current error on the basis of the power reaching law. The exponential term has a fast time-varying convergence rate, which effectively solves the problem that the original reaching law takes too long to reach the modal. The stability of the system is analyzed by Lyapunov stability criterion. Control experiments were carried out with the new sliding mode observer and the traditional sliding mode observer, the results show that the proposed new sliding mode observer has faster convergence speed, it can weaken highfrequency chattering and enhance load carrying capacity, and the adaptive sliding mode gain can effectively control the current error amplitude, Finally, the observation accuracy is improved by more than 20 percent.

Abstract:Path planning is one of the key technologies for reconfigurable robots to accomplish tasks quickly. In order to improve the driving efficiency of the reconfigurable robot and shorten the driving path, an improved A* path planning algorithm based on the idea of Bresenham straight line algorithm is proposed to achieve path point reduction and inflection point elimination of the reconfigurable robot and improve the smoothness of the pathfirstly. On this basis, considering the volume of the reconfigurable robot itself and the reconfigurable characteristics of the robot, the reconfigurable robot configuration library is established, and the relationship between the volume of the reconfigurable robot and the surrounding obstacles is discussed to reduce the probability of collision between the robot and obstacles in the process of walking. The improved A* path planning algorithm is simulated by using MATLAB simulation platform to verify the effectiveness of the algorithm, which can be applied to robot path planning in complex environment. The path planning problem after robot reconfiguration is analyzed, and the running distance of robot can be shortened by using reconfigurable characteristics, which reflects the superiority of reconfigurable robot.

Abstract:Dam monitoring data is an important guarantee of dam safety. Anomaly detection and recovery of dam monitoring data can effectively avoid the wrong estimation and judgment of dam status, which has important practical significance. In recent years, there are extensive studies on anomaly detection of dam monitoring data based on deep learning methods. However, the existing methods have some drawbacks such as insufficient data utilization and insufficient information mining. Therefore, a multi-dimensional LSTM anomaly detection and recovery method is proposed in this paper. The dam monitoring data of multiple monitoring points are fed into LSTM to predict the data of single monitoring point, and the relevant information between different monitoring points is effectively utilized. Finally, anomaly detection is performed on the data of the target detection point using Pauta criterion. In this paper, the laser collimation monitoring data of Fudougou Hydropower Station in Dadu River are used for case verification. By comparing with the single dimension LSTM anomaly detection and recovery algorithm, it is verified that the performance of proposed method is effective both in anomaly detection and data recovery, which is an effective method for dam monitoring data anomaly detection and recovery.

Abstract:Aiming at the high error rate of speech transcription text, proposes a text error detection and correction model based on MacBERT to correct the text after speech transcription. In the error detection stage, the MacBERT-BiLSTM-CRF model is used to check whether the text is wrong and where it is. In the error correction stage, starting from the two dimensions of confidence and phonetic similarity, a curve of "confidence-phonetic similarity" is delineated to determine whether candidate words are to be corrected for errors. The confidence of the candidate words is calculated using the MacBERT language model, and a phonetic similarity calculation method based on pinyin code is proposed. Experiments were conducted on the public speech dataset Thchs-30 by calling Baidu speech recognition API. Compared with the existing methods, the precision rate, recall rate and F1 value in the error detection stage and error correction stage have been improved. Among them, the error correction stage The accuracy rate reaches 83.32%, which improves the accuracy of the transcribed text.

Abstract:In view of the low detection accuracy, missed detection, and the wrong detection of road traffic signs in a weak light environment, a detection algorithm based on the improved YOLOX is put forward. Light weight network named Mobile Vi T Block module is adopted, meanwhile CNN is combined with Transformer to raise the network’s ability to learn local and global features of objects. By adding the adaptive feature fusion pyramid ASFF, the improved algorithm performs weighted fusion on the effective feature layers in order to accelerate the convergence speed of network training. The binary cross-entropy loss function is replaced by a Focal Loss, so as to solve the problem of inaccurate classification due to the small samples size. As shown by the experimental results, the mAP value of the improved YOLOX algorithm is increased by 2.89% than that of the YOLOX algorithm, and the number of parameters is reduced by 6.23 M. The visualization and other experiments further verify that the improved YOLOX algorithm can effectively avoid the phenomena of missing and wrong detection caused by weak light.

Abstract:Aiming at the problem that the tracking algorithm based on the fully convolutional siamese network is easy to tracking drift in the face of complex environments such as analog interference and illumination changes, this paper proposes the following strategies to optimize features on the basis of analysis and experiments. First, the deep convolutional neural network VGG16 is introduced into the tracking framework to improve the feature extraction ability of the model. Then, aiming at the problem that a single feature cannot adequately describe the target information and is sensitive to interferences, this paper designs a feature enhancement module, which integrates different levels of semantic information from shallow to deep to improve the expressiveness of features. Finally, a lightweight triple attention is proposed to help the model adaptively focus on dominant features and further improve the robustness of the model in complex environments. Applying the above strategies to the fully convolutional siamese network algorithm has achieved remarkable results. On the OTB100 dataset, compared with benchmark algorithm, the area under the success rate curve of the algorithm in this paper is increased by 15.1%, and the distance accuracy is increased by 16.3%, and the target can also be effectively tracked in complex environment.

Abstract:Aiming at the problem that traditional hyperspectral image clustering algorithms are difficult to effectively deal with hyperspectral images with rapidly increasing data volume, a hyperspectral clustering algorithm based on hyper-pixel anchor layer convergence point selection is proposed. SuperPCA is used to reduce the dimension of original data based on super pixel cutting. Selecting representative anchor points by K-means, and constructing an adjacency matrix based on anchor points. The similarity graph is constructed by the method of non-nuclear neighbor assignment to avoid the adjustment of thermonuclear parameters. Finally, spectral clustering analysis is carried out to obtain clustering results. The simulation experiments on Indian Pines and Pavia Centre hyperspectral data sets show that the classification map obtained by this algorithm contains fewer false points, and the distribution of ground objects is smoother. Compared with the current hyperspectral image clustering algorithms, this algorithm has a better clustering effect.

Abstract:Aiming at the problems of single feature extraction method and insufficient feature extraction of middle layer in existing image super-resolution reconstruction techniques, a channel-separable residual network based on attention mechanism is proposed. Firstly, a diverse branch block is designed by using the idea of multi-scale convolution to fully extract the low-frequency information of the image. Secondly, channel compression is used to reduce dimension to simplify feature information, and coordinate attention mechanism is introduced to enhance local fusion features. The trunk network is focused on extracting high-frequency features while accelerating convergence through long and short jump connections. Finally, the high-resolution image is reconstructed by upsampling layer. The proposed algorithm is compared and analyzed on the public data sets of Set5, Set14, BSD100 and Urban100 in the super-resolution reconstruction field. On set5 dataset of ×2 reconstruction task, compared with DBPN, parameters is 2/5 of DBPN, the PSNR and SSIM are improved by 0.09 dB and 0.001 6 respectively. Experimental results show that the proposed algorithm can fully extract image features and achieve similar or even better reconstruction results than other large-scale models with fewer parameters.

Abstract:In order to realize unsupervised detection of wafer surface defects, an unsupervised wafer surface defect detection model with improved generative adversarial network was proposed. The model detected the defects by the difference between the target image and the reconstructed image. In this method, an encoder-decoder convolutional neural network with two layers of skip connections and memory module was used to build the generator. The skip connections were used to capture multi-scale input image features, and the memory module was used to constrain latent characteristics to enlarge the distance between real defect samples and reconstructed samples. The method also makes the model lightweight by improving the discriminator network structure. Experimental results show that the proposed model can accurately distinguish the defective wafer samples, and the area value under the ROC curve reaches 0.934, which is better than the existing unsupervised learning detection methods, and the parameters and flops of the discriminator network is reduced to less than 1 M and 60 M.

Abstract:Aiming at the problems of poor effect and inaccurate defect location of traditional steel plate surface defect detection methods, a deep learning detection algorithm based on improved RetinaNet-GHM was proposed. Firstly, the path aggregation feature pyramid network is introduced to fuse shallow and deep semantic information to improve the detection effect of the network on small targets. Then, GHMC and GHMR loss functions are used to classify and locate defects. Finally, the soft-non maximum suppression algorithm in Gaussian form is introduced to improve the detection accuracy. The experimental results show that the average accuracy of the improved RetinaNet-GHM algorithm is 76.7%, and the average accuracy of crazing, inclusion, patches, pitted surface, rolled-in_scale and scratchs is 45.2%, 88.2%, 94.2%, 86.1%, 65.1% and 87.4% respectively. Compared with other classical algorithms, the improved RetinaNet-GHM algorithm has better detection effect.

Abstract:In order to avoid the occurrence of traffic accidents caused by fatigue driving and to safeguard urban road traffic as well as occupant safety, this project addresses the core problems in traditional fatigue driving detection methods, such as low accuracy, elaborate parameters and poor generalization, by using the MTCNN and infrared-based rPPG to accurately extract driver’s facial and physiological information in complex driving environments with changing light, partial occlusion and head deflection. At the same time, after deep mining the specific fatigue information of multi-modal modes, combined with the multi-loss reconstruction(MLR) feature fusion module to use the complementary information between each mode are employed to further construct the multimodality feature integration model, which breaks the limitation of single-mode detection methods and improves its the accuracy and robustness in complicated driving environments. Finally, by using the time-series nature of fatigue, a fatigue driving detection system based on the Bi-LSTM model is established. Experiments were conducted on a home-made dataset FAHD, which demonstrated the reliability of the infrared physiological feature extraction model. In addition, the accuracy of multimodal input increased by at least 5.6% compared to the single-modal input, while the correlation coefficient improved by 5.6% and the root mean square error was reduced by 25% compared to existing fusion methods, achieving an accuracy of 96.7%. While promoting the development of intelligent transportation, it also has a good positive significance for the maintenance of traffic safety.

Abstract:In order to find the defects of transmission lines accurately and timely, the three-dimensional inspection mode of transmission lines based on artificial intelligence image recognition technology is studied. Specifically, with the support of artificial intelligence image recognition technology, K-means algorithm is used to cluster the threedimensional inspection images. At the same time, artificial neural network is used for intelligent identification of transmission line defects in the image. According to the test, under the same workload, the transmission line defect identification without the use of artificial intelligence image recognition technology requires 5 image analysts to work continuously for 15 d, with an average of 2~3 images per minute and an image recognition speed of 20~30 s/piece; Using artificial intelligence recognition technology, recognition speed up to 0.25 s/sheet, only 3.6 h can complete the recognition task. In the low-light environment, the edge of the image is clearer after the enhanced force, and the target image is clearly separated from the background. At present, the recognition accuracy is higher than 90% by using 7 images from different angles of the same parts in the transmission line. In addition, by comparing the influence of constant learning strategies, AdaDec and AdaMix learning strategies on the convergence of the deep learning model under the same conditions, it is found that the reconstruction errors of the three strategies all show a tendency of gradually decreasing, and eventually tend to be stable. The results show that the artificial intelligence image recognition technology has a certain universality in the three-dimensional inspection image defect detection of transmission lines, and is conducive to the significant improvement of work efficiency.

Abstract:Automatic identification of bills is one of the important means of bill digitization and improving bill information processing ability. Considering the uniform specifications, the same structure and a large amount of duplicate information of the same type of bills, a bill recognition method based on template and content separation is proposed. In this method, the structure and inherent text of the bill are extracted as templates through color segmentation, and the rest is used as the content of the bill. Combined with the improved siamese neural network and template alignment, the bill template to be tested is matched with the existing bill in the template database, and then the new bill is reconstructed. The results show that compared with the original method Baidu OCR, the text detection time and recognition time of this method are reduced by 68% and 91.13% respectively, and the overall prediction time is reduced by 88.62%, reaching 3.45 seconds/piece.

Abstract:The study of cervical nucleus segmentation is of great significance for cervical cancer screening and diagnosis, but it brings great challenges to the segmentation task due to the influence of blurring edge and interference. In response to this problem, a nuclear segmentation method based on the DeepLabv3+network was proposed. Firstly, utilizing the output of the backbone network for multi-scale feature fusion and introducing an attention mechanism, a cell mass segmentation model was built to reduce the effects of interferences in the background on nuclear segmentation. Based on this, a two-path feature extraction module combining Transformer and ResNet50 was designed, which takes into account the sensitivity of the model to global information acquisition and low-level context features, and improves the discrimination ability of the model to nuclei and interference information. The experimental results show that the algorithm has achieved good segmentation results in the task of cervical cell nuclei, and MIoU is 0.832 9, which has increased by 2.33% and obtained better performance indicators compared to other methods.

Abstract:For the online measurement of size deviation in industrial pipelines, this paper proposes a pipe diameter measurement method integrated with multigroup biocular visual system. This method uses image acquisition, 3 D reconstruction and coordinate fusion through multiple laser markers projected on the pipeline measurement section, obtaining dimensional parameters such as outer tube diameter and ellipticity by ellipse fitting. This method realizes the automatic matching of feature points through the affine distance transformation algorithm, and adopts the optimized point cloud registration algorithm with normal vector constraints to ensure the accuracy of coordinate fusion. To verify the feasibility of this method, a multi-group binocular visual experiment system was constructed on three pipelines with nominal external diameter of 285, 299 and 325 mm. The experimental results show that the relative error of the systematic measurement is in the ±0.570% range, with the maximum standard deviation of repeatability of 0.551 mm and the longest measurement time consumption of 1.5 s. The method and system basically meet the needs of industrial pipeline online detection, and have good application prospects.

Abstract:User Perception Analyzation System (UPAS) plays important supportive role in maintaining every step of network operation and maintenance for communication carrier. However, the consistently changing of network and its relevant businesses bring forth great challenge to the perception analyzation. Based on the UPAS theory, this research first investigated its possible drawbacks on benefit and cost, and proposes a two-stage user perception measurement and analysis framework. In the first stage, a general quantitative evaluation model is constructed for evaluating single service quality. In the second stage, a quasi-unsupervised machine learning model is constructed, so that the satisfaction evaluation method has the ability of self-learning to adapt to the dynamic changes of the network. It can reduce the shortcomings of the network to specific cells and specific services that can greatly improves the practicability and usability of the method. The analysis from the existing network shows that the recall rate and precision rate of this method are much higher than those of the traditional method. Finally, the future evolution of the system in capacity, efficiency and operation are viewed.

Abstract:To address the problem of 16-line lidar sparse point clouds and poor perception algorithm due to the distant port sea surface targets, a dynamic and static target detection method IMU is proposed. Firstly, we propose an improved Ray Ground Filter algorithm for sea clutter filtering to address the problem of false detection caused by the point clouds of unmanned vessel wake stream; then, we propose a target clustering algorithm for different distances to address the problem of poor clustering due to different sparsity of target point clouds; finally, we realize inter-frame projection of lidar point clouds by fusing IMU to complete the dynamic-static target detection and key point prediction. By using the unmanned vessel experimental platform and the simulation platform for target detection experiments, the algorithm in this paper has high detection efficiency and stable robustness, which can better achieve the perception of port environment by unmanned vehicles.

Abstract:As the mask will greatly reduce the features available for face recognition, the recognition performance of the previously proposed face recognition algorithm will be greatly reduced in the existing external environment. Therefore, given the shortcomings of the existing face recognition technology in the current application scenarios, this study uses MobileNet v2 lightweight convolutional neural network to replace the Inceptionresnet-V1 network as the backbone network to improve the FaceNet face recognition method, which simplifies the model parameters and improves the operation speed of the model. In addition, a lightweight mixed attention module is introduced into the Mobilenet v2 network, and the weighted fusion of Softmax Loss and Triplet Loss is used as the joint Loss function of the network model, which is trained as the Loss function after the adjustment of weight reaches the optimal value to improve the recognition accuracy of the network. The experimental results show that the face recognition network proposed in this study achieves 92.1% recognition accuracy in face mask masking, which is significantly improved compared with the original face recognition network, and the recognition speed is also significantly better than the original network.

Abstract:Aiming at the problem of low accuracy in calculating the width of wall cracks, a sub-pixel based crack width calculation method is proposed. On the basis of extracting the crack area, this method uses the crack width measurement method based on the vertical line of the central axis, combined with polynomial fitting the crack edge, and extracts the sub-pixel coordinate points of the left and right edges of the crack, so as to calculate the sub-pixel crack width by using the Euclidean distance method, and compares it with the pixel level calculation method. The experimental results show that the algorithm is more accurate and can be used to measure the width of various types of wall cracks. The average relative errors of longitudinal, transverse and cross cracks are 3.02%, 2.44% and 3.72% respectively. Compared with the pixel level method, the average errors are reduced by 1.87%, 1.95% and 2.07% respectively, it has strong generalization ability and stability.

Abstract:Surface electromyography is the bioelectricity generated by the contraction of human surface muscles,whichcan reflect the activity of human muscles,It has important research value in the medical field.Because the surface EMG signal is a weak signal, the smaller noise can produce larger interference. Aiming at this weakness,In this paper, a surface EMG signal acquisition system with small volume, low power consumption and strong anti-interference ability is designed.AD8232 is designed as the acquisition chip, combined with amplification, filtering and analog-to-digital conversion circuits,and STM32 is used for data processing.Finally, the data is sent to the host computer through low-power Bluetooth.The test results show that the energy of surface EMG signal is mainly concentrated in 50~150 Hz,Above 300Hz, it obviously attenuates,and is greatly disturbed by low-frequency signals.After hardware filtering and wavelet denoising algorithm,Interference is well inhibit,Especially for low frequency interference.The system has strong anti-interference ability and reliability, and has a good market prospect.

Abstract:Aiming at deeply exploring the analysis of the recording performance of the distribution terminal, this paper proposes an improved dynamic time warping algorithm for detecting the terminal recording platform of the primary and secondary deep fusion equipment. Complete the waveform preprocessing by framing and windowing, and calculate the shortterm energy entropy ratio of the source signal waveform and the terminal recorded waveform affected by noise interference; The short-time energy entropy ratio sequences of the two waveforms were used as the input test vector, and the path planning and similarity calculation are carried out by using DTW for the two waveforms. And by computing two groups of short-time energy entropy than public substring sequence length defined optimization matching waveform similarity coefficient of correction. Experimental simulation and measured data analysis results show that the DTW algorithm combining energy-entropy ratio and common substring can accurately improve the computational efficiency and accuracy of the algorithm. Experimental results show that this method are conducive for the evaluation of recording performance of power distribution terminals.

Abstract:Aiming at the problems of privacy protection, fall detection and low recognition rate in home behavior recognition of the elderly, a new human behavior recognition algorithm based on WiFi signal is proposed in this paper. Firstly, 10 kinds of daily life behaviors of the elderly (drinking water, falling, sitting down, etc.) were collected in the simulated home environment; Then, the extracted WiFi channel state information is denoised by Butterworth filter, and the dimension is reduced by principal component analysis; Finally, the processed CSI signals with clear features are input into the attention based bidirectional long short-term memory model for behavior classification. The efficient bi-directional structure and attention mechanism not only produce more informative features, but also improve the generalization performance of behavior recognition; Experimental results show that, compared with some benchmark methods, the proposed algorithm can achieve the best recognition performance for all activities on both public data sets and self-collected data sets, and the accuracy rates are 98% and 96% respectively.