Research On Object Recognition And Heating State Detection Method Of Parts Based On Deep Learning

When the parts is continuously working,it is easy to cause overheating and damage of internal parts,resulting in a decrease in the working efficiency of the parts,and in severe cases,the parts will not work or even be damaged.It is very important to realize accurate and timely detection of the heating status of the parts.In recent years,inspection robots equipped with vision systems have solved the problems of high labor cost,low work efficiency,and unstable inspection quality in manual inspections.However,inspection robots are limited by the hardware level when calculating and analyzing images.The detection speed and accuracy of the target detection model cannot meet the requirements.To this end,this paper takes the YOLO v3 algorithm as the research object,and proposes a device target recognition method;for the problem of device heating detection,a device heating state detection method based on pruning optimization YOLO v3 is proposed,which is used in the realization of device heating state detection.Based on the design of the device heating status detection system,the proposed method is more usable and portable.The main research results of this paper are as follows:1.The infrared image data set of parts is constructed,and the classification and location of parts are realized based on the YOLO v3 model.Infrared thermal imager was used to collect infrared images of gearbox,motor and other parts,and the original image data set was preprocessed to enhance the image of the data set.Based on Label Img,the segmented training set and test set were annotated.Based on the K-means clustering algorithm,the appropriate anchor box is calculated for the infrared image of the parts,and the recalculated anchor box accelerates the convergence of the model.The training environment of YOLO v3 algorithm is built,and the initial parameters of the model training are set in detail.After training,the parts object recognition model is obtained,and the performance of the trained model is analyzed.The m AP of the model is 92.1 %,and the detection speed is 31 fps.The results show that the fast and accurate identification of motor and gearbox is realized.2.Based on the scale factor of BN layer,the pruning optimization training was carried out on the model,and the pruning optimization model was proposed.On the basis of achieving rapid and accurate identification of parts object,in order to reduce the amount of model calculations and reduce the hardware configuration required for model operation,the YOLO v3 algorithm has been pruned and optimized.Based on the BN layer after convolution layer in YOLO v3,the scale factor is sparsely trained by L1 regularization.After sparse training,the scale factor with smaller values tends to 0.By setting different pruning thresholds,the channels corresponding to the scale factor with the value tending to 0 are pruned.Through comparative experiments,it is determined that the pruning optimization effect of the model is the best when the pruning threshold is 80 %,and a smaller pruning optimization model is obtained,which creates conditions for the transplantation of the model.3.Based on the pruning optimization model,the multi-directional identification of the parts and the detection of the heating state of the parts are realized.The critical heating temperature of gearbox and motor is determined respectively.The parts state is divided into different categories according to the critical heating temperature,and the infrared image data set of parts is relabeled.Based on the pruning optimization model,the heating state detection model of the parts is obtained,and the accurate identification of the heating state of the parts is completed.The infrared image data of the parts in different directions and distances are collected again.The data set is enhanced and annotated.Based on the pruning optimization model,the multi-directional data set is trained to realize the multi-directional recognition of the parts.4.The parts heating state detection system was designed,and the detection software was developed based on Py Qt5.Based on the pruning optimization model,the parts heating state detection system is designed.The system is divided into three modules: acquisition module,detection module as well as demo module.Based on the human-computer interaction interface design software Py Qt5,the parts heating state detection software is developed.The software interface shows the parts detection information concisely and intuitively,which meets the needs of parts heating state detection well.Research on the basic theory,model training and test experiment of parts heating state detection method,development of detection system to realize the detection of parts heating state,it not only enriches the research on the field of parts condition detection and object detection engineering application,but also has important theoretical significance and practical value for promoting the development of intelligent and efficient parts detection management technology in intelligent manufacturing.