| With the continued development of intelligent production models such as energy meters and the widespread use of big data analytics in the manufacturing sector,the use of control charts as an important method for quality control of the energy meter manufacturing process must also be adapted.This topic is derived from the Huali Technology’s smart manufacturing project for electrical energy meters.With the voltage correction bias value of electrical energy meters as the research object,an intelligent identification method for control chart patterns is investigated with a high classification accuracy,low time-consuming and model-lightweight control chart abnormal pattern classification method as the research objective.The project proposes a control chart pattern recognition method based on Dense Net network,which can make use of big data from manufacturing sites to achieve high accuracy and fast recognition of control chart out-of-control discriminatory patterns and ensure efficient control of process quality.The main contents of this topic are as follows.(1)Based on the study of the national standard GB/T 17989.2-2020 "Control chart Part 2: Conventional control chart",a mathematical model for control chart pattern classification in accordance with the national standard for control chart anomaly pattern classification is established,on which the system technical scheme is designed,voltage bias value data and collection are carried out,for the subsequent voltage bias value control chart The system was designed on the basis of the technical scheme,and the voltage bias value data and collection were carried out to provide high-quality samples for the subsequent control chart anomaly pattern classification algorithm.(2)In order to meet the needs of control chart anomaly classification,the image classification algorithm is improved and applied to the control chart anomaly classification,Inception V3,Shuffle Net V2,Res Ne Xt and Dense Net networks are modified for the convolution and pooling layers,and the above four image classification networks are applied to the one-dimensional energy meter voltage bias value control chart in a multi-label classification model for anomaly patterns.On a test set consisting of 257,469 energy meter voltage bias data,Dense Net had the best classification but had the highest computation size of 584 M,model size of 881.5MB and CPU utilisation of 92.77%.Res Ne Xt had the lowest computation size of 238 M and the shortest test time of 667 s,thanks to the reduced parameters and increased network depth;Inception V3 achieved the lowest parametric size of 43.8M by replacing the fully connected layer with an average pooling layer,convolution and bottleneck layer design;Shuffle Net V2 used convolution to improve the efficiency of the network and achieved the lowest CPU utilisation of 50.8M.Shuffle Net V2 uses convolution to improve network efficiency to obtain a minimum CPU utilization of 50.04%.Therefore,this paper takes the Dense Net network as the basis,combines the advantages of Res Ne Xt,Inception V3 and Shuffle Net V2 networks,reduces the number of hyperparameters and fully connected layers,and increases the network depth by setting up convolution and bottleneck layers.(3)To address the problems of low parameter efficiency and huge consumption of video memory in Dense Net networks,we propose a solution that combines the advantages of Res Ne Xt,Inception V3,and Shuffle Net V2 networks for feature extraction,migration learning and FCBF algorithm for feature enhancement,and SVM classifier for feature classification.The quantitative design goals of high classification accuracy,small model size,low memory consumption and low time consumption are derived from the modeling of the total weighted score function of the design goals.In feature extraction,the data is trained through the structural features of Dense Net dense connectivity mechanism and feature reuse,and features are refined in the Dense Net network through bottleneck layer design and model compression to improve the efficiency of model parameters and shorten training time;in feature selection,the voltage biased data features extracted from the fully connected layer are enhanced through migration learning to increase the control chart anomaly pattern classification The FCBF algorithm is then used to select the most effective feature vector to improve the model performance and reduce the time complexity;in terms of feature classification,the SVM classifier is used as the model output,which can ensure robust model convergence.The training results show that the improved network achieves the best convergence with a maximum overall score of 2.3.The classification performance of the improved network is better than that of the pre-improvement network,the testing time is reduced to one-third,the number of parameters is only half that of the full model,and the computer running speed is almost doubled.It also achieved a small misclassification rate for multiple labels,with an average accuracy of 98.56% for the classification of eight anomalous patterns,achieving good classification.the average AUC was at 0.9,with excellent correct classification rate.(4)Field deployment of a multi-label classification algorithm for anomaly patterns of energy meter voltage bias value control graph based on improved Dense Net network.In this paper,a fully automated architecture capable of classifying energy meter voltage bias value control charts is proposed and software development is carried out.The software automatically annotates and divides the training set from the data on the production line via a Python script.The control chart classification results and the current detection speed can be displayed in real time,and the number of anomalous patterns in the classification process is summarised and counted.The test results show that the total false alarm rate is less than 5%,the false alarm rate for each anomaly pattern is less than 2%,the average CPU utilisation is 37.49%,the F1-score is 0.97,and the classification time is 70s/k,which meets the design objectives. |
PDF Full Text Request