Research On Lightweight Distributed Neural Network Technology For Fault Diagnosis

The mechanical fault diagnosis technology aims to monitor the operating status of the equipment in real time to make sure that the machinery can run safely and smoothly.The emergence of fault diagnosis technology has replaced the manual periodic inspection method,but the accuracy of traditional fault diagnosis technology is low,although the accuracy of deep learning fault diagnosis technology is higher,but the high cost of calculation and capacity makes it difficult to apply embedded device,and both of them need to transmit a large amount of data collected by the sensor to the server,so real-time fault diagnosis cannot be realized in a scenario where the transmission rate is limited.This article discusses a mechanical fault diagnosis method based on distributed deep learning,which aims to achieve real-time fault diagnosis with low latency,low power consumption and high precision.The main work and contributions are as follows:1.The thesis proposed a fault classification algorithm based on one-dimensional residual convolutional neural network called 1D-Res CNN.This thesis systematically analyzes and compares the application of the classic multi-dimensional time series data classification algorithms based on deep learning in the field of fault diagnosis,and compares them with the proposed 1D-Res CNN in the wind turbine fault diagnosis data set and the public data sets.The 1D-Res CNN proposed in this thesis achieves the best accuracy,can be competent for most time series data classification tasks and is suitable for small data sets.The proposed 1D-Res CNN solves the low-precision problem of traditional methods and provides a reference baseline for the follow-up research of the article.2.The thesis proposed DFECNN,a fault classification algorithm based on distributed feature extraction.Aiming at the shortcomings of the original method in low-bandwidth scenes,the DFECNN is designed to realize the distributed feature extraction of each sensor,and the features that reach the order of magnitude compression ratio are transmitted to the classifier for classification,so that the transmission volume of data is compressed extremely.At the same time,it ensures that the effect of the model does not decrease,so that we can achieve the purpose of high-precision and low-latency fault diagnosis.3.The thesis lightweight the fault classification algorithm based on distributed feature extraction,and the embedded application-level DFECNN Lite is obtained,finally it is deployed and verified in embedded devices.In view of the high computational consumption and high memory consumption of deep learning methods,this paper adopts two mainstream model compression technologies of lightweight structure design and parameter quantization to complete the lightweight transformation of the model,reducing the amount of model computation and parameters,and reducing the cost of embedded deployment.The operation pressure can achieve the purpose of real-time fault diagnosis with high precision,low delay and low power consumption,and finally the model is implemented in the embedded device.