Research On Tool State Recognition Based On Acoustic Emission And Deep Learning

With the rapid development of machine tools in the practice of automation,integration and unmanned, how to ensure both product quality and production efficiency shows great urgency. The tool condition recognition technology can ensure the quality of the product and achieve the efficient use of the tool at the same time, so the study of tool condition recognition technology is very important.According to the characteristics of the tool. In this paper, the acoustic emission signal is monitored. The acoustic emission signal can avoid the influence of low frequency noise because of its high frequency. However, due to the high frequency of the acoustic emission signal and the different working conditions of the tools, the collected signal is not only complex, but also very large. At present, using signal processing algorithm to extract signal feature and then recognizing the tool's condition with machine learning method is the most popular way. However, The features of the signal are closely related to the tool's condition, and the key to recognize tool condition is to extract the appropriate features. But the methods of extracting feature often require a large amount of heuristic knowledge, rich signal processing theory and practical experience. Meanwhile, there are many manual factors in the process of extracting feature, so the processing has strong uncertainty.Also, it always takes a lot of time and energy.This paper introduces the deep learning Method and use Stacked Denoising Auto-encoder and Convolutional Neural Network to recognize the condition of tool.(1) Auto-encoder network is an unsupervised algorithm which can map high-dimensional data into low-dimensional space,and the Stacked Denoising Auto-encoder is stacked by a number of Auto-encoder. In this paper, using Stacked Denoising Auto-encoder with two hidden layers, getting examples and labels with frequency signal of acoustic emission signal and training network to get feature.After that, training Softmax with features and labels. In experiment, we make a comparative analysis of the parameters, such as the number of hidden layers, the hidden layers' nodes and the learning rate,etc,so as to find a better model.(2) Convolutional Neural Network is a supervised network with convolution structure. In this paper, the short time Fourier transform is used to transform the one-dimensional signal into a two-dimensional time-frequency spectrum as network input. Using time-frequency spectrum to make a large number of examples with labels and ensuring the diversity of samples. After that,training the Convolution Neural Network with pretreated samples to achieve recognition of tools' condition.In experiment,we make a comparative analysis of the parameters,such as the learning rate and the batchsize, so as to find a better network parameter.Results show that the two methods have gotten rid of the dependence of a large number of signal processing technology and actual diagnostic experience,and can extract characteristics of acoustic emission signal and recognize the condition of tool on themselves.