Research On Deep Learning For Acoustic Emission Detection Of Rotor Fault

With the maturity of industrialization and the advancement of modern science and technology,large-scale rotating machinery plays a vital role in many production and living fields such as petrochemical industry,aerospace,transportation,new energy and civil engineering.Therefore,the regular maintenance and fault detection of rotating machinery is of great significance.Acoustic emission detection is a kind of non-destructive testing technology,which can detect equipment failure by analysing the acoustic signal changes caused by internal defects of the material without destroying the components and ensuring the continuous and normal operation of the machine.However,most of the current acoustic emission signal processing and recognition still rely on traditional signal processing methods.Acoustic emission fault detection still has great difficulty in extracting signal features and resisting noise interference.In this paper,the emerging deep learning algorithm is used to identify and process the rotor acoustic emission signal of rotating machinery,and a good classification effect is obtained.The main work of this paper are as follows:(1)The research background,research history and significance of acoustic emission are described.The characteristics and advantages of acoustic emission technology are introduced,as well as the commonly used methods in acoustic emission signal analysis and recognition.Finally,the development history of machine learning and deep learning as well as representative algorithms is introduced in general.(2)The theory of deep learning algorithm is introduced and the forward prediction and back propagation mechanism are analyzed in detail,as well as the activation functions and optimization algorithms,which lays a theoretical foundation for network model training.A network model of AutoEncoder-Softmax is introduced.The self-encoder aims to extract the effective features of the input data,and Softmax is used for classification and recognition.The error descent curves of SGD and Adam algorithms in the training process are compared,and the confusion matrix on AE signal classification is given in this paper.(3)An acousitc emission signal recognition method based on improved traditional convolutional neural networks is proposed.Traditional convolutional neural networks usually adopt a hierarchical stack structure,and it is difficult to extract multi-scale features for a particular network layer.The Inception structure is introduced in this paper,which makes it possible to extract image features of different scales on the same convolutional layer.1*1 convolution is introduced to provide cross-channel information interaction capabilities and reduce the number of feature maps.The STFT spectrogram of the rotor AE signal and its firstorder difference along the time axis and frequency axis are used as the network input,and the global average pooling is used instead of the fully connected layer,which improves the generalization ability of the network while reducing the parameters.The validity of the model is verified by comparison with the traditional acoustic emission signal recognition method.In addition,the influence of machine speed on the detection of rotor fault is explored.It is found that increasing the speed of the machine within a reasonable range can improve the recognition accuracy of the classifier.(4)An acoustic emission signal recognition method based on the feature fusion of bidirectional long short-term memory network and convolutional neural network is explored.A BiLSTM-Attention and CNN parallel algorithm model architecture is constructed.In BiLSTMAttention,the bidirectional recurrent neural network is used to obtain the context information between the frame-level feature sequences,and the long short-term memory unit is used to learn the long-term dependencies between the frame-level features.The Attention mechanism is used to effectively weight the hidden state of each time step.The CNN network is used to extract the spatial features of the input data.Finally,the hidden layer state output of BiLSTM-Attention and the spatial features obtained by CNN are merged,and input into Softmax for recognition.The experimental results show that the model can effectively mine the semantic information contained in different rotor AE signals,and the classification accuracy of the model is effectively improved compared with the traditional RNNs and BiLSTM.