Study On Feature Extraction And Intelligent Recognition Method Of Rolling Bearing Based On Acoustic Signal

Vibration signal monitoring is one of the most commonly used methods in fault diagnosis and condition monitoring of rotating equipment.But under certain extreme conditions,it is not suitable to use contact measurement for signal acquisition.Acoustic signals are collected in a non-contact manner and contain a wealth of equipment status information.But the signal-to-noise relative vibration signals are relatively low,which restricts the application of acoustic signals in the field of fault diagnosis.In order to explore how to effectively use acoustic signals for fault diagnosis,this issue combines traditional algorithms and deep learning algorithms to carry out research on fault extraction and intelligent recognition methods based on acoustic signals.The main contents are as follows:(1)A sound signal feature extraction method based on CEEMDAN and deconvolution algorithm is proposed.In the traditional modal decomposition method,it is necessary to analyze the generated empirical modal components one by one.For the deficiency of the traditional modal decomposition algorithm,the application of the maximum steepness criterion can effectively achieve the adaptive screening of the optimal modal components.At the same time,combined with the deconvolution algorithm,the periodic impact in the acoustic signal is enhanced,and the reliability of the maximum steepness criterion in the model is improved.Through simulation signals and experimental signals,the effectiveness of the proposed model is verified,and compared with a single mode decomposition algorithm,the amount of redundant analysis is significantly reduced,and adaptive fault feature extraction is achieved.(2)A fault recognition model based on TextCNN is constructed.Based on the advantages of convolutional neural networks in the field of image pattern recognition,combined with the timing characteristics of one-dimensional signals,one-dimensional data segments are constructed into two-dimensional feature maps.Using the time series feature extraction ability of one-dimensional convolutional neural network,one-dimensional convolution kernel is used to replace the two-dimensional convolution kernel in the field of image recognition.In traditional convolutional neural networks,only single-size convolution kernels exist in a single channel.To address the shortcomings of the narrow field of view of single-size convolutional neural networks,the types of convolution kernel sizes in the model are enriched to construct a multi-size convolutional network model.The experiments of the vibration signal data and the acoustic signal data set were used to verify the applicability of the TextCNN network model and the acoustic signal in the fault field.At the same time,the input of different size data sets verified the effect of the data type on the network training.Compared with one-dimensional single-size convolutional neural network,two-dimensional single-size convolutional neural network and Gaussian kernel function support vector machine(RBF-SVM),the advantages of the proposed model are effectively verified.(3)Developed a fault identification method based on LSTM-FCN network.When converting a one-dimensional signal into a two-dimensional data feature map,it is easy to destroy the time feature of the one-dimensional time series.The acoustic signal can be regarded as a time series.To retain the characteristics of the one-dimensional time series of data,the overlap sampling strategy is adopted to strengthen the internal characteristics of the data,and the one-dimensional data slice is directly used as the data set.Although the one-dimensional convolutional neural network can process the time series,the extracted feature matrix cannot guarantee to contain the effective time feature information.The long-term and short-term memory neural network can effectively retain the time series information,and superimpose the time displacement information on the basis of local features to solve the problem Non-coherence of features extracted by product neural networks.Construct the LSTM-FCN network model,use multi-speed data sets and multiple data length data sets to perform experiments based on the performance of the LSTM-FCN model,optimize the reliability of hidden features,and test the performance of each channel of LSTM-FCN separately,proving that LSTM-FCN Applicability of FCN model to fault diagnosis with acoustic signals.