Intelligent Diagnostic Method For Key Components Of Rotating Machinery Based On Convolutional Neural Network

Rotating machinery has been widely used in aerospace,electric power,metallurgy,petroleum and petrochemical fields.Its safe,reliable and stable operation is closely related to the economic and social benefits of enterprises.In order to improve the accuracy and reliability of fault intelligent diagnosis of rotating equipment,two typical components of rolling bearings and rotor systems are taken as the research object,and three theories,including signal processing,data fusion and deep learning,are comprehensively used.Partial research is carried out around the adaptive mining of time-frequency image features of non-stationary signals,the adaptive construction of deep diagnosis models,and the multi-source heterogeneous information data-level fusion diagnosis method.The main contents are as follows:(1)Adaptive mining method of time-frequency image features for non-stationary signals of rotating machineryIn the era of "big data",feature extraction and feature selection are very dependent on the personnel experience and prior knowledge in shallow intelligent classification models which have poor generality and nonlinear representation ability,and is difficult to deal with massive data.What’s more,the pattern recognition methods based on the time-domain and frequency-domain characteristics suffer the problems of stability and accuracy degradation when facing complex non-stationary monitoring signals.In view of this,an adaptive time-frequency image feature mining method for non-stationary signals based on the combination of Hilbert-Huang Transform and convolutional neural network is proposed.Firstly,the Hilbert-Huang Transform is used to process the low signal-to-noise one-dimensional time-domain signals to obtain time-frequency images in different states,retaining richer fault information;Secondly,a deep diagnosis model is constructed to excavate the fault-sensitive features of time-frequency images adaptively by using the powerful non-linear representation ability of convolutional neural network.Finally,the effectiveness of the proposed method are verified by simulation test data of rolling bearing fault.(2)Adaptive construction of convolutional neural network based on particle swarm optimizationAiming at the inherent shortcomings of convolutional neural networks,such as the need to manually set their structural parameters,hyper-parameters,which are time-consuming and labor-intensive,and low model reusability,an adaptive construction method for improved convolutional neural networks based on particle swarm optimization is proposed.First of all,the Hilbert-Huang Transform is used to obtain the time-frequency images,and the number of network layers and the pooling kernel size of the convolutional neural network are initialized.Then,the particle swarm optimization algorithm is used to adaptively optimize seven key parameters,such as the convolution kernel size of the convolutional neural network,which has strong self-adaptability.Finally,the application of this method in the fault diagnosis of rolling bearing and rotor system shows that this method can improve the adaptability of the diagnosis model.(3)Fault diagnosis method of rotor system based on data-level fusion of multi-source heterogeneous informationThere exists diagnostic uncertainly in fault diagnosis based on single type sensor information and difficulty of data-level fusion of multi-sensor heterogeneous information.Aiming at problems,a rotor system fault diagnosis method for multi-source heterogeneous information at the data-level fusion is proposed.First,the vibration signal is processed into a time-frequency image in the same dimension as the infrared image by using Hilbert-Huang Transform,and the data-level fusion is performed with infrared image to obtain the multi-channel fusion input;Secondly,these are input into adaptive multi-channel convolutional neural network for training to build a multi-source fusion diagnostic model.1 × 1 convolutional layer and global average pooling layer are used instead of CNN fully connected layer to reduce the risk of over-fitting and improve calculation efficiency.Finally,the proposed method is verified by the simulation fault test data of the rotor system,which improves the accuracy,robustness and noise immunity of the fault diagnosis,and achieves good diagnostic performance under small training sample background.