Fault Diagnosis Approach Based On Deep Convolutional Neural Network And Multi-source Data Fusion

To solve the problem of pattern recognition for complex multi-working machine and eliminate the influence of experts' experience on feature extraction and data fusion,and improve the generalization ability of the fault diagnosis model,a research on intelligent fault diagnosis method based on deep learning model and information fusion was carried out.Paper proposed intelligent diagnosis methods based on multi-dimensional image feature fusion and convolution neural network,and constructed an adaptive multi-source information fusion strategy via deep learning network,and developed the intelligent diagnosis method of transfer learning based on multi-source data fusion.Paper realized the intelligent pattern recognition and transfer learning fault diagnosis of typically mechanical faults and composite fault statuses.The main contents and research results are presented as follows:(1)An intelligent diagnosis method based on improved convolutional neural network and multi-dimensional image feature fusion was proposed.The traditional facility fault feature graphs used for convolutional neural network analysis and recognition mostly rely on experts' experience,and the tedious engineering feature screening and image preprocessing process are time-consuming and labor-intensive with poor universality.To construct the sample feature map with obvious features and not too much parameter setting,paper proposed the methods of signal to color feature map and gray scale feature map.The convolutional neural network frameworks are established based on the features of the two feature maps respectively so as to extract deep-level information of signal-to-color feature map and gray scale feature map.By placing a bottleneck layer at the top or bottom of the network,the information of the feature map can be significantly enriched without changing the size of the feature map,which is conducive to improving the accuracy of network model recognition.The size of the grayscale feature map is compressed efficiently via the convolution kernel of a specific size fusing the multi-source sensor data in the feature map.The algorithm of Adam optimizer optimized the weight of the network model,and the convergent network model achieved to realize the intelligent recognition of multi-source data.In addition,the t-distributed stochastic neighbor embedding algorithm visualizes the data of the full connection layer to obtain the clustering graph for evaluating the performance of the network,which effectively verifies the high accuracy and stability of the proposed network in the diagnostic missions.(2)An adaptive multi-source information fusion strategy via deep learning network was constructed.An intelligent diagnosis network framework for multi-source data fusion is proposed,considering that traditional multi-source data analysis needs the guidance of experts,some effective information gets lost in the process of data fusion,and the problem that non-intelligent fusion algorithm usually lacks universality.To improve the universality of the data fusion strategy,a self-adaptive size convolution kernel is proposed for fusion of multi-source data.The multi-source data is analyzed by proposed data fusion strategy without tedious preprocessing and restrictions of the number of data sources.The atrous convolution kernel with the optimal sensing field extract feature from the one-dimensional optimal representation sequence of the multi-source data and efficiently compress the fusion sequence,which improves the stability of the system.The features are further explored through one-dimensional convolution,pooling and Parametric Rectified Linear Unit.The global average pool layer is used to collect features and map them to sample labels,which reduces the complexity of the network.Moreover,the addition of batch normalization ensures a same data distribution and speeds up the process of network convergence.The diagnosis model based on the original data fusion has obtained stable and efficient recognition results.(3)The intelligent diagnosis method of transfer learning based on multi-source data fusion was developed.The pre-training network model is constructed in the source domain for transfer to improve the convergence speed and pattern recognition accuracy of the target domain model.To solve the problems of weak generalization ability and slow convergence,the pre-training model was obtained through the convolutional neural network diagnosis method based on the multi-source data self-adaptive fusion strategy.The network structure framework divides the transfer module combination and determines multiple transfer strategies.The effect of partial network weight freezing and transferring is verified when the numbers of sensor sources and pattern recognition categories for target domain and source domain are different.The transfer of the convolution module at the bottom of the network ensures that the network learns from fault features conducive to pattern recognition.The transfer of the convolution module at the last layer can effectively improve the convergence rates of the models in different domains.The reasonable initializing of model weight by the optimal transfer strategy ensures that the network model of multi-source data fusion can effectively diagnose facility faults in the target domain.