Research About Intelligent Fault Diagnosis Method Based On Convolutional Neural Network

In the industrial production,to conduct the condition monitoring and fault diagnosis of mechanical equipment was an important means to ensure the safe and reliable operation of the equipment,discover the hidden danger in time,and avoid the consequences of disaster.In recent years,under the background of intelligent manufacturing and big data,the technology of fault diagnosis was developing towards intelligence.Intelligent fault diagnosis based on deep learning was a hot topic in current and future research.Convolutional neural networks(CNN)was one of the classic models of deep learning,which had strong ability of data mining and information fusion.In this paper,gear and rolling bearing,the key parts of rotating machinery,were taken as the research objects.Based on CNN model algorithm,the influence of parameters setting of one-dimensional CNN model(1DCNN)on the model performance,multi-sensor information fusion,feature extraction under variable speed conditions and domain adaptability were studied respectively,and three intelligent fault diagnosis were proposed.The feasibility and superiority of the proposed models were verified by experimental data.The main research contents of this paper were as follows:In the field of fault diagnosis,there was short of systematic analysis on the influence of1 DCNN model parameter configuration on diagnosis accuracy.On the basis of studying the core idea,basic structure,model training and optimization method of CNN model,this paper established a single-layer 1DCNN model and expanded it according to the dimensions of the analysis parameters,to explore the impact of architectural components on the model performance.Through the diagnosis results of each model on the rolling bearing fault data set provided by the bearing data center of Case Western Reserve University(CWRU),the influence of the number of convolution cores(network width)and size,pooling operation,batch normalization layer,network depth and sample length on the diagnosis accuracy of the model were systematically analyzed.At the same time,the interpretability analysis of the features extracted by the first layer of convolutional layer was carried out through continuous wavelet transform.For the problem of incomplete diagnostic information in intelligent fault diagnosis methods based on a single sensor,the channel attention mechanism was combined to construct a multi-channel input CNN model(MCFCNN)in this paper.The signals collected by sensors at different measuring points were constructed into a signal set with multiple channel attributes,and the channels correspond to the signals collected by sensors one by one.The MCFCNN model took the constructed signal set as an input,established multiple channels in the input layer,and added the channel attention mechanism to learn the sensitivity of each channel to faults automatically,and to assigned different weights to the channels,then used the feature extraction between the convolution kernel channels to achieve multi-sensor information fusion.Experiments were carried out on the experimental platform of planetary gearbox fault diagnosis and CWRU rolling bearing data set.The results showed that MCFCNN model can diagnose the fault with high accuracy,stability and speed.Aiming at the problem that the gearbox fault features were difficult to extract under variable speed conditions,the idea of multi-branch structure and multi-scale feature extraction of the Inception model was introduced into the variable-speed fault diagnosis,and a CNN intelligent fault diagnosis model based on multi-scale kernels(MKCNN)was proposed,to achieve end-to-end feature extraction and fault diagnosis under variable speed conditions.The sample division strategy was used to solve the problem of imbalance in the proportion of fault feature information under variable speed conditions.A cascaded multi-convolution kernel unit containing multiple convolution kernels of different scales was used to improve the nonlinearity of the model,and make the extracted features more abstract and rich.Experiments were carried out under the speed-up condition of the planetary gearbox,and the results showed that MKCNN model had strong feature learning ability under the condition of raising speed in unknown speed changes,and the average diagnosis accuracy on the test set reached 97.95%.At present,most of the fault diagnosis methods based on transfer learning were transferred to devices wiht high similar or similar equipment under different working conditions.On the MCFCNN model,a transfer learning framework TL?MCFCNN was designed based on model transfer,combined with appropriate parameter adjustment strategies,to realize knowledge transfer between different devices.The transfer tasks of 18 kinds of the same equipment under different working conditions and 7 kinds of different equipment under different working conditions on the planetary gearbox and CWRU rolling bearing data set were tested.The results showed that the proposed framework had strong self-adaptive ability in the case of cross-device remote fault diagnosis.