Bearing Fault Diagnosis Of Rotary Kiln Reducer Based On Convolutional Neural Network

The stability and safety of industrial systems is particularly important in the production process.Healthy equipment operation determines the yield efficiency in industrial production.Therefore,it is necessary to establish an effective equipment stability monitoring system.As data information capacity improvement,the conventional signal processing of feature extraction and state diagnosis cannot meet the requirement of the data.It takes a long time for expert experience mode to process capacity enlargement information,so the universality cannot be effectively guaranteed.With the development of big data,we can make use of data information to optimize our own structure and learn more data features.Combined with the optimization strategy of big data,this thesis provides an idea for data fault diagnosis with transforming the influence of data capacity.Based on rotary kiln speed reducer bearing as the research object,this thesis putting forward vision of using convolution bearing fault diagnosis model of neural network framework DSCNN-GRU.For the first time,the depthwise separable convolution of image processing is applied to the study of one-dimensional vibration signal.According to the model structure,the design criterion of parameters is proposed and the training parameters are optimized.The structural operation is to take the data of the original vibration signal directly through the data partition,coding and standardized input model,to extract the features of the long sequence data by using one-dimensional two-layer wide convolution and four-layer depth separable fine convolution,to input the short sequence into the GRU layer after sampling,and finally output the classification results.After the training,the model was tested by variable load test and measured anti-interference test,aiming to analyze the model's generalization ability under the condition of variable load and noise interference.A DSTCN model using a temporal convolutional network framework is proposed.The training speed can be improved and network parameters can be reduced.In the structure,the TCN residual blocks are stacked and changed,the perceptive field is hanced by the dilated convolution,and the multilayer convolution structure is simplified without affecting the accuracy.By comparing the training and testing effects of the two models,the characteristics of the models are analyzed and summarized.In order to explore the performance of the model,the thesis uses visualization techniques to understand the structure of the model and the learning process,graphically represent the data,and clearly express the hierarchical information of the convolutional neural network.Including the visual analysis of convolution kernel,convolution layer output activation and classification output,so as to obtain the intuitive expression of the structure flow.