Research On Rotating Machinery Fault Diagnosis Based On Deep Learning Framework

Rotating machinery and equipment are widely used in modern industry.In actual operation,fault will inevitably occur,often leading to major economic losses and even casualties.Due to the operating speed,working environment,etc.,the key components of rotating machinery make it the main source of fault of rotating machinery equipment.Therefore,research on fault diagnosis of key components of rotating machinery is of great significance.Feature extraction is a key part of fault diagnosis,and its characterization ability directly affects the diagnosis results.However,most of the current fault diagnosis techniques based on signal processing or using shallow machine learning methods are manual feature extraction,which leads to the inability to accurately extract if there is lack of professional signal processing technology and sufficient prior knowledge.Then,it is difficult to guarantee the accuracy of fault diagnosis.Funded by the National Natural Science Foundation Youth Project(51505311),the General Project(51875375)and the Suzhou Key Industry Technology Innovation Project(SYG201802),this article aims at the problems of fault diagnosis of rotating machinery at this stage.Research on fault diagnosis of key components of rotating machinery is carried out under the framework of deep learning.Based on an in-depth understanding of the basic theories of deep belief networks and convolutional deep belief networks,improvements and optimizations are made to make it more suitable for the needs of mechanical fault signal feature learning,so that the model can automatically learn the deep-level features of the data from mechanical state signals.The qualitative and quantitative diagnosis methods of rotating machinery faults based on these two networks are proposed and verified.First,for the problem of fault identification of key train components(bearings),an adaptive fault diagnosis method based on optimized deep belief network is presented.The original vibration signal is converted to a frequency domain signal as the model input.Considering that over-fitting or missing optimal points may occur during the training of the model,a correction factor is added to the standard momentum method to give the momentum the ability to predict.Then,the learning rate adjustment strategy is used to adaptively select the step size during the gradient update process to optimize the parameter update process Finally,global fine-tuning is done through the top-level classifier to form the entire model.The performance of the raised method was certified by making use of the train wheel pair bearing data set,Case Western Reserve University bearing data set,and the bearing data set of different health conditions collected by the self-made bearing fault test bench.The results indicate that the model can complete the identification of fault type,fault degree and compound fault and has better convergence and higher test accuracy than standard deep belief network and support vector machines.It enhances the generalization ability of the deep belief model while stably and effectively improve the convergence of model trainingSecondly,the increase of the sampling time to obtain a better detection effect when using the deep belief network model for fault detection,the dimension of the input signal is increased.As a result,the training speed and forward calculation speed of the model are reduced,and parameter adjustment is more difficult.In view of the above problems,the convolutional deep belief network is improved and optimized,and an end-to-end fault diagnosis model is proposed.After the frequency domain signal is folded into an image,a band-pass operation is performed to filter out high-frequency noise and low-frequency color patches in the image.Then,the Adam optimizer is introduced in the model training process to speed up model training and improve model convergence speed.Finally,in order to give full play to the feature representation capabilities of each layer of the model and optimize the model structure,a multi-layer feature fusion learning structure is put forward to enhance the generalization ability of the model.Comparing the proposed model with traditional stack autoencoders,artificial neural networks,deep belief networks,and standard convolutional deep belief networks,the test results reveal that the proposed model has higher diagnostic accuracy and effectively realizes qualitative and quantitative diagnosis of bearing faultsFinally,what have been done in this study is generalized and several perspectives are put forwarded for the follow-up study.