Research On Faults Diagnosis Based On Improved Particle Filter And Deep Belief Networks

With the continuous development of machine learning,its application fields extend from simple image recognition to medical imaging,target locating and tracking,fault diagnosis and other fields.At the same time,the large-scale and automation of industrial equipment makes traditional fault diagnosis methods unable to adapt to this change.In this context,this paper takes rolling bearing as the research object,and proposes a fault diagnosis method based on improved particle filter and deep confidence network for the three difficult problems of traditional fault diagnosis method,such as difficulty in feature extraction,low classification accuracy and long classification time.In this paper,the typical faults of rolling bearings and their failure modes are analyzed,and the formula for calculating the characteristic frequency of fault signals is obtained.According to the experimental requirements,the rolling bearing test bench and data acquisition system are designed,the bearing defects of different degrees are designed and processed,the experimental steps are determined,and the bearing vibration data acquisition is completed.Aiming at the low signal-to-noise ratio(SNR)of the original signal and the over-fitting of the standard particle filter,the unscented Kalman algorithm is used to improve the particle filter to improve the signal-to-noise ratio and complete the original signal denoising.Firstly,CEEMD is used to decompose the original signal and reconstruct the noise signal.The ARIMA model approximate particle filter state space model is established,and the state function and observation function of the particle ? wave are obtained,and then the UPF is used for noise reduction processing.The method utilizes particles to reconstruct the original signal,so there is no periodicity requirement for the data,and the method has strong adaptability and is convenient for generalization to other types of data filtering.In the simulation analysis,RMSE is used as the reconstruction evaluation index.Compared with the standard PF,UPF has better performance.In the analysis of the actual bearing signal,the noise reduction signal can better reflect the fault frequency characteristics,which proves that UPF is suitable for Bearing vibration signal noise reduction.The fault classification uses the DBN model,and the output layer uses Softmax as the classification function.The DBN has a supervised fine-tuning phase.The BP algorithm is used to adjust the network parameters,and an improved gradient descent algorithm is proposed to make the learning rate adaptively adjust to meet the requirements of rapid convergence of the network while avoiding the “gradient explosion”.The MNIST handwritten data set is used to conduct preliminary experiments on the feature extraction ability and classification ability of DBN.The influence of parameters such as RBM iteration number,number of nodes and learning rate on feature extraction ability and the influence of network depth and BP finetuning on classification ability are determined.Subsequent use of the Western Reserve University bearing dataset and self-test bearing data to explore the DBN classification effect.Firstly,the original vibration signal and the filtered signal are used as DBN input.The classification accuracy of the latter is above 98%,and the training time is about 350 seconds.To solve the problem of too long training time,10 time domain features are extracted artificially.Three frequency domain features are used to construct the feature matrix.As a DBN input,the training time is shortened to 15 seconds.The classification accuracy of the three data sets is 100%.Only one sample identification error occurs in ten experiments;and PSO-SVM Compared with the three classification models of BPNN and DBM,the results show that the DBN classification model proposed in this paper has improved in both classification accuracy and training time.Finally,the paper summarizes the work of this paper,and considers the inadequacies of this method,and proposes the direction of subsequent improvement.