Research On Fault Diagnosis Method Of Rolling Bearing Based On Improved DQN Network

Rolling bearings are widely used in many industrial fields,requiring the machine to have high reliability and safety during operation,because the downtime caused by the failure may cause economic losses or even catastrophic accidents.In actual operation,the data distribution of the"normal state"samples and the"fault state"samples of rolling bearings is imbalanced,and the model recognition accuracy rate is not high,and the generalization is not strong.Therefore,studying the fault diagnosis method of rolling bearing under imbalance has strong application value and practical significance for processing actual industrial data.Traditional imbalance diagnosis algorithms perform poorly when the data distribution is extremely imbalanced,and cannot be applied to more general imbalance scenarios.Based on deep reinforcement learning,this paper starts with data preprocessing and model improvement,and proposes an improved deep Q network（DQN）rolling bearing fault diagnosis method to realize the fault diagnosis under the imbalance of rolling bearing data.In terms of data preprocessing,for the problem of less data set samples,the method of data enhancement is proposed to increase the sample set;for the problem that the model is not sensitive to the original one-dimensional vibration signal data,the short-time Fourier transform is proposed to transform the one-dimensional data into two-dimensional image data to improve the fault diagnosis rate of the model.In terms of model establishment,fault diagnosis is simulated as a sequential decision-making process in deep reinforcement learning.Aiming at the problem of insufficient pertinence of the reward function in the original DQN,a method of combining k-means and spatial distance to construct the reward function in the DQN network is proposed:taking the imbalance ratio as the benchmark,and taking the distance from the sample to the center in the k-means algorithm as the bias of the reward value to construct a personalized reward function for the training set.Aiming at the problem of insufficient feature extraction of the original DQN network,it is proposed to build the network part of the DQN model through the deep residual shrinkage module,and at the same time,the output layer part is improved to improve the stability of the model and the accuracy of diagnosis.And through theoretical analysis and multiple sets of comparative experiments,data imbalance and variable load issues are studied deeply.Experiments show that the improved DQN model can well solve the problem of imbalanced data distribution between normal and faulty vibration data,and the G-mean_total score under imbalanced and variable load conditions is above 0.982,which is excellent.When the data is extremely imbalanced,the accuracy rate reaches97%～99%,and some reaches 100%.Compared with the traditional imbalance classification method,the accuracy rate is increased by 5%～8%.