Deep Reinforcement Learning Networks For Gearbox Fault Diagnosis

The gearbox is a key moving part of the transmission system of mechanical equipment.Its failures may cause great economic losses or casualties.Gearbox fault diagnosis is a necessary measure to ensure the safe and reliable operation of equipment.At present,it is practical and effective to use vibration signals with high fault sensitivity for fault diagnosis.Traditional vibration-derived methods require manual denoising,feature extraction and selection,and fault identification is realized by fault frequency or shallow machine learning algorithms.The diagnosis process is tedious and lacks intelligence.Especially in the face of the diversity of working conditions,noise interference,and massive data,manual feature extraction becomes increasingly impractical.The accuracy and adaptability of such methods can not meet the needs of practical fault diagnosis.Recently,deep learning(DL)methods such as convolutional neural networks(CNN)can adaptively extract discriminative features from high-dimensional states by powerful feature expression ability and have significant effects on fault diagnosis.However,traditional DL methods lack decision-making ability,and the accuracy,interpretability,and universality of fault diagnosis under complex conditions need to be improved.As an emerging DL method,deep reinforcement learning(DRL)combines both the perception ability of DL and the decision ability of reinforcement learning(RL)and has the general intelligence to solve complex problems.It provides a new idea to solve fault diagnosis of the gearbox.Thus,this thesis takes gearbox vibration signals as the research object,and DRL as the method to study gearbox fault diagnosis under scenarios such as data imbalance,variable conditions,and strong noise.The main works are as follows:(1)To improve the accuracy and universality of fault diagnosis,a deep convolutional Qlearning RL network-based gearbox end-to-end fault diagnosis method is proposed.Based on the deep Q-learning RL algorithm(DQN),a CNN with powerful feature extraction ability is used to design the structure of the DQN algorithm.Through the classification Markov decision process,the model can automatically learn discriminative features and diagnosis strategies and finally diagnose gearbox fault types.Experimental results show the effectiveness of this method and provide a research basis for gearbox fault diagnosis under complex conditions.(2)Aiming at the problem of low generalization of fault diagnosis caused by multiple conditions and uneven data distribution,a multiscale feature deep attentional RL network-based gearbox imbalanced fault diagnosis method was proposed.Based on the DQN algorithm,multiscale convolution and channel attention are used to adaptively capture low-level multi-scale features from vibration signals,and suppress fault-independent redundant features.Then multiscale sparse structure and residual network are utilized to learn deep discriminative features,which enhances the stability of feature learning under multiple conditions.The class deviation degree is defined to construct the reward strategy,which makes the model pay attention to the minority classes automatically in interactive learning,and learns effectively imbalanced fault diagnosis strategy.Experimental results show that this method has high accuracy and flexibility in gearbox imbalanced fault diagnosis.(3)For the poor robustness of fault features under variable conditions,as well as the low diagnostic accuracy caused by domain differences and small sample size,a health adaptivity parameter transfer DRL network-based gearbox fault diagnosis method was proposed.Based on the dueling DQN algorithm,a health self-adaption representation layer was constructed by using multi-scale convolutions to give the model the ability to extract time-scale features,and a multi-scale residual network was designed by using dilated convolution to extract rich deep features while keeping the model lightweight,to accurately realize fault diagnosis under variable-speed source domain.A parameter transfer strategy was used to fine-tune the welltrained source model to realize end-to-end gearbox fault diagnosis under variable conditions and limited samples.Experimental results show that this method can improve the accuracy of gearbox fault diagnosis under variable working conditions.(4)Aiming at the problem of low robustness and interpretability of fault diagnosis in strong noise,a gearbox fault diagnosis method based on a deep temporal shrinkage interpretable RL network was proposed.Based on the dueling DQN algorithm,an interpretable SincNet convolution was built by incorporating the bandpass filter knowledge to enhance the robustness and interpretability of the output features of the model’s first layer.Considering the temporal correlation of gearbox vibration signals,multi-scale temporal convolution is used to capture the rich intrinsic time-dependent features.Then,residual learning and soft thresholding were used to build a multi-scale temporal-shrinking residual network to adaptively reduce the noise in feature maps and improve the robustness of deep-level features.Finally,gearbox fault diagnosis was realized by the interactive learning between the model and the environment simulation.Experiments show that this method has high noise immunity,accuracy,and interpretability.The proposed method effectively improves the accuracy and universality of the gearbox in end-to-end feature learning and recognition,as well as the fault diagnosis in the scenarios of class imbalance,variable conditions and strong noise.It not only expands the application scope of deep reinforcement learning algorithms,but also provides a new idea for the intelligent fault diagnosis of gearbox in complex environments.