基于改进DQN燃气轮机转子故障诊断方法  

Fault Diagnosis Method for Gas Turbine Rotors Based on Improved DQN

作　　者：崔英杰 王红军[1,2,3] 张顺利[4] 王星河 CUI Yingjie;WANG Hongjun;ZHANG Shunli;WANG Xinghe(School of Mechanical and Electrical Engineering,Beijing Information Science and Technology University,Beijing 100192,China;Beijing International Science and Technology Cooperation Base of High-level Equipment Intelligent Perception and Control,BISTU,Beijing 100192,China;Beijing Key Laboratory of Measurement and Control of Mechanical and Electrical Systems,BISTU,Beijing 100192,China;AECC Xi'an Aero-Engine Group Co.,Ltd.,Xi′an 710021,China)

机构地区：[1]北京信息科技大学机电工程学院,北京100192 [2]北京信息科技大学高端装备智能感知与控制北京市国际科技合作基地,北京100192 [3]北京信息科技大学机电系统测控北京市重点实验室,北京100192 [4]中国航发西安航空发动机集团有限公司设计所,西安710021

出　　处：《噪声与振动控制》2023年第4期109-115,199,共8页Noise and Vibration Control

基　　金：国家自然科学基金资助项目(51975058)。

摘　　要：燃气轮机转子系统作为燃气轮机关键部件,由于难以获取敏感故障特征导致故障诊断精度不高,影响设备的安全服役。针对以上问题,提出一种改进深度Q网络(DQN)深度强化学习燃气轮机转子系统故障诊断方法。首先,以采集的一维工况原始振动信号为输入,该DQN模型的环境状态采用故障样本集组成,转子故障类型为当前模型输入的动作集合;然后,DQN模型的智能体使用一维宽卷积神经网络(WDCNN)拟合得到Q网络,并使用ε-贪婪策略做出决策动作,反馈奖励和下一状态并存储到经验池内;智能体内采用时间差分误差(TD-error)优先经验回放,使得算法更加稳定和训练收敛;智能体与环境不断交互决策出最大奖励,输出最优策略故障诊断结果。将该模型应用于西储大学轴承数据集与燃气轮机试车台数据集中,分别达到99.2%与98.7%的准确率,可以用于快速有效地进行故障诊断。结果表明该改进DQN模型具有较高的故障诊断准确性与通用性。To solve the problem of low fault diagnosis accuracy of gas turbine rotor systems,a fault diagnosis method based on improved Deep Q-Network(DQN)for deep reinforcement learning is proposed.In this method,the onedimensional original vibration signal is collected and used as the input of the DQN model,the fault sample set is applied to form the environment status of the DQN model,and the fault type of the rotor is considered to be the action set to input into the model.Then,the intelligent agent of the model uses the one-dimensional wide convolutional neural network(WDCNN)to fit a Q network,and uses anε-greedy strategy to make the decision actions,feedback rewards and next states,and save them into the experience pool;the temporal difference error(TD-error)is used in the intelligent agent to give a priority for experience playback,which makes the algorithm more stable and training convergent;successive exchange between the intelligent agent and the environment can determine the maximum reward and yield the output of the optimal fault diagnosis result.The model is applied to analyze the bearing data set of Western Reserve University and the gas turbine test-bench data set,and the accuracy rates are 99.2%and 98.7%,respectively,and the training time is reduced by half of the original one.The improved DQN model can quickly and effectively diagnose faults,and has high fault diagnosis accuracy and versatility.

关 键 词：故障诊断 燃气轮机转子 深度强化学习 DQN 

分 类 号：TK47[动力工程及工程热物理—动力机械及工程] TP206.3[自动化与计算机技术—检测技术与自动化装置]