电子测量与仪器学报

刘小峰,徐全桂,金燕,柏林.噪声干扰环境下的深度强化学习故障诊断方法[J].电子测量与仪器学报,2024,38(12):145-154

噪声干扰环境下的深度强化学习故障诊断方法

Deep reinforcement learning fault diagnosis method undernoisy interference environment

DOI：

中文关键词: 故障诊断深度残差网络阈值软化深度Q学习竞争网络经验优先回放

英文关键词:fault diagnosis residual network efficient channel attention deep Q-learning competition network experience priority replay

基金项目:国家科技重大专项（J2019-IV-0001-0068)项目资助

作者	单位
刘小峰	重庆大学高端装备机械传动全国重点实验室重庆400044
徐全桂	重庆大学高端装备机械传动全国重点实验室重庆400044
金燕	重庆大学高端装备机械传动全国重点实验室重庆400044
柏林	重庆大学高端装备机械传动全国重点实验室重庆400044

Author	Institution
Liu Xiaofeng	State Key Laboratory of Mechanical Transmission for Advanced Equipment，Chongqing University, Chongqing 400044,China
Xu Quangui	State Key Laboratory of Mechanical Transmission for Advanced Equipment，Chongqing University, Chongqing 400044,China
Jin Yan	State Key Laboratory of Mechanical Transmission for Advanced Equipment，Chongqing University, Chongqing 400044,China
Bo Lin	State Key Laboratory of Mechanical Transmission for Advanced Equipment，Chongqing University, Chongqing 400044,China

摘要点击次数: 300

全文下载次数: 221

中文摘要:

针对深度强化学习在噪声干扰环境下故障诊断鲁棒性差问题，提出了一种噪声干扰环境自适应的强化学习故障诊断方法。该方法以高效通道注意力机制深度残差收缩网络为Q网络基本模型，避免Q网络结构复杂导致的梯度消失现象。采用高效通道注意力机制对深度残差收缩网络中软化阈值进行自适应调整，并在残差收缩单元的卷积层引入了膨胀卷积，以获取噪声环境下的不同尺度的故障特征信息，同时采用指数线性单元SELU作为激活函数，进一步提升网络对噪声的鲁棒性。设计了基于信噪比的量化奖励函数，结合双重Q网络竞争学习机制与优先经验回放机制方法，进行智能体的自主学习，生成智能体的最优诊断策略，并运用于干扰环境下的设备故障状态识别。实例分析结果表明，采用所提方法对轴承与齿轮箱故障的识别准确率分别能到达98.13%和93.45%，且对不同强度噪声具有较好的鲁棒性与环境自适应性。

英文摘要:

Aiming at the poor robustness of deep reinforcement learning for fault diagnosis in strong noise interference environments, a reinforcement learning fault diagnosis method with noise interference environment adaptation is proposed. The efficient channel attention mechanism based deep residual shrinkage network (ECA-DRSN) is taken as the basic framework of Q-network to avoid the phenomenon of gradient vanishing caused by the complex structure of Q-network. In the ECA-DRSN，the efficient channel attention mechanism is used to adaptively adjust the softening threshold，and the dilated convolution is introduced in the convolution layer of the residual shrinkage unit to obtain the fault characteristics in different scales under the noise environment. Meanwhile, the exponential linear unit is used as the activation function to further enhance the noise robustness. A quantized reward function based on signal-to-noise ratio is designed to stimulate self-directed exploratory learning of Agent. Combining the dueling Q network learning mechanism with the prioritized experience replay mechanism, the optimal diagnostic strategy of agent is generated and applied to identify the equipment fault states under noise interference environments. Example analysis results show that the recognition accuracy of bearing and gearbox faults using the method of this paper can reach 98.13% and 93.45%, respectively, and has better robustness to different intensity noise and adaptability to the environment.

查看全文查看/发表评论下载PDF阅读器