基于深度宽卷积Q网络的行星齿轮箱故障智能诊断方法

Intelligent fault diagnosis for the planetary gearbox based on the deep wide convolution Q network

针对行星齿轮箱故障诊断常依赖较强的专业知识,诊断模型通用性差的问题,基于深度强化学习,提出一种深度宽卷积Q网络的行星齿轮箱故障智能诊断方法。首先将行星齿轮箱的故障诊断分解为序贯决策问题,采用分类马尔科夫决策过程进行描述,并建立故障诊断模拟环境;其次设计深度宽卷积神经网络作为深度Q网络模型中的动作值网络,增强对环境状态的感知能力;最后模型通过与环境间的不断交互,并依据环境反馈的奖励,自主学习最佳诊断策略,从而完成行星齿轮箱的状态辨识。试验及案例结果表明:该方法能够在多个工况下均可有效、准确地实现行星齿轮箱的智能诊断,诊断准确率均超过99%,增强了诊断模型的泛化性和通用性。

The fault diagnosis of the planetary gearbox often relies on strong professional knowledge, and the universality of the diagnosis model is poor. Based on deep reinforcement learning, an intelligent fault diagnosis method of the planetary gearbox using the deep wide convolution Q network is proposed. Firstly, fault diagnosis of the planetary gearbox is resolved into a sequential decision problem, which is described by the classification Markov decision process. The fault diagnosis simulation environment is established. Secondly, a deep wide convolutional neural network is designed as an action-value network in the deep Q network model to enhance the perception ability of the environmental state. Finally, the model learns the best diagnostic policy autonomously by interacting with the environment and according to the reward of the environment. In this way, the state identification of the planetary gearbox can be achieved. Experiment and case results show that this method can effectively and accurately realize the intelligent diagnosis of the planetary gearbox under multiple working conditions. The diagnostic accuracy is more than 99%, which enhances the generalization and universality of the diagnosis model.