可解释机械智能诊断技术的研究进展与挑战

Research Progress and Challenges of Interpretable Mechanical Intelligent Diagnosis

基于深度神经网络的机械智能诊断方法近年来得到快速发展,各式各样的模型方法相继涌现。然而,此类方法的性能彰显主要是在实验室环境,实际工业场景鲜有应用。究其原因主要在于模型内部非线性变换复杂,特征提取机理不清晰,进而引发使用者对决策的信任问题。特别是对于一些关键设备,倘若不能提前知晓模型得到诊断结论的原由,贸然采取措施会隐藏相当大的风险。近年来,可解释机械智能故障诊断研究越来越多地被关注,一些学者已开始对该领域进行探索,并初见成效。为促进深化研究,推动该领域发展,首先对深度神经网络可解释性研究的各种范式进行分类和讨论,然后详述了可解释性机械智能诊断的当前发展,最后讨论了现有的挑战以及未来的研究方向。

The intelligent diagnosis method of machinery based on deep neural networks has developed rapidly in recent years,and various model methods have emerged one after another.However,the performance of such methods is mainly in the laboratory environment,and there are few applications in actual industrial scenarios.The main reason is that the nonlinear transformation inside the model is quite complex,and the feature extraction mechanism is difficult to understand,leading to the user's untrust in decision-making.Especially for some key equipment,if the reason why the model gets the diagnosis conclusion cannot be known in advance,taking measures rashly will hide considerable risks.In light of this,more and more attention has been paid to the interpretability of intelligent fault diagnosis recently,and some scholars have reached some preliminary conclusions.In order to deepen research and promote the development of the field,various paradigms of interpretability of deep neural networks are categorized and discussed,then,current developments in interpretable intelligent fault diagnosis of machinery are detailed,and finally,existing challenges and future research directions are discussed and summarized.