基于深度学习的故障检测方法

Fault Detection Method Based on the Deep Belief Network

针对传统故障诊断方法中特征提取技术难度大、故障样本获取困难等问题,在深度学习计算框架下提出了一种半监督训练的故障检测方法,利用深度信念网络中的受限波茨曼机堆栈结构实现了数据高层特征的自动提取,结合支持向量数据描述方法实现了异常数据检测,只需利用正常工况的数据样本进行网络训练和模型拟合,无需故障样本数据,也无需人工干预进行信号特征提取,即能实现对故障数据进行的实时检测和判别;经采用标准轴承实验数据的三组故障数据进行验证,故障识别率达到100%,具有很强的工程应用价值。

This paper presents a semi-supervised fault detection method based on deep learning framework,which utilizes the stack of Restricted Boltzmann Machines in Deep Belief Network to abstract high-level features from original signal data automatically,and apply Support Vector Data Description model to implement fault data detection.This method only needs normal status data as training samples and no any labeled fault data is required.Meanwhile,real-time detection and auto-recognition of fault data can be carried out without expert intervention.As result,the fault recognition rate achieves 100%in treating the standard bearings experiment data,which shows a significant effect and strong application value.