基于深度度量学习的电机故障诊断

Motor Fault Diagnosis Based on Deep Metric Learning

深度学习以其强大的自适应特征提取和分类能力在机械大数据处理方面取得了丰硕的成果,由于电机结构的复杂性,其信号表现出的非平稳、非线性和复杂多样等特点,使得传统分类方法中的Softmax分类器+交叉熵损失函数对电机故障诊断力不从心。根据电机信号非平稳、数据量大等特点,结合短时傅里叶变换(STFT)与深度学习中的卷积神经网络(CNN)算法和Triplet Loss三元组思想,提出了深度度量学习电机故障诊断方法。该方法能将电机故障信号转换成时频谱图,同时构建CNN,将预处理后的样本用于CNN的训练,采用Triplet Loss作为损失函数度量故障数据高维特征间的距离,并结合标签有监督地微调整个网络,从而实现准确的电机故障诊断。实验表明该方法在处理复杂数据时能够度量特征在高维空间中的距离,高效完成故障诊断任务,弥补了交叉熵函数的不足。

Deep learning has made great achievements in mechanical big data processing due to its strong adaptive feature extraction and classification capabilities. Because of the complexity of the motor structure,its signals show the characteristics of non-stationary,non-linear and complex diversity,which makes the Softmax +cross entropy loss function in the traditional classification method inadequate for motor fault diagnosis. According to the characteristics of non-stationary motor signals and large data volume,combining with short-time Fourier transform( STFT) and convolutional neural networks( CNN) algorithm in deep learning and Triplet Loss triple-tuple idea,a deep measurement learning motor fault diagnosis method is proposed. In this method,STFT is used to convert motor fault signal into time-frequency spectrum,and CNN is constructed at the same time.Preprocessed samples are used to train CNN. Triplet Loss is used as loss function to measure the distance between high-dimensional features of the fault data,and a supervised micro-adjustment network is used to realize accurate motor fault diagnosis. Experiments show that the method can measure the distance of features in highdimensional space when processing complex data,and efficiently complete fault diagnosis task,and make up for the deficiency of the cross entropy function.