基于平移不变CNN的机械故障诊断研究

Machinery fault diagnosis based on shift invariant CNN

常规机械故障诊断方法需要信号预处理、特征提取、特征选择、模式识别等多个步骤,过程复杂,通用性差。卷积神经网络(Convolutional Neural Network, CNN)是一种自学习性能好、抗干扰能力强的深度神经网络。为了简化步骤、提高效率,将CNN引入到机械故障诊断,直接使用传感器测得的原始数据进行故障识别。由于机械振动信号的特征具有典型的时移性,CNN需要大量数据才能自我学习到这种特性。结合故障信号的冲击特点和CNN的不足,提出了权值求和和大尺度最大值池化策略,有效解决了特征的平移不变性,增强了小样本时的泛化能力。通过对单点和多点故障的轴承进行诊断,验证了平移不变CNN的有效性。与常规故障诊断方法和其他机器学习算法对比显示,平移不变CNN不仅准确率高,而且使用简单,为故障诊断提供了一种新的途径。

Traditional machinery fault diagnostic method usually requires multiple processing steps including signal preprocessing,feature extraction,feature selection and pattern recognition to lead to complicated process and poor universal application.Convolutional neural network (CNN) is a deep neural network with strong self-learning and anti-interference abilities.To simplify steps and improve efficiency,here,a novel machinery fault diagnosis method based on CNN was proposed directly utilizing raw data measured with sensors to identify faults.Due to features of mechanical vibration signals having a typical time shift property,CNN needed a lot of data to self-learn this property.Combining fault signals’ impact feature and shortcomings of CNN,the strategies of weight summing and large-scale maximum value pooling were proposed to effectively solve CNN’s shift invariance to enhance small samples’ generalization ability.Through diagnosing single point fault and multi-point faults of bearings,the effectiveness of the shift invariant CNN was verified.Compared with traditional fault diagnosis approaches and other machine learning algorithms,it was shown that the shift invariant CNN not only has a higher accuracy but also is simple to use,and it provides a new way for fault diagnosis.