改进一维深度卷积神经网络的轴承性能退化指标构建方法

Construction Method of Bearing Performance Degradation Index Based on One-dimensional Deep Convolution Neural Network

为更好构建能够表征轴承退化过程的性能退化指标,提出一种改进一维深度卷积神经网络轴承性能退化指标的构建方法。首先,构建一维深度卷积神经网络,利用其对原始时域信号自适应提取特征优势,以深度挖掘全寿命时域信号的退化特征;其次,设计一种组合损失函数,在均方误差函数上引入退化特征相邻点正负微分累积值,使得网络在训练过程中退化特征相邻点正微分值不断增大,负微分值不断减小,以提高性能退化指标单调性;最后,通过全连接层将高维特征转化为低维特征,实现性能退化指标构建。通过在公开和实测的数据集上进行实验,验证了所提方法的有效性和可行性。

In order to construct a performance degradation index that can better characterize the bearing degradation process,a constrion method concerning improved one-dimensional deep convolutional neural network was proposed.Firstly,a onedimensional deep convolutional neural network was built to extract features adaptively from original time domain signals,and the degradation characteristics of full-life time domain signals were described in detail.Secondly,a combined loss function was designed with combing the cumulative value of the positive and negative differential of degradation feature at adjacent points based on the mean square error function.Consequently,the positive differential value increases while the negative decreases in the training process,leading to the improved monotonicity of performance degradation index.Finally,the high-dimensional features are transformed into low-dimensional ones through the full-connection layer to realize the construction of performance degradation indicators.The effectiveness and feasibility of the proposed method were verified by conducting experiments on public and measured data sets.