改进TCPN变工况轴承故障诊断方法

Improved TCPN Method for Fault Diagnosing of Bearings under Variable Working-conditions

针对变工况条件下传统方法仅提取局部特征导致准确率低以及样本维度过高导致训练耗时巨大等问题,提出一种改进的时域卷积池化网络模型(Temporal Convolutional Pooling Neural Network,简称TCPN)。首先,将原始轴承振动信号经过短时傅里叶变换得到二维时频矩阵,再将二维时频矩阵平铺为一维时频序列,并取绝对值;其次,虽然膨胀卷积可以大幅度扩大感受野,但是对于轴承振动信号等高维特征仍需要较深的网络结构,因此对膨胀卷积进行结构优化,用于挖掘全局特征,同时降低模型复杂程度。再次,为模糊不同工况下相同故障的分布差异,对残差块进行结构优化,使用小卷积核提取局部特征,以拼接的方式与膨胀卷积进行特征融合,兼顾轴承振动信号中的全局特征与局部特征;最后,为了解决训练样本维度太高导致的训练成本过高的问题,对原始数据进行降采样研究,在保持较高准确率的情况下大量节省了训练时间。将所提方法与传统卷积神经网络及时域卷积神经网络(Temporal convolutional neural network,简称TCN)相比,实验结果表明,提出的模型准确率提高约5%,模型训练耗时降低约30%,并且收敛速度更快,训练模型的迭代次数更少,具有很强的鲁棒性。

Aiming at the problem that traditional methods can only extract local features under variable working conditions,which leads to low accuracy,high sample dimensions and long training time,an improved time-domain convolutional pooling neural network(TCPN)model is proposed.Firstly,the short-time Fourier transformation is used to the original bearing vibration signal to obtain a two-dimensional time-frequency matrix,which is then laid into a onedimensional time-frequency sequence,and its absolute value is taken.Secondly,although the expansion convolution can greatly expand the receptive field,but for high-dimensional features of bearing vibration signals,a deeper network structure is still required.Therefore,the structure of the expanded convolution is optimized to detect the global features and reduce the complexity of the model.Thirdly,in order to blur the distribution difference of the same fault under different working conditions,the residual block is optimized,the small convolution kernel is used to extract the local features.The feature fusion is performed with the dilatational convolution in a splicing manner,and the global bearing vibration signal features and local features are considered.Finally,in order to solve the problem of high training costs caused by high training sample dimensions,the original data is analyzed by using down-sampling method,which saves a lot of training time while maintaining a high accuracy rate.The proposed method is compared with the traditional convolutional neural network and time domain convolutional neural network.The experimental results show that the accuracy of the proposed model is raised by about 5%,the model training time is reduced by about 30%,and the convergence speed is faster,the training model has fewer iterations times and strong robustness.