关键词: 双向传感器信息融合, 认知不确定性, 概率切片累计特征, 双通道并行卷积神经网络, 故障程度不均衡数据集

Abstract: Since the collected bearing vibration signals are susceptible to many uncertainties caused by the environment,a fusion fault diagnosis method based on probabilistic slicing cumulative features of bi-directional sensor information of bearings was used for the qualitative analysis of bearing faults.Probability box theory (P-box) method was applied to model the time domain signals from horizontal and vertical sensors separately to reduce the negative impact of cognitive uncertainty and fully characterize the multi-directional fault vibration information.Further,the probability box was divided into focal element slices to extract the model probability slice cumulative feature matrix.Subsequently,a dual-channel Parallel Convolutional Neural Network (PCNN) was established to fuse the bi-directional feature information by constructing a fusion layer before the fully connected layer of the network.A normalized exponential function was an input to achieve the fault site identification.The analysis results of wheel-to-wheel bearing data of a railroad bureau showed that the proposed method had high accuracy and stability when dealing with imbalanced data sets of fault degree and had certain robustness under different noise conditions.

Key words: bidirectional sensor information fusion, cognitive uncertainty, cumulative feature of probability slice, dual-channel parallel convolution neural network, imbalanced data set of fault degree