MLPC-CNN: A multi-sensor vibration signal fault diagnosis method under less computing resources

Abstract

This paper proposes a fault diagnosis method for multi-sensor vibration signals under few computing resources, called multi-level feature fusion convolution neural network based on multi-layer pooling classifiers (MLPC-CNN). First, MLPC-CNN introduces the single-sensor-to-single-channel (STSSC) convolution to comprehensively extract features from multi-sensor data grayscale image that integrates all sensor information. This design can adopt more targeted filtering strategies for the samples from different sensors, and avoid the risk of extracting conflicting evidence. Second, MLPC-CNN uses a bypass branch structure based on average pooling layer. This design fuses different levels of signal features extracted by different layers without increasing the network learning parameters, which can extract high-level features while retaining more information from low-dimensional features. Third, MLPC-CNN introduces a multi-layer pooling classifier to replace the fully connected layer in traditional CNN. The pooling layers with different scales are used to achieve multiple functions, which greatly reduces the number of network parameters and the risk of overfitting. The measured data collected by the fault simulation test stand and the bearing fault dataset produced by case western reserve university are used to verify the performance of MLPC-CNN. Experimental results show that MLPC-CNN has reached 100% accuracy on both two datasets. In addition, to explore the fault diagnosis mechanism of MLPC-CNN, this paper uses multiple visualization methods to analyze the function of the convolution kernel in the STSSC convolution layer, the maximum activation feature signal of different convolution channels, and the evolution process of features generated from different fault samples.