Adaptive information fusion for multi-fault diagnosis using multi-stream convolutional neural network

Abstract

The occurrence of multiple faults simultaneously in industrial machines is very complex. If maintenance operation is not carried out promptly, it leads to financial losses. Moreover, the unavailability of the multi-fault dataset imposes a problem for automated diagnosis model development. The model developed with a single sensor could not work with multiple sensor data and needs a separate fusion process. To tackle these problems, a robust methodology involving an Adaptive Information Fusion (AIF) process along with a Multi-Stream Convolutional Neural Network (MSCNN) is developed in this work. Initially, a large acoustics-based dataset is made by experimentally simulating 225 multi-fault conditions of motor, mechanical unbalance, system bearing and shaft misalignment at varying speeds. The acoustic signatures show distinct features with different fault conditions, passing through the AIF process for generating a time–frequency spectrum dataset using Wavelet Synchrosqueezed Transform (WSST). The spectrums are eventually used for training the optimised MSCNN model. A maximum overall classification accuracy of 97.19% is achieved. The performance of the methodology is rigorously tested under various background noise, data imbalance, speed conditions and with multi-sensor data fusion. The proposed method is useful for single as well as multiple-sensor data and can be helpful for incipient fault diagnosis in industries.