Gearbox Multiple Faults Diagnosis under Stationary and Non-Stationary Operating Conditions Using Convolutional Neural Networks

Abstract

High accuracy in gearbox fault diagnosis is of paramount importance for keeping industrial systems safe and working normally. Concerning various single or multiple faults diagnosis using classical machine learning algorithms, the feature extraction and selection methods are time-consuming and labor-intensive processes requiring expert knowledge of the relevant features related to the system. To mitigate this problem, a deep learning convolutional neural network (CNN) is proposed which enables automatic feature learning from the time-frequency domain image representations data input. The proposed approach employs a CNN model which uses max-pooling and batch normalization between each convolution for training acceleration and reduction of generalization error. The proposed model performance in multiple faults diagnosis is evaluated using gearbox vibration data obtained under stationary operating conditions. The proposed model classification performance is also evaluated using engineered non-stationary operating vibrations data sets. To achieve this, data sets were engineered using gearbox stationary operating conditions vibration data obtained using five operating speeds (30 Hz, 35 Hz, 40 Hz, 45 Hz, and 50 Hz) under low and high load conditions. The non-stationary operating conditions data sets were developed based on three operational conditions namely; constant operating speed and variable load, variable operating speed and constant load, and variable operating speed and variable load. This methodology was applied to PHM2009 gearbox vibration data sets which consist of multiple component faults. Time domain, frequency domain, and time-frequency analysis methods (short-time Fourier transform (STFT), continuous wavelet transform (CWT), empirical mode decomposition (EMD), Wigner Ville distribution (WVD), and wavelet synchrosqeezed transform (WSST)), were investigated for their effectiveness in multiple faults diagnosis. The proposed CNN architecture exhibited high classification performance as high as 99.9% under stationary operating conditions in multiple faults diagnosis as compared to other architectures. The use of image input to the CNN model gave better performance compared to feature vector input. The application of the developed diagnostic model for multiple fault diagnosis under non-stationary operating conditions gave satisfactory classification performance as high as 86.5% using scalogram images. This work provides the possibility that stationary operating conditions-based diagnostic models, can be deployed for multiple faults diagnosis on industrial equipment operating under non-stationary conditions.