Effect of Time-Frequency Representations for Fault Classification of Rolling Bearing in Noisy Conditions Using Deep Learning

Intelligent fault diagnosis techniques such as deep learning methods have gained immense attention recently for identifying bearing faults with higher accuracy and adaptability based on vibration analysis. This work proposes a fault diagnosis method for rolling bearings under extreme noise conditions based on the time-frequency representation (TFR) of signals with deep learning techniques. The bearing signals are masked with -5dB and -10dB of white Gaussian noise to create a noisy environment. Short-time Fourier transforms (STFT), Continuous wavelet transforms, and Gabor transforms (GT) are utilized to obtain spectrogram, magnitude scalogram, and constant-Q transform images to visualize the time-frequency relationship of the signal from one-dimensional vibration signals. These TFR images are directly given as input to VGG16-CNN deep learning architecture to classify the bearing faults. The effectiveness of each TFR is measured and compared based on classification accuracy. This work also studies the effect of TFR with and without overlapping the segmented signals. The case western reserve university (CWRU) standard-bearing dataset is used in this work and has achieved a satisfactory result. The result suggests that the magnitude scalogram of the vibration signal is an effective TFR that works efficiently with deep learning for bearing fault classification under different conditions, such as original signal with no added noise, -5dB and -10 dB of added noise, with an accuracy of 99.42%, 90.2 % and 85.03%, respectively when trained with 80% of the sample with 75% of the overlapping index.