BEARING FAULT DIAGNOSIS USING DEEP LEARNING NEURAL NETWORKS WITH INPUT PROCESSING

The roller bearings are widely used in aviation cargo systems, engines, agriculture, heavy equipment and machinery, solar panels, medical equipment, automobile industry, powerhouses, and many others. Bearing faults during the operation process will result in downtime, economic loss, and even human injury. To prevent these from happening, rolling bearing fault diagnosis has become a mature discipline. Deep learning networks have been known as effective methods for bearing fault diagnoses. Deep learning neural networks such as the convolutional neural network (CNN) use the images as inputs. In contrast, the others, such as long-short term memory (LSTM), may apply data sequences as inputs.
This thesis research work focuses on performance evaluations of deep learning networks according to the classification accuracy by utilizing various signal transforms to form the network inputs. CNN and LSTM are adopted as our deep learning network structures. Besides raw data, the algorithms for processing input signals include short-time Fourier transform (STFT), Cepstrum, wavelet packet transform (WPT), and empirical mode decomposition (EMD). In addition, this paper also applies three commonly used machine learning algorithms for comparison, namely K nearest neighbor (KNN), support vector machine (SVM), and random forest (RF). Finally, a one-dimensional CNN structure is designed and implemented.
Our simulations validate the effectiveness for each network input formulation based on the Case Western Reserve University (CWRU) bearing dataset.