Fault Diagnosis Of Roller Bearing Using Deep Learning

Roller bearing fault diagnosis plays a key role in machinery operation.To Mentioning a few,roller bearing fault diagnosis is important in ensuring safety is guaranteed in working environment,prevent breakdown of machineries,eliminate other risks associated with roller bearing components failure and most importantly helps the engineers and other authorities in decision making for timely preventive maintenance.Therefore,there is a strong need to invest more in research in the field of roller bearing fault diagnosis to improve the reliability of roller bearing fault diagnosis techniques.Fault diagnosis process can be summarized into four phases: vibration signal acquisition phase,feature extraction phase where important features are extracted from the signal,dimensionality reduction phase that reduces high dimensional features to a manageable size and finally classification to diagnose the condition of the component.The methods of classification in fault diagnosis of roller bearings have been categorized into two: statistical and unsupervised deep learning.The deep learning methods i.e.stacked autoencoder and stacked denoising autoencoder bear more advantages such as automatic feature extraction and high performance therefore have rendered them more preferred in selection of suitable methods for roller bearing fault diagnosis.In this thesis work,ensemble empirical mode decomposition(EEMD)has been used to extract features in statistical model,shallow learning,before being fed to singular value decomposition(SVD)for dimensionality reduction of the features and finally to support vector machines classifier(SVM)for condition recognition of the bearing.In the case of deep learning,simple autoencoder(AE),stacked autoencoder(SAE)and stacked denoising autoencoder(SDAE)are objectively employed to diagnose fault.These deep learning methods are purely unsupervised.Importantly,these methods do not require preprocessing of input data instead raw data is fed directly into the model followed by data training,adding noise in the case of stacked denoising autoencoder,then train every layer of autoencoder is trained and finally introduced to support vector machine for classification process.The outcome of both statistical model and deep learning model are then compared in terms of performance,stability and efficiency.