Data Segmentation and Augmentation Methods Based on Raw Data Using Deep Neural Networks Approach for Rotating Machinery Fault Diagnosis

Abstract

Intelligent fault diagnosis has been widely used for mechanical fault diagnosis. Most intelligent diagnostic methods extract fault features from the frequency domain or other domains, instead of from raw data. Given that converting raw data to other domains will cause a partial loss of information, this study is based on feature extraction from raw vibration data. However, there are some difficulties in extracting features from time domain data: 1) raw vibration data lacks regularity compared to the frequency domain data signal, making it difficult to extract features from it; 2) the number of labeled samples is so small that the deep neural networks can easily be over-fitted, making their generalization ability excessively poor; and 3) too many parameters in the neural networks need to be adjusted, such as learning rate, and the convergence speed is slow. To overcome the aforementioned difficulties, this study proposes the following three methods: 1) a sample segmentation method to effectively improve the feature extraction from raw data; 2) a data augmentation method for raw vibration data, which can increase the number of samples; and 3) the use of the scaled conjugate gradient (SCG) algorithm in the networks to quickly learn sample features without other parameters such as learning rate. Bearing and rotor datasets are used to validate the performance of the proposed methods. The results indicate that the methods obtain superior performance for feature learning and classification compared with the existing ones in the field of induction rotating machinery fault diagnosis.

Keywords

• Data segmentation
• data augmentation
• raw data
• deep neural networks
• fault diagnosis