Vibration analysis process based on spectrogram using gradient class activation map with selection process of CNN model and feature layer

Abstract

This paper presents a vibration analysis process using a convolutional neural network (CNN) model and a gradient class activation map (Grad-CAM), based on a model and feature layer selection process. The main problem associated with conventional bearing fault diagnosis is the need to calculate defect frequencies, which requires that all bearing-related information should be known beforehand. The extraction of information regarding the mechanical equipment and the subsequent generation of a database requires significant time and cost. The proposed methodology can detect frequencies without any prior information by visualizing the activation area of the CNN model. The proposed approach helps overcome the drawbacks of the conventional diagnosis using Grad-CAM, and it indicates the rationale of classification, thus explaining the cause of vibration data for a defect diagnosis. The paper presents a process that visualizes and analyzes the activated region through the conversion of vibration signals into spectrogram images and use of CNN models. The proposed method further serves as a process of selecting the appropriate CNN model and its feature layers among several models. The vibration data were collected from a motor using an accelerometer and IoT module. The feasibility of the proposed methodology was verified using the recorded data. Spectrogram images for five situations (normal, inner fault, outer fault, ball fault, and cage fault) were generated from the acquired data. A VGG-19 model with an accuracy of 99.92% was selected based on transfer learning of three CNN models (ResNet50, VGG-16, VGG-19) using these spectrogram images. Finally, the feature layer of dimension 28 × 28 is selected for the Grad-CAM activation model. The visualization model was analyzed and compared to conventional defect frequency analysis methods. The results analyzed through the activation area near the defect frequency region of the conventional frequency analysis show that the use of deep learning in the proposed model helps evaluate the frequency intervals required to determine defects without any prior information regarding mechanical specifications.