Predictive Method for Machinery Fault Detection Using Deep Learning and Vibration Images

Abstract

Purpose: This study investigates fault identification in rotating machinery using different fault classes from an open-access database. The research aims to evaluate and compare machine learning approaches for fault classification and to assess the effectiveness of a novel vibration-image-based method, particularly for incipient fault detection. Methods: Several machine learning approaches were evaluated. Convolutional Neural Networks (CNNs) were first applied directly to unprocessed vibration signals in both time and frequency domains. Statistical feature extraction methods were also investigated. Subsequently, a vibration image technique using frequency-domain signals was proposed and tested. The developed models were further evaluated using a database containing low-intensity (incipient) faults. Results: CNN-based classification using frequency-domain acceleration signals achieved an overall accuracy of 89.4%, while the statistical feature approach reached 60.5%. The proposed vibration image method yielded a classification accuracy of 99.4% in the frequency domain, outperforming the corresponding time-domain implementation (97.0%). For incipient fault detection, the proposed approach achieved satisfactory accuracies of 94.0% and 84.3% in the frequency and time domains, respectively. Conclusion: The proposed vibration image technique demonstrated excellent fault classification performance, particularly when applied to frequency-domain signals. The results indicate its strong potential for predictive maintenance applications, online machinery monitoring, Industry 4.0 environments, and Internet of Things frameworks due to its high accuracy and favorable computational cost–benefit ratio.