Faults' Diagnosis of Time-Varying Rotational Speed Machinery Based on Vibration and Acoustic Signals Features Extraction, and Machine Learning Methods

Purpose Many industrial fields count, these days, on data collection as a backup source for machinery condition monitoring to increase their equipment safety, reduce maintenance costs, avoid eventual damages and minimize operational downtime. In electromechanical systems, generated signals such as vibrations, acoustics, electrical current, and thermal images are considered a sensitive and a loaded source of information about the equipment's health condition. Despite its acquisition, these data had shown a direct dependence to their nature and need further steps to be adequately analyzed and ready to be used as a decision making reliable support. Hence, signal processing methods are usually used to identify the current state of the system. However, these techniques had shown some limitations, especially for varying rotating speed, noisy environment where filtering is required or when defects are in their early stages. These constrains led to the research community to consider further developed methods to deal with such issues. Methods Feature extraction based on the maximal overlap discrete wavelet packet transform MODWPT is known for its ability to decompose the raw signal into several intrinsic ones on which features will be extracted. That is why it is employed as a first expertise step on both vibration and acoustic signals. Machine learning techniques (KNN, DT, ET, RF, and SVM) are then implemented to diagnose and differentiate between different types of faults in time-varying rotational speed machines and to compare the outcomes of vibration and acoustic signals in this matter. Results The obtained results demonstrated the superiority of vibration signal's outcome over acoustic signal. Machine learning classifiers had yielded higher stability (STD = 0) and accuracy (100%) for ET and RF when fed with vibration signal's features. However, an STD of 0.005976 and an accuracy of 99.9433% were the best obtained results from SVM when fed with acoustic signal's features. Conclusion Machine learning methods had proven their efficiency in time-varying machines fault diagnosis when taking vibration and acoustic signals extracted features as inputs. However, the use of vibration signal's features demonstrated a higher robustness and a remarkable superiority.