Predictive Maintenance Systems Using Data-Driven Diagnostics for Industrial Rotating Machinery

Abstract

The recent development in the digitalization of industries has brought about the creation of large volumes of sensor data from engineering systems, hence the rapid use of machine learning (ML) in making intelligent decisions regarding maintenance in such industries. This research centers on industrial rotary equipment, especially electric motors and bearings that experience extensive damage and sudden breakdowns during production. Traditional maintenance procedures have been ineffective in detecting and analyzing faults within such industrial equipment. To overcome this challenge, a diagnostic approach based on past historical sensor data including vibrations, acoustic sensors, and temperatures is proposed. Z-Score normalization technique has been suggested as a means of normalizing sensor data by eliminating any variation in scale. Extraction of features is done through the use of Discrete Wavelet Transform (DWT) to detect time-frequency features of non-stationary signals of vibration, acoustic, and temperature sensors for proper fault detection and classification. MBO-kernel-SVM is recommended as a technique for finding optimal SVM hyperparameters for fault detection. The MBO algorithm is a nature-inspired metaheuristic approach that mimics the migratory behavior of butterflies to search for optimum solutions. This was achieved through the use of balance in exploration and exploitation to attain efficiency in solving complex optimization problems. Test results obtained from simulations done using Python software showed that the accuracy of the algorithm is 98% while the precision, recall, and F1 score values are all 0.98. The proposed approach uses a combination of diagnostics and prognostics for early detection of faults.