With the development of sensor technologies in manufacturing systems in the last two decades, data gathered from manufacturing tools such as pressure, temperature, and vibration are constantly monitored. This has led to the more widespread use of intelligent systems such as machine learning algorithms to detect failures that will occur as early as possible in the production process to save time and expense. The maintenance method that allows the prediction of potential faults using real-time data and machine learning and deep learning algorithms is called predictive maintenance. This paper describes an algorithm that can be used for bearing fault classification. The algorithm includes the wavelet scattering network to convert the raw vibration signals into a much smaller feature set to distinguish different classes of failures and then utilizes the ensemble machine learning method to classify healthy, transitive and faulty states. Moreover, the necessary hyperparameters for the machine learning method were optimized with the Bayesian model during developing the machine learning method, and the performance of the machine learning method was evaluated by the accuracy score, precision, recall and sensitivity. The developed method was able to accurately detect all faults by achieving 93.15% and 91.90% accuracy scores for the training and test datasets, respectively.
Anahtar Kelimeler: Machine learning, Predictive maintenance, Fault classification, Vibration signal, Wavelet scattering
|
