A Compacted Object Sample Extraction (COMPOSE)-based method for fault diagnostics in evolving environment

In this work, we consider the problem of developing a fault diagnostic system for an industrial component operating in an evolving environment characterized by continuous modifications of the working conditions. The main difficult to be addressed in this context is that the labeled data available to train the empirical classification model do not cover all the possible working conditions that will be experienced by the component during its life. We propose a diagnostic model based on the COMPacted Object Sample Extraction (COMPOSE) algorithm, which has been developed in the domain of streaming data learning. The key idea of this algorithm is to learn the concept drift caused by the modification of the working conditions. This is done by i) aggregating the labeled training dataset with unlabeled new data collected in a different working condition, and ii) identifying the data drift trend by performing a shrinkage of the aggregated dataset. In the present work, the original COMPOSE algorithm has been modified by introducing an automatic method for dynamically setting the COMPOSE internal parameters, taking into account the characteristics of the new collected data. The overall method has been verified with respect to the Case Western Reserve University Bearing dataset, considering the classification of bearing defects in an evolving environment characterized by modification of the motor load. The obtained performance is satisfactory: bearing defects at different motor loads from that considered for training the diagnostic model are classified with good accuracy. Furthermore, the performance of the proposed model overtakes that of the original COMPOSE method.