Streaming-based feature extraction and clustering for condition detection in dynamic environments: an industrial case

As relevant aspects of PHM, feature extraction and diagnostics represent the focus of many paper related to predictive maintenance. Feature extraction is a fundamental part in PHM, as the accuracy of diagnostic models strongly depends on the goodness of the features. In particular, it is performed when components are provided with many sensors, in order to extract relevant and non-redundant information from the raw dataset. Diagnostic is a typical pattern recognition problem, in which the data are classified according to the operating or fault condition they refer. Thus, to obtain a real-time Remaining Useful Life (RUL) of the component, usually an offline analysis, including feature extraction and diagnostics, is performed, whose results are then used for degradation modelling and finally RUL prediction. However, when systems do not operate in strictly controlled environments, as in case of industrial contexts, it is very challenging to obtain information about the operating condition at the moment of signal acquisition. This results in unlabeled datasets, which cannot be used by supervised learning algorithms, as ANNs and SVMs. In addition, operating conditions change over time and it is not always possible to know a priori all possible conditions. These considerations suggest to resort to streaming applications, in which models can directly learn from new incoming data. As the degradation rate may vary according to the operating condition, influencing the RUL prediction, one should always know in which condition the machinery is operating, or should recognize if a new condition is occurring. In addition, it is not possible to extract good features that distinguish different conditions, if a condition is not known. Therefore, features should be extracted in an unsupervised manner and incrementally, so that if a new condition occurs, eventually better features can be extracted. Furthermore, an incremental clustering should be conducted so to always recognize the condition under which the system is operating, if known, or to detect a new condition. In this paper, a streaming-based procedure for feature extraction and clustering is proposed, which is validated on a real industrial case study. A batch and supervised feature extraction and diagnostics are also performed on the same dataset, to demonstrate that the two approaches have similar results, in terms of accuracy with respect to the known conditions. In addition, thanks to the incremental clustering, the proposed approach is also able to detect and automatically label new machinery operating conditions.