A progressive self-supervised learning framework: From fault diagnosis to electric motor wear prediction

This paper presents a wear prediction framework for systems based on contrastive self-supervised learning, with a focus on knowledge transfer from fault diagnosis to prognosis. Specifically, the study examines a direct current (DC) electric motor with the aim of predicting the extent of its wear. The methodology includes a pre-training phase using contrastive learning, which extracts temporal features from multivariate time series collected from sensors. These features are constructed using labelled data, creating representations that are correlated across sequences that share the same label. These representations are then used for long-term wear prediction, integrating CNN (Convolutional Neural Network) and LSTM (Long Short-Term Memory) layers. This approach effectively discriminates between normal and failed states in an upstream task, thereby facilitating failure prediction in a downstream task. A progressive learning strategy was incorporated into the pretext task to accelerate the convergence of the contrastive loss function, although it remains optional within the methodology. The results show that the Root Mean Square Error (RMSE) of the proposed method is twice as low as that of the same model without the pretext task on extrapolated data, even in the absence of labels (RMSE 0.78 vs. 1.44).