Failure monitoring of industrial systems is imperative in order to ensure their reliability and competitiveness. This paper presents an innovative hybrid modelling approach applied to DC electric motors, specifically the Kollmorgen AKM42 servomotor. The proposed Cooperative Hybrid Model for Classification (CHMC) combines physics-based and data-driven models to improve fault detection and extrapolation to new usage profiles. The integration of physical knowledge of the healthy behaviour of the motor into a recurrent neural network enhances the accuracy of bearing fault detection by identifying three health states: healthy, progressive fault and stabilised fault. Additionally, Singular Value Decomposition (SVD) is employed for the purposes of feature extraction and dimensionality reduction, thereby enhancing the model’s capacity to generalise with limited training data. The findings of this study demonstrate that a reduction in the input data of 90% preserves the essential information, with an analysis of the first harmonics revealing a narrow frequency range. This elucidates the reason why the first 20 components are sufficient to explain the data variability. The findings reveal that, for usage profiles analogous to the training data, both the CHMC and NHMC models demonstrate comparable performance without reduction. However, the CHMC model exhibits superior performance in detecting true negatives (90% vs. 89%) and differentiating between healthy and failure states. The NHMC model encounters greater difficulty in distinguishing failure states (83.92% vs. 86.56% for progressive failure). When exposed to new usage profiles with increased frequency and amplitude, the CHMC model adapts better, showing superior performance in detecting true positives and handling new data, highlighting its superior extrapolation capabilities. The integration of SVD further reduces input data complexity, and the CHMC model consistently outperforms the NHMC model in these reduced data scenarios, demonstrating the efficacy of combining physical models and dimensionality reduction in enhancing the model’s generalisation, fault detection, and adaptability. This approach has the advantage of reducing the need for retraining, which makes the CHMC model a cost-effective solution for motor fault classification in industrial settings. In conclusion, the CHMC model offers a generalisable method with significant advantages in fault detection, model adaptation, and predictive maintenance performance across varying usage profiles and on unseen operational scenarios.

The generalisation of Deep Learning (DL) models under extrapolated usage conditions, particularly in the context of fault diagnosis, remains a significant challenge. Traditional methods encounter difficulties in terms of adaptability, necessitating extensive retraining when confronted with evolving conditions and unanticipated fault patterns. Moreover, existing DL approaches often demonstrate a lack of robustness in dynamic environments due to their limited capacity for generalisation. To address these limitations, this work introduces a novel Curriculum Learning (CL)-based approach, the Residual Singular Value Decomposition Curriculum Learning algorithm (ReSVD-CL), which integrates a Long Short-Term Memory (LSTM) model to capture temporal dependencies in fault patterns and Singular Value Decomposition (SVD) preprocessing to enhance feature representation. A complexity-based criterion segments the dataset into subsets, which are gradually incorporated into training via a pacing function. A comparison of ReSVD-CL with existing methods, such as Baby-Step Curriculum Learning (BS-CL), One-Pass Curriculum Learning (OP-CL) and baseline approaches of fault diagnosis, demonstrates that ReSVD-CL attains superior classification performance, particularly in extrapolated scenarios. The approach facilitates continuous adaptation to new usage profiles without the necessity of full retraining, thereby reducing computational costs and accelerating convergence. The study maintains a strong diagnostic performance with minimal usage profiles, thus demonstrating the potential of combining CL with dimension reduction and temporal modelling to develop more efficient and adaptable fault diagnosis systems.