Signal processing- and physics-informed neural network for explainable bearing condition monitoring

• Addressing lack of trust and automation in existing bearing fault detection methods. • Combining generalisability of analytical methods with automatability of NNs to a holistic bearing fault detection network. • Fully interpretable network, suited for scarce data environments and varying operational conditions. • Application to a numerical and experimental dataset with full physical interpretation of the results. Machine Learning (ML) approaches have significantly advanced the automation of machine condition monitoring (MCM) for bearings by leveraging vast amounts of data to derive solutions, reducing the need for expert knowledge and intervention. However, these methods often fall short in providing the trust and interpretability of traditional analytical techniques and lack generalisability. Recent studies have demonstrated that certain analytical, knowledge-based methods are already optimal and thus should not be replaced by ML techniques. On the contrary, a comprehensive approach that integrates both analytical and ML components is the most reasonable way forward, leveraging the automation capabilities of ML and the interpretability and generalisability of analytical approaches. In this paper, a holistic physics-informed neural network is proposed for the diagnostics of rolling element bearings, called the BearingNet. The proposed approach incorporates established signal processing methods within a neural network (NN) framework able to tune them and integrate them into a self-adaptive condition-monitoring strategy. The BearingNet can perform traditional knowledge-based signal processing when data is scarce but also has the capability to optimise itself when more and more signals become available for training. A carefully designed network structure ensures that trained parameters are meaningful and interpretable in signal processing terms, and that they remain within physical boundaries, thereby maintaining trust and explainability. The effectiveness of the approach is demonstrated using three bearing datasets: a numerically simulated dataset and two experimental bearing dataset. The results of the numerical tests allow confirming the network’s functioning against a known and controllable challenge, while the application to real bearings show how the BearingNet is valuable in practice. The network is compared to six state-of-the-art NNs. It is shown that these methods are incapable to generalise and perform consistently lower than the BearingNet. The findings highlight that the proposed BearingNet is effective even with limited training data, while always remaining easy to probe and interpret due to its physically explainable design. Each layer of the network can be explained with one or more analogous signal-processing steps, offering complete transparency. While this paper specifically targets rolling element bearings, the innovative approach represents a general new paradigm for utilising NNs in MCM, providing a trustworthy and explainable solution that leverages existing physical understanding. This leads to ideal and robust classification results for bearing diagnostic tasks, setting a new benchmark for upcoming automated bearing diagnostic methods.