Layer-wise relevance propagation for interpreting LSTM-RNN decisions in predictive maintenance

Predictive maintenance (PdM) is an advanced technique to predict the time failure (TTF) of a system. PdM collects sensor data on health system, processes information using analytics, and then establishes data-driven models that can forecast system failure. Deep neural networks are increasingly being used as these owing their high predictive accuracy efficiency. However, deep often criticized “black boxes,” which multi-layered non-linear structure provide little insight into underlying physics monitored nontransparent untraceable in predictions. In order address this issue, layer-wise relevance propagation (LRP) applied analyze long short-term memory (LSTM) recurrent network (RNN) model. The proposed method demonstrated validated for bearing monitoring study based vibration data. obtained LRP results insights how model “learns” from input demonstrate distribution contribution/relevance classification space. addition, comparisons made with gradient-based sensitivity analysis show power interpreting RNN models. proved have promising potential improving efficiency PdM.