LogicLSTM: Logically-driven long short-term memory model for fault diagnosis in gearboxes

This article introduces LogicLSTM, a hybrid neuro-symbolic model obtained by logically guiding a pretrained Long Short-Term Memory (LSTM) network with the support of a customized Logic Tensor Network (LTN). The model is further optimized by explainable AI techniques, for a refined fault classification of time-series data coming from industrial gearboxes. The framework leverages the intrinsic strengths of LSTMs deep recurrent networks for temporal data processing with logical reasoning capabilities, to improve prediction accuracy and interpretability of the classification. Our approach addresses the challenges of extracting relevant data features and integrating connectionist and symbolic methodologies to form a cohesive predictive model. Results from extensive testing show that our model significantly outperforms traditional LSTM models, particularly in complex fault scenarios where conventional methods may fail. Specifically, the hybrid model demonstrates a 16.03% average improvement in accuracy over standard LSTM models under conditions of sufficient data availability, and a 8.56% improvement in scenarios where data is scarce. This research not only demonstrates the potential of hybrid models in industrial applications but also highlights the importance of explainability in AI systems for critical decision-making processes. The proposed model’s ability to interpret and explain its predictions makes it a valuable tool for advancing predictive maintenance strategies within the Industry 4.0 framework.