Enhanced Predictive Maintenance through Explainable Multimodal Deep Learning and Human-in-the-Loop Systems

Predictive maintenance (PdM) is vital for ensuring the reliability and safety of complex systems in an industry. This work proposes an advanced predictive maintenance model based on explainable multimodal deep learning integrated with human-in-the-loop systems which focuses on condition-based maintenance. This study uses the NASA C-MAPSS dataset, which contains multivariate time series data from industry-type engines under various conditions and fault regimes, the study addresses the challenge of Remaining Useful Life (RUL) prediction. The methodology of this study includes accurately modeling succinct temporal sequences that represent computationally traced strong dependencies that help predict the RUL and incorporating XAI and Human inclusion. Deep learning models often lack robust explainability; hence, the incorporation of XAI components using SHAP is theoretically detailed in this work. This approach is complemented with expert validation to enhance the interpretability of the outcomes. The experimental work is conducted to determine the best-suited model for an efficient time-series analysis. It points out that another potential challenge for the industry in modeling complex systems means model complexity needs to be found in balance with time efficiency. The study weighed the error versus the computational cost through the use of an LSTM architecture. The LSTM is used for the sequence data because it requires less computational cost to run compared to other architectures. The Optuna algorithm has been used systematically in hyperparameter tuning so that optimal model performance can be guaranteed. It balances complexity with the need for power through the streamlined layers of the architecture. This is resource-friendly. Metrics used to evaluate this model are MAE and Weighted Loss Functions.