Explainable AI-Based Predictive Maintenance Model for Industrial Equipment Using Sensor Fusion and Deep Learning

DOI

10.2991/978-94-6239-789-7_30How to use a DOI?

Keywords

Predictive Maintenance; Explainable Artificial Intelligence; Sensor Fusion; Deep Learning; Convolutional Neural Network; BiLSTM; SHAP; Remaining Useful Life; Industry 4.0; Fault Detection

Abstract

The speedy growth of the Internet of Things (IoT) in the industrial settings has spawned accumulating amounts of multi-modal, heterogeneous, time-series sensor data, which has led to a real chance of shifting away reactive maintenance practices including data-driven predictive maintenance (PdM). The practical implementation of deep learning-based PdM models has not yet become a broad-scale application, in part due to their lack of opacity, which erodes the trust and confidence of operators and complicates the compliance with the regulations. This paper proposes a novel Explainable AI-Based Predictive Maintenance (XAI-PdM) framework that integrates attention-based adaptive multi-modal sensor fusion with a hybrid Convolutional Neural Network–Bidirectional Long Short-Term Memory (CNN-BiLSTM) architecture for accurate and interpretable fault classification and remaining useful life (RUL) estimation. The framework is designed around four functional sensor modality groups thermodynamic, structural-load, speed, and efficiency which capture the diagnostic roles of vibration, temperature, acoustic emission, and current signals, respectively. Crucially, this functional grouping is sensor-configuration agnostic: on the CMAPSS turbofan benchmark, which lacks physical vibration or acoustic transducers, the 14 retained sensor channels are partitioned into four analogous functional groups that replicate the diagnostic roles of those modalities in rotating machinery. This mapping is physically and analytically justified in Section III-B. Post-hoc interpretability is achieved through a dual-layer XAI mechanism combining Shapley Additive explanations (SHAP) for global and local feature attribution with Gradient-weighted Class Activation Mapping (Grad-CAM) for temporally resolved saliency analysis. Validated on the NASA CMAPSS and FEMTO-ST PRONOSTIA benchmark datasets, the proposed framework achieves an RMSE of 12.47 cycles for RUL estimation and a fault classification accuracy of 98.3%, outperforming six state-of-the-art baselines. An attention-based adaptive sensor fusion module dynamically assigns diagnostic weights to sensor modalities in real time, substantially improving robustness under noisy and missing-data conditions. The model features excellent predictive quality and practical and human readable explanations, enhancing scalable and reliable maintenance intelligent deployments in Industry 4.0.

Copyright

© 2026 The Author(s)

Open Access

Open Access This chapter is licensed under the terms of the Creative Commons Attribution-NonCommercial 4.0 International License (http://creativecommons.org/licenses/by-nc/4.0/), which permits any noncommercial use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license and indicate if changes were made.

Download article (PDF)

TY  - CONF
AU  - Abhilash Kusa
PY  - 2026
DA  - 2026/10/08
TI  - Explainable AI-Based Predictive Maintenance Model for Industrial Equipment Using Sensor Fusion and Deep Learning
BT  - Proceedings of theInternational Conference on Innovation and Practices in Artificial Intelligence (ICIPAI 2026)
PB  - Atlantis Press
SP  - 428
EP  - 451
SN  - 1951-6851
UR  - https://doi.org/10.2991/978-94-6239-789-7_30
DO  - 10.2991/978-94-6239-789-7_30
ID  - Kusa2026
ER  -