Explainable gradient boosting model for on-device mechanical anomaly detection trained and validated using data from a petrochemical process simulation facility

Fault diagnosis in petrochemical plants increasingly relies on artificial intelligence (AI), prompting a growing demand for explainable AI (XAI) to ensure transparency, trustworthiness, and safe decision-making in complex industrial environments. In this study, we developed an explainable, on-device AI framework for detecting mechanical anomalies using vibration signals collected from a petrochemical process simulation facility capable of generating data that closely reflect real industrial conditions. A total of 27 handcrafted time- and frequency-domain features were extracted, and Shapley additive explanations (SHAP) were applied to quantify each feature’s contribution to model decisions. Based on SHAP importance rankings, a sequential feature selection process identified a compact subset of 14 features that maintained high discriminative capability while reducing computational complexity. A gradient boosting model (GBM) trained on the selected features achieved an accuracy of 88.28%, demonstrating competitive performance relative to convolutional neural networks. Additionally, the GBM yielded the fastest inference time (0.018 s) among the evaluated AI models. To assess its deployability, the GBM was implemented on an STM32 microcontroller, where it required 106.76 KiB of Flash memory and only 2.04 KiB of RAM. These results indicate that the explainable GBM provides an effective balance of accuracy, interpretability, and computational efficiency, making it well-suited for real-time, on-device fault detection in resource-constrained industrial settings.