TreeSHAP-Guided Iterative Compression of Random Forests for Interpretable Edge Fault Detection and Failure-Mode Classification

Edge-deployed machine learning systems for industrial fault detection face two practical constraints: limited computational and memory resources on edge devices, and the demand for transparent, interpretable decision support. Tree ensembles such as Random Forest perform strongly on industrial tabular data, but their size impedes edge deployment, and existing compression methods often sacrifice reliable probability ranking for compactness. We design a TreeSHAP-guided iterative compression framework for Random Forest classifiers that uses exact Shapley scores to drive sequential feature elimination and retraining in a closed loop. On the AI4I 2020 and SECOM benchmarks, the compression strategies are compared with two node-extraction heuristics and several alternative feature-selection methods under a common protocol. Results show that the combined strategy substantially reduces model size while preserving ranking quality, and that TreeSHAP-based selection achieves a favorable accuracy–compactness trade-off over statistical and unsupervised alternatives; the framework also extends to multi-class failure-mode classification and to an external high-dimensional dataset. Edge inference latency on Raspberry Pi 4 and Jetson Nano is estimated by clock-frequency scaling from desktop measurements; the estimates suggest millisecond-level portable inference with a negligible storage footprint, providing a lightweight and interpretable option for edge fault detection.