HAMA: A Hierarchical Adaptive Multi-Agent Architecture for Industrial IoT Predictive Maintenance

Industrial IoT predictive maintenance demands real-time anomaly detection under tight resource and interpretability constraints, while monolithic LLM-based systems remain impractical for on-site deployment. We introduce HAMA (Hierarchical Adaptive Multi-Agent Architecture), in which ‘‘adaptive’’ refers strictly to online adjustment of policy parameters, not of agent structure. Edge agents perform lightweight statistical pre-filtering; K=3 Fog nodes run a five-model detection ensemble with consensus voting and federated-style parameter aggregation; Cloud agents adapt consensus weights and thresholds online via Proximal Policy Optimization (PPO) and audit alerts with SHAP attributions, which ground a locally hosted small language model (Llama-3.2-1B) that generates operator-facing explanations. We evaluate HAMA against static and rule-based adaptive baselines on a corrected Boiler Emulator benchmark, a full-length chronologically splitWind Turbine SCADA series with predictive 60-minute-ahead labels, and NASA C-MAPSS with genuine run-to-failure RUL labels (LSTM regressor: MAE 11.2 / RMSE 16.0 cycles, comparable to published results). Across five seeds, PPO-based adaptation is statistically indistinguishable from both baselines on detection F1 (p > 0.5 in every pairwise test) - a null result we report directly - while end-to-end detection latency (0.8–2.0 ms, measured on commodity CPU hardware) stays well within the 100 ms real-time budget, and robustness under noise, missing data, and reduced prevalence matches the static baseline. HAMA thus demonstrates that tiered deployment, real online adaptation, multi-node aggregation, and locally hosted explainability can be integrated into one measured, reproducible system without sacrificing detection quality - an architectural contribution whose statistics and resource footprints are all regenerated from released artifacts. Implementation: https://github.com/HySonLab/AgentIoT.