Confidence-Aware Tiny Machine Learning Orchestration for Vibration-Based Predictive Maintenance in Automotive Powertrains

Tiny Machine Learning (TinyML) enables real-time diagnostics on low-cost microcontrollers deployed inside vehicles. However, aggressive model compression—especially 8-bit quantization—can introduce a reliability gap that is risky for safety critical automotive monitoring. This paper presents a confidence aware TinyML orchestration framework for vibration-based predictive maintenance in automotive powertrains. The system dynamically balances edge and cloud inference using a lightweight uncertainty signal computed from normalized SoftMax entropy. High-confidence vibration windows are processed locally on an ESP32 microcontroller, while ambiguous samples are selectively offloaded to a cloud backend for higher-fidelity analysis. The framework exploits the ESP32 dual-core architecture to keep monitoring non-blocking during communication. Experiments on benchmark bearing vibration data show that the proposed approach recovers 85.7% of the accuracy lost due to 8-bit quantization, achieving 96.8% accuracy while offloading only 12% of samples. Compared to a cloud-only baseline, wireless energy is reduced by about 88% and average end-to-end latency remains 82.8ms. These results suggest that confidence-aware orchestration is a practical pathway to deploy TinyML predictive maintenance in connected vehicles.