An edge-deployable adaptive threshold-guided random forest framework for end-of-line NVH fault detection in new energy vehicle drive motors

This research addresses the challenge of deploying computationally intensive machine learning models for real-time, end-of-line (EOL) noise, vibration, and harshness (NVH) quality control in new energy vehicle (NEV) drive motors. Traditional fixed statistical thresholds struggle with non-stationary signals and transient production-line interference, while complex deep learning models are often too heavy for resource-constrained edge devices. The paper proposes an adaptive threshold-guided random forest (AT-RF) framework to optimize this trade-off. The methodology integrates multi-domain NVH feature extraction with a robust median absolute deviation (MAD)-based screening stage. Normal signal windows are accepted immediately via statistical screening, while suspicious windows are selectively routed to a Random Forest classifier for refined classification. According to

the study, this dual-stage approach reduces the RF invocation rate to 32%, achieving a 99.6% window-level fault detection rate, a 0.4% false alarm rate, and an average processing latency under 10 ms on an ARM-based industrial PC with memory usage below 200 MB. Limitations include a reliance on binary normal/faulty classification rather than multi-class fault isolation, and the evaluation is constrained to the specific motor-disjoint protocol splits provided.

Message: Hi FE, I need help with “An edge-deployable adaptive threshold-guided random forest framework for end-of-line NVH fault detection in new energy vehicle drive motors” in “Mechanical Engineering”