Abstract
Abstract High-throughput production of Brushless DC (BLDC) motors requires reliable, low-latency detection of nonconforming units (Abnormal) before integration into final products. This study, conducted in the scope of the R-PODID project, addresses embedded anomaly detection for an AI-enabled motor control board with a dual-processor architecture. The proposed pipeline is tailored to on-device constraints, where only on-board electrical measurements are available, and deliberately focuses on the three-phase current signals. More specifically, it (i) characterizes the current signals in the frequency domain to motivate segmentation and down-sampling, (ii) performs preprocessing focused on the steady-state regime, and (iii) deploys an LSTM-based autoencoder on a resource-constrained STM32H563 microcontroller using X-CUBE-AI. Experimental results on factory datasets comprising nominal (Normal) and faulty (Abnormal) motors show that frequency-informed decimation to 625 Hz preserves discriminative content while enabling on-device inference with sub-second latency and modest memory usage ( $$\approx $$ 70 kB RAM, 490 kB Flash). The proposed embedded configuration achieved a mean accuracy of 96.04% across runs, while maintaining compatibility with the memory and timing constraints of the target platform. Design trade-offs relevant to TinyML deployment are also discussed, including steady-state selection, anti-aliasing, percentile-based thresholding, and the impact of window size on accuracy and runtime. Overall, the results support the feasibility of embedded reconstruction-based anomaly detection for BLDC motor production testing under resource constraints.