Memory-Oriented Structural Pruning for Anomalous Sound Detection System on Microcontrollers

Anomalous sound detection (ASD) is crucial for the early identification of machine irregularities, preventing potential failures and ensuring smooth operations. Deploying ASD systems on edge devices enhances real-time monitoring and decision-making, reducing latency and dependence on centralized data centers. However, implementing neural network systems on microcontroller units (MCUs) is challenging due to limited Flash and SRAM resources. To address this, we designed a memory-oriented structural pruning algorithm utilizing a reinforcement learning agent that incorporates memory constraints to determine the optimal compression ratio. Our method effectively compresses model parameters with minimal performance impact, enabling efficient deployment on resource-constrained edge devices. The results show that we can compress the model parameters by 2 x and reduce the peak memory usage by 4.5 x to meet the memory constraints. Compared to prior work, our method improves AUC by 1.59%.