利用微控制器實現高效率異常聲音檢測系統

Abstract

在工業4.0中，預測性維護將人工智慧（AI）和物聯網（IoT）整合在一起，形成工業物聯網（IIoT）。預測性維護的核心目標是及早發現潛在的機器故障，以減少計劃外停機時間並將中斷降到最低。異常聲音檢測（Anomalous sound detection）系統非常重要，因為它們能夠偵測出問題機械所產生的異常聲音，從而在問題發展成重大故障之前進行預警。因此，開發能夠在邊緣設備上運行的高效且可靠的異常聲音檢測系統，有助於提升預測性維護策略的可靠性和效能。 然而，在邊緣設備上部署高性能異常聲音檢測系統面臨兩大挑戰。首先，在吵雜的環境中準確預測異常聲音具有高運算複雜度。其次，在微控制器 (MCU) 部署異常聲音檢測系統需克服嚴格的記憶體限制。為了打造輕量且高效的異常聲音檢測系統，我們在模型訓練階段導入了不同的資料處理方法並且加入新的訓練策略。此外，為了符合微控制器的嚴格記憶體限制，我們設計了一種記憶體導向的自動模型壓縮(Memory-oriented automated model compression)算法，該算法使用強化學習(Reinforcement learning)方法來計算最佳壓縮比。最後，我們詳細介紹了在微控制器上完整部署異常聲音檢測系統的過程。 透過解決有限的異常數據、環境中的噪音干擾和微控制器上記憶體資源不足等問題，我們提供了一種適用於邊緣裝置的即時異常聲音檢測方案。這種方法提升了工業4.0中預測性維護系統的可靠性和效率。

Predictive maintenance in Industry 4.0 integrates Artificial Intelligence (AI) and the Internet of Things (IoT), forming the Industrial Internet of Things (IIoT). A crucial aspect of predictive maintenance is the early detection of potential machine failures to reduce downtimes and minimize disruptions. Anomalous sound detection (ASD) systems are essential because they detect irregular sounds that signal mechanical issues before they develop into major problems. Therefore, developing an efficient and robust ASD system capable of operating on edge devices is essential for enhancing the reliability and effectiveness of predictive maintenance strategies. However, deploying high-performance ASD systems on edge devices faces two main challenges. First, the high complexity of accurately predicting anomaly sounds amidst noisy environmental disturbances. Second, deploying ASD systems within the tight memory constraints inherent in microcontrollers (MCUs). To create a lightweight and efficient ASD system, we implemented robust processing methods and reformulated the training strategy during the model training stage. Additionally, to comply with the strict memory limitations of MCUs, we designed a memory-oriented automatic model compression algorithm that uses a reinforcement learning agent to determine the optimal compression ratio for model deployment. Finally, we detailed the process of deploying the ASD system on MCUs to achieve end-to-end deployment. By solving challenges such as limited abnormal data, environmental interference, and MCU memory constraints, we provide a practical solution for real-time anomaly sound detection systems on resource-constrained edge devices. This approach enhances the reliability and efficiency of predictive maintenance systems in Industry 4.0.