在微控制器上实现在设备端训练的异常检测
Anomaly Detection on Microcontroller with On-Device Training

在当前嵌入式系统与人工智能技术融合的前沿领域，文章聚焦于一种基于单类支持向量机(One-Class SVM)的异常检测算法，并提供了一套完整的MCU友好的工程实现，不需要依赖于动态内存分配以及文件系统，特别适合于在资源受限的边缘设备上进行高效、实时的训练与预测。我们的方法不仅可以实现在MCU上训练和高效存储机器学习模型，还支持增量学习，从而在几乎不增加计算负担的前提下，持续改进模型对实际工况的适应能力。我们的实验装置是安装了三轴加速度传感器的震动源(如风扇)，以模拟在工作期间发出振动的工业设备。文章的方法也可以通过替换传感器和特征计算的预处理算法来实现对其它设备的监控，以适应不同的工况环境和应用的需求。
This paper introduces an anomaly detection algorithm based on one-class support vector machines (SVMs) and an MCU-friendly engineering implementation. It does not rely on dynamic memory and file systems, it is particularly suitable for efficient, real-time training and prediction on resource-constrained edge devices. Our method not only enables training and efficient storage of machine learning models on MCUs, but also supports incremental learning, thus continuously improving the model’s adaptability to actual operating conditions without increasing the computational burden. Our experimental setup is a vibration source (such as a fan) with a triaxial acceleration sensor installed to simulate industrial equipment that emits vibrations during operation. The method in this paper can also be used to monitor other devices by replacing sensors and computing features.