基于TinyML的设备用电分类系统研究与设计

Research and Design of an Appliance Classification System Based on TinyML

摘要:

对基于TinyML的低功耗设备用电分类系统进行了研究与设计；针对传统阈值法与简单统计法难以适应复杂用电场景、深度学习方法依赖云端资源的问题,提出了一种适用于计算与内存受限嵌入式设备的高效电器分类算法；该方法基于家用电器运行时的电流-电压(V-I)轨迹特征,采用轻量化卷积神经网络(CNN)实现设备识别；针对电网中多设备并联运行的复杂场景,结合Stockwell变换(ST)与卡尔曼滤波(KF)技术,有效捕捉负载切换引起的瞬态特征并实现信号分解；利用PLAID-III数据集,通过ST变换提取非平稳信号的时频特征,利用KF模型检测信号突变并跟踪误差；将提取的V-I轨迹输入CNN-V-I网络模型进行分类,并通过模型压缩技术将训练好的网络部署于STM32微控制器；结果表明,优化后的算法在保持分类精度的同时显著降低了计算开销,在PLAID-III数据集上的识别率达到99.18%；该工作为嵌入式设备上的非侵入式负荷监测提供了有效解决方案,展现了TinyML技术在资源受限场景中的应用价值。

Abstract:

A low-power appliance classification system based on TinyML was designed and implemented；to address the limitations of traditional threshold-based and statistical methods in complex electricity scenarios and the reliance of deep learning approaches on cloud resources,an efficient appliance classification algorithm suitable for embedded devices with constrained computing and memory resources was proposed；the method utilized current-voltage(V-I)trajectory features generated during appliance operation and adopted a lightweight convolutional neural network(CNN)for device identification；for scenarios involving multiple appliances operating in parallel,Stockwell transform(ST)and Kalman filter(KF)techniques were integrated to capture transient characteristics induced by load switching and to accomplish signal decomposition；using the PLAID-III dataset,time-frequency features of non-stationary signals were extracted through ST transform,while signal mutations were detected and errors were tracked via the KF model；the obtained V-I trajectories were input into a CNN-V-I network for classification,and the trained network was deployed on an STM32 microcontroller through model compression techniques；experimental results demonstrated that the optimized algorithm significantly reduced computational overhead while maintaining classification accuracy,achieving a recognition rate of 99.18% on the PLAID-III dataset；this work provides an effective solution for non-intrusive load monitoring on embedded devices and illustrates the application value of TinyML technology in resource-constrained environments.