Mechanical Anomaly Detection on an Embedded Microcontroller

This paper explores machine learning on an embedded device to detect anomalies with sophisticated low-power neural networks. We leverage this deep learning approach to detect mechanical anomalies as they occur on a top-load washing machine. We collect normal data from balanced laundry loads and abnormal data from unbalanced laundry loads, as they are being washed by the machine. The normal data is then used to train two different neural network models: autoencoder and variational autoencoder. This model is ported to an Arduino Nano microcontroller mounted to the washing machine. Using the autoencoder model, the microcontroller detects unbalanced washing machine loads with 92% accuracy, 90% precision and 99% recall. The battery life for this autoencoder model is 20 hours on 5 V lithium batteries, which is only 14.9% less than the life of a basic LED-blink application on the same platform.