TinyML, Anomaly Detection

When manufactured parts of industrial machines fail in the field, the result can be quite costly. Rather than incur the cost of repairs and replacement parts, an anomaly detection AI can be used to constantly monitor machine health and notify operators whenever any disfunction is detected. This can help identify anomalies before they become severe enough to alert machine operators, or before a complete system failure occurs. In this thesis, we explore machine learning on an embedded device to detect anomalies with sophisticated low-power neural networks. We employ the state-of-the-art TinyML framework, which enables the development and deployment of Edge AI. With TinyML, deep learning can execute on the same low-power computing platforms which collect sensor data in the field. We leverage this deep learning approach to detect physical anomalies as they occur, using a battery-powered embedded device with no network connection. Our embedded platform is the Arduino Nano 33BLE, mounted to a Kenmore top-load washing machine. We read the Arduino accelerometer sensor to collect normal data from a balanced laundry load and abnormal data from an unbalanced laundry load, as they are washed by the machine. Then, we use normal data to train two different neural network models: autoencoder and variational autoencoder. These neural networks are used to detect accelerometer anomalies as unbalanced washing machine loads. After generating the model, we use TensorFlow Lite to load the model on the Arduino Nano 33BLE board. The Arduino device detects and indicates unbalanced washing machine loads with 92% accuracy, 90% precision and 99% recall using the autoencoder model. When we use variational autoencoder model, we experience 66% accuracy, 74% precision and 80% recall. We also measured the battery life for the TinyML anomaly detector with autoencoder model as 20 hours with 5 V lithium batteries.