Anomaly detection with on-device learning with Rtone

Rtone, an ST authorized partner, is a product development studio that supports its customers in the study, design, industrialization, production, and maintenance of intelligent, connected solutions.

As integrators specialized in the field of IoT, Rtone has developed its skills in embedded machine learning and turned to our NanoEdge AI Studio tool to quickly develop a demonstration of anomaly detection on an engine and a pump.

The tool allowed a rapid and guided implementation of a ready-to-use ML model for deployment without advanced knowledge in data science.

This demo, running on a STEVAL-STWINKT1B, is based on a machine learning model that detects vibrational anomalies from a 3D accelerometer. Vibrations are generated by two types of actuators:

• An engine equipped with a magnetized device allowing the creation of a variable unbalance.
• A pump equipped with an air hose whose flow is controlled by a tap.

First, the engine is started and a first training phase is launched under normal operating conditions. The learning is done in real-time on the STM32L4R9 MCU and includes 50 signals. When a fault is simulated on the running engine, the anomaly is immediately detected.

This configuration is also applied to the pump. After training under normal operating conditions, the generated vibrational disturbances also trigger anomaly detection.

It is also possible to train the model to detect an anomaly in the engine or the pump when both are in operation.

3D accelerometer (ref: ISM330DHCX)