,

Anomaly detection in a motor running at different speeds

Contact us to learn more

Are you looking for ways to improve the reliability and efficiency of your equipment motors?

Industrial equipment, such as compressors, spindles, or water pumps, produce different vibrations during operation. Some vibrations may be perfectly normal. Others may be the sign of malfunctioning.

Condition monitoring is the first step in achieving an effective predictive maintenance solution. However, most solutions today send raw data to the cloud for further processing, which is costly and energy consuming. By moving data processing closer to the machine that is being monitored, edge computing reduces infrastructure cost, power consumption, and network bandwidth, since data does not need to be sent to and from the cloud.

In this use case, we used a smart sensor node with a machine learning capability that adapts to the system being monitored and detects early signs of equipment failure. As soon as an anomaly occurs, the equipment user receives alerts through the Bluetooth Low Energy radio embedded in the System On Chip (SoC), allowing him to plan maintenance actions.

The approach

The goal is to detect anomalies in the 1kHz range, such as a misaligned shaft or a friction on the motor disk, by collecting vibration data from the ISM330DHCX embedded in the STEVAL-PROTEUS1 board.

We used the following method to reach our goal:

  • Create a dynamic “anomaly detection” model in the NanoEdge AI studio tool.
  • Perform a first phase of “on-device learning” to adjust the model and then start the anomaly detection model on the engine.
  • Use the FP-AI-PDMWBSOC firmware package and STBLE sensor Mobile App to collect data and test the embedded NanoEdge AI machine learning model on the STEVAL-PROTEUS1 board.

We have also developed a test benchmark on which users can generate two different anomaly detection models using the push buttons: anomalies from shaft misalignments and anomalies from magnet interferences. During the “on-device learning” phase, the operator can use up to three motor speeds (low, medium and high), all of which are considered as normal operation. The device learning and sensing phases run on the STM32WB5M microcontroller module hosted on the STEVAL-PROTEUS1 board. They are controlled by a mobile application remotely.

Sensor

3-axis accelerometer featured on the STEVAL-PROTEUS1 wireless smart sensor node (reference: ISM330DHCX)

Data

Regular and abnormal signals:
– Regular signals: nominal behavior, 830 signals per speed (low, medium, high)
– Abnormal signals: anomaly behavior, 830 signals per speed per fault (magnet anomaly and shaft anomaly)
Signal length: 768 (256 per axis, 3 axis)
Data rate: 1.6 kHz; full scale: 2g

Results

Anomaly detection classes:
99.45 % accuracy, 5.7 Kbytes RAM, 6.9 Kbytes flash memory

UC NEAIS Detection of anomalies on an engine results
       Blue points correspond to normal signals, red points to abnormal ones.
The signal numbers are on the abscissa and the confidence of the prediction is on the ordinate.