Pump anomaly detection based on vibrations 

Predictive maintenance is a maintenance strategy that uses machine learning to predict when equipment is likely to fail.

Industrial equipment, such as water pumps, produce different signals during operation. By placing sensors on these pieces of equipment for data collection, it is possible to use machine learning to recognize normal and abnormal behavioral patterns. This proactive approach aims at minimizing downtime, increasing efficiency, and extending equipment lifespan.

We implemented the following approach on a water pump: we collected and analyzed its vibrations and used a machine learning model to detect anomalies. This approach can easily be adapted to many other industrial machines. 


Project goals:

  • detect anomalies in a hydraulic circuit equipped with a pump by using the vibrations of the pump.
  • show that our model can be used on a completely different integrated circuit and is still able to detect anomaly thanks to our edge AI learning technology.

In this project, we used a first circuit to log the vibration data of the pump using an accelerometer. We gathered both normal and abnormal signals, which were created by closing the taps that diverted the flow of water.

We then created an anomaly detection model in the NanoEdge AI Studio. The software tool generated the most optimized model for our learning data.

The model created was loaded on a Proteus board and connected to a first circuit. This prototype allowed us to detect both normal and abnormal data in real time.

We then moved the Proteus board containing the model to a second circuit. We used on-edge learning to retrain the same model with the signals of the new circuit in few seconds. The model performed well on the new circuit, even if it was originally optimized for the first one.


Accelerometer (3-axis): ISM330DHCX


Regular and abnormal signals:
– Regular signals: 640 signals of pump vibrations with the first circuit in a normal state
– Abnormal signals: 440 signals of pump vibrations with one of both deviations blocked
Signal length: 768 (256 per axis, 3 axes)
Data rate: 6667 Hz, range: 2g


Anomaly detection classes:
100% accuracy, 7.8 Kbytes RAM, 6.1 Kbytes Flash

Blue points represent the normal signals, red points the abnormal ones. 
The signal values are on the abscissa axis and the confidence level of the prediction is shown on the ordinate