AI-powered algorithms for Anomaly Detection on Industrial and Medical Applications on IoT devices

A machine learning platform for anomaly detection and prescriptive maintenance in industrial and research infrastructures. The developed technology can enhance system reliability, reduce downtime, and enable cost-efficient, data-driven maintenance by leveraging IoT devices and federated learning.

 

INTRODUCTION

Large-scale particle accelerators often experience downtime due to technical infrastructure anomalies. In this scenario, prescriptive maintenance is a crucial tool for optimizing availability and reducing costs. In the case of CERN’s cryogenic system, one of the main critical infrastructure is the helium compression system.

Cmopressor Organization

The classical vibration monitoring approach of helium compressors shows limitations over time, due to sparse data sampling and variability in assessments. This work introduces a machine learning-based framework for anomaly detection and classification using vibration data from helium compressors, demonstrating improved consistency and effectiveness in monitoring and estimating the Remaining Useful Life (RUL) of the analyzed assets.

 

PRESCRIPTIVE MAINTENANCE PLATFORM

The use of machine learning, leveraging autoencoder architecture, has demonstrated the capability of consistently analyze the acquired data to successfully spotting and labeling compressors issues. Further processing of the obtained results also allowed the creation of a framework that, based on the historical data of compressor problems, enables the estimation of their Remaining Useful Life.

ML pipeline

 

RESULTS

The model anomaly detection accuracy was first validated by comparing the outputs with the expert's labels. The model’s RUL estimate was then compared with the evaluation made by experts.

Results

CONCLUSIONS

The developed models were designed to be optimized for deployment on IoT devices, to enable Edge Machine Learning. This paves the way for the implementation of Federated Learning, using the already developed CERN platform CAFEIN FL.

CAFEIN_Pic

The creation of such an infrastructure allows for the inclusion of more facilities in the network and increases the accuracy and robustness of predictions. The improved model trained without sharing any data, will benefit all the involved facilities without compromising data privacy. These features also make the platform particularly well-suited for medical applications, due to the ensured high privacy standards.