Phase 2 of the DETECTA project addresses the development and implementation of anomaly detection models, with significant improvements compared to Phase 1.
One of the fundamental keys to this progress has been the incorporation of more advanced sensor technology, integrated directly into the production machinery. This has allowed data to be collected automatically through the PLC and PLC, providing a more accurate and detailed view of the machine status.
Unsupervised learning has played a crucial role, offering the ability to identify deviations from normality without the need for prior tags, speeding up the process of data analysis and identification of expected results. This is especially valuable in an environment where each workpiece is unique and does not conform to a standard batch.
The current semi-supervised approach is capable of picking up subtle and complex patterns that may indicate an emerging anomaly, giving operators the opportunity to proactively intervene and avoid disruptions or damage to production. To this end, there is a Novelty Module that identifies new data patterns that the AI system has not seen before, and a Point and Pattern Anomalies Module that allows differentiation between anomalies that are one-off incidents and those that follow a specific pattern, likely indicative of more systemic problems.
The models have been tested on data from the real production environment, using data collected during the months of February and March, thus reflecting the true behavior of the milling machine.
Thanks to this semi-supervised approach it has been possible to make an Artificial Intelligence system where the anomalies detected are based on utilization, combined with atypical positions or movements In addition, it highlights the identification of novelties that could indicate emerging patterns and not necessarily anomalies, which represents a significant advance in the flexibility of irregularity detection.
This methodology has succeeded in reducing the false positive rate to 10%, a remarkable achievement compared to the previous phase, which had a rate of 44%. However, the false negative rate is currently 20%, which underlines the complexity of the new sensorics and the importance of accumulating more historical data to improve the inference of the supervised models. The expectation is that as more data is collected, the accuracy of the models will continue to improve substantially.
The advances presented not only improve predictive capability and operational efficiency, but also strengthen the cybersecurity posture by detecting anomalies that could be indicative of potential cyberattacks.
The Detecta F2 project is funded by the Ministry of Industry and Tourism through the line of support grants for Innovative Business Clusters, in its 2023 call for proposals under the Recovery, Transformation and Resilience Plan.
Results published in January 2024.
