GMV PitIA®

Space missions continuously generate large volumes of telemetry. Detecting anomalous behavior in this data is essential for supporting operations teams, reducing false alarms, and anticipating potential system degradations.

GMV PitIA^® enables us to address this challenge by learning the normal behavior of multivariate systems and detecting significant deviations in telemetry data. In addition to generating alerts, the tool helps explain the source of the anomaly by indicating which parameters contribute most to the detected error. 

To validate the solution, GMV PitIA^® was applied to the ESA Anomaly Detection Benchmark, which is based on real, annotated data from European Space Agency missions. The objective was to compare GMV PitIA^® against benchmark algorithms and demonstrate a methodology that can be applied in an operational setting. 

The rows in the table represent different approaches to anomaly detection, ranging from simple statistical methods, such as Global STD3/STD5, to machine learning models such as iForest, KNN, or Telemanom ESA. 

GMV PitIA^® offers a solid balance between model performance, scalability, usability, and robustness compared to benchmark methods such as Global STD, iForest, KNN, and Telemanom ESA. It is particularly well-suited when the volume of telemetry data is high, tagged data is limited, and alerts must be explainable.

In industrial settings, plant analyzers are essential for monitoring critical process variables. However, continued use may entail operating costs, maintenance requirements, frequent calibrations, and the risk of downtime due to failures.

GMV PitIA^® makes it possible to address this challenge using predictive models trained on historical operational data. The tool learns the relationship between process variables and generates real-time predictions, maintaining visibility on critical variables even during periods when the analyzer is turned off.

The goal of this use case is to minimize the startup time of the plant analyzers by using predictions from GMV PitIA^® while they are turned off. This reduces the use of the analyzer without losing control over critical process variables.

The process is conceived as a continuous cycle: GMV PitIA^® is initially trained using the analyzer's historical data. When the analyzer is turned off, the model provides predictions. When the analyzer is turned back on, the new actual data is used to retrain the model and reduce the prediction error.

During power-off periods, GMV PitIA^®acts as a virtual simulation of the analyzer, allowing the process to continue to be monitored. During operational periods, the system incorporates new real-time measurements to keep the model up to date and improve the accuracy of subsequent predictions.

GMV PitIA^®proved capable of predicting critical process variables with high accuracy based on historical operating data, while maintaining a reliable estimate during periods when the analyzer was shut down. The results confirm that it is feasible to significantly reduce the analyzer's startup time without compromising process control.

GMV PitIA^® has been designed for different profiles:

• Modeling and analysis: Data scientists can develop and maintain predictive models in an agile manner, leveraging available historical and operational data.