Context
As part of its naval operations, the multi-mission frigate is equipped with a high-performance sonar, an electronic instrument essential to maritime defense missions. Unlike radars, which use electromagnetic waves, sonar perates by transmitting and receiving acoustic waves, which have the particularity of propagating well in water. This system enables us to detect and analyze the underwater environment, by capturing the echoes returned by submerged objects.
Multi-mission frigate sonar plays a crucial role in anti-submarine warfare, particularly in detecting the presence of enemy units, such as hostile submarines seeking to approach discreetly. It is also used to map the seabed to identify landforms, obstacles, mines and wrecks. This type of tracking enables the frigate to navigate in complete safety, even in shallow or unfamiliar areas.
Thanks to this system, the multi-mission frigate benefits from a very high level of underwater surveillance, detection and analysis capability, making it a strategic player in the protection of sensitive maritime zones.
Needs
Given the extreme conditions in which it operates: prolonged immersion, mechanical shocks, salinity, temperature variations, interference. Sonar is exposed to the risk of premature aging, performance drift or partial failure. It is therefore crucial to be able to detect signs of degradation at an early stage, whether they relate to on-board electronics, acoustic sensors, mechanical actuators or signal processing software.
What’s more, breakdowns in this type of equipment do not always occur suddenly. They often set in gradually, over time, in the form of discrete anomalies, difficult to distinguish from slightly disturbed normal behavior. The system therefore needs mechanisms capable of identifying these atypical behaviors in temporal data, and linking them to phenomena of failure or loss of performance.
Finally, to optimize maintenance and avoid unplanned sonar unavailability that could compromise the frigate’s mission, we need intelligent monitoring of the system’s state of health, based on actual operating data and not just on fixed thresholds or planned maintenance intervals. This would make it possible to anticipate technical interventions before failure becomes critically apparent.
Solutions
The DiagFit solution is based on an innovative approach to unsupervised learning, based on the exclusive use of healthy system data to model its normal operating state. This method makes it possible to build a robust repository of nominal behavior, without requiring failure data for learning.
When the system deviates from this normal state, DiagFit is able to automatically detect anomalies, indicating possible drift or degradation. This principle makes the solution particularly well-suited to industrial environments, where failures are rare and difficult to label.
To validate the effectiveness of this approach, DiagFit uses fault labels that are available to measure the accuracy of detection. These validations confirm the model’s ability to identify relevant anomalies in real-life conditions.
DiagFit analyzes system health at two levels:
- Globally, by assessing the system as a whole,
- Locally, sensor by sensor, to pinpoint sources of degradation.
Results
Thanks to this approach, DiagFit has demonstrated its ability to detect 100% of labeled faults, while also identifying non-labeled anomalies, thus confirming the robustness and sensitivity of the algorithm to unanticipated degradations.
These performances are achieved without the use of fault data for training, making the solution particularly suitable for industrial contexts where fault events are rare or poorly documented.
In addition, the increased availability of data, particularly in complex or multivariate environments, will further improve model performance, by refining the characterization of the normal state and reducing false positives.
In this way, DiagFit provides a solid basis for the early detection of anomalies, helping to improve predictive maintenance and enhance the reliability of monitored systems.
