The challenge of monitoring in isolated and confined environments
Some industrial sectors do not allow for regular human inspection. Our AMAD (Acoustic Monitoring for Anomaly Detection) method was specifically designed to ensure the reliability of critical devices where humans cannot, or should not, go:
- Isolated and autonomous devices: wind turbines (onshore and offshore), dams, electrical transformer substations, etc.
- Confined and hazardous environments: nuclear power plants, chemical plants, or petrochemical refineries where process monitoring is vital.
- Robotic fleets: integrating hearing into inspection robots or drones, complementing computer vision, to optimize inspection rounds.
In these contexts, sending entire audio streams to a remote Cloud is inefficient due to bandwidth limits, cybersecurity risks, and latency. That is why we prioritize embedded acoustic AI: signal processing is done directly at the source. Only qualified alerts are transmitted to your control centers.
How does our solution work? The unsupervised approach
In a complex industrial environment, it is impossible to model all the acoustic signatures of potential failures in advance. Simple detection based on sound level thresholds (dB) is useless given the natural variability of noise (impacts, passing vehicles, process variations).
To overcome this problem, Metravib Engineering deploys an unsupervised learning architecture based on specific neural networks, which operates in four steps:
- Baseline Learning: the model is trained on healthy acoustic data (defect-free) from the real environment, incorporating so-called normal process variations whenever possible.
- Feature and Indicator Extraction: signals are analyzed using spectrograms and/or acoustic descriptors. These elements can and often must be adapted to each use case based on our domain expertise.
- Detection by Reconstruction Error: the neural network attempts to reconstruct the input signal. If the signal is normal, the error is low. If the signal contains an anomaly (never seen by the model), the reconstruction error is higher, flagging a defect.
- Retraining Mechanism: in the event of a new variation in the “normal” environment, it is possible to integrate new data into the baseline using a guided method. This intervention requires operator action.
Pre-deployment validation: robustness through virtual defect injection
How can you be certain that the AI will detect an anomaly masked by a factory’s ambient noise? Before any critical implementation, we test the robustness of our algorithms by injecting “virtual defects” (real or simulated failure recordings) into your site’s audio Baseline.
We vary the emergence levels to assess the model’s sensitivity. If a failure (such as a bearing defect) has low emergence and is masked by ventilation noise, our teams adapt the sound capture: advanced frequency filtering, directional listening, or even spatial recording (ambisonics) to isolate the source of the problem, if known.
Tangible benefits for the reliability and integrity of your assets
By merging our historical mastery of vibro-acoustic phenomena with artificial intelligence, Metravib Engineering’s hybrid approach delivers measurable results for your maintenance campaigns:
- Electromechanical reliability: early detection of bearing wear, gear defects (multipliers, reducers), pump cavitation, turbulence, or pressurized leaks.
- Structural integrity: continuous monitoring of turbines, alternators, compressors, transformers, and circuit breakers.
- Drastic reduction in false positives: thanks to customized monitoring criteria (time delays, defect persistence), the system only alerts you when absolutely necessary.
- Cost optimization: you limit unplanned downtime, target your inspection rounds, and free your technicians from tedious monitoring tasks.
Conclusion
Every industrial environment has its own acoustic identity. Generic solutions quickly show their limits when faced with the realities of the field. By combining a deep physical understanding of materials and signals with the power of Machine Learning, Metravib Engineering transforms the noise of your infrastructures into strategic data.
Connect with Metravib Engineering
Are you looking to improve the reliability of your isolated equipment or equip your inspection robots with acoustic intelligence? Our engineers are available to evaluate the potential of your data.
