The issue
How can the reliability and integrity of critical equipment be ensured in areas that are difficult to access, confined, or hazardous to humans? In sectors such as wind power (onshore/offshore), hydraulics, hydrogen infrastructure, dam monitoring, the nuclear industry, or process industries, human surveillance is limited or even dangerous. It is crucial to continuously monitor electromechanical assemblies (bearings, gears, turbines, etc.) and structures to detect early-stage failures (cavitation, leaks, electrical faults, imbalance, friction, etc.) before they lead to critical shutdowns.
The challenge
The primary challenge lies in the variability of the acoustic environment and the impossibility of knowing a priori all the acoustic signatures of potential failures. Simple threshold-based sound level detection is ineffective due to natural ambient noise (processes, vehicles, impacts). The system must be capable of distinguishing a genuine anomaly from a normal process variation, while detecting low-emergence defects masked by background noise, without generating high rates of false alarms.
The solution
Implementation of the AMAD (Acoustic Monitoring for Anomaly Detection) method, an approach based on unsupervised artificial intelligence:
- Learning Normality (Autoencoder): The model learns the healthy acoustic signature of the environment (Baseline) and its normal variations, without the need for a historical database of faults.
- Detection via Reconstruction Error (RMSE): The AI analyzes signals in real-time. If the signal reconstruction error exceeds a threshold for a sustained period, an anomaly is flagged.
- Validation via Virtual Injection: The system’s robustness can be validated by digitally injecting fault signatures into the real ambient background, allowing detection calibration even for low-emergence signals (+0/+5 dB).
Guided Relearning: Following human intervention and environmental verification, the system can evolve over time by integrating new validated environmental variations, thereby reducing human intervention to the strict minimum.
Conclusion
The unsupervised learning acoustic monitoring method, validated by the injection of virtual faults into a representative database, confirms the feasibility of anomaly detection in remote environments without requiring prior knowledge of all possible defects. It is significantly more effective than simple detection based on global sound levels or specific frequency bands. Adapting monitoring criteria (time delays, frequency filtering, and dynamic thresholds) is essential to guarantee an operational balance between detection responsiveness and the minimization of false alarms. Intelligent management of relearning over time makes the system increasingly robust while limiting human involvement in the process.
