Advanced Acoustic Monitoring in Remote and Harsh Environments

Context and Target Applications

The AMAD (Acoustic Monitoring for Anomaly Detection) method focuses on the deployment of AID (Acoustic Intelligent Detection) surveillance systems. These systems are designed to enhance the reliability of critical assets located in remote, hard-to-reach, or hazardous environments. This approach is particularly relevant for the following sectors:

  • Isolated and Autonomous Assets: Including onshore and offshore wind turbines, hydropower plants, hydrogen infrastructure, transformer stations, dam monitoring, and conveyor systems. The goal is to ensure the reliability of electromechanical assemblies and structural integrity.
  • Confined Industry Environments: Such as nuclear, chemical, or petrochemical plants, where access is restricted and process reliability monitoring is critical.
  • Robotics: For inspection and manufacturing applications, serving as a vital complement to computer vision.

The target applications include, but are not limited to:

Reliability of Electromechanical Assemblies in Confined or Restricted-Access Environments:

  • Bearing and Gearbox Failures: (Step-up/Speed increasers, gear reducers).

  • Mechanical, Electrical, and Power Electronics Faults.

  • Cavitation, Turbulence, and Leakage Defects.

  • Imbalance and Friction/Rubbing.

Inspection Round Optimization:

  • Prioritization: Identifying which assets require urgent attention.

  • Frequency/Periodicity: Optimizing the timing of maintenance checks.

  • Mitigating Human Limitations: Reducing risks associated with limited knowledge, subjective interpretation, and lapses in human attention.

Structural Integrity:

  • Wind, Gas, and Steam Turbines.

  • Electrical Generators and Alternators.

  • Industrial Electric Motors.

  • Industrial Pumps and Compressors.

  • Circuit Breakers and Contactors.

  • Transformers.

  • Gearboxes and Mechanical Reducers.

  • Ventilation Systems and Blowers.

Methodology: Unsupervised Approach

In these complex and variable environments, defining all possible acoustic failure signatures a priori is a significant challenge. Consequently, the chosen method relies on unsupervised learning of the environment’s “acoustic normalcy,” leveraging Artificial Intelligence and Machine Learning techniques.

The method uses a specific neural network model for unsupervised learning as follows:

  • Baseline Learning: The model is trained on healthy acoustic data collected from the real environment, integrating normal process variations whenever possible.
  • Feature and Indicator Extraction: The 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 via 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, indicating a fault.
  • Retraining Mechanism: In the event of a new variation in the so-called normal environment, new data can be integrated into the baseline using a guided method. This intervention requires action from an operator.

Pre-implementation Method Analysis on Virtual Signals 

To validate system sensitivity before critical deployment, an analysis based on virtual fault injection can be performed. To test detection robustness against ambient noise, real fault signatures (pre-recorded or simulated) are digitally injected into the baseline recordings. For conveyor systems, this typically involves mechanical faults such as bearing issues or friction.

The injection algorithm proceeds as follows:

  1. Selection of a representative background noise data range.
  2. Periodic insertion of a short-duration typical fault.
  3. Variation of Emergence Levels (Gain): The fault is added at different intensity levels relative to ambient noise: +0 dB, +5 dB, +10 dB, and +15 dB. These gains can be applied to global levels or specific frequency bands depending on the nature of the targeted faults.

Analysis of these virtual signals allows for the following conclusions on performance:

  • High Emergence (+10/+15 dB): Faults are almost systematically detected across all use cases, with the significantly exceeding the alert threshold.

Low Emergence (+0/+5 dB): Detection is more complex. Certain faults whose spectral energy is masked by background noise (e.g., low frequencies dominated by ventilation noise) may fall below the detection threshold if appropriate frequency filters are not applied. In such configurations, implementing directional listening and/or spatial filtering via Ambisonic recording methods can partially overcome background noise. This depends on the trade-off between the detection rate and the expected False Alarm Rate (FAR).

Adaptation of Monitoring Criteria

Studies demonstrate that simple threshold-based sound level detection is ineffective in harsh environments due to natural noise variability (shocks, vehicles, processes, etc.). The unsupervised approach yields far superior results by dissociating sound events in complex environments. To ensure reliable monitoring, certain criteria must be adapted to client constraints:

  • False Alarm Rate (FAR) Management.
  • Time-delay and Persistence settings.
  • Operator-validated model retraining management.

Acoustic monitoring via unsupervised learning, validated by virtual fault injection on representative databases, confirms the feasibility of anomaly detection in remote environments without requiring a priori knowledge of every possible fault. It significantly outperforms simple global or band-specific sound level detection. Adapting monitoring criteria (time-delay, frequency filtering, and dynamic thresholds) is essential to guarantee an operational balance between detection responsiveness and false alarm mitigation. Intelligent retraining management over time makes the system increasingly robust while minimizing human intervention in the process.