AI-Based Leak Detection in Compressed Air Systems

We treat your compressed air system as a rich source of information. Instead of relying on costly additional sensors, we leverage the data already provided by your compressors and sensors, selecting the most suitable learning approach based on the available data.

Anomaly detection scenarios: Selecting the right strategy for your data

Unüberwachtes Szenario (Der Regelfall): Es liegen keine Labels vor. Unsere Systeme, basierend auf Architekturen wie Autoencodern oder Transformern, lernen selbstständig das Normalverhalten aus den Daten und identifizieren jede signifikante Abweichung als Anomalie. Dieser Ansatz ist ideal für die Praxis, da der manuelle Annotationsprozess entfällt.

• Unsupervised scenario (standard case): No labeled data is available. Our systems – based on architectures such as autoencoders or transformers – learn normal behavior directly from your data and identify any significant deviation as an anomaly. This approach is well suited for practical applications, as it avoids manual annotation.
• Supervised scenario (ideal case): When labeled measurement data is available, models such as decision trees learn to associate specific signal patterns with particular leak locations. This enables high accuracy for known fault types.
• Semi-supervised scenario (practical compromise): We combine a small amount of labeled data with a large volume of unlabeled data or use time-series reconstruction methods. Even a limited number of labels can significantly improve detection performance.

Event detection and feature extraction

• Leaks typically appear as unexpected pressure drops or persistent base loads. Our change-point detection algorithms identify significant shifts in pressure and flow patterns. We then extract characteristic signatures, such as pressure decay curves and load gradients. Baseline analysis during idle periods provides additional insight into the overall tightness of the system.

Detection alone is not sufficient. The key challenge – and our core expertise – lies in accurately localizing leaks to enable targeted repairs.

• Physical network modeling: We model the physical network of your compressed air system and correlate measurements from different locations.
• Pressure wave analysis: A leak generates a small pressure wave. By analyzing time-delayed signals from multiple distributed pressure sensors, we can localize the source of the disturbance.
• Zone-based consumption analysis: Through intelligent segmentation of your network, we assign losses to specific pipeline sections or consumer groups.

To further improve the accuracy of localization and quantification, we apply one of the most advanced approaches in AI research: Physics-Informed Neural Networks (PINNs).

• AI meets fluid mechanics: A PINN goes beyond purely data-driven models by integrating fundamental physical laws – such as the continuity and momentum equations of fluid dynamics – directly into the loss function of the neural network. This ensures that the model is optimized not only to fit sensor data, but also to produce physically consistent predictions.
• Key benefits:
  • Higher accuracy with less data: By incorporating physical principles, the model requires significantly fewer data points to generate reliable results.
  • Robustness and interpretability: Results remain within physically plausible limits, reducing false alarms and improving transparency.
  • Reduced sensor requirements: PINNs combine data-driven and model-based approaches, reducing the need for extensive and costly sensor deployment.

Our algorithms are not developed in theory alone – they are designed and validated under demanding, real-world conditions. At our unique WanDa facility (Wandlungsfähige Druckluft-Forschungsanlage), we integrate prototypes into an industry-relevant test environment to evaluate their performance in reproducible scenarios.

In our demonstration setup, compressed air flows either through intact hoses or through hoses with barely visible leaks, kinks, or faulty connections – the most common causes of leakage in industry. Regardless of the flow path, these differences are not visible to the naked eye: the actuators continue to operate as usual.

Our system, however, precisely measures what remains hidden. It captures pressure variations, flow rates, actuator positions, valve states, and even ultrasonic signals. With up to 20 individually controllable actuators, we simulate complex industrial load profiles and generate large volumes of data at sampling rates of up to 100 Hz – amounting to as much as 300 GB per day of operation.

This extensive, high-quality dataset forms the foundation for developing AI models that operate reliably and accurately, even in complex industrial environments.