Automatisierte Fehlererkennung und -diagnose für den Betrieb von raumlufttechnischen Anlagen durch den Einsatz künstlicher Intelligenz

Automated fault detection and diagnosis (FDD) methods in building systems can effectively identify operational faults, thereby improving energy efficiency and occupant comfort. However, FDD systems are rarely implemented in practice due to their high implementation effort. This work develops a method for automated FDD of air handling units (AHU) using machine learning (ML) techniques to reduce the complexity of parameterization and improve transferability.
The proposed residual-generating approach involves multi layer perceptron (MLP) neural networks (NN) and Bayesian optimization for hyperparameter tuning to estimate the nominal, fault-free behaviour of AHU systems. Residuals from the observed data provide information about the faults for FDD. A key advantage of this approach is that readily available, fault-free data from historical operations can be used for training. By modelling the data points individually, the system automatically adapts to specific con-ditions and can be applied by non-ML experts.
The method’s performance was evaluated in 17 case studies using real operational data from three different buildings. The results show that the approach generally provides good estimates, but the accuracy depends on the specific operational behaviour of the AHU systems. In particular, the control strategy and the number of possible operating modes influence the estimation quality.
To improve fault detection, a scoring method for residual evaluation was developed, which uses a variable tolerance band. It adjusts to the system's operating mode and reduces false alarms caused by model uncertainties, while maintaining high sensitivity in well-learned operating modes. Additionally, the residuals are normalized in fault-prone ranges through scoring, allowing the calculation of an overall score for threshold-based fault detection that accounts for the significance of individual residuals.
A decision tree approach was used for fault isolation, which assigns fault patterns in the residuals to specific fault types with high accuracy. However, a key limiting factor is the availability of historical faulty operation data required for training. It was demonstrated that systems with identical configurations exhibit similar fault patterns, indicating potential for transfer learning processes.
The developed method presents a promising approach to improving FDD in building systems and offers new opportunities for maintenance planning in facility management.