Adaptive fault detection in solar thermal for industrial processes based on conditional quantile regression

Solar thermal systems are increasingly used in industrial processes, making early and accurate fault detection essential for efficient system operation. One of the most common faults in such systems is the opacification of solar collectors, which limits solar transmission and thus reduces thermal performance. However, detecting this fault remains challenging due to the scarcity of labelled fault data. To address this problem, this study proposes an adaptive fault detection method to detect abnormalities without requiring labelled fault data. The proposed method, formulated as a statistical deviation problem, uses conditional quantile regression to automatically learn adaptive thresholds from historical normal operations, which are then applied to detect abnormal deviations in subsequent operations. The data used are based solely on a physics-based simulation model of a solar thermal plant for dairy production. Healthy data were obtained by simulating under normal operating conditions, while faulty data were generated by injecting faults (sudden or gradual, and of different intensities) into the model. The results indicate that the proposed method effectively detects faults, alarming about 80-90% of faulty patterns, of which 90-96% are the highest severity level (e.g. opacification levels around 30%). For faults that appear gradually, the detection performance remains consistent with the limited deviation from nominal system behaviour, reflecting the inherent difficulty of early-stage fault detection. Besides, the proposed method maintains a low false alarm rate of less than 5% under normal operating conditions. To enhance stability, a rule-based time validation layer is introduced, relying on the frequency and persistence of daily detections to minimize false alarms. These findings suggest that the proposed method is a promising approach for detecting faults in solar thermal systems for industrial processes. As future work, the method could serve as the first layer in a multi-layer fault detection framework that combines physics-informed and artificial intelligence-based approaches.

Work In Progress