A physics-informed autoencoder method with automatic weighted loss for chiller water system sensor fault detection

Machine learning (ML) methods have been increasingly applied in sensor fault detection and diagnosis (FDD) of heating, ventilation and air-conditioning (HVAC) systems. However, most current ML based sensor FDD methods overly focus on extracting features from sensor data and using combined ML methods to improve detection efficiency, which often neglects the inner physical correlation information among HVAC sensors. To address this issue, this study proposes a physics-informed autoencoder (PIAE) method that fully utilizes the physical correlation information between sensors. PIAE can efficiently extract complex relationships among input variables by integrating the physical correlation information into the autoencoder structure through the combination of the decoder’s output and the loss function. This integration approach effectively avoids results from a single autoencoder output that may deviate from the physical correlation information, thus improving the accuracy and reliability of sensor fault detection. The ASHRAE RP-1043 data was used for validation. The results show that, compared with a single autoencoder, PIAE significantly improves the fault detection accuracy. The average Youden’s index increases from 0.28 to 0.69 and the detection accuracy of the temperature sensor increases from ±3.25 °F to ±1.75 °F. Furthermore, the analysis indicates that an increase in training data amount can effectively enhance the accuracy of the proposed model. The proposed method contributes to providing a concise and reliable approach for sensor fault detection in HVAC systems.