Sensor fault detection, diagnosis, and data reconstruction strategy for chiller based on an improved tree-based pipeline optimization tool framework

Machine learning is widely applied to fault detection, diagnosis, and data reconstruction for chiller sensors but often requires domain expertise and manual intervention. Tree-based Pipeline Optimization Tool (TPOT), an automated machine learning framework, shows promise in fault detection, diagnosis, and reconstruction by automating model optimization and parameter tuning. Although the TPOT framework includes automated data preprocessing functions, it lacks the ability to automatically handle outliers. Outliers in sensor data can adversely affect the quality of the modeling process. By leveraging TPOT's capability for automated modeling, an ensemble fault diagnosis model can be developed. However, this model is prone to misdiagnosis when the sensor variables exhibit high correlations. Therefore, this study proposes an improved TPOT framework by incorporating a sliding window strategy to enhance TPOT's ability to handle outliers. The ensemble fault diagnosis model based on TPOT incorporates a Euclidean distance strategy, which identifies faulty sensors by quantifying the difference between the input data and the predicted results. Results show that the improved TPOT framework enhances fault detection, diagnosis, and data reconstruction. In the detection of sensor bias, drift, and precision degradation faults, the fault detection rates increased by a mean of 3.11 %, 4.64 %, and 8.62 %, respectively. The diagnostic strategy incorporating Euclidean distance reduced the number of misdiagnoses by one in the diagnosis of nine different sensor faults. In sensor data reconstruction, the RMSE was reduced by a mean of 68.26 %.