Knowledge mining for chiller faults based on explanation of data-driven diagnosis

Data-driven model is considered to be an efficient and convenient diagnosis method for refrigeration systems, especially in Big Data Era, but the black box characteristic has always been a criticism and hindered its application. For this problem, SHapley Additive explanation (SHAP) method is explored to unveil the black box and the knowledge of system-level chiller faults has been mined. Two ensemble models with excellent diagnostic performance have been investigated including Random Forest (RF) and Light Gradient Boosting Machine (LightGBM). Detailed analysis shows that the sub-cooling at the condenser outlet (TRC_sub) and the condenser approaching temperature (TCA) have a great contribution to the detection of refrigerant leakage and overcharge with different margin, which is consistent with but standing out from the prior knowledge. In practical applications, temperature and pressure sensors for lubricating oil are recommended for a refrigeration system with lubricant due to the sensitive variation with system change (failure or condition altering), which can be a supplement to the traditional experience. The data-driven model is, indeed, an effective way to discover the hidden knowledge of a data set from all dimensions to upgrade the traditional experience. Furthermore, the local explanation by individual samples and the explanation comparison between the RF and the LightGBM model show that the primary features on which the model(s) depends to diagnose the same fault are basically the same, while the auxiliary features may vary slightly. The consensus further proves that the primary features and the related rules mined out in this study can be trusted.