Improving the Performance of PCA-Based Chiller Sensor Fault Detection by Sensitivity Analysis for the Training Data Set

An improved approach of fault detection for chiller sensors is presented based on the sensitivity analysis for the original data set used to train the Principal Component Analysis (PCA) model. Sensor faults are inevitable due to the aging, environment, location and so on. Meanwhile, because of the wide range of operational conditions, the fault of a certain sensor is very difficult to be directly detected by its own historical data. PCA is a multivariate data-based statistical analysis method and it is very useful for the sensor fault detection in HVAC&R. The undetectable zone of a certain sensor by Q-statistic is derived from the definition of Q-statistic which is usually employed as a boundary to detect the sensor fault situation. Due to the similar style between Q-statistic and Hawkins’ 2 H T , the undetectable zone by Hawkins’ 2 H T is also obtained. Undetectable zone is a predictive index to indicate the detectability of different sensors by different statistics. Since undetectable zone is the character of the original training data set, it can indicate the quality for the selected training data. One field data set is employed to validate the presented approach. Results show that the undetectable zone of a certain sensor by Q-statistic is quite different from that by Hawkins’ 2 H T . Therefore, the undetectable zone can be used to improving the performance of PCA-based chiller sensor fault detection by choosing different fault detection statistics with less undetectable zone for different sensor.