An enhanced PCA-based chiller sensor fault detection method using ensemble empirical mode decomposition based denoising.

Highlights • An enhanced PCA sensor fault detection method was presented for screw chillers via EEMD. • EEMD was applied for signal decomposition and noise elimination of raw sensor measurements. • 11 sensors were chosen for PCA modeling by quantifying the chiller energy performance. • Field operating data from real screw chillers at summer seasons were used for validation. • EEMD significantly improved the PCA-based sensor fault detection performance. Abstract In heating, ventilating and air conditioning (HVAC) systems, sensor faults cause improper control strategy resulting in both energy penalty and service costs. As the first and foremost step of an entire on-line fault-tolerant control strategy, sensor fault detection is crucial for maintaining system operation performance. Principal component analysis (PCA) is a widely studied sensor fault detection method in HVAC area. However, the noise information contained in field sensor measurements may confuse the PCA model training process and deteriorate the sensor fault detection performance. To address the problem, this study presented an enhanced PCA-based sensor fault detection method using ensemble empirical mode decomposition (EEMD) denoising. The proposed EEMD-PCA method includes two steps: (1) EEMD is adopted to decompose and denoise the raw measured data, namely eliminating the possible noise information and extracting the possible useful information; (2) PCA is employed to establish the Q -statistic with threshold for sensor fault detection. A case study on a real screw chiller system was conducted to validate the proposed method. 11 sensors were chosen for modeling by qualifying chiller energy performance. Field operating data were used for validation with various magnitudes of added biases. Results revealed that EEMD-PCA showed better detection performance than PCA for 8 critical sensors. The fault detection ratios of the 8 sensors were increased by 22% at average. After pre-processed by the EEMD method, the denoised data were smoother than the raw data. Data distributions in the PCA residual subspace indicated that EEMD-PCA is sensitive to the smaller biases since the denoised data can separate faulty data from the normal data. [ABSTRACT FROM AUTHOR]