Wavelet Packet and Neural Network Methods for Fixed Bias Fault Diagnosis for Air Handling Unit Sensors

摘要

Sensor faults may result in wasted significant energy, shortened equipment life and even degraded indoor air quality for air conditioning systems (ACS). As one of the most common sensor faults, fixed bias fault (FBF) is difficult to be identified due to the same change trends for sensor measurement and actual value. The automated fault detection and diagnosis (FDD) methods, such as model-based, datadriven, and knowledge-based, have been successfully applicable to ACS. The approach performance, however, depends strongly on the accuracy and reliability of the sensor measurements. In addition, a wide range of training data may be entailed for data-driven methods. The wavelet packet and neural network (WPNN) fault diagnosis methods present more favorable benefits because of fully combining advantages from wavelet packet transform (WPT) and back-propagation neural network (BPNN). WPNN can process the original measurement data to get characteristic information from wavelet decomposition, and then train these characteristic data as BPNN inputs. Based on transient system (TRNSYS) simulation program, the WPNN diagnosis strategies for ACS sensor faults are validated in the simulator. Three sensors, including flow rate, temperature and pressure sensors, are selected to analyze the FBF generated by a fault generator which is incorporated into the simulator. Then, a required fault is introduced to the given sensor in the simulator at a certain moment. The signal decomposition can be acquired by the WPT coefficients of third resolution level. The fault diagnosis results show that the output data are considerably consistent with the original ones, and the deviation from set value is lower than 0.01.With appropriate training, WPNN-based method can be applicable to diagnose the FBF in ACS. Compared with single BPNN, WPNN presents less epochs for convergence precision with the value of 0.001.Since WPT considerably reduces the number of training samples, the WPNN diagnosis comes to be more efficient and the accuracy is improved desirably at the same time. The validation results indicate that average diagnosis efficiency approaches to 91% for 300 groups of pre-diagnosing data. It suggests that the WPNN-base method is successful for fault diagnosis. Also, further study should be focused on the multiple faults diagnosis based on the methods such as joint information.