Chiller fault transfer diagnosis based on TrAdaBoost

Chiller faults can be diagnosed by machine learning, but it needs a lot of training data, and it is difficult and costly to obtain effective fault data. Traditional fault diagnosis is mainly based on the existing data of a single unit, which is difficult to cover all working conditions, and the diagnostic performance deteriorates under new working conditions. In this paper, a multiple data processing method of data space extrusion is proposed to reduce the differences between different distributions. Moreover, the information transfer ability of the TrAdaBoost algorithm for different data distributions is utilized to build a chiller fault diagnosis model combined with different base classifiers, which realizes effective fault diagnosis in new working conditions and is expected to reduce the experiment cost. The diagnosis results of seven typical faults of chillers show that when the data of new working condition is only 20 groups, the overall accuracy increases by 22.00%, 2.50% and 32.33%, respectively, compared with the case without transfer diagnosis. By adding two working conditions data to verify the effectiveness of migration diagnosis for chiller fault diagnosis under different modes, the performance of transfer diagnosis under single mode is improved by 18.39% to 22.43% compared with conventional diagnosis, and the performance under full mode is improved by 1.21% to 2.55%. Parameter optimization can help improve single-mode migration diagnosis (3.06%), but can decrease the performance of full-mode migration diagnosis (-4.23%) due to overfitting. It can be seen that the transfer diagnosis model based on the source working condition knowledge and a small amount of target working condition data is an effective way to solve the lacking problem of the new working condition data.