A significant amount of unlabeled data are idle when labeled data with known operating statuses are used in the fault diagnosis of conventional refrigeration systems. Therefore

a semi-supervised fault diagnosis method is proposed in this study for refrigeration systems based on the Tri-Training method

which uses the information in unlabeled data to improve fault diagnosis performance. Seven typical faults in a 316 kW centrifugal chiller are used for verification

and the results show the effectiveness of the method. The semi-supervised fault diagnosis model based on Tri-Training can mine unlabeled data information and afford significantly improved performances compared with three typical supervised diagnosis models

i.e.

support vector machines

K-nearest neighbor

and random forest. The overall diagnostic accuracy of the proposed system is 99.43%; meanwhile

the diagnosis accuracy of its system-level fault is higher by 1.73%–3.90%

and its false

neglectful

and wrong alarm rates are improved compared with those of the abovementioned three models. Meanwhile

the performance and diversity of the three initial classifiers in the fault diagnosis model are the main factors affecting the use of unlabeled data in refrigeration systems.