It mainly aims at a fault diagnosis model based on self-attention deep cyclic neural network SA-DRNN was proposed to solve the problem that the feature parameters of air conditioning unit fault operation data are coupled and the time series features are difficult to extract. The model divides variables into multiple sub-blocks according to the physical relationship between the air conditioning unit sensor and the system structure, mining the feature information of the air conditioning unit operation data in the sub-blocks, strengthening the influence weight of key features on the fault diagnosis results through the self-attention layer, and then using the self-attention mechanism to weight and fuse the local features output by each sub-block model to build the global features. Normalized exponential function Softmax is used for classification. Compared with the control method, the DRNN-LSTM method with the function of short and long time memory has better performance in the common fault diagnosis of air conditioning units, and has stronger ability to identify system faults.