Abstract: To solve the problems of strong time-varying, multiple types, complex faults in the air data, a fault diagnosis method was proposed based on the multi-head self-attention mechanism. The long short term memory（LSTM） was used to extract the time domain features of the faulty data, and the multi-head self-attention（MSA） was combined to extract the spatial location features between different types of data. The multilayer perceptron（MLP） was used to improve the generalization ability of the model, and the troubleshooting process was given. By the method, the fault mutual judgements between different types of data were obtained to realize multiple fault identification, and the method was fully verified by the atmospheric data. The experimental setup was given, and the grid search method with k-fold cross-validation was used to determine the optimal model parameters. To verify the performance of LSTM-MSA model, four deep learning models of MSA-LSTM, LSTM-MSA-P, LSTM-CNN and RNN-MSA were constructed for comparison experiments. To verify whether the diagnostic model can pinpoint faults, the fault classification confusion matrix was constructed using the predicted and true labels of the validation set. To further verify the diagnostic capability of the method, the visualization experiments were conducted based on t-SNE. The results show that the fault recognition accuracy of the proposed method is 96.696% with F₁ of 96.777%, and the misclassification rates of all kinds of faults are controlled below 10%, which illuminates that the diagnosis model has high robustness.