A Novel ResMTN-Based Methodology for HVAC AHU Fault Diagnosis

The Backpropagation Multidimensional Taylor Network (BP-MTN) classifier improved based on MTN may face issues such as overfitting and gradient vanishing when solving complex classification problems. To address these issues, this paper proposes a novel classifier, the Residual Multidimensional Taylor Network (ResMTN) classifier. The classifier introduces the idea of Residual Network (ResNet) in the MTN, by adding a direct connection between the input and the fully connected layer beside the polynomial layer of the MTN, that is, using the skip connection to skip the polynomial layer, thus improving the generalization ability of ResMTN. In addition, in order to improve the parameter learning ability of ResMTN, this paper uses the Adaptive Moment Estimation (Adam) algorithm for parameter estimation to improve the stability of training and learning effect. In experiments, we use the ASHRAE RP-1312 public dataset to verify the performance of the proposed ResMTN-based fault diagnosis method for air handling units. Experimental results show that ResMTN has higher average accuracy than some other existing methods, and the training time of the model is relatively short.