Fault Diagnosis Method for HVAC Sensors Based on Improved 1-D CNN Model and Wavelet Clustering Analysis

Lei Wang^1*Ruoxiao Hu²Jianli Liang²

• ¹Department of Engineering Management, Sichuan College of Architectural Technology, Chengdu, China
• ²Sichuan College of Architectural Technology, Deyang, China

The fault diagnosis method for Heating, Ventilation and Air Conditioning (HVAC) sensors has significant shortcomings in feature scale adaptability, temporal dependence modeling, and small sample generalization, and is sensitive to Gaussian noise disturbance. Therefore, a fault diagnosis method on the basis of improved one-dimensional convolutional neural network and wavelet packet clustering is built. A multi-scale convolution module is introduced and multi-scale temporal features are extracted through 3/5/7 parallel convolution and residual connections. The signal is decomposed into 8 frequency bands using wavelet packet transformation and energy vectors are constructed. After unsupervised clustering using K-means, weight fusion is performed with Softmax output to achieve end-to-end diagnosis. The accuracy was 97.84%, the F1-score was 0.97. Under 30% Gaussian white noise disturbance, the area under the model curve only decreased by 4%, and the instantaneous robustness drop only increased by 0.01 under 10%-30% noise. The fault diagnosis method balances small sample learning, high noise robustness, and low computational deployment, providing a feasible new paradigm for real-time and reliable operation and maintenance of intelligent buildings.