Abstract
In HVAC systems, various failures increase energy consumption and maintenance costs, highlighting the need for reliable fault diagnosis. This paper proposes a novel approach by integrating time-domain and frequency-domain features into fault diagnosis. The proposed network architecture involves segmenting 1D sensor time series data into 2D variables based on periodic signals derived from the frequency domain. These variables are then processed using the ConvNext feature extractor to capture intrinsic inter-sensor information. Subsequently, deep features are extracted through a multi-layer perceptron layer. The effectiveness of this method surpasses leading-edge machine learning and deep learning algorithms such as XGBoost (XGB), Random Forest (RF), lightGBM (LGBM) and attention-based models, utilizing the ASHRAE RP-1043 dataset. The algorithm achieves a remarkable classification accuracy of up to 100% for individual fault types and an average accuracy of 99.47% in nested cross-validation. However, XGB, LGBM, RF, and FEDformer only achieved 99.29%, 99.20%, 99.02%, and 99.27% accuracy respectively.
Published Version
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call