Breaking the data scarcity barrier in HVAC fault diagnosis via feature-sample collaborative augmentation

Data-driven heating, ventilation, and air conditioning (HVAC) fault detection and diagnosis (FDD) plays a vital role in enhancing building energy efficiency; however, the extreme scarcity of real-world fault samples hinders model training and remains the primary bottleneck to practical implementation. Current research largely relies on sample-level oversampling or generative augmentation, which struggles to produce stable and informative data under limited-sample conditions. Meanwhile, feature-level enhancement remains underexplored, which limits the deeper representation of fault mechanisms. To overcome these limitations, this study proposes a feature–sample collaborative augmentation (FSCA) framework that integrates feature enhancement with sample augmentation in a unified manner, thereby enlarging the effective learning space and extracting the maximum value from limited datasets. On the feature side, a spatially correlated feature mapping-based enhancement method is proposed to achieve continuous mapping from correlation modeling to spatial visualization, yielding more discriminative feature representations. On the sample side, a stable augmentation strategy based on an improved denoising diffusion probabilistic model is proposed, incorporating conditional generation and squeeze–excitation attention to enhance the stability, controllability, and fidelity of generated samples. Experimental results demonstrate that FSCA significantly improves FDD accuracy under severe sample scarcity, achieving performance gains of up to 8.63% and 15.18% on two building HVAC systems, and clearly surpassing state-of-the-art methods. These findings highlight the effectiveness and application prospects of FSCA for HVAC FDD, providing a promising solution for data-driven FDD deployment in real-world small-sample environments.