The Heating, Ventilation and Air Conditioning (HVAC) systems are complex and prone to failures during operation, often leading to significant energy waste. Timely and accurate Fault Detection and Diagnosis (FDD) can enhance energy efficiency. The HVAC system operates under diverse conditions, data-driven models trained under existing conditions may experience performance degradation when faced with new conditions. Transfer learning offers an effective solution to this issue. This study proposes a novel transfer learning ensemble model based on adaptive weights, leveraging different transfer learning strategies to improve diagnosis performance under new conditions. Multiple cross-condition transfer learning tasks were implemented to test the proposed method, and its effectiveness was validated through multiple experiments to minimize the impact of randomness. Results showed that, compared to fine-tuning (FT), domain-adversarial neural network (DANN), and baseline models, the proposed method outperforms the other models. The average accuracy of multiple experiments improved by 0.21 % to 2.34 % compared to FT. Additionally, modifying DANN to utilize a small amount of labeled information from the target domain has led to greater overlap between the feature distributions of the source and target domains, resulting in improved performance that is close to that of FT. Finally, we analyzed the impact of target domain data volume on the performance of the four methods. The performance of the baseline model improved significantly with the increase in data volume, while the other models showed less improvement. Meanwhile, the diagnostic results of the baseline model were significantly influenced by experimental randomness when there is less training data, whereas the FT diagnostic results were relatively more stable.

中文翻译：

---

基于自适应权重分配的集合方法的冷水机组跨条件故障诊断


供暖、通风和空调 （HVAC） 系统很复杂，在运行过程中容易出现故障，通常会导致大量能源浪费。及时准确的故障检测和诊断 （FDD） 可以提高能源效率。HVAC 系统在各种条件下运行，在现有条件下训练的数据驱动模型在面对新条件时可能会出现性能下降。迁移学习为这个问题提供了一个有效的解决方案。本研究提出了一种基于自适应权重的新型迁移学习集成模型，利用不同的迁移学习策略来提高新条件下的诊断性能。实施了多个跨条件迁移学习任务来测试所提出的方法，并通过多次实验验证了其有效性，以最大限度地减少随机性的影响。结果表明，与微调 （FT） 、域对抗神经网络 （DANN） 和基线模型相比，所提出的方法优于其他模型。与 FT 相比，多次实验的平均准确率提高了 0.21 % 至 2.34 %。此外，修改 DANN 以利用来自目标域的少量标记信息，导致源域和目标域的特征分布之间有更大的重叠，从而获得接近 FT 的性能改进。最后，我们分析了目标域数据量对四种方法性能的影响。基线模型的性能随着数据量的增加而显著提高，而其他模型则表现出较少的改进。 同时，当训练数据较少时，基线模型的诊断结果受到实验随机性的显著影响，而 FT 诊断结果相对更稳定。