Accurate fault diagnosis of chillers is essential for extending equipment lifespan and reducing energy consumption. Currently, data-driven diagnostic models for chillers exhibit impressive performance. However, the outstanding performance is only guaranteed in condition of sufficient and high-quality data, i.e., data is often uncertain and insufficient. To resolve this problem, this study proposes a novel multimodal co-learning framework based on infrared thermography (IRT) and operational state parameters. State parameters, referred as evidence source 1, undergo data augmentation using conditional Wasserstein generative adversarial networks (CWGAN) before classification by a base classifier. IRTs referred as evidence source 2, are enhanced through a method called self-attention BAGAN with gradient penalty (SA-BAGAN-GP). Self-attention mechanisms is integrated in the encoder layer to capture critical features to produce high-quality samples. Then, generated IRT samples are then classified using the self-attention convolutional neural network (SA-CNN) model. Finally, Dempster-Shafer (D-S) evidence theory is utilized for the fusion of decision information from both modalities. By simultaneously capturing and integrating data from diverse sources, the model improves generalization and robustness. Experimental validation conducted on actual chillers demonstrated an average accuracy of 92.75% across four cross-condition tasks, with noise test accuracy ranging from 89.2% to 99.6% and outlier test accuracy between 98.5% and 99.4%.

中文翻译：

---

一种基于多模态数据融合的数据驱动冷水机组故障诊断数据不确定性和不足性的新解决方案


冷水机组的准确故障诊断对于延长设备使用寿命和降低能耗至关重要。目前，冷水机组的数据驱动诊断模型表现出令人印象深刻的性能。然而，只有在数据充足和高质量的情况下才能保证出色的性能，即数据通常是不确定和不足的。为了解决这个问题，本研究提出了一种基于红外热成像 （IRT） 和运行状态参数的新型多模态协同学习框架。状态参数，称为证据源 1，在通过基本分类器进行分类之前，使用条件 Wasserstein 生成对抗网络 （CWGAN） 进行数据增强。IRT 称为证据来源 2，通过一种称为梯度惩罚的自我注意 BAGAN （SA-BAGAN-GP） 的方法得到增强。编码器层中集成了自注意力机制，以捕获关键特征以生成高质量样品。然后，使用自注意力卷积神经网络 （SA-CNN） 模型对生成的 IRT 样本进行分类。最后，Dempster-Shafer （D-S） 证据理论用于融合来自两种模式的决策信息。通过同时捕获和集成来自不同来源的数据，该模型提高了泛化性和稳健性。在实际冷水机组上进行的实验验证表明，在四个跨条件任务中，平均准确率为 92.75%，噪声测试准确率在 89.2% 到 99.6% 之间，异常值测试准确率在 98.5% 到 99.4% 之间。