基于深度信念网络的制冷系统故障诊断方法与特征学习研究

Fault diagnosis is critical to eliminate faults in refrigeration systems and achieve energy-efficiency and energy-saving goals. Since the refrigeration system works under a wide range of operating conditions and often changes operations, the system operating data show high nonlinear, multi-mode and low-value density characteristics. Due to such data characteristics, the traditional shallow fault diagnosis method could fail to describe the refrigeration system operating data precisely, showing poor diagnosis performance and weak feature extraction ability for some complex system faults under less severity levels. These are key issues of restricting the development of fault diagnosis method and theory for refrigeration systems. .Hence, from the interdisciplinary perspective of data science and refrigeration discipline, the new methods and new theories of fault diagnosis for refrigeration systems are explored in this project. A novel fault diagnosis method based on deep belief networks is proposed to improve the diagnosis performance and feature extraction ability. First, model input data are prepared using data pre-processing methods, i.e., multi-source information integration, missing data reconstruction and data quality improvement. With these methods, information integrity and data quality of the input data are improved. Combining the fault dependence information structure intervention technique and deep network parameter optimization strategy, an optimized diagnosis model is established and can improve the diagnosis performance. The deep fault features are explored layer by layer in the deep belief network. The least dimensional deep features are extracted. These deep features should be strongly sensitive to various faults even under less severity levels but weakly sensitive to different operating conditions. Visualization of deep features should be performed using the neuron node unit sampling method to analyze the deep feature diagnosis mechanism. The variable importance of key sensors in refrigeration systems are evaluated. Impacts of data quality, the deep fault features on the fault diagnosis performance are analyzed and validated. Finally, a novel general fault diagnosis method with deep network structure, high diagnostic accuracy and strong feature extraction ability is put forward suitable for describing the the refrigeration system operating data. The fault diagnosis theories for refrigeration systems will be enriched by the project. This project will provide a set of effective and reliable featured data and methods to address the critical issues on fault diagnosis researches for refrigeration systems.

故障诊断对保持制冷系统高效、节能运行十分关键。因制冷系统多变工况运行，数据呈非线性、多模态、低价值密度特性，传统浅层故障诊断方法存在微小程度复杂系统故障诊断性能弱、特征不显著等关键问题。为提高诊断性能和特征提取能力，从数据科学、制冷学科交叉视角开展制冷系统故障诊断新理论、新方法研究，提出一种基于深度信念网络的新型故障诊断方法。基于多源信息集成、缺失数据修复、质量提升方法提高建模信息完整性和数据质量；结合故障相关信息结构干预和深层网络参数寻优，建立优化诊断模型，提高诊断性能；模型网络结构逐层提取故障强敏感、工况弱影响且维度最小的深层特征，进行特征可视化，解析深层特征诊断机制；评价系统关键测点重要度，探明数据质量、故障特征对诊断性能的影响规律，验证研究并形成符合制冷系统数据特性的、诊断精度高、特征提取能力强的通用故障诊断方法，为突破制冷故障诊断关键问题提供有效、可靠的特征数据基础和方法指导。

故障诊断对保持制冷空调系统高效、节能、可靠运行十分关键。本项目尝试从故障诊断模型的底层数据和特征角度出发，针对传统浅层故障诊断方法存在微小程度复杂系统故障诊断性能弱、特征不显著等关键问题，致力于探索高效、可靠的深度神经网络故障诊断与特征学习新方法。形成如下创新成果：第一，针对制冷空调系统数据特性，提出一种基于可解释深度神经网络的故障诊断模型。设置合理的激活函数和信息卷积过程，在网络传递过程中，避免制冷空调数据中负值有用信息丢失，同时保证输入变量特征顺序稳定一致与物理意义可溯源。不仅获取较高诊断准确率，还有助于解析深层网络模型的诊断过程与决策机制，提高模型的可靠性。第二，提出一种故障导向、基于绝对梯度类激活映射的故障深层特征提取方法。对不同类型、不同水平制冷空调系统故障，逐层提取有效激活的故障强敏感深层特征子集。深层特征专业知识解释合理、数据区分度高。在建模数据体量偏小的少样本情形下，通过特征组合优化深度神经网络模型后，可增强微小程度制冷剂工质充注量异常故障的诊断效果。第三，开发1套基于深度神经网络的制冷空调系统故障诊断与特征学习分析软件，展示模型建模过程与诊断结果，实现深层特征学习过程和提取结果的可视化。第四，构建一种虚拟传感器与贝叶斯推理相结合的现场传感器数据校准方法，可修复低质量的异常数据，有效保证故障诊断建模的数据基础。最后，基于已取得的研究成果，在研究计划内容之外，将所提取的故障诊断用深层特征与预训练-微调的迁移学习方法相结合，可有效提升跨不同工况、跨不同系统场景下的制冷空调故障诊断准确率。深度神经网络的特征学习方法还应用于建筑用能预测模型的优化。本项目很好地完成了既定目标，共发表论文20篇，授权发明专利2项。研究成果不仅丰富了制冷空调系统故障诊断与特征学习方法理论，也为促进故障诊断方法应用提供有效、可靠的特征数据基础，同时为实现国家“3060”双碳目标贡献力量。