在综合能源系统（Integrated energy system, IES）中，大量能源耦合设备的接入形成了多能互补，促进了工业园区低碳、高效运行。与此同时，IES规模快速扩大，结构日益复杂，也为故障诊断带来了新挑战。利用分布式多传感器采集的数据能够学习关于故障的抽象特征表示。然而采样过程难以避免受到噪声和数据随机丢失的影响，且多域互联特性使多变量受到故障信号的扰动。面对以上问题，本文利用不完备数据展开故障诊断方法研究，提出了衡量样本重要度的超限学习算法，构建了具有时空特征学习能力的深度学习框架。本文主要研究内容如下： (1) 针对IES在冗余采样下的样本类别不平衡问题，提出了样本重要度衡量机制下的超限学习算法，并通过多分类器推断融合解决采样过程中存在的多源不确定性，提升了IES故障诊断的准确性。利用分类模型对输入样本的敏感性分析计算样本重加权向量，并推导了基于最小二乘优化的模型权重更新算法，因此同样适用于动态实时更新快的采样场景。 (2) 针对IES在非理想采样下的连续数据缺失与节点特征缺失问题，提出了高斯混合模型表征与注意力机制下的时空图模型，实现了不依赖数据填补的端到端学习。在空间特征提取模块通过邻接矩阵引入了IES的拓扑结构知识，并利用高斯混合模型表征缺失值；通过将图数据空间特征提取嵌入门控循环单元实现了时空特征融合；并引入焦点损失联合优化数据表征与故障分类的模型参数。 (3) 针对IES部分区域的故障数据稀缺问题，提出了跨域场景迁移下小样本目标域数据的元度量模型，通过故障知识迁移提升了新环境新区域的故障诊断性能。基于不同区域故障信号趋势的相似性，在元学习框架下利用源域数据构建丰富的小样本分类任务；以距离度量模型作为基分类器，学习类别恒定特征表示方法。 (4) 针对IES故障信号跨域传播特性影响下复合故障样本稀缺问题，构建了具有多尺度故障特征学习能力的二阶多标签分类框架，给出了IES多区域复合故障以及区域内故障的识别策略。考虑不同区域故障信号的耦合特性，以多尺度时空图模型作为基分类器构建成对标签推断融合机制，根据复合故障和相关单一故障样本在特征空间分布的相似性解耦复合故障成分。 最后以陕鼓集团能源互联岛和首钢集团钢铁生产园区作为研究对象，通过多组不完备数据下的故障诊断案例说明了本研究所提方法的有效性。

Nowadays, a substantial number of integrated energy systems (IES) are built with various energy-conversion equipment such as machines of combined heat and power generation, which promotes the low-carbon and efficient operation of industrial estates. At the same time, the rapid expansion of scale and increasingly complex structure of IES also bring new challenges to fault diagnosis. Exploiting the data collected by distributed multi-sensors, the abstract feature representation of faults can be learned by learning algorithms. However, it is difficult to avoid the impact of noise inference and random missing data in the sampling process. In addition, the connection characteristic of multi-region in IES also results in multiple regional state variables being affected by the propagated fault signal. In view of the above difficulties, in this dissertation, the fault diagnosis methods of IES are investigated utilizing incomplete process data. The main research contents of this dissertation are listed as follows: In order to solve the IES problem of class imbalance with redundant sampling, a weighted extreme learning machine is proposed under a mechanism of sample importance measurement, and multi-classifier inference fusion is accomplished to handle the multi-source uncertainties in the sampling process, which improves the accuracy of IES fault diagnosis. Based on the sensitivity analysis of the classifier output to the input samples, the sample reweighting vector is obtained. The updating algorithm of the weight matrix corresponding to the model structure is derived based on the least square optimization. The proposed method is also suitable for dynamic scenarios with real-time sampling and fast updating. For the massive missing data situation of random time series and node features, a spatial-temporal graph neural network model is proposed based on Gaussian mixture model representation and attention mechanism to realize end-to-end learning independent of data imputation. In the module of spatial-feature extraction, the topological structure knowledge of IES is integrated into an adjacent matrix, and the missing values are represented by the Gaussian mixture model. The spatial-temporal feature fusion is realized by embedding the spatial-feature extraction into a gated recurrent unit. The focal loss is introduced to jointly optimize the model parameters of missing-data representation and fault classification. Aimed at the scarcity of fault samples in some regions of IES, a meta-metric model is proposed to improve the fault diagnosis performance through fault knowledge migration with a small sample in the target domain which is a new region with a new environment. Considering the similarity of fault-signal trends in different regions, we utilize the source-domain data to construct a rich amount of small-sample classification tasks under the framework of meta-learning. Taking the distance measure model as the base classifier, the constant-feature representation manner of each class is learned. For the compound-fault diagnosis with scarce compound-fault training samples under the cross-domain propagation influence of fault signals in IES, a second-order multi-label classification framework with the capability of learning multi-scale fault features is constructed. A complete diagnosis strategy for IES multi-region compound faults and intra-region faults is proposed. Considering the coupling characteristic of fault signals in different regions, we construct a paired-label inference fusion mechanism with a novel multi-scale spatial-temporal graph neural network model as the base classifier. The compound faults are decoupled exploiting the distribution similarity of a compound fault and its concomitant single-fault components in the feature space. In this dissertation, the distributed energy system of ShaanGu Group as well as the iron and steel production estate of ShouGang Group are taken as the research objects. The effectiveness of the proposed methods are illustrated by several fault diagnosis experiments with incomplete process data.