针对当前研究在处理复杂指标体系中指标间相关性不足以及指标筛选机制缺乏明确性的问题，从整体指标体系转化与逐个指标删减2个维度出发，采用统一流形逼近与投影（uniform manifold approximation and projection，UMAP）以及卷积神经网络（convolutional neural networks，CNN）技术，开展复杂水工闸门安全评价指标体系降维研究。构建水工闸门安全评价的初始指标体系，在分析指标相关性的基础上，使用UMAP对原始指标体系进行降维处理；提出一种基于指标值离散化的CNN训练样本生成方法，引入相对变化幅度和敏感度两个指标，以定量评估指标本身及其相对变化对闸门安全综合评价结果的影响，筛选关键评价指标；基于沙坪二级水电站中孔闸门的监测数据，对评价指标体系降维前后的结果进行对比验证，并从多角度探讨2种评价指标体系降维方法的差异及其各自的优缺点。提出的基于UMAP、CNN技术的2种方法不仅实现了复杂水工闸门安全评价指标体系的有效降维，而且为水工闸门安全综合评价与预警提供了重要的前置工具，为相关领域的研究提供了新的视角。

The safety assessment of hydraulic gates is crucial for the regulatory capacity of hydropower stations, as it is closely related to factors such as water levels upstream and downstream of the station, structural stress, gate vibration, and the status of opening and closing mechanisms. Many of the hydraulic gates in operation in China were built in the 1960s and 1970s, and due to the harsh working environment, there are numerous safety risks. Therefore, conducting safety evaluations of hydraulic gates is essential to prevent potential accidents. However, existing studies on gate safety assessments have two main shortcomings in terms of indicator selection and system construction. Firstly, many studies overlook the impact of correlation between indicators in the evaluation system on the assessment results. High correlation between indicators can lead to redundant and interfering information, potentially distorting the evaluation results. This is especially evident in hydraulic systems like hydraulic gates, which have numerous parameters and complex structures. Additionally, as the number of measurement points increases, the scale of the indicator system also grows exponentially, complicating the evaluation process. Secondly, existing studies generally lack the identification and selection process of indicators' influence on the comprehensive safety assessment results. The impact of indicators on assessment results mainly lies in two aspects: the different magnitudes of changes in each indicator and the varying responses in comprehensive assessment values caused by relative changes in each indicator, indicating varying levels of sensitivity. Therefore, in conducting safety evaluations of hydraulic gates, both aspects need to be considered to ensure the rationality of the indicator selection process, the effectiveness of selected indicators, and the simplification of the indicator system. Two approaches, overall transformation and gradual reduction of indicators were employed in this study, utilizing uniform manifold approximation and projection (UMAP) and convolutional neural networks (CNN) to reduce the dimensionality of the evaluation indicator system. Initially, an initial indicator system for the safety assessment of hydraulic gates was constructed. Based on the analysis of indicator correlations, uniform UMAP was used to reduce the dimensionality of the original indicator system. A CNN training sample generation method based on indicator value discretization was proposed, introducing two indicators, relative change magnitude, and sensitivity, to quantitatively evaluate the impact of indicators and their relative changes on the comprehensive safety assessment of gates, thereby selecting key evaluation indicators. By comparing and verifying the results before and after dimensionality reduction of the evaluation indicator system based on monitoring data from the middle hole gate of Shaping II hydropower station, the differences and respective advantages and disadvantages of the two dimensionality reduction methods from multiple perspectives were disscussed. The results indicate that the dimensionality reduction strategy based on UMAP demonstrates significant advantages in reducing inter-indicator correlations and improving computational efficiency, while the CNN-based dimensionality reduction strategy shows more pronounced superiority in maintaining the accuracy of the indicators' physical meanings. These proposed methods not only advance the theory and practice of reducing complex indicator systems but also enhance the rationality of the indicator selection process, the effectiveness of selected indicators, and the simplicity of the indicator system, providing professionals with a quantitative analytical tool. Furthermore, based on the findings of this study, future research will focus on how to more effectively coordinate the weight distribution between indicators for specific applications such as safety assessments of hydraulic gates, propose a more reasonable comprehensive safety assessment method for hydraulic gates, and further develop safety classification identification and warning technologies to offer more comprehensive and in-depth theoretical and practical support in this field.