为了能够尽早发现燃气轮机的压气机轴承在运行期间出现的故障问题，本文在传统门控循环单元（Gated Recurrent Unit, GRU）神经网络的基础上，采用改进麻雀搜索算法（Improved Sparrow Search Algorithm, ISSA）进行超参数优化，并结合核密度估计(Nuclear Density Estimation , KDE)，提出了一种基于ISSA-GRU-KDE的故障预警方法。首先对压气机相关历史数据进行预处理和特征筛选得到高质量数据集，以此建立ISSA-GRU的压气机正常轴承温度预测模型，利用预测残差作为故障预警阈值的选取标准，采用KDE拟合残差以确定预警阈值，并通过滑动窗口分析消除干扰，最终实现了故障预警。结果表明，该方法比其他预测算法拥有更高的预测精度和泛化能力，且有效地监测了潜在故障隐患，能够提前数小时对压气机轴承进行故障预警，对燃气轮机运维具有一定的指导意义。

In order to be able to detect the failure of compressor bearings of gas turbines during operation as soon as possible, this paper uses the Improved Sparrow Search Algorithm (ISSA) to perform hyperparameter optimization on the basis of the traditional Gated Recurrent Unit (GRU) neural network, combined with Nuclear Density Estimation. KDE), a fault early warning method based on ISSA-GRU-KDE is proposed. Firstly, the relevant historical data of the compressor is preprocessed and the feature screening is carried out to obtain a high-quality data set, so as to establish the normal bearing temperature prediction model of the compressor of ISSA-GRU, use the prediction residual as the selection standard of the fault warning threshold, use the KDE fitting residual to determine the early warning threshold, and eliminate the interference through sliding window analysis, and finally realize the fault warning. The results show that this method has higher prediction accuracy and generalization ability than other prediction algorithms, and effectively monitors potential fault hidden dangers, and can warn the compressor bearing fault several hours in advance, which has certain guiding significance for gas turbine operation and maintenance.

国家自然科学基金项目（52006131）；国家电网公司华东分部科技项目（H2021-111）