针对船用燃气轮机关键运行参数实时监测的难题，采用某型号燃气轮机80,475组历史数据（训练集42,197组，测试集31,112组，验证集7166组），基于深度学习方法（一维卷积神经网络1DCNN、门控循环单元GRU）和机器学习方法（反向传播神经网络BPNN、支持向量回归SVR），开展燃气轮机关键参数软测量建模对比试验。为有效捕捉参数的非线性耦合关系，使用SHAP值驱动的随机森林特征筛选，构建关键运行参数的最优软测量参数组合。建立1DCNN、GRU、BPNN和SVR模型，并对比不同模型的预测性能。研究结果表明：1DCNN模型凭借其优异的局部特征提取能力和时序建模优势，在预测精度、鲁棒性和泛化性方面均展现出显著的优势，对8个关键参数的平均预测均方根误差（RMSE）较次优的GRU模型降低20.3%，平均绝对误差（MAE）较次优的GRU模型降低20.4%，平均决定系数（R2）高达0.994，是解决船用燃机复杂参数软测量问题的有效方案。

To address the challenge of real-time monitoring of key operating parameters in marine gas turbines, a comparative study on soft sensor modeling was conducted using 80,475 sets of historical data from a specific gas turbine model (42,197 sets for training, 31,112 sets for testing, and 7,166 sets for validation). The study employed both deep learning (One-dimensional Convolutional Neural Networks 1DCNN, Gated Recurrent Unit GRU) and machine learning methods (Backpropagation Neural Network BPNN, Support Vector Regression SVR). To effectively capture the nonlinear coupling relationship of parameters, SHAP value-driven random forest feature screening was?utilized to identify the optimal parameter combination for soft sensing. Models based on 1DCNN, GRU, BPNN, and SVR were developed and their predictive performance was compared. Results demonstrate that the 1DCNN model, leveraging its superior local feature extraction and temporal modeling capabilities, exhibited significant advantages in prediction accuracy, robustness, and generalization. It achieved a 20.3% reduction in average root mean square error (RMSE) and a 20.4% reduction in mean absolute error (MAE) across eight key parameters compared to the suboptimal GRU model, while also attaining an average coefficient of determination (R2) of 0.994. This establishes the 1DCNN model as an effective solution for the soft measurement of complex parameters in marine gas turbines.

先进船舶发动机技术全国重点实验室