安全与环境学报

为解决天然气阀室管道泄漏扩散预测中由特征混叠带来的扩散边界信息提取难题，提出了一种物理约束动态选择核网络的特征提取方法。首先，设计动态选择性卷积核，以子网络预测不同卷积核权重，并采用双层加权机制将卷积核权重与通道注意力融合，聚焦关键特征；然后，以动态选择性卷积核构建编码器并联合长短时记忆(Long Short-Term Memory，LSTM)网络处理编码特征时序依赖，解码时引入残差连接，增强特征传递；最后，设计物理约束的损失函数，提高特征物理关联性。以两点泄漏为例，通过物理约束动态选择核网络捕捉温度场时序演变规律以验证预测方法，相比于卷积神经网络(Convolutional Neural Network，CNN)与长短时记忆网络联合模型(CNN–LSTM)、物理信息神经网络(Physics-Informed Neural Network，PINN)，所提方法决定系数(R²)分别提升13.3%、1.1%，均方误差(MSE)分别降低65.3%、19.0%，平均绝对误差(MAE)分别降低35.4%、14.3%，能够有效提取泄漏混叠特征，提高了扩散预测精度。

In natural gas valve chambers， accurately predicting leakage diffusion is often hindered by the aliasing of global diffusion trends with local temperature gradients， particularly in multi-point leakage scenarios. To address the challenge of extracting diffusion boundary information obscured by this feature aliasing， this paper proposes a Physically constrained Dynamic Selection of Kernel Network (P–DSKNet). First， a Dynamic Selective Kernel (DSK) module is designed to achieve adaptive multi-scale feature extraction. By employing parallel branches with varying kernel sizes (e.g.， 3×3 and 5×5)， this module utilizes a sub-network to predict convolutional kernel weights. A two-layer weighting mechanism is then applied to deeply fuse these kernel weights with channel attention vectors， allowing the network to dynamically filter out redundant information and focus on critical temperature gradient features. Second， the architecture integrates the DSK module as a spatial encoder with a Long Short-Term Memory (LSTM) network to rigorously model temporal dependencies. In the decoding phase， residual connections are strategically incorporated to superimpose the input features onto the decoded outputs， effectively mitigating information loss caused by upsampling and enhancing the transfer of deep semantic features. Third， to ensure physical rationality， a hybrid physically constrained loss function is formulated. By integrating spatial gradient continuity and temporal consistency constraints， the training process reinforces physical correlations， ensuring that the predicted temperature fields adhere to fluid dynamics principles. Experimental validation on Computational Fluid Dynamics (CFD) simulated two-point leakage scenarios indicates that P–DSKNet achieves superior prediction performance compared to Physics-Informed Neural Networks (PINN)， LSTM， Convolutional Neural Network–LSTM (CNN–LSTM)， and Transformer models. Quantitatively， compared to CNN–LSTM and PINN， the proposed method increases the coefficient of determination (R²) by 13.3% and 1.1%， reduces the Mean Squared Error (MSE) by 65.3% and 19.0%， and decreases the Mean Absolute Error (MAE) by 35.4% and 14.3%， respectively. These results confirm that P–DSKNet effectively decouples aliased features and substantially improves the accuracy and physical consistency of leakage diffusion predictions.