Electrical leakage diagnostics is central to source-oriented industrial site management. In both conventional workflows and modern machine learning (ML) frameworks, large simulation ensembles are necessary in finite element method (FEM) forward modelling, mainly used for calibration, sensitivity analysis, inversion, and generation of synthetic training data. This requirement imposes substantial computational demands and delays site deployment. To address this limitation, this study provides the first demonstration of designing a physics-informed neural network (PINN) surrogate for the forward step, enabling mesh-free, physics-constrained, high-throughput three-dimensional electrical modeling to support leakage diagnostics. By incorporating physics-based constraints into the loss function, the PINN generates continuous and physically rigorous electrical field predictions while being trained on only 48 samples of the lined facility. Optimized through quasi-static sampling and interior H¹ regularization strategies, the model reduces the validation error by >90 %, achieving a mean absolute error (MAE) of and relative error to 0.2 %. Inference results indicate that the 128-neuron configuration of the PINN achieves computational speeds nearly three orders of magnitude faster than FEM, while reaching high accuracy after only a few hours of training. This PINN framework represents a new paradigm for forward modeling in electrical diagnostics, decoupling downstream detection and monitoring workflows from the FEM bottleneck.
扫码关注我们
求助内容：
应助结果提醒方式：
京公网安备 11010802042870号