Manifold Regularization for Time Domain Fault Location

The patent replaces the mechanical/computational process of solving differential equations with a machine learning model that has been trained offline. The trained model directly estimates fault location from impedance measurements without requiring real-time differential equation solving, thus substituting a complex computational mechanism with a simpler, faster inference process while maintaining estimation accuracy.

The patent performs the complex differential equation solving and model training in advance (offline phase) to create a pre-trained machine learning model. During actual fault detection (online phase), the pre-trained model is used for rapid estimation without requiring real-time differential equation solving, thus preparing the solution beforehand to eliminate computational delays during critical fault events.

The patent segments the training process into distinct phases: data preprocessing, feature extraction, model training, and validation. The heterogeneous high-dimensional data is processed through multiple stages including clustering, dimensionality reduction, and feature selection before being fed to the machine learning model. This segmentation allows complex tasks to be broken down into manageable steps, improving both accuracy and training efficiency.

The patent applies manifold regularization with adaptive weighting that dynamically adjusts parameters during training based on data characteristics. The regularization parameter and weight coefficients are optimized to balance model complexity and generalization performance, allowing the model to effectively handle high-dimensional heterogeneous data without requiring excessive computational resources or training time.