Arc Detection in Bow Networks Based on MTF-TLBKA-DarkNet-GRU-MSA and Transfer Learning

Abstract

Investigating the impact of high-speed airflow on the contact pressure and arc state between the carbon slide plate of the pantograph and the contact network during high-speed train operation. By calculating the contact pressure and arc state models that are more aligned with the actual conditions, an experimental model of bow network arc considering the influence of high-speed airflow field is established. A MTF-TLBKA-DarkNet-GRU-MSA Transfer Learning fault detection model is proposed. A Markov transformation field (MTF) is applied to convert a one-dimensional contact voltage signal time series into a two-dimensional image graphically. A pre-trained DarkNet19 model is used as a starting point, and a gated recurrent unit network along with a multi-head self-attention mechanism layer are incorporated to enhance the model's recognition accuracy. For the challenging parameters in the model, such as the learning rate and the number of neurons in the gated recurrent unit network layer, an enhanced Black-winged Kite Algorithm is integrated to optimize the parameters and make the model more robust. Combining the weight sharing characteristics of the DarkNet19 model, migration learning is incorporated to enhance the convergence speed and generalization ability of the classification model. This approach is referred to as TLBKA-DarkNet-GRU-MSA Transfer Learning model is constructed. Finally, the generated 2D images are input into the proposed model for testing. At the same time, the proposed model and the other three models are tested against three groups of bow network arc models with different experimental conditions to verify that the proposed model demonstrates strong robustness and high accuracy.