Weld penetration state identification based on time series multi-source data fusion

Abstract

Addressing issues such as low efficiency in feature extraction, suboptimal feature quality, and low identification accuracy, this paper proposes an innovative Time Series Identification Method (TSIM) of weld joint penetration states based on multi-source data fusion. Firstly, a convolutional block, based on the architecture of convolutional neural networks, is designed. This forms part of an image representation network composed of three such blocks, incorporating a channel attention mechanism for high-quality image representation. Additionally, welding current and voltage data are integrated to create a comprehensive multi-source dataset. Secondly, an innovative atrous convolutional block is introduced, incorporating Efficient Channel Attention Networks (ECANet) to enhance the processing of multi-source data. Leveraging Temporal Convolutional Networks (TCN), an Efficient TCN (ETCN) with an advanced attention mechanism, is proposed. It is designed to extract global spatial features from the time series multi-source data, while concurrently feeding these data into a Transformer encoder to complete the extraction of time series features. Ultimately, a novel network utilizing a cross-attention mechanism is developed to amalgamate the time series and spatial features of the multi-source data, facilitating the prediction of Back-Side Bead Width (BSBW) and identification of the subsequent joint penetration state. Experimental findings indicate that irrespective of variations in welding current, the proposed algorithm achieves an RMSE of 0.17 mm. This represents a reduction in root mean squard error(RMSE), mean absolute error(MAE), and relative absolute error(RAE) values, along with an increase in R-square(R\(^{2}\)) value, surpassing the performance of existing methods such as CNN, ResNet, CNN-LSTM, and AE-GRU.