An Improved Vision Transformer Model for Ship Plating Welding Defect Detection Based on Multi-Source Information Feature Fusion

Welding quality is directly related to the strength and service safety of ship structures. Due to complex imaging environments, severe noise interference, and diverse defect forms, traditional non-destructive testing methods for welds still have shortcomings in detection accuracy and automation. Considering that the formation of welding defects follows a distinct temporal sequence and exhibits specific physical characteristics, this paper introduces a multi-source information fusion strategy to jointly model image features and time-domain signals during welding. A welding defect detection model is proposed that combines the multi-source convolution enhancement module MS-SCM (Multi-source Split Convolution Module) with the information feature fusion module TA-WTM (Temporal-Aware Windowed Transformer). By extracting multi-source features in parallel to enhance the model’s ability to characterize defects of different sizes, and by introducing temporal information fusion and a sliding-window strategy within the attention mechanism, the model’s ability to model the defect space and temporal features is improved while reducing computational complexity. The study conducted experimental verification on multiple types of weld samples, including cracks, pores, incomplete penetration, and no obvious defects. The results showed that the proposed model exhibited certain advantages in detection accuracy and stability.