Research on a deep learning-based automatic identification system for the fracture fibrous ratio in impact fractures of 9Ni steel

Abstract

As a critical structural material for liquefied natural gas storage and transportation, 9Ni steel requires precise assessment of its impact toughness, which hinges on the accurate measurement of the fracture fibrous ratio (FFR) from impact tests—a process currently reliant on manual visual inspection that is subjective, inefficient, and inconsistent. To address this, we propose an automated system integrating an improved recursive kernel mean distance (RKMD) denoising algorithm with a Yolov5s-MobileNetV3-convolutional block attention module (CBAM)-gradient-weighted class activation mapping (Grad-CAM) deep-learning model for high-precision FFR calculation. First, an enhanced adaptive RKMD algorithm is introduced to effectively denoise impact fracture images while preserving essential details across varying noise levels. Subsequently, the MobileNetV3-based network achieves a mean average precision (AP) of 95.27% in feature recognition and exhibits low relative deviation in predicting the FFR under four macroscopic fracture conditions. By incorporating the CBAM attention mechanism and Grad-CAM visualization, the system’s capability to recognize micro-features in fractures with medium-to-low fibrous ratios (65% and 75%) is notably enhanced, leading to a significant reduction in overall prediction deviation. Overall, this system establishes a complete automated workflow from image preprocessing and feature extraction to FFR calculation, offering a reliable and efficient solution for precise toughness evaluation of 9Ni steel.