The corrosion behavior of copper in atmospheric environments is significantly influenced by regional climate. To achieve rapid, in-situ identification and provenance tracing of corrosion status, this study proposes an intelligent analysis method integrating real-field images and explainable machine learning. Departing from planar specimens, this work employs three-dimensional pure copper tubes for greater engineering representativeness. These were exposed outdoors for 2–24 months in three typical environments: Wenchang (marine), Chongqing (urban), and Qingdao (industrial–marine) of China. A dataset of 1500 macroscopic corrosion images was constructed using smartphones, addressing challenges like uneven illumination on curved surfaces. A set of 24 color and texture features was extracted, and six machine learning models were employed for regional classification. The Support Vector Machine (SVM) achieved optimal performance, attaining 99.67 % test accuracy using data mixed from all exposure periods. All models achieved test set accuracies of 96 % or higher, confirming the method's capability to capture regional corrosion visual features that are resistant to temporal variations. Explainability techniques, including feature importance analysis and SLIC superpixel heatmaps, revealed the mean blue and red channel values, blue channel standard deviation, and image entropy as the most critical discriminative features. These features possess clear physical correlations with the chemical composition, spatial distribution, and morphological complexity of corrosion products, constructing an interpretive bridge between macroscopic image features and microscopic corrosion mechanisms. This study validates the feasibility of using smartphones and machine learning for high-accuracy corrosion provenance tracing in complex, real-world scenarios, providing a scientific paradigm and practical foundation for intelligent corrosion diagnosis.


扫码关注我们
求助内容:
应助结果提醒方式:

