基于多模态的管道非接触式磁记忆检测方法研究

Research on Multimodal Non-Contact Magnetic Memory Method for Pipeline Defect Detection

摘要: 针对管道缺陷检测中非接触式磁记忆技术存在的数据量大、人工分析效率低及易受外部干扰等问题，本文提出一种结合异常检测算法与图像分类模型的多模态数据处理方法。研究采用孤立森林算法联合局部分析策略对磁信号进行初步筛选，快速定位疑似异常点，并结合图像分类模型识别外部环境干扰，辅助判断并剔除其中可能由外部干扰引起的异常数据。依托自主研发的检测装置实现磁信号、位置信息与图像的同步采集，并在标准化实验场地完成了30组检测任务。结果显示，本方法在铁质干扰与焊缝识别中的准确率分别为80.7%与90%，图像分类准确率达94%，融合分析后，外部干扰剔除率较未引入图像模型提升48%。该方法在特定实验条件下有效提升了磁检测数据的处理效率与准确性，为复杂工况下管道检测中异常数据的识别提供了可行思路。

Abstract: To address issues such as large data volume, low efficiency of manual analysis, and vulnerability to external interference in non-contact magnetic memory testing for pipeline defect detection, this study proposes a multimodal data processing method that combines anomaly detection algorithms with image classification models. The method employs an isolation forest algorithm integrated with local analysis strategies to preliminarily screen magnetic signals and quickly locate suspected abnormal points. It then incorporates an image classification model to identify environmental interference, assisting in the judgment and removal of abnormal data potentially caused by external factors. A self-developed detection device is used to synchronously collect magnetic signals, positional data, and images, and 30 standardized experimental tests were conducted. Results show that the proposed method achieves accuracy rates of 80.7% and 90% in identifying iron-based interference and weld seams, respectively, with an image classification accuracy of 94%. After multimodal fusion, the external interference elimination rate improved by 48% compared to the method without image assistance. Under specific experimental conditions, this method effectively enhances the processing efficiency and accuracy of magnetic detection data, providing a feasible solution for abnormal data identification in pipeline inspection under complex working conditions.