首页|期刊导航|能源研究与管理|CNN与Point-Net在不同训练数据比例下辐射异常识别分析

CNN与Point-Net在不同训练数据比例下辐射异常识别分析OA

Analysis of Radiation Anomaly Identification Using CNN and Point-Net with Different Training Data Proportions

中文摘要英文摘要

为了解决航空伽马能谱测量难以直接反映地面异常细节的问题,本文将Point-Net与CNN引入核环境辐射异常识别任务.基于福岛第一核电站事故17月后的地面与航空测量数据,评估2类算法在不同训练数据比例下的性能.5次独立训练后的预测结果表明:随训练比例由10%增至70%,Point-Net的平均预测准确率由58.22%升至88.03%,较CNN分别提高10.91%、3.71%、7.75%和9.70%,且平均损失值均低于CNN.10%比例下2类算法均易将环境本底误判为低值异常;30%比例下Point-Net预测结果已接近真实标签,而CNN对高值异常区域的预测范围偏大;50%比例下,Point-Net对高值异常区域已取得较好的识别效果,其高值异常的正确识别率较CNN提高21.13%,CNN仍将38.51%的环境本底样本误判为低值异常;70%比例下,Point-Net高值异常的正确识别率达到97.86%,较CNN提高6.26%.证实Point-Net更适用于处理不规则辐射监测数据,可为核应急环境下辐射异常识别的算法选择及训练数据规模的确定提供参考.

To address the difficulty of directly reflecting ground anomaly details in airborne gamma-ray spectrometry measurements,this paper introduces both Point-Net and convolutional neural networks(CNN)into the task of nuclear environmental radiation anomaly identification.Based on ground and airborne measurement data acquired 17 months after the Fukushima Daiichi Nuclear Power Plant accident,the performance of the two types of algorithms is evaluated under different training data proportions.The results show that,after five independent training sessions,as the training proportion increases from 10%to 70%,the average prediction accuracy of Point-Net increases from 58.22%to 88.03%,outperforming CNN by 10.91%,3.71%,7.75%,and 9.70%under the four respective proportions,while its average loss values are consistently lower than those of CNN.At the 10%proportion,both algorithms tend to misclassify environmental background as low-value anomalies.At the 30%proportion,the predictions of Point-Net are already close to the ground-truth labels,whereas CNN exhibits an over identification of high-value anomaly regions.At the 50%proportion,Point-Net achieves good identification results for high value anomaly regions,with a correct recognition rate for high value anomalies that is 21.13%higher than that of CNN.Meanwhile,CNN still misclassifies 38.51%of environmental background samples as low-value anomalies.At the 70%proportion,the correct recognition rate of Point-Net for high-value anomalies reaches 97.86%,which is 6.26%higher than that of CNN.These findings confirm that Point-Net is more suitable for processing irregular radiation monitoring data and can provide a reference for anomaly identification and the required training data proportion in nuclear emergency environments.

吴囿昂;吴和喜;梁益

东华理工大学核科学与工程学院,南昌 330013东华理工大学核科学与工程学院,南昌 330013东华理工大学核科学与工程学院,南昌 330013

能源科技

异常识别卷积神经网络(CNN)Point-Net福岛第一核电站

anomaly identificationconvolutional neural networks(CNN)Point-NetFukushima Daiichi Nuclear Power Plant

《能源研究与管理》 2026 (3)

45-53,9

国家自然科学基金(12265003)

10.16056/j.2096-7705.2026.03.005

评论