基于XGBoost特征重要度的储罐缺陷ANN面积量化模型

Quantitative model of ANN area of tank defects based on XGBoost feature importance

针对超声波检测的储罐缺陷的面积量化问题,提出一种改进的储罐腐蚀缺陷面积量化模型。该模型利用XGBoost的特征重要度对人工神经网络(ANN)的参数进行先验初始化实现ANN模型的改进。该模型可以更快的收敛,并且提高准确率。按照国家标准设计实验平台,获取实验信号,并提取信号的统计特征得到特征数据集,利用数据集训练和测试改进的模型,并与传统模型进行对比。通过实验验证得出,改进的ANN模型能够更快的收敛,并且准确量化缺陷面积,相比于ANN量化模型,在训练集上准确率提高了17.9%,达到了98.3%,在测试集上提高了16.6%,达到了92.2%。

In order to solve the problem that quantifying the area of tank defects detected by ultrasonic wave, an improved quantitative model of tank corrosion defect area is proposed. This model uses the feature importance of XGBoost to initialize the parameters of artificial neural network(ANN) a priori to improve the ANN model. The model can converge faster and improve the accuracy.Design an experimental platform according to national standards, obtain experimental signals, and extract the statistical features of the signals to obtain a data set. Use the data set to train and test improved models, and compare them with traditional models. The experimental results show that the improved ANN model can converge faster and quantify the defect area accurately. Compared with the ANN quantization model, the accuracy in the training set has been improved by 17.9%, reached 98.3%.and increased by 16.6% on the test set, reached 92.2%.