基于穷举搜索策略与逻辑回归算法的RC柱地震破坏模式判别模型

为了准确判别钢筋混凝土(RC)柱的地震破坏模式，基于穷举搜索策略和逻辑回归算法，提出了一种RC柱地震破坏模式判别的两阶段逻辑回归模型。基于穷举搜索策略，分别遴选了判别弯曲破坏与非弯曲破坏以及弯剪破坏与剪切破坏的最优特征参数；结合最优特征参数和逻辑回归算法，建立了RC柱地震破坏模式判别的两阶段逻辑回归模型；通过与传统方法进行对比分析，验证了该模型的有效性。分析结果表明：该模型不仅构建了特征参数与地震破坏模式之间的显式函数关系，克服了传统“黑盒”机器学习判别方法存在的预测结果解释性较差的缺陷，而且通过合理遴选最优特征参数，在保证判别精度的前提下合理简化了判别模型函数形式，解决了传统机器学习判别方法存在的判别模型复杂程度高、计算效率低的问题；对于RC柱的三种地震破坏模式，该文模型的总体判别准确率均达到90%以上，比经典逻辑回归算法提高5%左右，比传统经验判别方法提高20%左右。

DISCRIMINATION MODEL OF SEISMIC FAILURE MODE OF RC COLUMNS BASED ON EXHAUSTIVE SEARCH STRATEGY AND LOGISTIC REGRESSION ALGORITHM

In order to accurately classify the seismic failure modes of reinforced concrete (RC) columns, a two-stage logistic regression model was proposed based on an exhaustive search strategy and on a logistic regression algorithm. The optimal characteristic parameters to classify flexure failure and non-flexure failure as well as flexure-shear failure and shear failure were selected respectively based on the exhaustive search strategy. A two-stage logistic regression model to classify the seismic failure modes of RC columns was established by combining the optimal characteristic parameters with the logistic regression algorithm. The classification accuracy of the proposed model was validated by comparing the new method with traditional methods. The analysis results show that the model not only constructs the explicit function relationship between the characteristic parameters and the seismic failure modes, but also overcomes the defect of poor interpretation of the prediction results in the traditional 'black box' machine learning discriminant methods. Moreover, through the reasonable selection of the optimal characteristic parameters, the function form of the discriminant model is reasonably simplified on the premise of ensuring the discrimination accuracy. It solves the problems of high complexity and low computational efficiency of the discrimination model in the traditional machine learning discrimination methods. For the three seismic failure modes of RC column, the overall discrimination accuracy of this model is more than 90%, which is about 5% higher than that of the classical logistic regression algorithm and 20% higher than that of the traditional empirical discrimination method.