基于改进支持向量机的超声参数与受载混凝土应力状态研究

Analysis on the relationships between ultrasonic parameters and the stress state in loaded concrete based on improved support vector machines

在研究支持向量机模拟受载混凝土5个超声参数(波速、首波幅值、主频幅值、非线性系数和超声波谱面积)与应力的关系时,为提高其运算效率和寻优结果。通过数值模拟,对比了不同的归一化方式和核函数对寻找超声参数与应力相关性效果的影响;并采用试验探究了经验参数、遍历算法优化参数和粒子群算法优化参数三种支持向量机的模拟效果。结果表明:提出的首项归一化方式更适用于超声参数与应力这一特殊问题,核函数选择径向基核函数效果更优;不同的支持向量机模拟效果差异明显。简单支持向量机模拟效果较差,在低应力阶段尤为明显;遍历算法优化的模型得到的结果效果较好,但是计算时间过长,在低应力阶段判断误差较大,预测的结果高于真实值;粒子群算法优化支持向量机相比于遍历算法大幅减少了计算时间,模拟的结果效果也最好。

The relationships between five ultrasonic parameters(wave velocity, first wave amplitude, main frequency amplitude, non-linear coefficient and ultrasonic spectral area) and the stress state in loaded concrete were simulated by support vector machines in order to improve their operational efficiency and optimisation results. The effects of different normalization methods and kernel functions on the effectiveness of finding the correlation between the ultrasonic parameters and stress state were compared through numerical simulations, and the simulation results by using three types of support vector machines, namely, empirical parameters, traversal algorithm optimized parameters and particle swarm algorithm optimized parameters, were investigated experimentally. The results show that the proposed first normalization method is more suitable for the specific problem of the ultrasonic parameters and stress state, and the radial basis kernel function is a better choice than other kernel functions. The simulation results of different support vector machines differ significantly, and the simulation results of the simple support vector machine are poor, especially in the low stress stage. The results obtained by the model optimized by the traversal algorithm are better, but the time consumption is too high, and the judgment error in the low stress stage is rather great. The particle swarm algorithm optimises the support vector machine to reduce the computation time significantly compared to the traversal algorithm and gives the best simulation results.