基于遗传算法优化神经网络权值的损伤识别

Damage Identifacation on the Base of Optimic Neural Network Weight by Genetic Algorithm

针对传统BP神经网络存在着容易陷入局部极小点、训练时间太长等缺点,采用基于浮点编码的遗传算法对BP神经网络的初值空间进行遗传优化。用基于浮点编码的遗传算法来优化BP神经网络的权值,得到最佳初始权值矩阵,并按误差前向反馈算法沿负梯度方向搜索进行网络学习。以弹簧质量系统作为算例,用结构的模态频率变化作为网络的输入向量,结构的损伤位置作为输出向量。对网络进行训练,仿真结果表明:遗传BP神经网络的收敛和诊断能力优于传统BP神经网络,可有效运用到结构的模态参数识别中。

The traditional BP neural network has the disadvantage of being misled to local minimal point, overlong time training etc. In this paper, a genetic algorithm （GA） based on floating-point coding was utilized to optimize the initial-value space of BP neural network. And the optimal initial weight-value matrix was obtained. The method of network learning was analyzed using the error-forward-feedback algorithm with negative gradient searching. Taking a spring-mass system as an example, with the variation of frequency as the input vector and the position of the structure damage as the output vector, the network training was carried out. The result shows that both the convergence speed and diagnose capability of GABP neural network are better than the traditional BP neural network. The method of this paper can also be used in modal parameter identification in structural dynamics.