改进BP神经网络模型及其稳定性分析

Improved BP neural network model and its stability analysis

针对传统BP算法抗干扰能力差、学习速率慢且易陷入局部极小点等缺点,提出一种基于变更传递函数倾斜度和动态调节不同学习速率的BP改进算法,并在此基础上采用Lyapunov稳定性原理分析改进算法的收敛性。该算法综合考虑网络训练方式和学习率的不足,设计新的复合误差函数,同时采用一种分层动态调整不同学习率的新方法,并采用批量样本进行训练,以加快传统BP算法的收敛速度和避免陷入局部极小值点。在此基础上,将该算法应用于带钢表面缺陷图像检测中,并比较改进算法与传统算法在缺陷检测中的性能参数。研究结果表明:该改进算法能够提高缺陷识别率,检测速度快,更能满足钢板表面质量检测的实时性要求,是一种行之有效的方法。

As for shortcomings of classical BP algorithm such as bad anti-jamming ability, low learning rate and easy plunging into local minimum, a new kind of improved BP algorithm was proposed with varying slope of activation function and dynamically adjusting different learning rates. Moreover, the convergence of this improved algorithm was analyzed based on the principle of Lyapunov stability. Considering the deficiency of network training and insufficient learning rate, a new composite error function was invented. A new method of dynamic adjustment of different learning rate was adopted to accelerate the convergence of classical BP algorithm, and to avoid plunging into the local minimum point. The proposed algorithm was applied to the inspection of the surface defective image of steel strips and compared with traditional algorithm with defect detection performance parameters. The results show that the improved BP algorithm has many merits such as high inspection speed, high discrimination and real-time capacity which can satisfy the demand of defect detection on steel plate surface, so it is an effective method.