基于加速梯度求法的置信规则库参数训练方法

Parameter Training Approach for Belief Rule Base Using the Accelerating of Gradient Algorithm

置信规则库(belief rule base,BRB)的参数训练问题实质上是一个带有约束条件的非线性优化问题,目前在求解该问题上主要使用FMINCON函数及群智能算法,但在算法的应用中存在移植性差,难实现,计算时间长等局限性。通过对这些问题的研究,结合现有的参数训练方法提出了基于加速梯度求法的置信规则库参数训练方法,并将其应用在多峰函数、输油管道泄漏检测的置信规则库的参数训练上。以收敛误差、收敛时间和皮尔森相关系数作为衡量指标,对新方法与其他传统方法进行了对比,实验结果表明,新算法在收敛精度和收敛速度上具有更理想的综合效益。

The problem of training parameters for belief rule base （BRB） is essentially a nonlinear optimization problem with constraints, which is mainly solved by the FMINCON function or the swarm intelligence algorithms. However, these approaches have many shortages, such as poor portability, difficult to be implemented and requiring a large amount of calculation. To solve these problems, this paper proposes a new parameter training approach for BRB using the accelerating of gradient algorithm, which is improved from the existing parameter training methods, and is applied to the parameter training of multimodal function and pipeline leak detection. The proposed approach is compared with other traditional approaches in terms of convergence error, convergence precision and Pearson correlation coefficient in experiment analysis. The results show the better comprehensive benefits of the proposed approach, including convergence accuracy and convergence speed.