人工神经网络在车辆声学分类中的应用

传统反向传播(BP,Back-Propagation)算法虽然解决了多层感知器的收敛问题,但是训练时间长、收敛速度慢。本文针对训练样本分布状态未知的问题,提出了一种有效的加速收敛方法,即对不同的训练样本选择不同的学习率。将这种改进的BP算法应用到履带车与轮式车的声学分类中,明显提高了算法的收敛速度、泛化能力及稳定性,并可根据需要调整两种车辆的识别率。

Application of Artificial Neural Networks in Acoustic Classification of Vehicles

Basically,the conventional back-propagation (BP) algorithm overcomes the convergence problem of the multilayer perceptron,but slower convergence rate and longer training time are some disadvantages as compared with other competing techniques.In this paper,with the consideration of unknown data distribution,an efficient technique is proposed to improve the convergence rate,in which the learning rate coefficients are taken to be variable for different training samples.Based on the improved BP algorithm, experimental results in the application of acoustic classification for tracked and wheeled vehicles indicate a superior convergence rate,generalization capability and steadiness, and that the classification accuracy of vehicles could be adjusted according to request.