PCA-改进RPROP方法的BP算法在音乐信号分类中的应用

Application of BP Algorithm of PCA-Improved RPROP Method in Music Signal Classification

针对标准的BP神经网络仅从预测误差负梯度方向修正权值和阈值,学习过程收敛缓慢,并且容易陷入局部最小值,导致泛化能力不足的问题,提出了一种基于学习经验变学习速率改进的RPROP方法作为BP神经网络权值和阈值更新方法,并与主成分分析法(Principal Component Analysis,PCA)相结合,形成了PCA-改进神经网络算法。同时,采用Matlab软件对四类音乐信号进行分类实验。实验结果表明,改进算法比标准算法的稳定识别率提高2.6%,当稳定识别率达到90%时,用时节省75%,表明该算法可以加快网络的收敛过程,提高泛化能力。

The standard BP neural network corrects the weights and thresholds from the negative direction of the prediction error,which causes the learning process to converge slowly and easily fall into the local minimum,resulting in insufficient generalization ability.In order to solve this problem,an improved RPROP method based on learning experience variable learning rate is proposed as BP neural network weight and threshold updating method,and combined with principal component analysis(PCA)to form PCA-modified neural network algorithm.By classifying the four types of music signals in Matlab software,the final classification results show that the improved algorithm is 2.6%higher than the standard algorithm,and it can save 75%of time when the stable recognition rate reaches 90%.The algorithm can speed up the network convergence process,and improve the generalization ability.