概率神经网络(PNN)因训练方法简单且具有较好的分类能力而广泛应用于储层参数预测、裂缝识别及地震属性模式识别.在勘探初期,往往会遇到小样本量的情况,为获得好的模式识别效果,有必要对平滑参数和训练样本的选取方法进行研究.在分析了平滑参数对网络分类符合率的影响后,利用取值试验得到样本归一化情况下平滑参数的最优取值区间.在此基础上进行训练样本选取的随机性、均匀性及数量试验,发现均匀选取各类训练样本时,小样本量能使网络获得较高的分类符合率,而大样本量则能得到更高的分类符合率.X工区的实际应用结果表明,概率神经网络在少井情况下具备一定的应用潜力,可作为勘探初期利用地震属性进行模式识别的一种选择.%The probabilistic neural network (PNN),with simple training method and good classification ability,can be utilized in pattern recognition of seismic attributes.However,the number of well logs is lack at the early stage of seismic exploration,which leads to the sample size of the pattern recognition using seismic attributes is small.In this case,the critical issues are also the smoothing parameter and the selection of training samples,which can determine the recognition accuracy of the network.Firstly we analyzed the influence of smoothing parameter on the classification coincidence rate of the network.Then,after a series of experiments with regard to selection of smoothing parameter and training samples,we proposed a method of training sample selection and an optimal interval of smoothing parameter to achieve the high classification coincidence rate in the context of sample normalization.Finally,the application of real seismic data in X area indicated that the PNN has application potential in the case of less well logs and can be used as an alternative in pattern recognition of seismic attributes at the early stage of exploration.
展开▼
