多层DAE协同LSSVM的瓦斯突出预测模型

为准确预测瓦斯突出,提出多层去噪自编码器(Multi-layer DAE)搭载最小二乘支持向量机(LSSVM)的瓦斯突出预测模型。多层DAE网络提取瓦斯突出原始数据的有效特征,并链接LSSVM进行突出分类。因突出影响因子边界存在一定的模糊性,从时间角度考虑将其分为动、静态影响因子,并依此对多层DAE网络按照交叉熵规则设计新的代价函数。利用收集的100组真实样本数据,多次实验确定最稳定的模型结构参数。对比分析PCA-LSSVM、LLE-LSSVM、BP神经网络模型,结果表明,该模型有更优越的特征提取能力和预测性能,更适用与瓦斯突出预测问题。

MULTI-LAYER DAE COLLABORATED LSSVM GAS OUTBURST PREDICTION MODEL

In order to predict the gas outburst accurately,this paper presented a gas outburst prediction model based on multi-layer denoising autoencoder(DAE) with least square support vector machine(LSSVM).The multi-layer DAE network was used to extract the effective features of the gas outburst raw data,and collaborated the LSSVM for gas outburst classification.As there was a certain ambiguity in the boundary of the outburst influence factors,it was divided into dynamic and static influence factors from the perspective of time,and we designed a new cost function for the multi-layer DAE network according to the cross entropy rule.In addition,we collected one hundred real sample data,and used them to test the most stable model structure parameters.The same data was utilized to compare and analyze PCA-LSSVM,LLE-LSSVM,LSSVM and BP neural network models to confirm the validity of our proposed model.The experimental results show that the proposed model has superior feature extraction ability and prediction performance,and it is more suitable for gas outburst prediction problems.