基于分布式并行分层极限学习机的大数据多模式质量预测

MULTI-MODE QUALITY PREDICTION OF BIG DATA BASED ON DISTRIBUTED PARALLEL HIERARCHICAL EXTREME LEARNING MACHINE

考虑分布式系统质量预测中的大数据处理问题,提出一种基于分布式并行分层极限学习机(distributed parallel hierarchical extreme learning machine, dp-HELM)的大数据多模式质量预测模型。根据Map-Reduce框架,将高效的极限学习机算法转化为分布式并行建模形式。由于分层极限学习机(hierarchical extreme learning machine, HELM)的深度学习网络结构在特征上具备的预测精度优势,结合深层隐藏层的ELM自动编码器,进一步开发了dp-HELM。通过dp-ELM和dp-HELM以同步并行方式进一步训练分布式并行K均值划分的过程模式,利用贝叶斯模型融合技术来集成用于在线预测的局部模型。将所提出的预测模型应用于预脱碳装置中残留的二氧化碳含量估算,实验结果表明了该方法的有效性与可行性。

Considering the problem of big data processing in distributed system quality prediction, we propose a multi-mode quality prediction model of big data based on distributed parallel hierarchical extreme learning machine(dp-HELM). According to the Map-Reduce framework, the efficient limit learning machine algorithm was transformed into the distributed parallel modeling form. The deep learning network structure of hierarchical extreme learning machine(HELM) had the advantage of prediction accuracy in features. And the dp-HELM was further developed by combining the ELM automatic encoder in the deep hidden layer. Dp-ELM and dp-HELM were used to train the process mode of distributed parallel K-means partition in a synchronous and parallel way. We adopted Bayesian model fusion technology to integrate the local model for online prediction. The proposed prediction model was applied to estimate the residual carbon dioxide content in the pre-decarbonization unit. And the results prove the effectiveness and feasibility of the proposed method.