基于稀疏表示剪枝集成建模的烧结终点位置智能预测

烧结终点位置(BTP)是烧结过程至关重要的参数,直接决定着最终烧结矿的质量.由于BTP难以直接在线检测,因此,通过智能学习建模来实现BTP的在线预测并在此基础上进行操作参数调节对提高烧结矿质量具有重要意义.针对这一实际工程问题,首先提出一种基于遗传优化的Wrapper特征选择方法,可选取使后续预测建模性能最优的特征组合;在此基础上,为了解决单一学习器容易过拟合的问题,提出了基于随机权神经网络(RVFLNs)的稀疏表示剪枝(SRP)集成建模算法,即SRP-ERVFLNs算法.所提算法采用建模速度快、泛化性能好的RVFLNs作为个体基学习器,采用对基学习器基函数与隐层节点数等参数进行扰动的方式来增加集成学习子模型间的差异性;同时,为了进一步提高集成模型的泛化性能与计算效率,引入稀疏表示剪枝算法,实现对集成模型的高效剪枝;最后,将所提算法用于烧结过程BTP的预测建模.工业数据实验表明,所提方法相比于其他方法具有更好的预测精度、泛化性能和计算效率.

Intelligent prediction of burning through point based on sparse representation pruning ensemble modeling

The burning through point (BTP) is a crucial parameter in sintering process, which directly determines thequality of the final sinter. Since the BTP is difficult to directly detect online, it is of great significance to realize the onlineprediction of BTP through intelligent learning modeling and adjust the operating parameters on this basis to improve thequality of sinter. Aiming at this practical engineering problem, a Wrapper feature selection method based on the geneticalgorithm is firstly proposed in this paper, which can select the feature combination that optimizes the subsequent predictivemodeling performance as much as possible. Secondly, in order to solve the problem of easy overfitting in intelligentmodeling of a single learner, a sparse representation pruning (SRP) ensemble modeling algorithm based on the randomvector functional-link networks (RVFLNs) is proposed, namely SRP-ERVFLNs. The proposed method uses RVFLNswith fast modeling speed and good generalization performance as individual base learners, and perturbs the parametersof the base learner to increase the difference between the ensemble learning sub-models. At the same time, in order tofurther improve the generalization performance and computational efficiency of the ensemble model, a sparse representationpruning algorithm is introduced to achieve effective pruning of the ensemble model. Finally, the proposed SRP-ERVFLNsalgorithm is used for prediction modeling of the BTP in the sintering process. Experiments using industrial data show thatthe proposed method has better prediction accuracy, generalization performance and computational efficiency than othermethods.