基于比例损失去噪自编码器的冷连轧轧制力预测分析

Rolling force prediction analysis of tandem cold rolling based on proportional loss denoising autoencoder

为了更加精确地预测冷连轧轧制力,设计了一种通过分层提取和目标相关特征来实现的比例损失堆叠去噪自编码器。首先,通过堆叠去噪自编码器(SDAE)构建深度网络并在SDAE顶层中加入输出层;然后,通过部分有标签样本实现网络权重变量的调节;最后,按照设定目标参数调节深度网络变量,从而降低网络预测值和目标值的偏差。本方法通过在训练过程加入目标值信息实现了特征提取有效性的显著提升,具有很好的预测稳定性。通过试验测试本算法,其预测结果在±3%误差内,可以满足实际生产控制要求。本算法能够从输入层内找到和目标值关联的特征,在预训练阶段完成目标值的整合。相比其他预测算法,本算法获得了很小的预测误差,能够更快完成收敛,表现出了更优的预测精度和效率。

In order to predict the rolling force of tandem cold rolling more accurately, a proportional loss stack denoising autoencoder(SDAE) was designed by the hierarchical extraction and target correlation features. Firstly, the output layer was added to the top layer of SDAE by the stack denoising autoencoder. And then, the weight variables of network were adjusted through some labeled samples. Finally, the depth network variables were adjusted according to the set target parameters, so as to reduce the deviation between network predicted value and target value. Furthermore, by the method of adding target value information in the training process, the effectiveness of feature extraction was improved significantly and had a good predictive stability. The experimental results show that the prediction result of this algorithm is within ±3% error through test, which can meet the actual production control requirements. The algorithm can find the features associated with the target values from the input layer and complete the integration of target values in the pre-training stage. Compared with other prediction algorithms, this algorithm has a small prediction error and can complete convergence faster which shows better prediction accuracy and efficiency.