基于DAE-BP神经网络的工业质量预测

Prediction of Industrial Quality Based on DAE-BP Neural Network

摘要:

BP神经网络因具有良好的非线性拟合能力，在建立预测模型中得到广泛应用。但化工过程数据不仅存在非线性特征，而且难以避免受噪声影响，造成数据波动从而影响预测模型准确性。为此，提出一种降噪自编码器融合反向传播算法（简称为，DAE-BP）的化工过程质量预测方法。首先，采用无监督学习模型降噪自编码器完成初始数据的噪声消除，其具有噪声鲁棒性的特点，在数据受到损坏的情况下可尽可能地恢复数据的原始状态，有利于进一步的质量预测。在此基础上，将获取的数据特征作为有监督学习模型BP神经网络的输入以获得可靠的预测结果。在脱丁烷塔化工过程实例上验证方法有效性。并与单一BP算法、主成分分析(PCA)及自编码器(AE)改进的BP算法作为对照。结果表明，经过DAE改进后的BP算法预测误差为1.2%，相比单一的BP算法提高了3.2%精度，较PCA-BP及AE-BP预测误差精度分别提高了2.3%、1.9%，表现出最好的预测性能。

Abstract:

BP neural network has been widely used in building prediction models because of its good nonlinear fitting ability. However, the chemical process data not only has nonlinear characteristics, but also is difficult to avoid the influence of noise, which causes data fluctuation and affects the accuracy of prediction model. Therefore, a DAE-BP method for chemical process quality prediction based on denoising autoencoder fusion was proposed. Firstly, the unsupervised learning model denoising autoencoder is used to eliminate the noise of the initial data, which has the characteristics of noise robustness and can restore the original state of the data as far as possible in the case of data damage, which is conducive to further quality prediction. On this basis, the obtained data features are used as the input of BP neural network of supervised learning model to obtain reliable prediction results. The effectiveness of the method was verified by an example of chemical process of debutanizer column. The results were compared with the single BP algorithm, principal component analysis (PCA) and autoencoder (AE) improved BP algorithm. The results show that the prediction error of BP algorithm improved by DAE is 1.2%, which is 3.2% higher than that of the single BP algorithm, 2.3% higher than that of PCA-BP and 1.9% higher than that of AE-BP, showing the best prediction performance.