Abstract:

To address the challenges of strong nonlinearity and time-varying characteristics in chemical industrial processes, this paper proposes a Newton-Raphson based optimizer (NRBO) driven stacked long short-term memory (SLSTM) network for operational state evaluation. The method enhances the expression ability of time series data by stacking multiple layers of LSTM networks and introducing the Dropout layer. At the same time, the second-order derivative optimization characteristics of the NRBO algorithm are used to effectively improve the convergence speed and classification accuracy of the model, avoiding the problem that the traditional LSTM evaluation method is prone to fall into the local optimum in the high-dimensional parameter space. In experimental validation on the Tennessee Eastman (TE) process, the proposed method achieved a prediction accuracy of 99.31%, which is significantly higher than several comparison methods. For identifying non-optimal states, a combined method based on principal component analysis and group lasso regularization contribution (PCA-GLC) is proposed. This method can effectively identify key variables, reduce misjudgments and interference, and provide accurate information for real-time adjustments in industrial processes. Experimental validation on the TE process demonstrated that the proposed method, compared to the PCA-based graph contribution method, more accurately identified key variables and reduced the interference from other variables.

Key words: chemical industrial processes, algorithm, operational status evaluation, long short-term memory network, principal component analysis

摘要：

针对化学工业过程中存在的强非线性和时变特性等问题，提出了一种基于牛顿-拉夫逊优化算法（Newton-Raphson based optimizer，NRBO）驱动的堆叠长短期记忆网络（stacked long short-term memory network，SLSTM）的运行状态评价方法。该方法通过堆叠多层LSTM网络并引入Dropout层，增强了时序数据的表达能力。同时利用 NRBO 算法的二阶导数优化特性，有效提高了模型的收敛速度和分类精度，避免了传统LSTM评价方法在高维参数空间中易陷入局部最优的问题。在Tennessee Eastman（TE）过程的实验验证中，所提方法的预测准确率达到了99.31%，显著优于其他几种对比方法。针对非优状态，提出了基于主元分析和组套索正则化贡献（principal component analysis and group lasso regularization contribution，PCA-GLC）相结合的非优因素识别方法，该方法能够有效识别关键变量，减少误判和干扰，为工业过程的实时调整提供准确依据。在TE过程的实验验证中，所提方法相对于基于PCA的图贡献法，对关键变量的识别更加准确，并且降低了其他变量对结果的干扰。

关键词: 化学工业过程, 算法, 运行状态评价, 长短期记忆网络, 主元分析