在化工生产过程中, 由于不同变量的采样频率不同, 产生了大量无标签数据, 难以有效利用, 造成了数据浪费。此外, 分散式控制系统(DCS)在数据采集过程中, 受环境干扰和测量仪器老化等因素的影响, 会产生大量噪声数据, 增加了软测量建模的难度, 为了解决这些问题, 获知关键产品的质量, 提高企业的效率, 该文提出了一种基于梯形网络的半监督建模及化工产品组分软测量方法, 在利用有标签数据的同时, 充分利用了大量的无标签数据提高软测量模型的预测精度和泛化能力, 针对数据中存在的噪声, 梯形网络在编码器中逐层加入噪声, 然后利用解码器和跳跃连接逐层协同去噪, 重建无噪声的特征表示和输入, 以达到去除噪声的目的。将该方法应用于甲醇制烯烃过程对产品乙烯组分进行软测量, R²达到0.899, 预测效果比常见的有监督学习方法和半监督学习方法更准确。

Objective: During chemical production, the different sampling frequencies of different variables generate a substantial amount of unlabeled data, which is challenging to use effectively, resulting in data waste. Additionally, distributed control systems frequently produce noisy data due to environmental interference and aging measurement instruments, complicating soft sensing modeling. Furthermore, in semi-supervised tasks, unsupervised components can undermine the accuracy of supervised tasks. To address these issues, this study proposes a semi-supervised soft sensing method for product quality based on a ladder network, enabling accurate, timely determination of key product quality and enhancing operational efficiency. methods: A two-step variable screening method—maximum mutual information (MIC) followed by minimum redundancy maximum relevance (mRMR)—was used to screen auxiliary variables. MIC was first applied to eliminate low-correlation variables, and mRMR was then used to remove redundant variables among the auxiliary set, yielding an optimal selection for modeling. The ladder network-based soft sensing method was then established, improving noise resistance by injecting disturbances into each encoder layer and reconstructing noise-free features layer by layer through the decoder. Skip connections were added between encoders and decoders to extract more information from unlabeled data, enhancing focus on supervised tasks and strengthening the model's robustness and generalization. Results: This method was applied to the methanol-to-olefin (MTO) process, termed DMTO. The MIC and mRMR screening reduced 203 auxiliary variables to an optimal 50. After preprocessing, several soft sensor models were established to compare outcomes. Results showed that unlabeled samples improved the effectiveness of supervised soft sensing tasks, with the proposed method enhancing various evaluation metrics. Residual analysis further indicated that the predicted residuals of the ladder network-based semi-supervised method closely aligned with a standard normal distribution, validating the method's superiority. Conclusions: Compared with supervised and other semi-supervised learning methods, the ladder network demonstrates superior prediction accuracy and generalization in soft sensing ethylene products in the DMTO process. The proposed approach offers promising applications for real-time monitoring and control of product quality in chemical production.