Abstract Deep learning-based soft sensor has been a hot topic for quality variable prediction in modern industrial processes. Feature representation with deep learning is the key step to build an accurate and reliable soft sensor model from massive process data. To deal with the limited labeled data and abundant unlabeled data, a semi-supervised pre-training strategy is proposed for deep learning network in this paper, which is based on semi-supervised stacked autoencoder (SS-SAE). For traditional deep networks like SAE, the pre-training procedure is unsupervised and may discard important information in the labeled data. Different from them, SS-SAE automatically adjusts the training strategy according to the given data type. For unlabeled data, it learns the shape of the input distribution layer by layer. While for labeled data, it additionally learns quality-related features with the guidance of quality information. The proposed method is validated on two refining industries of a debutanizer column and a hydrocracking process. The results show that SS-SAE can utilize both labeled and unlabeled data to extract quality-relevant features for soft sensor modeling, which is superior to multi-layer neural network, traditional SAE and DBN.

中文翻译：

---

一种新的深度网络半监督预训练策略及其在工业过程质量变量预测中的应用

摘要 基于深度学习的软传感器一直是现代工业过程中质量变量预测的热门话题。使用深度学习进行特征表示是从海量过程数据构建准确可靠的软传感器模型的关键步骤。针对有限的标记数据和丰富的未标记数据，本文提出了一种基于半监督堆叠自编码器（SS-SAE）的深度学习网络半监督预训练策略。对于像 SAE 这样的传统深度网络，预训练过程是无监督的，可能会丢弃标记数据中的重要信息。与它们不同的是，SS-SAE 根据给定的数据类型自动调整训练策略。对于未标记的数据，它逐层学习输入分布的形状。而对于标记数据，它还在质量信息的指导下学习与质量相关的特征。所提出的方法在脱丁烷塔和加氢裂化工艺这两个炼油工业中得到了验证。结果表明，SS-SAE 可以同时利用标记和未标记的数据来提取与质量相关的特征进行软传感器建模，优于多层神经网络、传统的 SAE 和 DBN。