Recently, deep-learning (DL)-based soft sensor has been widely applied to industrial processes, which plays a vital role for process monitoring, control, and optimization. However, most existing soft sensor models are established for only one quality variable or multiple quality variables with the same sampling rate. There are very few models focusing on prediction for multirate quality variables, especially with DL networks. To handle this problem, a novel DL strategy based on multirate stacked autoencoder (MR-SAE) is proposed. In MR-SAE, the network is composed of two parts: the cascade shared network for joint feature representations and the parallel quality-specific network for individual feature learning and quality prediction. The training procedure consists of three steps for MR-SAE. First, the available input data are used to pretrain the shared network. Then, all the multirate data are used to fine-tune the whole network. Finally, individual fine-tuning is further carried out for quality-specific subnetworks. The proposed MR-SAE method is used to build a unified soft sensor model for predicting both the 50% boiling point and cetane content of diesel oil. The results show that the performance of MR-SAE-based model is superior to SAE and deep belief networks. Moreover, the parameters and training time of MR-SAE model are less than the other two methods.

中文翻译：

---

工业过程中多速率质量变量数据建模的深度学习


近年来，基于深度学习（DL）的软传感器已广泛应用于工业过程中，对过程监测、控制和优化发挥着至关重要的作用。然而，大多数现有的软测量模型都是仅针对一个质量变量或具有相同采样率的多个质量变量建立的。很少有模型专注于多速率质量变量的预测，尤其是深度学习网络。为了解决这个问题，提出了一种基于多速率堆叠自动编码器（MR-SAE）的新型深度学习策略。在MR-SAE中，网络由两部分组成：用于联合特征表示的级联共享网络和用于个体特征学习和质量预测的并行质量特定网络。 MR-SAE 的训练过程由三个步骤组成。首先，可用的输入数据用于预训练共享网络。然后，所有多速率数据都用于微调整个网络。最后，针对特定质量的子网络进一步进行单独的微调。所提出的 MR-SAE 方法用于构建统一的软测量模型，用于预测柴油的 50% 沸点和十六烷含量。结果表明，基于 MR-SAE 的模型的性能优于 SAE 和深度信念网络。而且，MR-SAE模型的参数和训练时间都比其他两种方法少。