A multimode mechanism-guided product quality estimation approach for multi-rate industrial processes

Abstract

Discrete and delayed laboratory analyses of product quality restrict the operational optimization of industrial processes. However, it is challenging to build an accurate online estimation model for product quality because of complex process dynamics, multiple working conditions, and multi-rate characteristics. Therefore, a multimode mechanism-guided product quality variable estimation model is proposed in this study. First, representative features are extracted from high-dimensional and redundant process variables via both feature engineering and deep learning to describe the internal reaction state. Then, the representative features are used to partition the data samples which are used to train the multi-mode long short-term memory (LSTM) network to increase the adaptability of the estimation model. Finally, the LSTM units are cascaded to learn the variation in the quality variable against time to handle the multi-rate problem. The mechanism models are placed in parallel with the LSTM units to guide the learning process. The estimation model utilizes production data, mechanism knowledge and working condition information, which increases model interpretability and adaptability. A zinc fluidized bed roaster is used to illustrate the proposed estimation approach. The simulation results demonstrate the feasibility and effectiveness of the proposed multi-rate estimation approach.

Introduction

In process industries, the acquisition of quality variables plays an important role in process control, monitoring, and optimization [1], [2]. Owing to the high cost and difficulty of maintaining online analyzers, the measurement of some key quality indicators often relies on offline sampling and laboratory analysis. The low sampling frequency and large delay in the measurement of quality variables greatly restrict the implementation of advanced process control approaches. On the other hand, basic process variables e.g., flow rate, temperature, can be measured online using high-rate sensors, leading to the multi-rate characteristic of the industrial process (Fig. 1). The data imbalance between the quality-related variables (laboratory analysis variables) and the basic variables (online measured variables) makes it challenging to build an accurate estimation model for quality variables [3], [4], [5], [6], [7]. The challenges mainly include:

• Multiple working conditions: Affected by the inlet conditions and reaction conditions, an industrial process always runs under different working conditions. The single-mode estimation model with one group of determined parameters cannot satisfy the requirements of industrial application under different working conditions [8]. Besides, the high-dimensional and redundant characteristics make industrial data unsuitable for working condition classification and modeling.

• Complex process dynamics: A model capable of explaining underlying physicochemical laws can guide process operation. As a data-driven model is a black-box model, it is not interpretable, leading to the blindness in process optimization and control. A mechanism model based on domain knowledge has good interpretability and can be used to characterize the process dynamics effectively [9], [10]. However, due to the complicated process dynamics, a comprehensive mechanism model is too complex for process optimization and control. Hence, neither the data-driven model nor the mechanism model can meet the industrial requirements when they are used alone.

• Multiple sampling rates: Because of the multi-rate characteristics of industrial processes, a large number of online measured variables have no corresponding quality variable as the label, whereas traditional data-driven modeling approaches require the same sample size of input and output variables for modeling [11], [12], [13], [14]. If only limited labeled data are used to train the estimation model, then the information of a large amount of unlabeled data will be ignored. Some semi-supervised data-driven modeling methods have been developed to address the multi-rate problem [15], [16], [17], [18]. However, there are some strong assumptions regarding the training data, for example, the labeled data and unlabeled data must obey the same distribution.

Considering the above difficulties, a multimode mechanism-guided product quality variable estimation model (MMEM) is proposed for multi-rate industrial processes. In the modeling framework, a representative feature system is first constructed based on both feature engineering and deep learning to extract representative features from high-dimensional and redundant industrial data. Then, the industrial data is divided into different working conditions based on the pattern differences of representative features to improve the adaptability of the estimation model. Subsequently, data balancing is performed under different working conditions to ensure the training performance of the estimation model under extreme conditions. Finally, a multi-rate estimation model is established, in which cascaded LSTM networks are used to learn the sequential characteristics from representative features and parallel mechanism models providing the labels for each LSTM network. On one hand, it solves the quality variable estimation problem of multi-rate industrial processes. More importantly, it provides a framework for merging the mechanism and data-driven models, thereby improving the interpretability of the overall estimation model.

The rest of this paper is organized as follows. The basic LSTM network is introduced in Section 2. Section 3 details the modeling framework of the MMEM. A case study of a zinc fluidized bed roaster based on industrial data is conducted in Section 4, and the corresponding results are analyzed. General conclusions are given in Section 5.