A multimode mechanism-guided product quality estimation approach for multi-rate industrial processes
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:
- 1.
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.
- 2.
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.
- 3.
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.
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.
Section snippets
Long short-term memory network
The LSTM network is a variant of the standard RNN that can learn the temporal relationships of time-sequential data [19], [20]. The detailed structure of the LSTM network is shown in Fig. 2. Three gate controllers are included in the LSTM unit, namely the forget gate (), input gate () and output gate (), which control what information in time-sequential data should be remembered.
For the Nth LSTM unit, the three gates and the intermediate cell state are calculated as follows:
The mechanism-paralleled multi-rate LSTM estimation network
The modeling strategy of the MMEM is shown in Fig. 3, which mainly contains four parts: a representative feature system, working condition classification, mechanism model and LSTM network.
is the vector of the online measured industrial data, is the vector of the laboratory analysis data, is the estimation vector of the product quality calculated by the mechanism model, is the vector of representative features, is the label set of representative features . First, is extracted by
An industrial case study
To illustrate the proposed MMEM, a zinc fluidized bed roaster in a certain zinc metallurgical plant is studied as a multi-rate industrial process, the schematic of which is shown in Fig. 6.
Conclusion
In this study, a multimode mechanism-guided product quality estimation model for the multi-rate industrial process is proposed. In this estimation model, representative features are extracted from the original industrial data based on feature engineering and deep learning methods. These features are used to categorize working conditions and train the LSTM network. Then, a multi-rate estimation model is built by integrating the mechanism model into the LSTM modeling framework. Finally, an
CRediT authorship contribution statement
Zhenxiang Feng: Conceptualization, Methodology, Software, Writing - original draft. Yonggang Li: Supervision, Writing - review & editing, Validation. Bei Sun: Conceptualization, Writing - review & editing, Validation. Chunhua Yang: Resources, Project administration. Tingwen Huang: Writing - review & editing.
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgments
This project is financially supported by the National Key R&D Program of China (Grant Nos. 2019YFB1704703 and 2020YFB1713700), the National Natural Science Foundation of China (Grant Nos. 61860206014 and 61973321) and the Fundamental Research Funds for the Central Universities of Central South University (Grant No. 2019zzts064) and the Hunan Provincial Innovation Foundation for Postgraduate (Grant No. CX20190120).
References (32)
- et al.
Advances and opportunities in machine learning for process data analytics
Computers and Chemical Engineering
(2019) - et al.
A robust transfer dictionary learning algorithm for industrial process monitoring
Engineering
(2021) - et al.
A multi-rate sampling data fusion method for fault diagnosis and its industrial applications
Journal of Process Control
(2021) - et al.
Combining machine learning and process engineering physics towards enhanced accuracy and explainability of data-driven models
Computers and Chemical Engineering
(2020) - et al.
Information synergy entropy based multi-feature information fusion for the operating condition identification in aluminium electrolysis
Information Sciences
(2021) - et al.
An integrated prediction model of heavy metal ion concentration for iron electrocoagulation process
Chemical Engineering Journal
(2020) - et al.
Semi-supervised selective ensemble learning based on distance to model for nonlinear soft sensor development
Neurocomputing
(2017) - et al.
Robust semi-supervised mixture probabilistic principal component regression model development and application to soft sensors
Journal of Process Control
(2015) - et al.
Unsupervised deep clustering via adaptive GMM modeling and optimization
Neurocomputing
(2021) - et al.
A Gaussian mixture model based virtual sample generation approach for small datasets in industrial processes
Information Sciences
(2021)