**1. Introduction**

At present, the steel industry is facing unprecedented challenges including resource consumption, serious environmental pollution, substandard process and product stability, and low productivity [1]. Steelmaking is a typical process industry, with long production processes, complicated manufacturing processes, and many process control factors involved [2]. The changes in product types and raw materials of different companies will be different, and it is difficult for knowledge-based models to adapt to all changes, which makes the migration and maintenance of models difficult. Therefore, the deep integration of information technology and the steel manufacturing industry, as the entry point for industrial upgrading, is of grea<sup>t</sup> significance to the realization of intelligent and green steel production.

Continuous casting is the most critical part of steelmaking [3]. Stable and high-quality continuous casting production is the top priority of iron and steel enterprises. Continuous

**Citation:** Wu, X.; Jin, H.; Ye, X.; Wang, J.; Lei, Z.; Liu, Y.; Wang, J.; Guo, Y. Multiscale Convolutional and Recurrent Neural Network for Quality Prediction of Continuous Casting Slabs. *Processes* **2021**, *9*, 33. https://dx.doi.org/10.3390/pr9010033

Received: 15 October 2020 Accepted: 21 December 2020 Published: 25 December 2020

**Publisher's Note:** MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

**Copyright:** © 2020 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https://creativecommons.org/ licenses/by/4.0/).

casting is the process of solidifying molten metal into semifinished slabs and rolling them in a finishing mill [4]. As shown in Figure 1, the molten metal is transferred from the ladle to a tundish and slowly injected into the continuous caster. Then, the crystallizer in the continuous caster shapes the casting and rapidly solidifies and crystallizes. In this process, the mold level fluctuation will greatly affect the quality of continuous casting slabs (CCSs). With the sharp fluctuation of the liquid level in the mold, the content of oxide inclusions under the slabs will increase significantly [5]. However, mold level fluctuation is likely to cause slag entrapment of molten steel, which further leads to the deterioration of slab quality.

**Figure 1.** A schematic diagram of the continuous casting process.

Major steel producers are leveraging information technology such as the Internet of Things (IoT) and embracing big data to change the current state of the steel industry [6]. The use of sensor-based data acquisition systems in factories and the explosive growth of steel data make data modeling and analysis possible [7]. Furthermore, over the last decade, intelligent technologies, represented by data mining [8] and neural networks [9], have been developed from the theoretical research into their industrial applications. In the field of steelmaking, numerous scholars focus on the classification of steel surface defects [10,11]. Although continuous casting is the main process phase affecting the final quality of the steel products, the continuous casting system has a large number of complex input parameters; thus it is well adapted for big data analysis. Lei et al. have used machine learning methods to develop an offline system for continuous casting data collection and data mining [12], a small amount of research work involves the classification and prediction of continuous casting slabs quality. Nandkumar et al. [13] predicted and improved the quality of iron casting with the Six Sigma approach. A two-layer feedforward backpropagation neural network model was developed to predict the possibility of defects in foundry products [14]. The feedforward backpropagation neural net is out of practice currently, and the vanilla recurrent neural net performs poorly in engineering. Artur et al. designed a specific convolutional neural network (CNN) to detect stickers during continuous casting [15]. Although their method can reduce false alarms, when CNN is used alone for detection, the effect is not respectable. Indeed, we have incorporated two neural net architectures into our multiscale convolutional and recurrent neural network (MCRNN) to build one more robust and better network.

In this work, based on the process data acquisition system, a real-time prediction closed-loop control system was constructed to predict and improve the quality of CCS. In the system, a framework composed of an MCRNN is proposed for real-time quality prediction of CCS. Various conversions are made at different times and frequencies to obtain time series data for fluctuations in the level of the original mold. The CNN can apply

to time series analysis of sensor data well, and it can also be used to analyze signal data with a fixed-length period. Feature extractors based on the fully convolutional network (FCN) and long short-term memory (LSTM) are used to capture long-term dependencies and extract local features of time series, respectively, and we use the advantages of CNN to automatically learn features [16] in the downsampling transformation representation and frequency domain, extracting features of different time scales and frequencies and solving the limitations of many previous features that can only be extracted at a single time scale [17,18]. As a result, the proposed MCRNN enhances feature representation and improves the performance of quality prediction compared to traditional time series classification models. Moreover, the number of normal samples is much larger than the number of abnormal samples. Average production is 100 slabs, with production of only 5 abnormal slabs. We use the random undersampling (RUS) method to reduce the number of majority classes to address the class imbalance. We introduced expert knowledge into the system. When the predictive model detects an abnormal slab, the continuous casting process adjusts in real-time based on expert knowledge, which improves steelmaking efficiency and slab quality.

The organizational structure is as follows: In Section 2, we review the work related to time series classification. In Section 3, we describe our proposed MCRNN and established system in detail, which is the core section of the paper. In Section 4, we present the detailed process and experimental results of the method. Finally, in Section 5, we draw the main conclusions of this work.

## **2. Related Work**

In our real world, time series data are ubiquitous; examples include temperature, click volume, stock prices, and sensor data. They are sequential data of real value type with a large amount of data, high data dimensions, and constant updating of data. In the data-driven era, there is an increasing demand for information extracted from time series, the main task of which is time series classification (TSC). It is a long-standing problem involving a wide range of practical applications, such as the classification of financial time series [19], the judgment of individual agricultural land-cover types [20], and early churn detection [21].

Traditional time series classification methods are mostly based on distance measurement. Lines and Bagnall [22] proposed nearest neighbor classifiers with elastic distance measures to improve classification accuracy. In particular, the dynamic time warping (DTW) distance combined with the nearest neighbor classifier has proved to be a strong baseline [23]. Nevertheless, the performance could be rarely acceptable when it was applied to the engineering field with big data. There are other methods of distance measurement and spatial transformation for time series, such as information entropy [24], weighted dynamic time warping (WDTW) [25], and shapelet transformation [26]. Moreover, enhanced weighted dynamic time warping [27] and distributed fast-shapelet transform [28] were proposed to improve the performance of times series classification. Based on ensemble schemes and data conversion, Bagnall et al. not only aggregated different classifiers on the same transformation but also collected different classifiers in different time series representations [29]. However, these methods only have linear separability.

In recent years, deep learning has developed rapidly and achieved excellent results in classification tasks. Convolutional neural networks and recurrent neural networks are widely used in image recognition [30], video classification [31], machine translation [32], information extraction [33], and other fields. CNN can use convolutional layers to learn complex feature representations automatically, with the advantage of absorbing a large amount of data to learn feature representations. In recent years, many neural networks for time series classification, such as multilayer perceptron (MLP), fully convolutional network (FCN), and residual network (ResNet) [34], emerged. Convolutional neural networks (CNN) have been applied to time series applications, though CNN is mainly for the image field [35,36]. In the classification of high-dimensional time series, Zheng et al. proposed to

use a multichannel convolutional neural network for modeling [37]. The echo state network (ESN) is a time-warping invariant, limited to static patterns rather than temporal patterns, and was applied to time series classification tasks [38]. Joan et al. studied the use of a time series encoder and established a hybrid deep CNN with an attention mechanism [39]. For the quality prediction system, however, these present methods cannot meet the demands of overall continuous casting slab production pipelines.
