Article Preview
TopIntroduction
The blast furnace, as the cornerstone of the steel manufacturing process, stands as the world's largest chemical reactor, boasting a staggering capacity of up to 
. Inside the furnace, a complex interplay of up to 108 chemical reactions takes place (Li et al., 2017). The stable and seamless operation of blast furnaces not only guarantees safe production but also serves as a prerequisite for cost reduction, efficiency improvement, and quality enhancement. In the era of cutting-edge technologies, such as big data and artificial intelligence, harnessing blast furnace production data for comprehensive digitization and intelligence has emerged as an effective means of optimizing manufacturing processes for steel enterprises (Han et al., 2018; Kawahata et al., 1988; Zagoskina et al., 2019). Notably, the detection of furnace anomalies with the aid of data analysis has emerged as a focal point of research in the realm of blast furnace intelligence application.
The blast furnace is a vertical reactor that produces liquid pig iron from coke, iron ore (natural rich ore and sintered ore and pellets), and flux (limestone and dolomite) in a continuous process. The blast furnace has five sections: the throat, the furnace body, the furnace waist, the belly, and the hearth from top to bottom. The reliability of the furnace is decided by various factors such as raw materials, operation, environment, and equipment. Typical anomalies are hanging burden, slipping burden, channeling, chilled hearth, low stockline, scaffolding, and hearth accumulation.
Furnace anomalies can be detected and diagnosed by two main categories of methods: expert system-based methods and data-driven methods. Expert systems use process knowledge and historical experience to construct rules and knowledge bases for furnace anomalies. These methods can achieve high accuracy and meet process requirements. However, these methods are unscalable and show low tolerability to status changes in the furnaces. Moreover, they require high maintenance costs. Some examples of blast furnace expert systems are AGS system by Kawasaki Steel Corporation in Japan, SIMETAL BF VAiron by Primetals Technologies, AI-based expert systems by Baosteel (Dou et al., 2015), etc.
Data-driven anomaly detection algorithms have been a popular research direction in recent years (Abdel-Sayed et al., 2016; Shi et al., 2020; Zeberli et al., 2021). They can be categorized into four types: (1) reconstruction algorithms, (2) clustering-based methods, (3) multivariate statistical methods, and (4) improved PCA algorithms.
Reconstruction algorithms, such as Auto Encoder (AE) (Chen et al., 2018), Variational Auto Encoder (VAE) (An & Cho, 2015), Auto-Regressive Integrated Moving Average (ARIMA) (Yaacob et al., 2010), and Prophet (Thiyagarajan et al., 2020), use deep learning or machine learning techniques to detect anomalies based on the reconstruction errors between the original and the generated data. Clustering-based methods, such as Density-Based Spatial Clustering of Applications with Noise (DBSCAN), detect anomalies by identifying the samples that are isolated from the main clusters (Çelik et al., 2011). Multivariate statistical methods, such as Principal Component Analysis (PCA), reduce the dimensionality of multidimensional parameters and detect anomalies based on the deviation from the normal distribution (Ringberg et al., 2007). Improved PCA algorithms, such as Convex Hull PCA (B. Zhou et al., 2016) and PCA-Independent Component Analysis (PCA-ICA) (P. Zhou et al., 2020), enhance the performance of PCA by incorporating additional features or constraints. Data-driven methods usually have advantages in generalization and maintenance. However, the key to their successful application lies in how to utilize data effectively while aligning with the process requirements.