Siamese Neural Network-Based Quality-Related Abnormality Detection for Sintering Process

The sintering process serves as a crucial pretreatment unit in ironmaking, supplying the primary raw materials for the blast furnace. Accurately and real-time estimating the quality of the final sintered ore holds significant importance. Nonetheless, the sintering process data is marred by complexities such as nonlinearity, nonstationarity, non-Gaussianity, and dynamic behaviors, which complicate the application of traditional fault detection methods. In this paper, we introduce a novel approach employing Siamese neural networks for quality-related abnormality detection (SNN-QAD) for the sintering process, focusing specifically on detecting abnormalities in the ferrous oxide (FeO) content of sintered ore. Siamese neural networks are initially used to isolate quality-related stationary features, adeptly navigating the data's nonlinear properties. The Wasserstein distance is then applied to gauge the dissimilarity between distributions, tackling the non-Gaussian distribution challenge. These stationary features are subsequently leveraged as input for a regression model, with the FeO content acting as the output variable, to facilitate training. Hotelling's <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$(\mathrm{T}^{2})$</tex> statistic is employed to identify deviations and pinpoint fault occurrences. The paper culminates with the experimental validation of our method using actual sintering process data.