A comparative study of multiclass fault classification in industrial processes: Conventional, binary decomposition, and residual-based classifiers with interval-valued features

Accurate multiclass fault classification is a critical component of process monitoring in complex industrial processes, where numerous fault modes with overlapping signatures, class imbalance, and sensor noise present persistent challenges to classification performance. This paper presents a unified comparative study of a broad family of data-driven multiclass classifiers, spanning conventional single-boundary methods (Linear and Non-linear Support Vector Machines, Random Forests), binary decomposition-based extensions of these methods, interval-valued feature representations (centers, radii, and their concatenation), enhanced binary decomposition with per-pair data-type selection, and residual-based classifiers (Bayesian-optimized Interval Principal Component Analysis and Bayesian-optimized Gaussian Process). The complete framework is evaluated on three distinct case studies of increasing complexity: a nonlinear synthetic dataset, the Tennessee Eastman Process (TEP) with all 20 fault classes, and a bench-scale Gas-to-Liquid (GTL) Fischer–Tropsch reactor. Experimental results show a clear and consistent hierarchy in classification performance. First, conventional classifiers using scalar data achieve mean classification rates lower than interval-concatenated features. Second, binary decomposition further improves separability between features of different classes. Finally, residual-based Bayesian-optimized Interval Principal Component Analysis (BOIPCA) and Bayesian-optimized Gaussian Process (BOGP) classifiers achieve the highest mean classification rate across all case studies.