Steady-state parameter-space analysis for interpretable fault diagnosis of centrifugal pumps

This paper investigates a steady-state parameter-space approach for interpretable fault diagnosis of centrifugal pumps under controlled benchmark conditions. The analysis assesses whether physically interpretable coefficients derived from pump head and shaft-torque relations can provide an informative diagnostic representation within the tested steady-state scope. A semi-empirical model is formulated for the pump head and shaft torque, and six trial-wise coefficients are estimated by least squares from repeated steady-state measurements. These coefficients define a compact parameter space associated with hydraulic and mechanical characteristics of the pump. The resulting parameter space is evaluated using fault-dependent signatures, multivariate separability, feature redundancy, parameter repeatability, noise sensitivity, repeated-split stability, bootstrap variability, and feature contribution. The analysis indicates that several head-related coefficients and the hydraulic torque-transfer term contain dominant fault-dependent structure, whereas the speed-related parasitic torque coefficient provides comparatively limited incremental information in the evaluated dataset. This motivates evaluating both full and reduced parameter-space representations in the subsequent diagnostic benchmark. To examine diagnostic sufficiency, ridge-regularized multinomial logistic regression is compared with nonlinear coefficient-space classifiers and a raw-signal CNN baseline using full and reduced parameter spaces. Under the evaluated steady-state benchmark, the regularized linear models achieved performance comparable to the evaluated nonlinear coefficient-space benchmarks, suggesting that the main class-discriminative structure is largely accessible in the proposed parameter space without requiring a more complex classifier for this dataset. The reduced representation also preserved the observed performance of the full representation, but this reduction is interpreted as dataset-specific rather than as a general rule for other pumps or operating regimes. Overall, the results support the feasibility of using the steady-state coefficient-based parameter space as an interpretable diagnostic representation within the evaluated benchmark setting. At the same time, this conclusion should be interpreted within the controlled steady-state benchmark considered here, and broader validation is required for variable-speed, transient, field-noise, and compound-fault conditions.