Answers are grounded in this paper and cite their sources.
Shows enhanced wheel-flat detection accuracy in laboratory tests, indicating practical embedded application.
Wheel-flat detection models are commonly developed with limited defect data and later adjusted to meet embedded constraints, creating a gap between design and deployment. To address this, model selection is formulated as a deployment-aware multi-objective optimization problem balancing classification accuracy, reconstruction loss, and parameter count. Using a genetic algorithm (GA), axle-box vibration segments from a laboratory-scale test rig are used to jointly search multilayer perceptron (MLP) autoencoder and classifier architectures under four objective cases with different priorities. The trade-offs guide selection of practical operating points, which are checked by inference-time profiling and embedded experiments on a microcontroller unit (MCU). The selected configuration reduces parameters from 111K to 11.7K while maintaining 91% test accuracy, and receiver operating characteristic (ROC) evaluation yields an area under the curve (AUC) of 0.966. The proposed approach provides a lab environment to explore and narrow candidates for embedded use, to facilitate field-stage preparation.
No takes yet. Share an insight, caveat, or question.
Kim et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: