Optimized hybrid residual compensation feedforward neural networks for predicting field dependent hysteresis in magnetorheological valves under normal and degraded conditions

Smart damping systems in manufacturing and transportation rely on magnetorheological (MR) valves to adapt
responses under varying loads. Accurately predicting magnetic-field-dependent pressure hysteresis in these
valves is challenging due to strong nonlinearities, possible magnetic saturation, and long-term performance drift.
This paper introduces the first application of optimized Residual Compensation Extreme Learning Machines (RCELM),
Hybrid RC-Feedforward Neural Networks (RC-FFNNs), and a selective retraining strategy for predicting
the valve’s hysteresis under normal and degraded conditions. Sensitivity analysis shows that model depth,
hidden node number, and activation-function choice strongly affect accuracy and training time. The hybrid RCFFNN
embeds Levenberg-Marquardt Artificial Neural Network (ANN) blocks within the RC-ELM, combining
ELMs’ fast learning with ANN simplicity. The architecture provides consistently higher accuracies for both
training and unseen data while keeping a simpler model than RC-ELM. The proposed models achieve RMSE
below 0.4 MPa and R2 exceeding 0.95 on unseen data, outperforming standard ELM and ANN. A two-stage multiobjective
particle swarm optimization (PSO) automatically tunes depth and other hyperparameters, balancing
accuracy and complexity. In a degraded valve scenario, a selective-retraining strategy is proposed to update only
the compensator layers. Hence, it preserves the core valve dynamics and avoids the computational cost of fullmodel
retraining. The updated model is trained based on simulated sparse post-degradation data while maintaining
acceptable accuracy over the valve full range. These innovations provide a practical route for updating a
personalized model for an MR valve under normal or deteriorated conditions.