Deep Learning-Enhanced Intelligent Gearbox Health Assessment Framework Integrating Adaptive Feature Extraction and Operator-Guided Classification for Robust Multi-Speed Load Condition Monitoring

• Introduces a novel condition-invariant operator learning for gearbox monitoring. • Combines deep learning and operator-guided frameworks for robust diagnostics. • Achieves high accuracy (96.40%) under varying speed and load conditions. • Decouples degradation effects from operational variability in real-time. • Demonstrates strong noise robustness and generalization across unseen conditions. Gearbox condition monitoring is critical for ensuring the reliability and safety of rotating machinery operating under variable speed and load conditions. However, most existing data-driven fault diagnosis approaches rely on condition-dependent feature learning, which leads to performance degradation when operating conditions vary or when noise is present. Recent deep learning models improve feature extraction capability, yet they remain sensitive to nonstationary operating regimes and lack robust generalization. To address these limitations, this study proposes a novel Condition-Invariant Rotational Disturbance Operator (CIRDO) that reformulates gearbox fault diagnosis as a function-to-function learning problem. The proposed method integrates disturbance representation learning with a Deep Operator Network (DeepONet) based architecture, where operating conditions (speed and torque) are explicitly encoded to construct condition-invariant mappings of vibration responses. This design enables the model to decouple degradation-related dynamics from operational variability. The framework is implemented using Python with PyTorch, ensuring scalability and real-time feasibility. Experiments are conducted on the Multi-mode Fault Diagnosis Dataset of Gearbox Under Variable Working Conditions obtained from Mendeley Data, which includes multiple fault types and compound fault scenarios. The proposed approach achieves an accuracy of 96.40%, representing an improvement of approximately 4–8% over conventional machine learning and deep learning baselines under variable operating conditions. Additionally, the model demonstrates strong robustness to noise and superior generalization across unseen speed–load combinations. The results confirm that operator-based learning provides a powerful and interpretable solution for gearbox condition monitoring. Overall, this study offers a robust and transferable diagnostic framework suitable for practical industrial deployment, motivating further exploration of operator learning in intelligent maintenance systems.