Planetary gearbox diagnosis is challenged by nonstationary vibration, acquisition-condition variation, and noise. This study proposes a GNA-VMD temporal convolution–Transformer (GNA-VMD-TCT) method. The Geomagnetic Navigation Algorithm jointly optimizes the VMD mode count and penalty factor using a diagnosis-oriented fitness function that combines envelope entropy, modal redundancy, Fisher class separability, and a mode-count penalty. The resulting multi-channel intrinsic mode functions are processed by a temporal convolution–Transformer network integrating multi-scale convolution, gated temporal convolution, Transformer encoding, and cross-gated fusion. Under MAT-file-level splitting and source-validation-based checkpoint selection, five-seed cross-condition Macro-F1 was 31.33 ± 8.12% for Condition 1 → 2 and 29.68 ± 5.73% for Condition 2 → 1, with no consistent advantage over the baseline methods. Retrospective within-condition tests favored Raw-TCT, and operator-level ablations did not confirm consistent gains from every network component. Gaussian-noise tests using fixed seed-0 checkpoints and three noise realizations per SNR likewise showed no consistent robustness advantage. Although the revised runs excluded target data from GNA optimization, network training, and checkpoint selection, target results had been examined during earlier exploratory work, so the evaluation was not fully blind. These findings characterize the limitations of diagnosis-oriented VMD parameter selection under acquisition shifts and noise.
13 August 2026 in Optics and Lasers in Engineering
Aiming at the problem of insufficient multi-scale feature extraction and real-time detection ability of the current valve body surface defect detection model based on machine vision, this paper proposes a visual detection model GME-DETR for valve body surface defects. The model adopts the GMENet backbone network and enhances the adaptive feature extraction ability of the model by designing the GMELAN module. At the same time, the convolutional additive token mixer (CATM) is combined to replace the traditional multi-head self-attention structure to reduce the computational overhead. Through the MSC3 multi-scale feature fusion module, the detection accuracy of small target defects is improved. In addition, by constructing the shape-normalized wasserstein distance loss function, the fitting ability of the model to defects with different shape scales is enhanced. The experimental results show that GME-DETR achieves 94.8% accuracy, 88.9% recall, 92.2% mean average precision (mAP) 50 and 52.2% mAP95 on the valve surface defect dataset. At the same time, the number of parameters and computational complexity are reduced by about 65% compared with the benchmark model RT-DETR. In addition, online real-time inference test, robustness test under different illumination conditions and cross-dataset verification show that the GME-DETR model can meet the requirements of real-time detection in industrial field, and has good robustness and generalization ability in complex industrial environment, which verifies the potential application value of this method in valve body surface defect detection task.
Purpose: To address the non-stationarity of wind turbine gearbox vibration signals and the limitations of fixed-parameter variational mode decomposition (VMD) in fault diagnosis, this study proposes an IHO-VMD-Attention-Fusion-CDBN framework (IHO, improved hippopotamus optimizer; CDBN, convolutional deep belief network) that combines adaptively optimized modal information with the original vibration signal for robust fault representation. Methods: An IHO is developed by integrating enhanced initialization, adaptive nonlinear energy regulation, Lévy-flight exploration, hierarchical exploitation, elite refinement, covariance matrix adaptation (CMA)-style perturbation, and Powell local search. IHO jointly optimizes the VMD modal number K and penalty factor α using a composite envelope-entropy fitness function. The raw vibration signal and optimized VMD components are then fused as multichannel inputs and recalibrated by an input-channel attention mechanism before classification by a CDBN. To avoid information leakage, all diagnostic experiments adopt a strict file-level partition, and IHO-VMD parameter optimization uses training files only. Results: On the IEEE Congress on Evolutionary Computation 2017 (CEC2017) benchmark suite, IHO achieves the lowest average rank among the evaluated optimizers and shows competitive convergence and stability. Under the strict file-level diagnostic protocol, Raw-CDBN remains a strong baseline, and no statistically significant difference is detected between the proposed model and Raw-CDBN. However, under noise-augmented conditions, IHO-VMD-Attention-Fusion-CDBN consistently improves VMD-CDBN by 5.77–6.70 percentage points across the investigated noise levels, with statistically significant differences in all cases. Conclusion: The results indicate that optimized VMD components are more effective as complementary representations than as replacements for the raw vibration signal. The proposed framework provides an optimization-guided raw-modal fusion strategy that reduces the information loss associated with VMD-only modeling while maintaining competitive diagnostic performance under strict cross-file evaluation.