Introduction <p>Planetary gearboxes are widely used in high-load and high-precision technology such as wind turbines, electric vehicles, and aerospace systems. Assembly errors in planetary gears usually generate only weak and subtle vibration signals, making anomaly detection a challenging task.</p> Objective <p>This paper proposes a vibration signal classification framework that integrates multidomain feature fusion and dimensionality reduction techniques and applies it to the early diagnosis of planetary gear assembly errors.</p> Methods <p>Multiple features from both time and frequency domains are combined, and early and late fusion strategies are compared. To address the problem of high-dimensionality data caused by feature fusion, four dimensionality reduction methods are employed: principal component analysis, linear discriminant analysis, minimum redundancy maximum relevance, and an autoencoder. This study also proposes the energy-efficiency metric (EEM) for quantifying the trade-off between a model’s accuracy and training cost. An experiment was conducted with five conditions: a control condition with no assembly errors, and four conditions with planetary gear assembly errors. Additionally, 20% Gaussian noise was added to evaluate the model’s robustness. Results The results revealed that LDA combined with the early fusion strategy performed best in all evaluations. Despite retaining only four feature dimensions, this method achieved classification accuracy of 99.9% and obtained the highest EEM value. The t-distributed stochastic neighbor embedding visualization results confirmed that this method has excellent class separation ability.</p> Conclusion <p>This study provides a high-performance vibration signal processing solution suitable for edge computing and real-time applications.</p>