Table 1 Comparative analysis of the proposed method and three existing methods.

Dimensionality reduction based on VAE³⁰

Unsupervised learning, high adaptability, no need for large amounts of labeled data

Suitable for unsupervised learning tasks, particularly when labeled data is limited, enabling effective feature extraction through dimensionality reduction.

Q-Transform and HOG feature extraction³¹

Strong feature engineering, model diversity, high interpretability

Ideal for applications requiring high-quality feature extraction and model selection, especially when model decision processes need to be explained.

Data generated from 3D simulation³²

Enhanced synthetic data, benefits of deep learning

Appropriate for fault diagnosis tasks with limited labeled data, particularly when obtaining real-world data is challenging, by generating synthetic data through simulations.

Proposed

Capable of handling open-set problems, cross-domain fault diagnosis

Suitable for scenarios involving open-set problems, requiring cross-domain learning and fault detection, especially in complex environments.