Sample less meta-learning fault diagnosis based on ordered time–frequency features

To address the challenges of insufficient fault samples and incomplete feature acquisition that compromise diagnostic accuracy, this study proposes a meta-learning framework for fault diagnosis integrating temporally-ordered time–frequency feature representation and a self-attention-enhanced multi-scale network. The proposed methodology systematically extracts time–frequency features from raw vibration signals and organizes them into discriminative two-dimensional (2D) image representations through structured temporal sequencing. A neural architecture combining spatial self-attention mechanisms with multi-scale convolutional feature extraction is developed, enhanced by meta-learning strategies to optimize parameter initialization. The model undergoes primary training on publicly available benchmark datasets to establish generalized feature representations, followed by task-specific fine-tuning and evaluation using targeted diagnostic datasets. Comprehensive experimental validation demonstrates the efficacy of the proposed approach, with the ordered time–frequency feature extraction achieving superior precision of 98.28%. The trained network exhibits exceptional few-shot learning capabilities, and when only a single fault sample is available, it attains a maximum diagnostic accuracy of 81.69%, which significantly outperforms conventional methods. Comparative analyses reveal enhanced adaptability and generalization capacity across diverse operational conditions, confirming the framework’s robustness in addressing data scarcity challenges inherent in industrial fault diagnosis scenarios.