<p>In complex noise environments, conventional fault detection methods often suffer from severe interference, resulting in low identification accuracy and high misclassification rates, particularly for partial discharges and ground faults. To address these challenges, this paper proposes a real-time cable fault detection framework that integrates Temporal Convolutional Networks (TCNs) with a self-attention mechanism. The TCN employs dilated causal convolutions to capture long-range temporal dependencies, while the self-attention module enhances feature discrimination by adaptively weighting critical fault patterns and suppressing noise. To improve robustness under high-dimensional data and limited labeled samples, a joint dimensionality reduction strategy combining principal component analysis (PCA) and t-distributed stochastic neighbor embedding (t-SNE) is introduced, ensuring efficient feature extraction and improved interpretability. Furthermore, transfer learning is employed to migrate and fine-tune pre-trained parameters from related industrial domains, thereby enhancing model generalization in data-scarce conditions. Extensive simulations, covering scenarios such as insulation breakdown, overcurrent, ground faults, and partial discharges, demonstrate that the proposed method achieves superior detection accuracy with significantly reduced false alarms compared to traditional approaches. The framework also exhibits strong noise resistance, computational efficiency, and real-time performance, providing both theoretical support and practical value for intelligent cable monitoring in smart grids.</p>