Time series anomaly detection plays a pivotal role in ensuring the reliability of industrial and aerospace systems, where complex temporal dynamics and diverse anomaly types pose significant challenges. Reconstruction-based methods, which detect anomalies by reconstructing input data and identifying deviations, struggle to handle non-stationary data and complex temporal dependencies, resulting in poor detection of subtle anomalies. Recently, contrastive learning has rapidly advanced in this field, offering robust representation learning for anomaly detection. However, most contrastive learning approaches focus solely on the time domain, overlooking critical frequency-domain information, which limits their ability to capture periodic patterns and subtle anomalies in non-stationary data. To overcome this limitation, we propose FAD-CRL, a frequency-aware framework that leverages dual-domain representations and cross-domain contrastive learning. By integrating time- and frequency-domain insights, FAD-CRL enhances robustness and detection precision across diverse scenarios. Extensive experiments on five benchmark datasets, show that FAD-CRL outperforms traditional methods and achieves near state-of-the-art F1-scores, demonstrating its potential for real-world applications.