Time series analysis is a critical task in numerous high-stakes domains, yet the performance of deep learning models is often compromised by domain shift when encountering new data distributions. While Unsupervised Domain Adaptation (UDA) presents a potent solution, conventional methods for time series are often ill-suited, as they typically perform coarse-grained alignment and neglect the rich information within the frequency domain. To address these deficiencies and the challenges of fine-grained alignment, this paper proposes a Time-Frequency Multi-Adversarial Adaptation (TFMADA) framework. The framework designs a dual-branch encoder that simultaneously extracts information from the time and frequency domains via joint self-supervised learning to enhance feature transferability. Subsequently, it introduces a multi-discriminator adversarial mechanism to decouple and align the dual-domain features, achieving more robust domain-invariance learning. Finally, through a teacher-student model combined with a dual-screening criterion, it generates high-quality pseudo-labels to precisely guide the class structure alignment in the target domain. Extensive experiments on three real-world datasets demonstrate that the performance of TFMADA is significantly superior to that of various mainstream baseline methods, validating the method’s effectiveness and superiority in handling complex time-series domain shift problems.