Most data collected in real-world scenarios is contaminated, adversely affecting the performance of existing anomaly detection methods that typically rely on clean training data to learn normal patterns. While some approaches have made progress in addressing data contamination, they still face the following challenges: 1) Most methods use pixel-level constraints between the contaminated and reconstructed data as the training loss, and the contaminated data is preferentially learned due to the larger error; 2) Current anomaly detection methods based on frequency-domain do not fully utilize the complex spectral characteristics nor the cross-domain consistency of time series data, which limits the further improvement of the anomaly detection performance. This paper proposes a filter-enhanced reconstruction network with domain-adversarial training for multivariate time series anomaly detection (FERDA), which enables the model to extract normal patterns and high-level semantic information more accurately by training with the error of filtered and reconstructed data. FERDA designs a dual-domain contamination filtering module consisting of the multi-head memory module and the complex linear network. The module is trained to learn normal patterns and cross-domain information through adversarial learning of intra-domain semantic consistency and inter-domain distinguishability. Moreover, a two-stage contamination suppression strategy is proposed. The contamination degree of data is quantified by the difference between pre- and post-filtering, and the impact of contamination is further mitigated by two-stage training with different losses. The proposed method outperforms 23 state-of-the-art baselines in experiments on five benchmark datasets across different domains.