Distributed edge devices in the industrial, healthcare, and other application fields frequently contain time series data. Due to the often unrecognizable features it possesses, there are common issues in time series classification tasks based on real-world data, such as `data islands' and labeling errors. To address this difficulty in distributed data environments, a federated temporal filtering framework is proposed. It incorporates the advantages of self-supervised contrastive learning in extracting complex temporal data representations and is combined with the federated learning approach to tackle the privacy and security issues of distributed systems, while also reducing the communication cost. By maintaining a set of benchmark samples on the server, this paper employs a time-series augmented pre-supervised strategy that relies on distinguishing contrast loss and predicting contrast loss. A pre-supervised model with a high-capacity for generalizing time-series characterizations is achieved through a pre-training and fine-tuning methodology in this approach. Meanwhile, a new approach for label noise filtering is introduced, which utilizes pseudo-labels guided by the pre-supervised model to filter the noisy data in the device in concert with local dataset labels, and uses the clean dataset for the training of the global model. Finally, this paper validates the framework's effectiveness across different types of labeling noise, examines the impact of varying baseline data ratios on the constructed framework, and confirms the filtering effects of each loss in the pre-supervised model through ablation experiments.