The spatiotemporal data across WSN subnets exhibits strong non independent and identically distributed characteristics, and subnets are susceptible to targeted attacks such as label flipping attacks and backdoor attacks, which reduce the security of WSN. To address these issues, an anomaly detection method based on FL optimized joint learning has been proposed. Each WSN subnet generates a private spatiotemporal graph training set based on local data, and uses FL optimization joint spatiotemporal sliding average method to smooth the data to eliminate sensor noise interference, suppress data mutations and abnormal jitter. Normalize the maximum and minimum values of the data within the sliding spatiotemporal window to ensure uniform data distribution and improve the accuracy of subsequent anomaly detection. Build an anomaly detection framework based on FL optimized joint learning, aggregate the federated average parameters of edge nodes in WSN, establish the FL optimized joint learning objective for anomaly detection, and use the federated subdomain fine-tuning mechanism and encrypted parameter sharing, combined with differential privacy and dynamic weight aggregation, to adapt to non independent and identically distributed spatiotemporal data while suppressing targeted attacks, achieving secure and accurate cross domain anomaly detection. The experimental results show that the maximum values of node spatial distribution entropy under label flipping attack and backdoor attack modes are 0.8 and 0.6, respectively, and the simulation errors of RF signal identification are 0.7 and 0.3, which are consistent with the experimental indicators. This indicates that the detection results of this method are accurate and can effectively guarantee the transmission and storage of WSN privacy data.