The crowdsourcing task recommendation method addresses worker selection overload by intelligently matching workers with tasks. However, existing research often overlooks the inherent characteristics of workers and tasks, the long-term behavioral preferences of workers, and abnormal information about workers and tasks, which impacts the accuracy of recommendations. To address these issues, we propose a novel crowdsourcing task recommendation method based on anomaly detection using multi-attribute heterogeneous graph features. This method first establishes the connection relationships between crowdsourcing workers and tasks to construct a worker-task heterogeneous graph. DeepWalk generates initial node features, and the message-passing mechanism in the graph neural network is employed to explore the feature information of workers and task nodes. During message transmission, the graph convolutional autoencoder drives the detection of worker and task feature information. Finally, Bi-GRU captures the long-term behavioral preferences of workers, and link prediction constructs links based on node similarity to complete the recommendation of sensing tasks. We evaluate nine real datasets, and the experimental results demonstrate that the proposed method outperforms the baseline method.