Abstract: Mineral resources are of great significance to China's economic security and development. Geochemical anomalies are one of the important indicators for mineral exploration. Accurate identification of geochemical anomalies can effectively reduce the exploration scope and enhance the efficiency of mineral exploration. The traditional methods for identifying geochemical anomalies rely on statistical analysis, and the multivariate geochemical data often exhibit complex nonlinear characteristics, making it difficult to accurately identify anomaly data. Deep learning methods are frequently employed for the task of identifying geochemical anomalies due to their ability to automatically capture data features. This study proposes a geochemical anomaly identification method based on a convolutional autoencoder model. This model captures the feature information of different scales of the input data through multi-scale convolution operations. The encoder part can convert complex high-dimensional data into representative low-dimensional feature vectors, while the decoder part maps the low-dimensional feature vectors to the original data space through a series of up-sampling operations. The model utilizes the encoding and decoding processes to fully explore the deep patterns hidden within the data. Finally, anomaly areas are identified by calculating the error between the reconstructed data and the original data. This study trained the model through an unsupervised learning approach, without the need for manual label assignment, thereby avoiding the imbalance issue of positive and negative samples in the data as well as the uncertainty caused by human intervention. The experimental results from the gold deposits in the Hatu area of Xinjiang show that the anomaly areas identified by the convolutional autoencoder are highly consistent with the known gold deposits, and the AUC value is approximately 6.3 % higher than that of the optimal baseline method. Meanwhile, through ablation experiments, the effectiveness of the convolutional autoencoder component was further verified, demonstrating its reliability in the task of identifying geochemical anomalies.