<p>Open set supervised anomaly detection aims to use some abnormal class samples seen during training to detect anomalies that have been seen or not seen. In order to solve the problem of open set anomaly detection, a multi-level latent residual based anomaly detection method is proposed. By calculating multi-level residual features with reference samples, integrating the anomaly indices of these residual features in multi instance anomaly learning, and fully utilizing these residual features in anomaly binary classification networks, open set anomaly detection is effectively achieved. The experimental results show that this method can not only improve the detection performance of seen anomalies under different sample sizes, but also improve the recognition performance of multiple unseen anomalies when only one type of anomaly is seen. This verifies the feasibility, reliability, and accuracy of applying this method to open set anomaly detection problems. Code and data are available at <a href="https://github.com/Goolubo/MLR">https://github.com/Goolubo/MLR</a>.</p>