With the continuous development of information technology, security issues from internal networks are becoming more and more important. Many anomaly detection algorithms are designed to identify anomalies, but these algorithms usually do not care about temporal information, or only care about fine-grained temporal information. On the other hand, deep-learning-based algorithms must rely on complex neural networks to capture temporal information, which is difficult to deploy and maintain in actual internal network environments. Addressing the aforementioned problems, this paper proposes a multi-timescale analyzer using wavelet packet transform and semi-supervised learning, and a method that uses the analyzer to modify the original anomaly scores called Multi-timescale Modification (MTM) method. The method is validated using the network flow dataset CIC-IDS 2017 and a real-world dataset. Experimental evaluations show that using the proposed method can obtain better performance than the original algorithms.