Anomaly detection on time series data is an important research topic in various domains, which enables information systems to prevent unexpected failures and enhance their security. While conventional time series anomaly detection mainly focuses on single-scale models, it remains challenges to develop a multi-scale model to detect diverse anomalies in the forms of point outliers, short-term failures, and long sequence anomalies. In this paper, we propose MSEAD, a multi-scale anomaly detection method based on ensemble learning, which improves the detection accuracy through the combination of multiple scale detectors. The proposed model trains several prediction models on different scales based on the mechanism of multi-scale convolution for feature extraction, and calculates the anomaly scores based on the predicted values on multiple scales. Finally, ensemble polling is used to vote for each point in the series to form the results of multi-scale anomaly detection. Extensive experiments based on five datasets show that the proposed model outperforms the state-of-the-art anomaly detection methods.