This paper proposes an unsupervised deep machine learning (DML) method for anomaly detection in unlabeled time series data. The method uses autoencoder for data reconstruction, combined with temporal convolutional networks (TCN), long short-term memory networks (LSTM), and causal masking attention mechanisms (CMA) to model short, medium, and long-term dependencies. This method effi-ciently models time series data while fully exploring short, medium, and long-term dependencies under the premise of preserving the causal structure of the data. Additionally, the method directly uses the output of the causal attention mechanism for parallel decoding, which effectively reduces the model’s training time cost. After data reconstruction, Isolation Forest is used to analyze the reconstruction errors for anomaly detection. Experiments on multiple datasets of various sizes and types demonstrate that the proposed method performs better in anomaly detection.