Multivariate time series anomaly detection algorithms has important research significance in many application fields such as system state estimation, fault prediction and diagnosis, and network behavior anomaly detection. Due to the large noise, high dimensionality, and hidden features of abnormal data in various application data, it is difficult to learn abnormal features in multivariate time series data, which brings many challenges to anomaly detection. In response to these problems, we propose a deep neural network model utilizing convolutional network and Long short-term memory (LSTM) to achieve unsupervised anomaly detection algorithm. This method consists of two steps using feature fusion and timing prediction. The convolutional network automatically learns the correlation among features in multivariate time series data, abstracts the features of multiple timing data. And LSTM deep neural network can effectively identify anomalies in sequences and predict the offset value of time series data. We applied this method to server fault diagnosis and internal network traffic anomaly detection, and obtained significant application effects, realizing anomaly detection in complex scenarios. Extensive experimental evaluations over large multivariate time series data have shown that our method has outperformed other existing time series anomaly detection methods based on traditional data mining methodologies.