The multivariate time series (MTS) is observed in various fields nowadays so as to monitor the operating status of equipment like server machines, satellites, engines, etc. Tasks like anomaly detection and anomaly pattern classification of MTS, accordingly, are of great importance to fault warning and system stability. Data-driven methods are increasingly becoming prevailing regarding these problems with a huge amount of input information. In this paper, we propose a new architecture where MTS is processed as images and fed into an attention-based CNN-LSTM network (AC-LSTM). CNN-LSTM network is responsible for extracting both spatial and temporal features among and within each dimension of MTS, while the modified attention mechanism helps promote reasonable weight allocations without the introduction of additional parameters. The proposed model firstly gives the result of anomaly detection and makes a further decision on anomaly classifications if the input is detected as an abnormal one. With the modified attention mechanism and the cooperation of CNN and LSTM, experiments on both public datasets and the operation data collected from real-world applications prove the effectiveness and promotion of the proposed model.