This paper is a study on the failure prediction of vacuum pumps essential for semiconductor manufacturing processes. Since vacuum pumps suffer great economic losses such as Wafer loss and production suspension in the event of a failure, it is important to detect and prevent the failure in advance. However, since most of them show rapid anomalies just before failure, it is difficult to predict failures with short-term trend analysis. Therefore, this study conducted long-term time series data analysis using 12 types of sensor data from vacuum pumps that operated for about 31 months. Most vacuum pumps have normal data and a small number of abnormal data. Accordingly, it was confirmed that the failure can be diagnosed in advance using an abnormality detection technique to which deep learning is applied. This can provide a technical foundation to prevent production interruptions and economic losses caused by failures.