KLS-A: A Full-Life-Time Anomaly Detection Method

Anomaly detection is a significant field in the industry. The anomaly detection system could detect the possible abnormalities of equipment in time, and reduce the human and material resources loss caused by equipment abnormality. It is an effective method to analyze operating status of the equipment by collecting its status data. Anomaly detection usually relies on large amounts of equipment status data. After analyzing a large amount of normal data, a model of normal data is obtained. When abnormal data arrives, the data will be identified by analyzing if the data is different from the existing model. However, a device's Full-Life-Time must go through the cold-boot stage, when a device is cold-booted, the amount of data is usually small. As the operating time of the equipment increases, data gradually accumulates. So, the efficient anomaly detection research usually focuses on the equipment in smooth-running stage. Some traditional anomaly detection methods can be used when the amount of data is small, but they usually have limited anomaly detection capabilities. This paper is motivated by the problem, we have proposed an anomaly detection method, which could run during cold-boot and also fully explore the ability of large amount of data. We call it KLS-A. "KLS" means the three components of the algorithm: KMeans, LSTM and selector, "A" means adaptive, this means our algorithm can be easily transplanted in different equipment. Through the transformation of KMeans and LSTM, KLS-A has strong adaptability, and it can automatically adjust parameters according to the amount of data. By using a customized selector, KLS-A can freely switch algorithms under different data volumes to meet the anomaly detection at different stages of the equipment. Therefore, it can maintain a good anomaly detection effect under different stages. Meanwhile, we implemented KLS-A on Spark and applied it to the monitoring of SINOPEC's hydrocondensation tower. We monitored twenty-three indicators of the hydrocondensation tower and tested the anomaly detection ability from cold-boot stage to smooth running stage. We also used the same data set to do comparative experiments with traditional anomaly detection algorithms. The results show four abnormalities can be predicted at the earliest time by KLS-A and all stages can guarantee a high accuracy rate.