最後，我們藉由捕獲率(Catching rate)以及誤報率(False alarm)，並使用交叉驗證的方式，來評估系統的表現與穩定性。;Advances in technology have allowed us to quickly collect streaming time-series data using sensors in the machine. Data mining can help us find useful information from thousands of data and assist managers to make appropriate decisions.
In data mining, anomaly detection is one of the popular technologies. In the past research, point anomalies and subsequence anomalies have been widely discussed, and the whole time series of time series anomaly detection has been rarely mentioned.
In the semiconductor industry, in the process of the cutting silicon ingot, the occurrence of anomalies will damage the product and cause a lot of money and delay in delivery.
Therefore, it is necessary to find the anomalies as soon as possible and let the operators perform maintenance and stop.
Our research data is taken from the machine data of the semiconductor industry factory. In this study, we define the label presented in binary and probability. When the label presented in binary, we find anomaly point with the help of a classification model. Accumulation of sliding windows is performed through anomaly points. We proposed a new alarm rule, which is used to design a monitoring scheme. In the label presented in probability, we define the label as the non-anomaly probability. With the help of the regression model, we observe the difference between the normal and abnormal in the past, propose a corresponding monitoring scheme, and use it as issuing an alarm.
Finally, we using cross validation to evaluate the performance and robustness of the system by catching rate and false alarm.
