binaryAllNaturalPlusNormalVsAttacks

ZJT datasets: It was collected from the production line of China Tobacco Zhejiang Industrial Company. The data was sampled every two seconds for a week from 162 sensors deployed on a variety of production devices (e.g., paper cut-ting wheel, power supply, etc.). Since ZJT is a dataset from real-world production line, it does not contain serious anoma-lies from accidents or attacks. Thus, we treat the states of transforming between different producing modes as anoma-lies. The ratio of normal states to abnormal states is 4:1.

HAI dataset : The data was gathered from a practical Industrial Control System (ICS) test environment, which was enhanced with a Hardware-In-the-Loop (HIL) simulation system. This system was designed to mimic the processes involved in steam turbine power production and pumped-storage hydroelectric power generation. The dataset collection spans 11 days. It contains data collected every second from 84 sensors and actuators. The anomalies are generated from 50 cyber-attacks. The training set and the testing set are explicitly separated in HAI, where the training set is collected without anomalies and the testing set contains 1/40 abnormal samples.

PS datasets : It was collected by Mississippi State Uni-versity and Oak Ridge National Laboratory. It involves five types of anomalies, including short-circuit fault, line mainte-nance, remote tripping command injection, relay setting change, and data injection. The PS dataset was collected from 128 sensors. The ratio of normal states to abnormal states is 6:4.