Anomaly Detection Method Based on One-Class Random Forest with Applications

Aiming at the limitation that the random forest algorithm cannot deal with the problem of anomaly detection, a one-class random forest based on improved Graham scanning method is proposed, which realizes the classification application of the random forest with only a single class of samples. Following the principle and process of Graham scanning method, the concept of boundary softening ratio is introduced to increase the flexibility of the outer boundary of data points. Ray casting is used to generate data set with inverse distribution of input samples to make the traditional random forest model become a one-class random forest with fine decision boundary after training, which outputs the abnormal probability of the data to be tested. The effectiveness of this method for the condition monitoring of rolling bearings is verified on the XJTU-SY bearing data set. The results show that the one-class random forest can accurately separate normal data from degradation data. Adjusting the boundary softening ratio, the balance between outlier monitoring accuracy and true positive rate can be realized. The one-class random forest with hard boundary can achieve detection accuracy of 98.37% and recall rate of 0.972 with threshold for 0.5. At the boundary softening rate of 0.05, the minimum RMSE for post-degradation prediction can be obtained, which gets 1.01% lower than that of the hard boundary. With the increasing threshold, the boundary softening rate provides a strong recall rate guarantee. © 2020, Editorial Office of Journal of Xi'an Jiaotong University. All right reserved.