Tham khảo tài liệu 'aerospace technologies advancements fig part 6', kỹ thuật - công nghệ, cơ khí - chế tạo máy phục vụ nhu cầu học tập, nghiên cứu và làm việc hiệu quả | Evaluation of Anomaly Detection Capability for Ground-Based Pre-Launch Shuttle Operations 163 due to realizations being based on independent experiments and as such should still fall within the confidence bands formed in both Figs. 11 and 12. The light blue highlights superimposed over the failure injected scores in dark blue represent the ground truth time of failure injection and duration so as to give a feel for the false alarm and correct detection rates. Evidently there is a clear bifurcation between nominal and anomalous scores for both IMS and SVM and for Orca the same is true although it is less apparent. As can be discerned from Figs. 11 - 13 we have identified the fact that both in complexity and accuracy IMS seems to be the best choice among all of the algorithms investigated. However there is some overlap in the confidence intervals for IMS and SVM AUC values and the alert thresholds applied for both corresponding ROC curves yield almost identical true positive rates. 6. Conclusion and next steps We have provided a thorough end-to-end description of the process for evaluation of three different data-driven algorithms for anomaly detection. Through optimization of algorithmic parameters using the AUC we were able to choose parameters yielding the best detection capability. The respective ROC curves corresponding to these parameters were then used to inform alert threshold selection by enforcement of a maximum allowable false alarm rate. It was found that IMS was the best performing algorithm when considering both computational complexity and accuracy. However when evaluating the results based upon accuracy alone the OCVSM approach is competitive with IMS due to overlapping confidence intervals present in the accuracy results. In subsequent research studies we will provide results of unseen hold out test cases to which optimized parameters and thresholds will be applied in order to provide additional evidence demonstrating the superiority of a particular
