基于增量式孤立森林的液体火箭发动机异常检测方法

Anomaly detection method of liquid rocket engine based on incremental isolation forest

为解决液体火箭发动机故障标签缺失条件下流数据无监督检测问题,以及满足不同发动机台次和不同工况的自适应检测需求,基于增量学习思想,提出了基于增量式孤立森林的异常检测算法。设计了多工况流数据检测条件下的在线更新策略、异常分数表达式,并通过更新停止策略避免故障数据对模型的污染。利用多台次试车数据对该模型进行验证,并与传统方法进行比较,结果表明,该算法能够对样本异常程度进行量化评价,能够有效检测早期缓变故障,其F1指标较原始孤立森林算法提高了43%,检测及时性优于红线算法和自适应阈值算法。

In order to solve the problem of streaming data unsupervised detection of liquid rocket engine with the absence of fault label,and to enable adaptive detection of various engines and multiple working conditions,an anomaly detection algorithm of liquid rocket engine based on incremental isolation forest was proposed based on incremental learning.The online updating strategy and anomaly score expression for streaming data under various working condition were designed.And the update stop strategy was applied to avoid the pollution of fault on the model.Using a number of test data for analysis and comparison with traditional methods,the result showed that the algorithm can evaluate the anomaly degree of sample quantitatively,and detect the degree of fault effectively.The algorithm had 43%improvement on F_(1) score over original isolation forest algorithm.And its detection timeliness was better than the red line algorithm and adaptive threshold algorithm.