基于机器学习的测试数据异常检测方法研究

Research on Anomaly Detection Methods for Test Data Based on Machine Learning

摘要: 面对现代工业与制造业设备运行状态检测、环境监测的在线检测需求，文章提出了一套面向测试数据的异常检测方法。该方法构建统一数据门禁与特征工程流程，结合汉宁权滑动去噪、缺失修补、IQR软截断与稳健标准化，并以定长切窗提取时域与频域联合特征，经互信息筛选后输入孤立森林。模型采用子采样集成与固定污染率先验，阈值以得分0.5并联动分位进行轻量校准。基于温度传感器基准数据集的时间阻塞评估，孤立森林取得精确率0.90、召回率0.94、F1‑score 0.92，优于局部离群因子与3σ法则。在相对标准差5%与10%的噪声注入下，F1‑score分别为0.88与0.82，表现出良好的鲁棒性。结果表明，该方法能够在不依赖严格分布假定的条件下稳定识别漂移、跳变与噪声类异常，具备跨批次上线与滚动再训练的工程可用性。

Abstract: In response to the online monitoring demands for equipment operational status inspection and environmental monitoring in modern industry and manufacturing, this paper proposes an anomaly detection method for test data. The method establishes a unified data access control and feature engineering workflow. It integrates Hanning-weighted sliding denoising, missing value imputation, IQR soft truncation, and robust standardization. Fixed-length sliding windows are adopted to extract combined time-domain and frequency-domain features, which are filtered through mutual information and then fed into the Isolation Forest. The model adopts subsampling ensemble learning and a fixed contamination prior. The threshold is lightly calibrated with a baseline score of 0.5 combined with quantile adjustment. Evaluated via time-blocked validation on a benchmark temperature sensor dataset, the Isolation Forest achieves a precision of 0.90, a recall of 0.94, and an F1-score of 0.92, outperforming the Local Outlier Factor and the 3σ criterion. Under noise injection with relative standard deviations of 5% and 10%, the F1-scores reach 0.88 and 0.82, respectively, demonstrating excellent robustness. The experimental results indicate that the proposed method can stably identify drift, jump, and noise-based anomalies without relying on strict distribution assumptions, and it possesses engineering practicability for cross-batch deployment and rolling retraining.