Scholarlink

<p>Change point detection in sequential data becomes substantially more challenging when datasets exhibit both autocorrelated noise and a high concentration of transient outliers. We present a two-phase framework that first detects candidate change points using the DeCAFS algorithm under an AR(1) noise model, and then classifies each detection as a sustained structural shift or a recoiled transient outlier using a Fourier Probabilistic Neural Network (FPNN). The framework incorporates three methodological contributions beyond the base DeCAFS detector: (i) a Bayesian Online Change Point Detector (BOCPD) used as a labelling oracle during training to produce principled sustained/recoiled annotations; (ii) a local Extreme Value Index estimated via Generalized Pareto Distribution fitting, introduced as a fifth classification feature encoding tail behaviour; and (iii) a class-weighted FPNN with SMOTE-balanced training, evaluated under balanced accuracy, Matthews Correlation Coefficient, and other robust metrics. The framework is evaluated on three datasets: the <Emphasis FontCategory="NonProportional">well-log</Emphasis> benchmark, the <Emphasis FontCategory="NonProportional">oilwell</Emphasis> industrial sensor dataset, and US Industrial Production monthly growth rates with NBER recession ground truth. In Monte Carlo simulations (<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(B=200\)</EquationSource> </InlineEquation> synthetic replications drawn from a per-dataset empirically-calibrated synthetic prior), the FPNN achieves a mean balanced accuracy of 0.795 and AUC-ROC of 0.866 on the <Emphasis FontCategory="NonProportional">well-log</Emphasis> benchmark (with comparable performance on <Emphasis FontCategory="NonProportional">oilwell</Emphasis>: 0.804 and 0.845), outperforming logistic regression, isolation forest, one-class SVM, feedforward, and GRU baselines. The full pipeline executes in under 17 seconds for series of length <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(n \approx 4000\)</EquationSource> </InlineEquation>.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A Data-driven Approach to Detecting Change Points in High Concentration of Outliers and Autocorrelated Noise

  • 期刊论文

  • 发表日期：

    2026年5月28日

查看全文
  • Agniva Das,
  • Muralidharan Kunnummal

摘要

Change point detection in sequential data becomes substantially more challenging when datasets exhibit both autocorrelated noise and a high concentration of transient outliers. We present a two-phase framework that first detects candidate change points using the DeCAFS algorithm under an AR(1) noise model, and then classifies each detection as a sustained structural shift or a recoiled transient outlier using a Fourier Probabilistic Neural Network (FPNN). The framework incorporates three methodological contributions beyond the base DeCAFS detector: (i) a Bayesian Online Change Point Detector (BOCPD) used as a labelling oracle during training to produce principled sustained/recoiled annotations; (ii) a local Extreme Value Index estimated via Generalized Pareto Distribution fitting, introduced as a fifth classification feature encoding tail behaviour; and (iii) a class-weighted FPNN with SMOTE-balanced training, evaluated under balanced accuracy, Matthews Correlation Coefficient, and other robust metrics. The framework is evaluated on three datasets: the well-log benchmark, the oilwell industrial sensor dataset, and US Industrial Production monthly growth rates with NBER recession ground truth. In Monte Carlo simulations ( \(B=200\) synthetic replications drawn from a per-dataset empirically-calibrated synthetic prior), the FPNN achieves a mean balanced accuracy of 0.795 and AUC-ROC of 0.866 on the well-log benchmark (with comparable performance on oilwell: 0.804 and 0.845), outperforming logistic regression, isolation forest, one-class SVM, feedforward, and GRU baselines. The full pipeline executes in under 17 seconds for series of length \(n \approx 4000\) .