PaperlayerView exact revision on arXiv

Source-linked AI summary

Towards Universal Representation-Based Process Control

Jinmyeong Choi, Taesup Kim, Artur Dubrawski

arXiv:2609.30790v1cs.LG

TL;DR

Window-level monitoring often relies on tests of predefined statistical hypotheses, although practical process control asks whether each window matches an empirical reference regime. This paper combines learned representations with nonparametric reference modeling and conformal calibration; experiments report well-calibrated performance across deviations, including settings where stable references are cyclostationary.

  • Problem

    Classical window-level tests target predefined parametric hypotheses, limiting assessment of consistency with task- or domain-specific empirical reference behavior.

    Source: §1 · pp. 1–2

  • Method

    The framework embeds windows with a pretrained time-series encoder and applies nonparametric reference modeling with conformal calibration for finite-sample-valid consistency assessments.

    Source: §1 · p. 2

  • Results

    The framework achieves strong, well-calibrated performance across a wide range of deviations and accommodates stationary and cyclostationary regimes as valid operation.

    Source: §1 · p. 2, §5 · p. 8

  • Takeaways & Limitations

    Reference-relative diagnostics offer a flexible foundation for process control across stable regimes that may include structured non-stationarity.

    Source: §1 · p. 2, §5 · p. 8

    Effectiveness depends on whether the reference set represents the full support of normal operation; missing benign phases may elevate false positives.

    Source: §5 · p. 8

Abstract

from arXiv · show

Many temporal process learning and monitoring pipelines operate in local windows, making window-level decisions unavoidable in practice. In such settings, classical statistical tests can be applied to individual windows, but they typically evaluate predefined parametric hypotheses-such as unit-root or moment-based conditions-thereby limiting flexibility when reference behavior is defined empirically from task- or domain-specific data. In this work, we view window-level monitoring as a process control problem and reformulate it as reference-based hypothesis testing, where the null hypothesis is specified by an empirical reference distribution rather than a fixed parametric model. We operationalize this perspective through a representation-based, nonparametric framework that combines pretrained time series encoders, kernel density estimation, and conformal calibration, yielding finite-sample valid inference in learned representation space. Classical notions such as stationarity and cyclostationarity arise as natural instantiations of empirical reference sets within this framework. Through experiments, we demonstrate sensitivity to window-level distributional deviations while maintaining well-calibrated inference under stable reference regimes, highlighting the applicability of the proposed approach to a broad class of time series process control and monitoring tasks.

1. Introduction

Window-level monitoring is framed as reference-relative process control: the question is whether a window remains consistent with an empirical stable regime, not whether it meets a fixed statistical definition. The proposed representation-based framework uses conformal calibration to assess consistency and treats stationary and cyclostationary references as valid regimes.

1. Introduction

  • Short rolling windows make window-level monitoring a practical necessity in streaming systems.Examples include data-quality monitoring, distributional drift detection, and checking whether operation remains within expected conditions.Source: §1 · p. 1
  • Process control asks whether operation remains consistent with a designated reference, unlike anomaly detection’s focus on rare or unusual observations.A system can produce locally plausible patterns while drifting from stable operating conditions.Source: §1 · p. 1
  • Classical stationarity tests evaluate predefined hypotheses and may have low power, unstable calibration, or miss structural deviations on short windows.These tests address statistical definitions such as unit roots or deterministic trends rather than consistency with a learned baseline.Source: §1 · p. 1
  • The framework assesses whether individual windows remain consistent with an empirical reference regime, rather than testing fixed parametric hypotheses.It embeds windows in a learned representation space and uses conformal calibration for finite-sample-valid inference.Source: §1 · p. 2
  • Stationary and cyclostationary processes are both treated as valid stable reference regimes, so structured periodic variation need not count as abnormality.Cyclostationary behavior can be stable and repeatable while violating weak stationarity.Source: §1 · p. 2
  • The experiments report robustness to structural deviations and potential for diagnostic attribution beyond binary detection.The framework is presented as accommodating diverse temporal behaviors without fixed parametric assumptions.Source: §1 · p. 2

2. Related Work

Related work includes classical process-control methods, tests of stationarity and cyclostationarity, change-point detection, anomaly and distribution-shift detection, and learned time-series representations. These approaches address related problems, but differ from assessing window-level consistency against a fixed empirical reference regime.

Process Control and Statistical Monitoring

  • Classical statistical process control uses summary statistics and fixed control limits, but can be difficult to apply directly to high-dimensional representations and heterogeneous windows.Shewhart charts, CUSUM, and EWMA are examples of methods relying on low-dimensional statistics and parametric assumptions.Source: §2 · p. 2

Stationarity and Cyclostationarity Testing

  • Stationarity and cyclostationarity tests assess specified statistical properties, rather than consistency with a chosen empirical reference regime.Cyclostationarity methods often target periodic structure in second-order statistics through spectral or cyclic-correlation analysis.Source: §2 · p. 2

Change-Point Detection and Structural Breaks

  • Change-point detection localizes when sequence statistics change, whereas this work assesses window-level consistency with a fixed reference.The related methods include likelihood-ratio, cumulative-sum, kernel, spectral, and representation-based approaches.Source: §2 · p. 2

Anomaly Detection and Distribution Shift

  • Anomaly and distribution-shift methods generally target rarity or population-level deviation, while process monitoring may reject patterns that are common globally but inconsistent with a specific reference.The cited approaches include reconstruction, prediction-residual, and learned-feature density methods.Source: §2 · pp. 2–3

Representation Learning for Time Series Monitoring

  • Pretrained time-series representations have been used for downstream detection tasks, but their high-dimensional information may not align with sensitivity to statistical stability.The related work motivates orienting representations toward directions sensitive to distributional deviation.Source: §2 · p. 3

3. Methodology

The methodology frames monitoring as independent window-level consistency testing against an empirical reference regime, then implements it in learned representations with density scoring and conformal calibration. It also addresses variable window lengths and permits modular substitutions of framework components.

3.1. Problem Formulation: Reference-Relative Process Control

  • Each window is assessed independently against an empirical reference set that may represent stationary, cyclostationary, or other stable behavior.Changing the reference set changes the notion of normality without changing the underlying framework.Source: §3.1 · p. 3

3.2. Representation-Based Consistency Testing

  • Learned representations address the difficulty of comparing noisy short windows whose local statistics can vary under stable structured behavior.A fixed pretrained encoder maps reference and test windows into a common embedding space.Source: §3.2 · p. 3

3.3. Auxiliary Sensitivity Alignment via Generic Statistical Priors

  • Auxiliary synthetic windows with mean, variance, trend, and unit-root deviations support an LDA projection that emphasizes statistically sensitive embedding directions.The projection is fixed across experiments and generic rather than tailored to specific regimes; the reported ablation finds improved efficiency and robustness over alternatives.Source: §3.3 · pp. 3–4

3.4. Reference Modeling and Scoring

  • Projected reference embeddings are modeled with a nonparametric density estimator, and negative log-density scores measure window typicality relative to the empirical regime.Experiments use Gaussian KDE with bandwidth selected by Silverman’s rule of thumb.Source: §3.4 · p. 4

3.5. Calibration and Decision Thresholds

  • Length-conditional calibration restores consistent type-I error control across window lengths, while global calibration shows pronounced length-dependent deviations from nominal rates.Null p-value QQ plots further support near-uniformity across the significance range.Source: §3.5 · pp. 4–5

3.6. Practical Considerations

  • Reference and test windows may have different lengths when the encoder maps them to fixed-dimensional representations, and each window is evaluated independently.The framework is modular: encoders, projections, density estimators, and calibration procedures can be substituted without changing the formulation.Source: §3.6 · p. 5

4. Experiments

The experiments evaluate window-level consistency with empirical reference regimes, including stationary and cyclostationary operation. The framework performs strongly across these settings, while calibration and ablations examine operating-point control, mixed window lengths, and sensitivity to design choices.

4.1. Experimental Setup and Evaluation Protocol

  • The experiments use independent window-level decisions against stable reference sets, with synthetic deviations spanning shifts, trends, unit-root behavior, and periodic-pattern changes.The mixed-length protocol calibrates separately by length bucket and reports aggregate performance across lengths.Source: §4.1 · p. 5

4.2. Reference-Relative Monitoring under Cyclostationary Regimes

  • Adapted change-point detectors remain ineffective on isolated windows because they rely on sequential context to locate boundaries between regimes.The evaluation applies each detector independently within a window and uses its statistic or maximum response as a deviation score.Source: §4.2 · p. 6
  • The method achieves consistently high discrimination across stationary and cyclostationary reference regimes, including for frequency-based and composite deviations.The reported pooled ROC-AUC remains close to perfect in most settings.Source: §4.2 · p. 6, Table 1 · p. 7

4.3. Additional Results under Stationary Reference Regimes

  • At α = 0.01, the proposed method attains TPR 0.984 with FPR well below nominal, and it achieves near-perfect detection power at both tested operating points.Classical stationarity tests show trade-offs: ADF and KPSS have elevated or unstable false alarms, while PP is more conservative with lower power.Source: Table 2 · p. 7, §4.3 · p. 7

4.4. Ablation Study

  • Under mixed window lengths, pooled calibration can mask length-dependent miscalibration, whereas length-conditional calibration yields near-uniform null p-values for short and long windows.The figure attributes stable type-I error control across lengths to conditioning calibration on length.Source: Table 3 · p. 7, Figure 1 · p. 8
  • Across bandwidth rules and scaling factors, ROC-AUC, PR-AUC, and F1-score remain nearly unchanged, indicating little dependence on fine-grained KDE bandwidth tuning.The calibrated decision threshold varies smoothly with the degree of smoothing.Source: §4.4 · p. 7, Table 4 · p. 8
  • Larger reference sets improve performance, but the method remains effective at relatively small sizes and degrades gracefully in low-sample regimes while retaining false-alarm control.The reference-size results use length-conditional calibration at α = 0.05.Source: §4.4 · p. 7, Table 5 · p. 8

5. Conclusion

The method assesses window-level consistency against an empirical reference using learned representations, nonparametric modeling, and conformal calibration. Experiments report well-calibrated performance across deviations, while the conclusion identifies reference coverage, encoder quality, and real-world validation as boundaries.

5. Conclusion

  • The framework provides finite-sample-valid, reference-relative consistency assessments in learned representation space without deviation-specific assumptions.It models the reference distribution nonparametrically and applies conformal calibration.Source: §5 · p. 8
  • Experiments report strong, well-calibrated performance across deviations while accommodating stationary and cyclostationary regimes as normal operation.Classical stationarity diagnostics and change-point methods struggle when the reference regime has structured non-stationarity.Source: §5 · p. 8
  • Effectiveness depends on whether the reference set represents the full normal regime, including rare but benign cyclostationary phases.A mismatch may elevate false-positive rates, making careful reference construction necessary.Source: §5 · p. 8
  • Performance may falter under extreme domain shifts and remains contingent on the quality of the frozen embedding space.The concern is that encoder pretraining may diverge substantially from the monitored signal.Source: §5 · p. 8
  • Validation primarily uses controlled synthetic regimes; large-scale, noisy, unlabeled real-world benchmarks remain to be evaluated.The synthetic setup supports rigorous checks of statistical calibration and Type-I error control.Source: §5 · p. 8

Impact Statement

The paper frames its impact as methodological, connecting representation learning with calibrated, reference-dependent monitoring. Its appendices describe stationarity and cyclostationarity and explain the DG test implementation used in experiments.

Impact Statement

  • The work may bridge statistical process control and time-series representation learning, with explicit false-alarm control favored over ad hoc anomaly scores.The framework is intended for benign monitoring and quality-control settings; misuse without domain knowledge or sound reference construction could mislead.Source: Impact Statement · p. 9

A.1. Stationarity and Unit Root

  • Weak stationarity requires finite second moments, a constant mean, and second-order statistics that depend on time lag rather than absolute time.The appendix focuses on univariate time series; a unit root violates weak stationarity because variance grows unbounded over time.Source: §A.1 · p. 12

A.2. Cyclostationarity

  • Cyclostationary processes have statistics that vary periodically and remain stable across cycles, rather than requiring constant first- and second-order moments.The appendix notes that dedicated tests can exploit cyclic frequencies or compare statistics across corresponding cycle phases.Source: §A.2 · p. 12

A.3. Cyclostationarity Test of Dandawate and Giannakis

  • The DG test estimates a dominant cyclic frequency from each window’s periodogram, then uses the test as a secondary rescue after a stationarity test.A frequency above 2/L is required for resolvability, where L is window length; exceeding the χ2 critical value reclassifies a window as in-regime.Source: §A.3 · p. 13

B.1. Fixed Data Length Setting

In fixed-length experiments, the proposed method achieves high AUC across deviation types and outperforms classical stationarity and change-point baselines. Conformal calibration aligns false-alarm rates with the nominal level, while comparison with mixed-length results attributes pooled miscalibration to length heterogeneity.

B.1. Fixed Data Length Setting

  • The fixed-length experiment uses the same data-generating processes and reference-relative protocol as the main experiments, but removes window-length variation.Chronos2 with mean pooling and the same conformal procedure is used; exchangeability holds trivially, so per-length conditioning is unnecessary.Source: §B.1 · p. 13
  • The proposed method achieves consistently high AUC across deviation types, substantially outperforming classical stationarity tests and change-point detection baselines.The combined stationary–cyclostationary reference results are reported in Table 7, aggregated across lengths over 8 runs.Source: §B.1 · p. 13, Table 7 · p. 14
  • Under fixed window lengths, conformal calibration yields empirical false-alarm rates aligned with nominal α, without requiring conditional or stratified calibration.The results indicate that miscalibration under mixed lengths stems from length heterogeneity rather than the conformal procedure itself.Source: §B.1 · p. 13
  • Length-conditional calibration is a structural necessity when window lengths are heterogeneous, not a modeling choice.In the fixed-length setting, the method’s performance advantages persist without length mixing or calibration heuristics.Source: §B.1 · p. 14

B.2. Choice of Encoder Model

Chronos2 and MOMENT provide discriminative representations in the reported encoder comparison, while TOTEM performs poorly. The framework’s practical performance depends on representation quality.

B.2. Choice of Encoder Model

  • The framework compares Chronos2, MOMENT, and TOTEM as pretrained time-series encoders for representing local windows.The encoders supply representations in which distributional consistency is assessed.Source: §B.2 · p. 14
  • Practical performance depends on the quality of the learned representation, although the framework is modular and encoder-agnostic in principle.Encoders are used as frozen feature extractors without task-specific fine-tuning.Source: §B.2 · p. 14
  • Chronos2 and MOMENT produce highly discriminative representations under View 1, while TOTEM fails to separate reference and out-of-regime windows.The table caption describes the comparison under stationary and cyclostationary reference regimes.Source: Table 10 · p. 15

B.3. Choice of Projection Method

Projection choice affects window-level detection performance. For Chronos2, supervised LDA generally outperforms PCA and random projection on the reported metrics, especially in mixed-length regimes.

B.3. Choice of Projection Method

  • Projection choice has a non-negligible impact on ROC-AUC, PR-AUC, F1 score, precision, and recall across both encoders.The comparisons include random projection, PCA, and supervised LDA under identical experimental settings.Source: §B.3 · p. 15
  • For Chronos2, supervised LDA yields consistently higher ROC-AUC and PR-AUC across most settings, with particularly pronounced gains in mixed-length regimes.The passage suggests that emphasizing directions sensitive to known distributional deviations can improve subsequent density-based scoring, even when projection is learned independently of reference and test sets.Source: §B.3 · p. 15
  • For Chronos2, unsupervised PCA provides moderate improvements over random projection, while random projections show less consistent behavior.This comparison is reported across projection settings in the passage.Source: §B.3 · p. 15

C. Implementation Details

The implementation evaluates pretrained encoders and mixed-length windows against structured stationary and periodic regimes and their deviations. It also describes reference modeling by fitting a density estimate on projected reference representations.

C. Implementation Details

  • Mixed window lengths are drawn from 64, 96, 128, 192, 256, 336, 512, and 720, alongside a fixed-length setting.The lengths are described as providing multi-scale temporal coverage.Source: C. Implementation Details · p. 15
  • AR(1) deviations include mean and variance shifts, linear trends, combined shifts, and unit-root random-walk behavior.These departures preserve local temporal structure while violating distributional consistency with the stationary reference regime.Source: C. Implementation Details · p. 15
  • The periodic reference regime repeats an unknown zero-mean, unit-variance template, with its window distribution invariant to shifts by multiples of its period.The construction permits rich within-cycle shapes rather than restricting waveforms to sinusoids.Source: C. Implementation Details · pp. 15–16
  • Under View 2, Chronos2 remains robust, MOMENT degrades gracefully, and TOTEM performs poorly on cyclostationary deviations.The comparison is summarized in the encoder-ablation table caption.Source: Table 11 · p. 16
  • Periodic deviations include amplitude steps or drift, template distortion, frequency mismatch, and compound changes that may alter periodic organization.Compound deviations can preserve a fixed periodic structure, whereas compound structural deviations alter that organization.Source: C. Implementation Details · p. 16
  • Chronos2 is the primary encoder, with hidden states mean-pooled into one embedding per window; MOMENT and TOTEM are also used as pretrained representation extractors.The implementations use frozen pretrained models without task-specific fine-tuning or supervision.Source: C. Implementation Details · p. 16
  • Chronos2 demonstrates the best overall detection performance across encoders and sequence lengths, peaking at sequence length 336.The table caption also reports that TOTEM fails to extract meaningful anomaly features.Source: Table 13 · p. 17

1. Reference Modeling

  • Reference modeling splits the reference into fitting and calibration sets, then fits a density estimate on projected representations from the fitting set.The algorithm heading identifies this as reference-relative consistency testing with length-conditional conformal calibration.Source: §1 · p. 17, C. Implementation Details · p. 17

3. Length-Conditional Calibration

The supplied section details a reference-relative calibration step and describes synthetic regimes and baseline evaluation. The examples include stationary AR(1) windows and a stable periodic-template regime.

3. Length-Conditional Calibration

  • The algorithm selects a calibration set conditional on the test window’s length.This is the stated selection step for length-conditional calibration.Source: §3 · p. 17
  • The evaluation includes classical stationarity tests and change-point detection baselines, with results averaged over multiple runs.Baselines use standard configurations.Source: §3 · p. 17
  • Synthetic-data parameter ranges are chosen to produce diverse but controlled stationary and cyclostationary regimes and structured within-window deviations.Parameters are independently sampled per window unless otherwise stated.Source: Table 14 · p. 18
  • The illustrated data families include stationary AR(1) windows and a stable periodic regime generated by repeating an unknown template.The AR(1) examples contrast stationary windows with mean shifts, variance shifts, linear trends, unit-root behavior, and combined changes; the periodic figure caption identifies the stable regime but truncates its deviation description.Source: Figure 2 · p. 19, Figure 3 · p. 19
Loading 2609.30790v1…