PaperlayerView exact revision on arXiv

Source-linked AI summary

From Benchmarks to Production: Transferring Time Series Anomaly Detection Methods for Electricity Production Monitoring

Nicolas Vautier, Paul Caron, Nardi Xhepi, Félicie Bizeul, Manel Boumghar, Christophe Degouy, Paul Boniol

arXiv:2609.39257v1cs.LGcs.AIcs.DB

TL;DR

Detecting daily production-forecast patterns that deviate from historical behavior is important for reliable industrial monitoring. TAMIS scores and ranks daily cases for expert review, and the evaluation reports a superior accuracy-efficiency trade-off over baselines.

  • Problem

    The problem is identifying EDF production days whose intra-day profiles deviate from normal patterns learned from historical data, supporting reliable anomaly detection for automated forecasts.

    Source: §I · p. 1, A. Problem Definition · p. 3

  • Method

    TAMIS uses a lightweight, interpretable feature-based approach to score daily forecasts and rank the most abnormal cases in an automated report.

    Source: §I · p. 1

  • Results

    TAMIS balances accuracy and efficiency, with the compact two-detector configuration reported to achieve the best trade-off for scalable deployment while maintaining strong out-of-distribution performance.

    Source: §VIII · p. 12

  • Takeaways & Limitations

    TAMIS’s ranked daily reports let experts review the most abnormal cases and make decisions accordingly.

    Source: §I · p. 1

Abstract

from arXiv · show

Accurate forecasting of electricity production is essential for maintaining the operational efficiency and strategic planning of energy utilities. In industrial settings, such forecasts are generated daily to ensure supply-demand balance and optimal management of production assets. However, the increasing complexity of modern power systems and data flows poses significant challenges for ensuring the reliability and consistency of these forecasts. This paper addresses the problem of anomaly detection in short-term production forecasts at EDF, formulated as identifying atypical intra-day patterns that may signal data quality issues or operational irregularities. We introduce TAMIS, a scalable and interpretable system that analyzes daily production time series to automatically detect anomalous days based on deviations from historical patterns learned from past data. Designed for human-in-the-loop workflows, TAMIS surfaces top-ranked anomalies through an automated daily newsletter, enabling efficient expert review and continuous monitoring. An extensive experimental evaluation on real-world industrial data demonstrates that TAMIS achieves the best accuracy-efficiency trade-off compared to baseline methods. To foster further research and reproducibility, we publicly release the anonymized application datasets used in our study.

I. INTRODUCTION · II. INDUSTRIAL CONTEXT AND PROBLEM FORMULATION

The paper presents TAMIS as an interpretable, scalable anomaly-detection system for EDF production forecasts, designed around operational constraints and expert review. It reports strong benchmark performance and releases anonymized energy time-series datasets for further research.

I. INTRODUCTION

  • EDF’s day-ahead forecasts are fixed-length vectors of 48 half-hourly values over a 24-hour horizon and inform supply-demand balancing, dispatch planning, and energy-market activity.The paper frames forecast reliability as essential to these operational decisions.Source: §I · p. 1
  • The industrial setting requires online detection from historical data, interpretable results for daily expert analysis, and scalable low-latency operation on limited hardware.These requirements rule out many approaches that rely on deep learning models and exhaustive offline training.Source: §I · p. 1
  • TAMIS scores each daily forecast for abnormality, ranks the results, and returns only the most abnormal cases for expert review.The system is described as lightweight and interpretable, with results integrated into a daily report.Source: §I · p. 1
  • The paper identifies a tailored industrial architecture, the lightweight ASHES detector, the TAMISF feature set, and an end-to-end framework balancing interpretability, accuracy, and scalability.These are presented as designed-for-purpose components of the proposed solution.Source: §I · p. 2
  • The work releases three open-access datasets of real, labeled time series, described as among the first of their kind in the energy sector.The anonymized datasets are intended to encourage future anomaly-detection research.Source: §I · p. 2
  • TAMIS is reported to outperform state-of-the-art baselines while maintaining a superior accuracy-efficiency trade-off.The evaluation uses a large, multi-year dataset of EDF production forecasts; the paper also releases an anonymized version as a public benchmark.Source: §I · p. 2

II. INDUSTRIAL CONTEXT AND PROBLEM FORMULATION

  • EDF’s monitoring data span heterogeneous measurements—including power output, reservoir levels, flow rates, turbine states, and production targets—with differing scales, dynamics, semantics, and physical meanings.Their volume makes manual inspection infeasible and motivates automated, scalable monitoring.Source: §II · p. 2
  • A daily mailing-list report presents only the highest-ranked anomalies, with contextual information such as score and timestamp, to focus expert triage and mitigate alert fatigue.The workflow automatically scores and ranks potential anomalies before distributing the concise summary to experts.Source: §II · p. 2

A. Problem Definition

The section defines anomaly detection as online scoring of the latest daily production forecast against historical patterns, so anomalies can be identified as they emerge. The problem is framed for EDF operations, where deployment also requires scalable, interpretable detection on constrained hardware and heterogeneous data.

A. Problem Definition

  • The task is online detection of whether the most recently observed forecasted value deviates from historical behavior, enabling timely intervention rather than retrospective discovery.This real-time framing reflects EDF’s operational need to avoid planning errors or missed opportunities from undetected anomalies.Source: A. Problem Definition · p. 2
  • A window is a 24-hour profile of a variable c sampled every 30 minutes, represented by 48 measurements, one for each half-hour interval of a given day.Source: A. Problem Definition · p. 2
  • The anomaly function takes the time series and its last daily window as input, returning a score from 0 to 1, with higher scores indicating greater likelihood of anomaly.The function may represent a single detector or an automatic solution combining multiple detectors.Source: A. Problem Definition · p. 3
  • Industrial deployment requires low execution time, near-real-time operation on standard CPUs, interpretable outputs for expert review, and support for heterogeneous data with meaningful missing values and out-of-distribution shifts.Heterogeneity includes diverse data provenance and measurement types, which can produce shifts in trends and anomaly types.Source: A. Problem Definition · p. 3

III. TIME SERIES ANOMALY DETECTION: Foundations and Background · A. Raw-based time series anomaly detection

The section reviews time-series anomaly detection methods by their operating principles and assesses how they meet industrial constraints. It describes prediction-, distance-, and density-based approaches and their limitations.

III. TIME SERIES ANOMALY DETECTION: Foundations and Background

  • The review frames existing anomaly-detection methods through operational constraints and groups them into process-centric families based on prior benchmarks and taxonomy, –, [9].Figure 2 depicts a panel of methods grouped by these families.Source: §III · p. 3, A. Raw-based time series anomaly… · p. 3

A. Raw-based time series anomaly detection

  • Prediction-based methods flag observations that significantly deviate from a self-supervised model’s predictions, using prediction error as the anomaly score; they may be reconstruction- or forecasting-based,.Source: A. Raw-based time series anomaly… · p. 3
  • Distance-based methods assume anomalies are isolated from most data and identify outliers by measuring distances to nearest neighbors, with larger distances indicating abnormality.Source: A. Raw-based time series anomaly… · pp. 3–4
  • Density-based methods score time-series representations such as trees or graphs using criteria including tree depth or graph centrality,.Source: A. Raw-based time series anomaly… · p. 4
  • Prediction-based methods model temporal dynamics but often require complex, compute- and memory-intensive architectures, limiting high-velocity real-time use on constrained hardware and challenging interpretability.These limitations make them ill-suited to constraints C1, C2, and C3.Source: A. Raw-based time series anomaly… · p. 4
  • Distance- and density-based raw methods may be computationally lighter, but can struggle with missing values and lack contextual transparency for non-specialists, motivating methods designed for industrial constraints.These limitations concern constraints C4 and C3, respectively.Source: A. Raw-based time series anomaly… · p. 4

B. Lack of interpretability: Toward Feature-based approaches · C. No Universal Best: Toward Automatic Solutions

Feature-based detection turns time series into interpretable feature vectors for outlier detection, but existing feature sets do not fit the industrial use case. Because detector performance varies across series and anomalies, the paper motivates automatic strategies, ultimately prioritizing robust, low-cost ensembling for out-of-distribution settings.

B. Lack of interpretability: Toward Feature-based approaches

  • Feature-based methods expose which characteristics drive an anomaly and can handle missing values through feature design, balancing interpretability, operational feasibility, and detection performance.They transform raw series into structured statistical, structural, or domain-specific descriptors for density-based detection in a lower-dimensional space.Source: B. Lack of interpretability: Toward… · p. 4
  • A feature-based pipeline converts each time-series window into a fixed-length feature vector, then applies a classical outlier detector to identify abnormal vectors.Features can include basic statistics, frequency-domain information, and entropy; examples of detectors include Isolation Forest, LOF, and One-Class SVM.Source: B. Lack of interpretability: Toward… · p. 4
  • HCTSA extracts over 7,700 features, TSFRESH extracts 794 by default with hypothesis-testing selection, and CATCH22 offers 22 features tuned for z-normalized classification tasks.HCTSA, TSFRESH, and CATCH22 provide distinct feature-set options; CATCH22’s 22 features were selected for performance across 93 UCR/UEA datasets [28].Source: B. Lack of interpretability: Toward… · p. 4
  • Existing feature sets are unsuitable for the industrial use case: HCTSA and TSFRESH are too high-dimensional, while CATCH22 features miss anomalies detected by TSFRESH in the illustrated IOPS series.TSFRESH’s selection is not anomaly-specific, and CATCH22 was optimized for classification rather than anomaly detection; the feature comparison is shown in Figure 3.Source: B. Lack of interpretability: Toward… · p. 4, Figure 3 · p. 4

C. No Universal Best: Toward Automatic Solutions

  • Benchmarking studies report no single detector that performs best across heterogeneous time series, as performance varies with series properties and anomaly types,,,.This setting also requires one decision at the last timestamp, unlike algorithms that score each time point.Source: C. No Universal Best: Toward Automatic… · pp. 4–5
  • Ensembling combines multiple detectors’ scores by averaging or maximizing them and can outperform individual algorithms on public benchmarks, but adds computational overhead.Automatic model selection instead maps extracted features to a detector through a meta-model,, [32], but may generalize poorly to out-of-distribution series.Source: C. No Universal Best: Toward Automatic… · p. 5
  • The paper therefore motivates an ensembling strategy that prioritizes out-of-distribution robustness while keeping computational cost low.Source: C. No Universal Best: Toward Automatic… · p. 5

IV. RESEARCH QUESTIONS · V. PROPOSED METHOD: THE TAMIS SYSTEM

TAMIS is designed to balance detection performance with operational feasibility, addressing research questions about detector combinations, feature selection, and out-of-distribution conditions. Its production workflow computes features, combines complementary detectors in a supervised ensemble, and surfaces the most abnormal series for review.

IV. RESEARCH QUESTIONS

  • TAMIS asks how to combine detectors to balance accuracy and efficiency, since ensembling can improve robustness and accuracy but raises computational cost, while selection may falter out of distribution.Source: §IV · p. 5
  • TAMIS asks which feature set to use: smaller sets improve scalability, but small or non-dedicated features can reduce detection accuracy and practical usefulness.Source: §IV · p. 5
  • TAMIS asks how out-of-distribution conditions affect automatic solution settings, seeking low execution time alongside greater robustness in those scenarios.Source: §IV · p. 5
  • The study presents a data-based production anomaly-detection pipeline and a roadmap for transferring data-mining and data-management research findings into production systems.Source: §IV · p. 5

V. PROPOSED METHOD: THE TAMIS SYSTEM

  • The training pipeline computes TAMISF features from manually labeled data, applies two complementary detectors, and combines their outputs in the supervised ensemble SEASA for production inference.Source: §V · p. 5
  • In production, TAMIS reuses cached prior-day features to score the next day near real time, then presents the most abnormal time series through a visual interface.Source: §V · p. 5

A. TAMISF: A Novel Feature Set · 1) Expert-Based Feature Pool: · 2) TSAD-Based Feature Pool:

TAMISF combines an expert-derived pool with a TSAD-based pool to capture informative, generalizable characteristics of industrial time series. Its feature-selection pipelines address both expert knowledge and scarce labels.

A. TAMISF: A Novel Feature Set

  • TAMISF combines complementary expert-based and TSAD-based feature pools to describe daily production windows for anomaly detection.The pipeline aggregates raw measurements into daily windows and computes a feature vector for each day.Source: A. TAMISF: A Novel Feature Set · pp. 5–6

1) Expert-Based Feature Pool:

  • Energy-production specialists proposed 36 candidate indicators, and leave-one-out ablation, full-set baseline, and single-feature models assessed their contributions.Four recall-based curves per configuration measured detections by score and rank; qualitative and quantitative analysis retained the ten most informative features.Source: 1) Expert-Based Feature Pool · p. 6
  • The expert-based selection may favor historically labeled anomaly types and common anomalies, limiting generalization to unseen faults and risking neglect of rare critical ones.The evaluation dataset's anomaly distribution may not reflect operational frequency, introducing sampling bias.Source: 1) Expert-Based Feature Pool · p. 6

2) TSAD-Based Feature Pool:

  • Ten TSAD-based features are selected through a synthetic-data pipeline designed to address scarce labels and improve cross-domain generalization.Artificial point perturbations, temporal shifts, and amplitude scaling are injected into normal JO time series to create labeled data.Source: 2) TSAD-Based Feature Pool · pp. 6–7
  • The TSAD pipeline computes statistics and signal transformations using TSFRESH and CATCH22, then ranks features for synthetic-label discrimination using statistical tests and gradient-boosting importance.Relevance filtering uses p-values with Benjamini–Hochberg correction, ANOVA F-test, and gradient boosting importance scores.Source: 2) TSAD-Based Feature Pool · p. 6
  • The TSAD selection clusters top-ranked features by pairwise correlation and retains one representative per cluster, maximizing target correlation while minimizing average inter-feature correlation.Target correlation is assessed using the ANOVA F-test score.Source: 2) TSAD-Based Feature Pool · pp. 6–7
  • TAMISF can be integrated into supervised or unsupervised anomaly detection systems, beyond TAMIS itself.The paper evaluates traditional anomaly-detection methods applied to TAMISF and assesses the feature set's relevance separately.Source: 2) TSAD-Based Feature Pool · p. 7

B. Anomaly Detectors Pool · 1) KDE: Kernel Density Estimation-based anomaly detection: · 2) ASHES: Anomaly Scoring-based on Historical ExtremeS:

TAMIS combines two complementary detectors: KDE scores rarity from historical feature distributions, while ASHES scores extreme deviations in light of their magnitude. Together, they provide anomaly scores for the supervised ensemble while limiting computational overhead.

B. Anomaly Detectors Pool

  • The detector pool deliberately uses only KDE and ASHES to reduce computational overhead while maintaining high detection accuracy and production scalability.Each detector produces an anomaly score that is combined by the supervised ensemble.Source: B. Anomaly Detectors Pool · p. 7

1) KDE: Kernel Density Estimation-based anomaly detection:

  • KDE identifies observations with low density under a univariate Gaussian estimate of historical feature values, rescaling density by historical standard deviation to make scores dimensionless and account for scale differences.Its decreasing score preserves density ordering: rarer observations receive higher scores, and ranking is equivalent to ranking by negative estimated density.Source: 1) KDE: Kernel Density… · p. 7

2) ASHES: Anomaly Scoring-based on Historical ExtremeS:

  • ASHES defines extremity rank K as the minimum insertion rank of an observation in ascending or descending historical order; K = 1 marks a new minimum or maximum, and K > 10 yields score zero.This cutoff is described as a practical heuristic.Source: 2) ASHES: Anomaly Scoring-based on… · p. 7
  • ASHES uses relative amplitude ratio r to measure how much an observation expands the historical range after removing the K −1 largest and smallest values; r near 1 indicates a typical range, while smaller r signals greater deviation.The filtered set is denoted Hᵢₖ, with minK and maxK as its minimum and maximum.Source: 2) ASHES: Anomaly Scoring-based on… · pp. 7–8
  • ASHES combines K and r with exponential weighting so low ranks and small amplitude ratios receive higher anomaly scores, highlighting rare events that substantially alter time-series dynamics.The score accounts for both extremity and deviation magnitude.Source: 2) ASHES: Anomaly Scoring-based on… · p. 8
  • Unlike HBOS and LOF, which treat rarity and amplitude deviation equally, ASHES accounts for low-amplitude rare values that may not be operationally critical; its rank-based extremity complements KDE’s probabilistic rarity model.This complementary pairing motivates the detectors’ joint use in TAMIS’s supervised ensemble.Source: 2) ASHES: Anomaly Scoring-based on… · p. 8

C. Supervised Ensemble Anomaly Scores Aggregation · 1) Base Anomaly Score Generation: · 2) Meta-Learner and Prediction Strategy:

SEASA combines KDE and ASHES anomaly scores through supervised learning to produce a unified prediction, adapting detector contributions according to their reliability across anomaly types.

C. Supervised Ensemble Anomaly Scores Aggregation

  • SEASA is a stacking-based ensemble that combines KDE and ASHES outputs into a unified, more accurate prediction.Its supervised learning mechanism infers weightings and interactions among detector outputs to improve robustness.Source: C. Supervised Ensemble Anomaly Scores… · p. 8

1) Base Anomaly Score Generation:

  • For each daily time-series window, TAMISF features feed independent KDE and ASHES detectors, each producing a scalar anomaly score.The resulting scores serve as the input features for SEASA’s meta-learning stage.Source: 1) Base Anomaly Score Generation · p. 8
  • The KDE and ASHES scores capture complementary aspects of the data distribution and become SEASA’s meta-learning input features.Source: 1) Base Anomaly Score Generation · p. 8

2) Meta-Learner and Prediction Strategy:

3) Experimental setup: … 1) Datasets:

The evaluation uses EDF’s diverse annotated thermal and hydraulic production data to study TAMIS’s overall performance, feature choices, and cross-system transfer. Its setup also specifies held-out prediction and deployment procedures.

3) Experimental setup:

  • LOOCV trains XGBoost on all other labeled days for each held-out prediction, while deployment trains SEASA once on available labels before applying it to new daily measurements.Source: 3) Experimental setup · p. 8

VI. EXPERIMENTS

  • The overall evaluation compares TAMIS’s accuracy and throughput with baselines and examines learned detector selection versus ensemble combination.This experiment addresses R1.Source: §VI · p. 8
  • The feature-set experiment compares TAMISF with TSFRESH and CATCH22 for accuracy and efficiency.This experiment addresses R2.Source: §VI · p. 8
  • The out-of-distribution and transfer evaluation tests whether TAMIS generalizes across different energy-production systems.This experiment addresses R3.Source: §VI · p. 8

A. Experimental Setup

  • Experiments ran on a node with two Intel Xeon Gold 6234 CPUs, 16 physical cores, and 384 GB RAM; the implementation is publicly available.Source: A. Experimental Setup · p. 8

1) Datasets:

  • EDF’s industrial data spans thermal and hydraulic systems, with 6,931 annotated days and 376 anomalies; retained missing values reach 27.9% of labeled days.The series include physical behaviors such as flow rates and temperatures, as well as economic variables such as marginal costs.Source: 1) Datasets · pp. 8–9
  • The dataset targets validation of short-term production forecasts by comparing new values with historical profiles, unlike benchmarks focused on retrospective monitoring of continuous sensor streams.This forecast-validation structure is intended to anticipate future failures.Source: 1) Datasets · p. 9
  • The anonymized release covers THERM (4,515 sensors), HYDRAU (1,740), and JO (3,198), with JO representing EDF’s initial monitoring training database and disjoint from the other sets.The authors describe the collection as, to their knowledge, the most extensive energy-production time-series dataset for anomaly detection.Source: 1) Datasets · p. 9

2) Baselines: · 3) Evaluation Measures:

TAMIS is compared with raw-based, feature-based, and automatic anomaly detectors, while evaluation emphasizes AUC-PR and throughput. The sections also describe missing-value handling and broader TAMIS variants.

2) Baselines:

  • Raw-based baselines are OCSVM, HBOS, and LOF, selected for industrial constraints and operation on raw series or extracted features.The study compares TAMIS across raw-based, feature-based, and automatic detector categories.Source: 2) Baselines · p. 9
  • Feature-based comparisons apply OCSVM, HBOS, LOF, KDE, and ASHES to the same features, using the best of min, max, mean, or meta-learner M for aggregation.The OCSVM, HBOS, and LOF feature-based methods each produce scores per feature, which are aggregated.Source: 2) Baselines · p. 9
  • Automatic baselines are an unsupervised average ensemble and model selection (MS) introduced in.Because baselines do not natively handle NaNs, missing values are imputed by forward and backward filling to preserve local temporal consistency, piecewise-constant signals, and observed-value range.Source: 2) Baselines · p. 9
  • The evaluated TAMIS variants add all feature detectors or combine those detectors with OCSVM, HBOS, and LOF on raw subsequences.They are denoted TAMIS (ALL FEATURES) and TAMIS (ALL FEATURES AND RAW), respectively.Source: 2) Baselines · p. 9

3) Evaluation Measures:

  • The evaluation uses AUC-PR because rare anomalies can make AUC-ROC overstate detector accuracy, and throughput measures efficiency and scalability.Table IV reports AUC-PR for TAMIS and baselines on raw data and TAMISF features; Fig. 6 shows precision-recall curves and throughput versus AUC-PR for competitive baselines.Source: 3) Evaluation Measures · p. 9, Table IV · p. 10, Figure 6 · p. 10
  • The paper notes that VUS is a robust evaluation measure, but argues that its day-level abnormality task is unaffected by potential misalignment.Each day receives a single score, supporting their stated justification for using VUS.Source: 3) Evaluation Measures · p. 10

B. Overall Evaluation · C. Evaluating the Relevance of the Feature Sets

TAMIS achieves the best accuracy–efficiency trade-off among evaluated configurations, while TAMISF outperforms alternative feature sets in accuracy–throughput balance. Raw time series provide poor anomaly-detection accuracy, and supervised ensembles improve accuracy over unsupervised and selection-based approaches.

B. Overall Evaluation

  • Raw-series detectors perform poorly: even LOF, the most accurate raw-based detector, remains below 0.23 AUC-PR on JO and HYDRAU and is approximately 0.07 on THERM.The paper attributes this to unprocessed time series lacking discriminative structure for effective anomaly detection.Source: B. Overall Evaluation · p. 10
  • TAMIS (KDE and ASHES) delivers near-best accuracy while being 2.4× faster, whereas TAMIS (ALL FEATURES AND RAW) achieves the best overall AUC-PR of 0.856, +6.40 points over ASHES.The largest ensemble has the lowest throughput because of its high inference cost.Source: B. Overall Evaluation · p. 10
  • Supervised ensembles, including SEASA, outperform unsupervised ensembles and supervised model selection; ASHES achieves AUC-PR values of 0.792 on JO, 0.571 on HYDRAU, and 0.229 on THERM.ASHES is the highest-performing individual detector on JO.Source: B. Overall Evaluation · p. 10

C. Evaluating the Relevance of the Feature Sets

  • TAMISF offers the best accuracy–throughput balance, reaching approximately 0.8–0.9 AUC-PR at several-times-higher throughput than TSFRESH and comparable throughput to CATCH22.CATCH22 AUC-PR saturates around 0.15, while TSFRESH achieves high accuracy at prohibitively high computational cost.Source: Figure 7 · p. 11, C. Evaluating the Relevance of the… · p. 11
  • TAMISF’s efficiency advantage persists through detector computation and ensemble weight inference, maintaining high accuracy and throughput across all processing stages.Figure 7 compares feature computation, detector execution, and weight inference for TAMISF, CATCH22, and TSFRESH.Source: Figure 7 · p. 11, C. Evaluating the Relevance of the… · p. 11
  • TAMISF provides near-state-of-the-art detection accuracy at significantly lower computational cost than the alternative feature sets.The evaluation compares TAMISF, CATCH22, and TSFRESH across the TAMIS pipeline stages.Source: C. Evaluating the Relevance of the… · p. 11

D. Toward Production: an Out-of-Distribution test … VIII. CONCLUSION

TAMIS remains more robust than individual detectors under distribution shift and also generalizes to a heterogeneous public benchmark. Operationally, its ranked alerts and interpretable features support expert review and cold-start monitoring; the conclusion highlights efficiency, feature design, dataset release, and future extensions.

D. Toward Production: an Out-of-Distribution test

  • AUC-PR >0.5 for TAMIS versus ≈0.2 for KDE on JO→HYDRAU, while its JO→THERM loss is ∼34% versus ASHES’s 53%.TAMIS also outperforms KDE and ASHES on the more challenging JO→THERM transfer; the paper attributes its stability to ensemble and feature aggregation.Source: D. Toward Production: an… · pp. 11–12

A. Operational feedback

  • TAMIS’s ranking mechanism reduced alert fatigue by focusing daily reports on the most critical deviations, supporting more efficient expert triage.Operational teams also validated TAMISF’s explicit features as immediate proxies for physical root causes, enabling faster decisions.Source: A. Operational feedback · p. 12

VII. OPERATIONAL INSIGHTS AND EXTENSIONS

  • For cold-start monitoring, TAMIS begins with a generic model trained on corporate data, while validated alerts build a domain-specific labeled dataset for iterative retraining.The visualization interface supports both daily decisions and expert annotation of alerts.Source: A. Operational feedback · p. 12

B. Generalization on public Benchmarks

  • VUS-PR ≈0.38 on TSB-UAD lets TAMIS outperform the Average Ensemble, Best Detector, and Best Model Selection (VUS-PR ≈0.35).The evaluation compares TAMIS with NormA as Best Detector, an Unsupervised Average Ensemble, and the Best Model Selection strategy.Source: B. Generalization on public Benchmarks · p. 12, Figure 9 · p. 12

VIII. CONCLUSION

  • The conclusion identifies TAMIS’s compact two-detector configuration as the best accuracy-efficiency trade-off for scalable deployment while retaining strong OOD performance.It also reports that TAMISF outperforms CATCH22 and TSFRESH in accuracy–throughput balance.Source: §VIII · p. 12
  • The paper releases an anonymized energy-production anomaly-detection benchmark and identifies online learning and multivariate time series as future work.The conclusion describes the released benchmark as among the most extensive datasets of energy-production time series for anomaly detection.Source: §VIII · p. 12
Loading 2609.39257v1…