Source-linked AI summary
Generative multi-domain transfer learning for fault detection in data-scarce wind turbines
Stefan Jonas, Angela Meyer
TL;DR
The paper addresses unreliable wind-turbine fault detection when data-scarce targets lack representative fault-free training data. It uses StarGAN-based multi-domain mapping to translate target SCADA measurements toward several data-rich source turbines, and reports stronger fault-detection performance under severe scarcity while proposing a proxy for training-time model selection.
Problem
Normal behavior models require representative fault-free observations, but data-scarce target turbines may provide too little data to capture normal operation reliably.
Method
A StarGAN-based generative domain mapping model translates SCADA samples from one data-scarce target turbine into several data-rich source domains while preserving operational state.
Results
The model largely outperformed single-source fine-tuning and domain mapping when only 1–2 weeks of target training data were available.
Takeaways & Limitations
Multi-source mapping can improve fault detection under severe target-data scarcity, while performance depends on source-turbine choice and training-data representativeness.
Takeaways & Limitations
The proxy metric mainly separates poor from functional mappings, requires reliable candidates among the models considered, and cannot tune anomaly-specific components without real anomalies.
Abstract
from arXiv · showhide
Normal behavior models have shown promise for reliable fault detection in wind turbines. However, these unsupervised anomaly detection models require sufficient fault-free training data to learn the normal operation behavior of turbines. Under data scarcity, for example in newly deployed wind turbines, these models may result in poor fault detection performance. In this work, we propose a multi-domain generative domain mapping approach based on Star Generative Adversarial Networks (StarGAN) to improve fault detection on data-scarce wind turbines. Our model maps SCADA measurements from a data-scarce turbine to resemble those of several data-rich turbines. By preserving the operational state during translation, faults occurring in a data-scarce domain can be mapped and detected by reliable pre-trained normal behavior models of data-rich domains. Highlighting the benefits of an ensemble fusion strategy, we show that under severe data scarcity our method can produce anomaly scores comparable to models trained on large representative datasets. Our approach can consistently outperform models trained on scarce data when less than 2 weeks of training data are available. With just 2 weeks of accumulated training data, we achieve an anomaly score similarity that is, on average, +16% higher than conventional fine-tuning, and +10% higher than single-source domain mapping. As a step towards unsupervised model selection, we propose a proxy metric that detects poor performance at training time, despite an absence of anomalies. Our study presents the potential and challenges of multi-domain mapping for wind turbine fault detection under unrepresentative training data.
1 Introduction
Wind-turbine normal behavior models need representative fault-free data, which is often unavailable for newly commissioned turbines. The paper therefore extends generative domain mapping to multiple data-rich source turbines and introduces strategies for source and model selection.
- Motivation: Fault-free training data scarcity can leave normal behavior models with insufficient coverage of turbine operating conditions.Newly commissioned wind farms may have only a few weeks of historical data, and prior work reports detrimental effects on anomaly detection performance.
- Background: Domain adaptation transfers knowledge from data-rich source turbines to data-scarce target turbines, but turbine differences create domain shifts.Relevant differences include turbine specifications, environmental conditions, site characteristics, and operational histories.
- Research gap: Existing wind-turbine applications mainly use fine-tuning or single-source mapping, leaving multi-source adaptation under strong scarcity largely unexplored.Single-source approaches also create an unresolved source-turbine selection problem because representative target validation data and anomalies are unavailable.
- Approach: The proposed StarGAN-based framework maps one data-scarce target turbine to several data-rich source turbines simultaneously with one generative network.This contrasts with training separate models for each turbine-to-turbine pair and supports practical ensemble fusion across source domains.
- Contributions: The study introduces a proxy metric for identifying models likely to perform poorly during training despite the absence of anomalies.It also investigates ensemble fusion to combine detection results from all source domains and avoid selecting a single source turbine.
2 Related Work
Domain adaptation for unsupervised wind-turbine anomaly detection addresses scarce normal-only target data and distribution shifts by mapping target measurements to data-rich source domains. Prior work uses fine-tuning and generative mappings, while this study extends mapping to multiple sources.
- Domain adaptation: Domain adaptation transfers knowledge from data-rich source domains to data-limited target domains despite distributional differences between them.
- Problem setting: Unsupervised anomaly detection assumes only normal training data, with anomalies appearing at inference and target observations scarce relative to source data.
- Multi-source extension: Multi-source adaptation uses several source turbines simultaneously, avoiding separate WT-to-WT models and addressing source-pair and model-selection challenges.
- Domain mapping: Generative domain mapping transforms unpaired target samples into source-domain samples while preserving content, after which a source normal behavior model scores anomalies.
- Prior work: Prior wind-turbine studies primarily use fine-tuning or single-source generative mappings to improve fault detection with scarce data.
3 Dataset
The dataset contains SCADA measurements from seven onshore wind turbines and evaluates multi-source transfer learning across setups with one scarce-data target and several data-rich sources. Target-data scarcity is simulated from one week to two months of measurements.
- Dataset: Seven operational wind turbines provide multiyear, 10-minute averaged SCADA measurements from different onshore wind farms.
- Data processing: Each 12-hour sample contains 11 channels describing wind speed, rotor speed, power, stator temperature, and rotor temperature.
- Experimental setups: Six experimental setups assign one target turbine and several source turbines, selecting sources that differ operationally from the target.
- Artificial scarcity: Target training and validation data are restricted to 1 week through 2 months, corresponding to 1008–8064 SCADA samples before the unaffected test set.
- Source and target domains: The target turbine has limited training data, whereas source turbines supply representative training data within each setup.
4 Methodology
The method uses a single StarGAN-based generative network to map scarce-target SCADA samples across multiple data-rich wind-turbine domains, preserving operational content for downstream anomaly detection. Source-domain normal behavior models score the mapped samples, whose binarized outputs are fused; model selection uses normal-data proxy scores, while early stopping guards against target overfitting.
- Multi-domain mapping: The generative network translates fault-free SCADA measurements between multiple wind-turbine domains while retaining operational-state characteristics.The approach adapts StarGAN to N source turbines and one data-scarce target turbine.
- StarGAN for WT mapping: StarGAN uses one shared generator and discriminator to learn mappings among multiple domains rather than training separate networks for each turbine pair.The generator receives a sample and destination-domain label, while discriminator losses enforce realism and domain classification.
- State preservation: Cycle consistency and physics-informed zero and rated-power losses preserve content and operational states during SCADA translation.The zero loss preserves idle states, while the rated-power loss matches the destination turbine’s rated power at corresponding operating positions.
- Training strategy: Anomaly augmentation corrupts randomly selected input channels so the generator does not overfit to repairing samples into normal-looking outputs.The training loop maps corrupted batches across domains to help preserve anomalous states during operation.
- Fault detection: At inference, target test samples are mapped to every source domain, scored by pretrained source normal behavior models, binarized, and fused into final anomaly scores.This multi-source procedure is the downstream fault-detection pipeline illustrated in Figure 3.
- Unsupervised model selection: The proxy metric compares real and mapped anomaly-score tails on normal validation data, selecting configurations with the lowest average tail-mass difference.Because source-domain mappings can improve while scarce-target samples overfit, early stopping instead uses anomaly-based criteria.
5 Results and Discussion
The multi-source StarGAN framework improves fault detection most strongly under severe data scarcity, while ensemble fusion and proxy-based model selection offer practical benefits with important limits. Performance depends on source-turbine choice, training-data representativeness, and the amount of target data available.
- Ensemble fusion: Anomaly-biased majority voting outperforms consensus and minority voting when combining anomaly scores from multiple mapping directions.Majority voting is robust to poorly converged or unusually strong mapping directions and generally approaches average performance.
- Baseline comparisons: 91.7% of single-source domain-mapping cases exceeded data-scarce NBMs, and 79.2% exceeded fine-tuning when training data was limited to one month or less.With larger two-month training sets, negative transfer remained: data-scarce NBMs exceeded fine-tuning in 66.7% and single-source mapping in 50% of setups.
- StarGAN-based mapping: 83.3% of severe-scarcity cases were won by StarGAN, with one-week gains of +36.9%, +24.3%, and +22.9% over the data-scarce NBM, fine-tuning, and CycleGAN mapping.With two weeks of data, the corresponding improvements were +22.8%, +15.5%, and +9.7%.
- StarGAN-based mapping: Shared StarGAN representations across data-rich turbines are associated with better out-of-distribution performance and reduced dependence on scarce target samples.The authors attribute the gains to learning source-to-source mappings and shared normal-behavior patterns, which can improve training stability and reduce overfitting.
- Performance with more available training data: With three or more weeks of data, StarGAN exceeded the no-transfer baseline in 72.2% of cases, but its advantage over single-source methods weakened.It exceeded both domain-adaptation baselines in only 29.2% of cases; at two months, its F1 changes versus the three baselines were -3.5%, -0.3%, and -4.5%.
- Unsupervised model selection: The proxy metric ΔT mainly filters poor from functional mappings rather than reliably selecting the best target-turbine model.Its discrimination among high-performing models diminishes, and turbine-specific cycle-consistency weights can produce target F1 scores above 90% or around 65%.
6 Conclusions
The study introduces multi-source StarGAN domain mapping for fault detection under scarce wind-turbine training data. It reports strong performance under severe scarcity and proposes a fault-free proxy for filtering poor model configurations.
- The StarGAN-based model maps scarce target-turbine SCADA measurements to resemble multiple data-rich source turbines simultaneously.It preserves operational states, including idle and anomalous states, during translation.
- Mapped samples can be evaluated by pretrained normal behavior models from the source turbines.
- With 1–2 weeks of training data, the approach largely outperformed single-source fine-tuning and domain mapping.The authors attribute this to shared representations learned across multiple source domains.
- The multi-source benefit no longer held as target-turbine training data became more available.
- A fault-free proxy metric reliably filtered poorly performing configurations but could not sufficiently distinguish among well-performing candidates.
Appendix A: Dataset Overview
Appendix A documents the seven operational wind turbines used in the study and explains how rated power values support data processing.
- Table A1 lists details for all 7 wind turbines used in the study.The dataset was adapted from an earlier study using the same data.
- Observed rated power values were used for filtering and rated-power-loss calculations.
Appendix B: Training Details
The appendix describes autoencoder normal behavior baselines and a StarGAN-based multi-source mapping system trained with adversarial, classification, cycle-consistency, and reconstruction-related objectives.
- Normal behavior models: Each turbine used an autoencoder-based normal behavior model trained by minimizing mean squared reconstruction error on fault-free 12h-SCADA samples.Models covered both the full representative training set and 1–8-week scarcity scenarios.
- StarGAN-based domain mapping model: The StarGAN generator and discriminator were adapted to multiple source domains by increasing capacity and incorporating destination labels and classification.
- Implementation: Training uses Adam for the StarGAN model, PyTorch implementation, and a single NVIDIA GPU.The implementation is publicly released on GitHub.
- Architectures: Table A3 specifies the generator and discriminator architectures, including the residual-block structure illustrated in Figure A1.
- Discriminator architecture: The discriminator uses a 1D convolutional architecture with 11-channel, 72-point inputs, convolutional blocks, a critic, and a classifier.
- Training algorithm: The algorithm samples origin and randomly selected destination batches, maps them in both directions, and cycles them back to their origins.
- Training objectives: Generator training combines adversarial, classification, cycle-consistency, zero, and reconstruction-related loss terms.
- Training objectives: GAN-QP supplies the adversarial objective, while the discriminator also uses classification loss for real and mapped samples.The stated quadratic-potential weight is λ_QP = 1, with L1 distance d.
Appendix C: Results
Appendix C reports F1 scores across baselines, setups, and data-scarcity scenarios, with majority voting combining binarized anomaly scores.
- Table A4 reports F1 Scores [%] for all baselines, setups, and data scarcity scenarios.
- F1 scores combine N binarized anomaly scores using a majority voting strategy.
- Within each setup and scarcity scenario, bold marks the best score and underlining marks the second-best performance.
Appendix D: Detailed Results
Table A5 reports detailed F1 scores for all target-to-source mappings across experimental setups and data-scarcity scenarios. It also defines the abbreviations used for fine-tuning and the two mapping approaches.
- Table A5 reports F1 scores, expressed as percentages, for all target-to-source mappings.
- The table covers all experimental setups and data-scarcity scenarios.
- FT denotes fine-tuning, CG denotes CycleGAN-based single WT-to-WT mapping, and Ours denotes StarGAN-based WT-to-WT mapping.