Source-linked AI summary
A Hybrid Two-Stage Machine Learning Pipeline for Fault Detection and Classification in Power Transmission Systems
Sahil Manikshete, Atharva Gujarathi, Thanh Long Vu, Akhtar Hussain, Van-Hai Bui
TL;DR
Transmission fault diagnosis must handle severe class imbalance and fault signatures that overlap normal operation. The paper proposes a hybrid two-stage pipeline that separates recall-oriented detection from fault-type classification, using conditional supervised detection and physics-informed features. Across independent datasets, it improves difficult fault performance and achieves strong end-to-end accuracy, while showing that zero-sequence decision boundaries must account for system dependence.
Problem
Transmission fault datasets are imbalanced, and some fault signatures are nearly indistinguishable from normal operation, limiting single-model detection.
Method
A two-stage pipeline combines Isolation Forest anomaly detection with conditional supervised binary detection, then applies Random Forest classification to flagged samples using per-point engineered features.
Results
97.25% end-to-end accuracy was achieved on an independent single-point dataset, while Line-fault accuracy on TLFaultDataset rose from 31.3% to 95.8%.
Takeaways & Limitations
Zero-sequence features resolve the three-phase versus three-phase-to-ground ambiguity, while learned boundaries accommodate system-dependent signature direction.
Takeaways & Limitations
Both evaluation datasets are simulated, so validation on field data remains the primary next step.
Abstract
from arXiv · showhide
Rapid and accurate fault detection in high-voltage transmission networks is essential for grid reliability and equipment protection. Transmission fault datasets are frequently imbalanced, and certain fault types produce electrical signatures that fall within the normal operating envelope, causing single-model classifiers to fail on safety-critical cases. This paper proposes a hybrid two-stage machine learning pipeline that decouples detection from classification. Stage 1 combines an Isolation Forest anomaly detector with an optional supervised binary detector through an OR-fusion rule; the supervised branch is allocated automatically during training for any fault class the anomaly detector cannot resolve, and is omitted when no such class exists. Stage 2 applies a Random Forest multiclass classifier only to samples flagged by Stage 1. Feature engineering is expressed as a per-measurement-point operator mapping six raw channels to eighteen features, including zero-sequence symmetrical components derived from Fortescue's theorem, yielding 18L features for L measurement points. On the TLFaultDataset, the pipeline raises Line-fault end-to-end accuracy from 31.3% to 95.8%. On an independent single-point dataset, the same framework attains 97.25% end-to-end accuracy across all classes including normal operation, exceeding the TLFed federated benchmark of 94.84% without GPU or federated infrastructure, at 0.05 ms per sample on CPU. Ablation on both datasets shows zero-sequence features resolving the three-phase versus three-phase-to-ground ambiguity, raising the F1-score of that class pair from 0.39 to 0.997. The direction of the zero-sequence signature is found to be system-dependent, motivating a learned decision boundary in place of a fixed relay threshold.
1 Introduction
Transmission fault diagnosis is difficult because datasets are heavily imbalanced and some critical faults resemble normal operation. The paper addresses these limitations with a conditional two-stage pipeline, physics-informed features, and system-aware zero-sequence analysis.
- Motivation: 88.1% of TLFaultDataset samples represent normal operation, while the rarest fault type accounts for under 1%.This imbalance biases classifiers toward the majority class and can obscure rare faults.
- Motivation: 31.3% is the single-model end-to-end detection accuracy for Line faults whose signatures nearly overlap normal operation.The overlap makes simultaneous fault detection and fault-type classification difficult for one model.
- Proposed approach: The proposed pipeline decouples detection from classification, using an Isolation Forest plus an automatically allocated supervised detector before Random Forest fault classification.The supervised branch is fused through OR logic only for classes the anomaly detector cannot resolve.
- Results: 95.8% is the Line-fault end-to-end accuracy achieved on TLFaultDataset, compared with 31.3% for the prior single-model result.On the second dataset, the allocation rule identifies a different blind-spot class, indicating system-specific allocation.
- Feature engineering: 18L features are constructed by applying the same per-measurement-point operator to six raw channels at L measurement points.The features include energy, phase asymmetry, and zero-sequence components.
- Feature engineering: 0.39 to 0.997 is the ablation improvement in F1-score for the three-phase versus three-phase-to-ground pair after adding zero-sequence features.The zero-sequence signature direction is system-dependent, so the paper favors a learned boundary over a fixed threshold.
2 System Model
The framework treats fault presence and fault-type identification as independently optimized stages. It applies network-size-independent feature construction, recall-oriented hybrid detection, and focused multiclass classification to flagged samples.
- System model: The pipeline applies identical feature construction, detection, and classification stages to transmission systems with L measurement points.The per-point outputs are concatenated, without referring to a particular topology or number of lines.
- Stage 1: Hybrid fault detection: Stage 1 prioritizes recall by flagging samples when the Isolation Forest or an allocated supervised detector raises an alarm.The supervised branch is optional and is omitted when no blind-spot class exists.
- Stage 1: Hybrid fault detection: Isolation Forest detects anomalies using random-tree path lengths and is trained exclusively on normal samples.Shorter average path lengths indicate samples that are easier to isolate from the learned normal distribution.
- Stage 2: Fault type classification: Stage 2 applies a Random Forest multiclass classifier only to Stage 1 fault flags and trains exclusively on true fault samples.This lets the classifier specialize in distinguishing fault types rather than separating faults from normal operation.
- Feature engineering: 18 features are generated per measurement point: six raw channels, eight statistical descriptors, and four zero-sequence quantities.For L = 7, this yields 126 features; for L = 1, it yields 18 features.
- Feature engineering: The imbalance ratio is the most informative statistical feature for Stage 1 anomaly detection because it magnifies phase asymmetry during faults.Zero-sequence current and voltage features add information about ground involvement that raw phase measurements do not expose directly.
- Evaluation: Per-fault-type end-to-end accuracy equals Stage 1 detection rate multiplied by Stage 2 classification accuracy for that fault type.End-to-end accuracy includes normal samples and counts Stage 1 misses as errors.
3 Datasets
The study evaluates the pipeline on two independently produced transmission-fault datasets that differ in network, simulation, scale, balance, and labeling. The primary dataset is highly imbalanced and contains near-normal Line faults, while the second dataset provides independent validation without those characteristics.
- Evaluation datasets: Two independently produced datasets differ in network size, measurement scale, simulation software, class balance, and label encoding.The first supports comparison with TLFed; the second tests transfer to an independently developed system.
- Primary dataset: The TLFaultDataset contains 578,923 samples from seven transmission lines, with 42 raw measurements per sample and seven fault categories including normal operation.Each line contributes three-phase currents and voltages.
- Primary dataset: Normal operation comprises 88.1% of the TLFaultDataset, while the rarest fault is under 1% and Line faults comprise 1.4%.Line-fault measurements are nearly identical to normal operation, limiting threshold-based and unsupervised detection.
- Second dataset: The second dataset is an independent single-measurement-point simulation used to test transfer of the architecture and feature construction.Its voltage and phase-current values are reported in per-unit form and range up to ±900 A.
- Second dataset: After merging overlapping files, the second dataset contains 6,505 normal and 5,496 fault samples, for a 45.8% fault ratio.The merge prevents identical samples from appearing across train and test partitions.
- Dataset scope: The second dataset is close to balanced, lacks a near-normal Line-fault class, and provides no fault-location label for evaluating localization.Labels are re-encoded from phase and ground indicators into the standard fault taxonomy.
4 Results
Stage 1 improves detection by fusing Isolation Forest with a class-specific supervised detector when anomaly detection misses a fault class. On the primary dataset, this particularly addresses Line faults that overlap with normal operation.
- Hybrid fault detection: 0.9% Line-fault detection by Isolation Forest increases to 98.5% with the fused detector.The allocation rule selects Line fault because its Branch A detection rate is the only one below τ = 0.5.
- Hybrid fault detection: The fused Stage 1 detector improves detection across every fault type on the TLFaultDataset.The supervised branch is allocated specifically to the class that the anomaly detector fails to resolve.
- Hybrid fault detection: The Isolation Forest is blind to faults whose signatures overlap with normal operation, whereas a labeled supervised detector addresses that class-specific limitation.Increasing the Branch B threshold from 0.50 to 0.70 reduces false alarms but lowers Line-fault recall.
4.3 Stage 2: Fault Type Classification (Primary Dataset)
Stage 2 classifies faults detected by Stage 1, with zero-sequence features resolving the primary dataset’s three-phase versus three-phase-to-ground confusion. The complete pipeline reaches 95.8% Line-fault end-to-end accuracy.
- Stage 2 classification: 99% Stage 2 accuracy and 0.99 macro F1 are achieved on true fault samples detected by Stage 1.L-L-L and L-L-L-G each reach F1 of 0.99 in the filtered classification input.
- End-to-end performance: The complete pipeline reaches 95.8% Line-fault end-to-end accuracy, up from 31.3% after adding the dedicated detector and zero-sequence features.Isolation Forest alone achieves 19.4% for Line faults, and the dedicated detector alone raises this to 31.3%.
4.5 Validation on the Second Dataset
The unchanged pipeline transfers to an independent single-point dataset, where the allocation rule selects L-L rather than Line fault. Overall end-to-end accuracy reaches 97.25%, with remaining errors governed mainly by Stage 1 detection.
- Independent validation: The architecture, feature operator, allocation rule, fusion rule, and classifiers transfer unchanged to the second dataset.Only the input adapter changes to identify the three-phase channel groups; L = 1 produces 18 features.
- Stage 1 allocation: 45.3% L-L detection falls below τ = 0.5, so the allocation rule instantiates one supervised detector for L-L and raises its detection to 100%.All other classes exceed 88% Isolation Forest detection on the validation split.
- Overall results: 97.25% overall end-to-end accuracy is achieved across 2,401 test samples, exceeding the TLFed benchmark of 94.84%.Stage 1 fault recall is 95.00%, while Stage 2 accuracy on detected faults is 99.33%.
- Error distribution: End-to-end accuracy is the product of the two stage-wise values, and detection rather than classification governs overall performance on the second dataset.Further improvement would therefore come from raising Stage 1 recall rather than using a stronger classifier.
4.6 Cross-Dataset Feature Ablation
Feature ablation shows that zero-sequence components specifically resolve the L-L-L versus L-L-L-G ambiguity on the independent dataset, while other classes are already separable from raw channels. The same effect appears across systems, supporting a physics-linked rather than dataset-specific interpretation.
- Stage 2 feature ablation: F1-scores rise from 0.400 and 0.375 to 1.000 and 0.993 for L-L-L and L-L-L-G after adding zero-sequence features.Raw channels leave the pair near random assignment, while statistical descriptors raise performance only to approximately 0.80.
- Stage 2 feature ablation: DLG, L-L, and SLG remain resolved from raw channels and vary by less than one percentage point across feature sets.The engineered features neither help nor harm classes already separable from the raw measurements.
- Interpretation: The zero-sequence benefit is confined to the L-L-L and L-L-L-G pair rather than acting as a generic increase in representational capacity.This selectivity is consistent with zero-sequence features encoding ground involvement specifically.
- Cross-dataset interpretation: The same feature group resolves the three-phase versus three-phase-to-ground pair on two systems differing in size, scale, and simulation software.The cross-system repetition indicates that the effect follows from the underlying physics rather than one dataset.
4.7 System Dependence of the Zero-Sequence Signature
The zero-sequence signature separates three-phase from three-phase-to-ground faults, but its direction depends on fault balance and grounding conditions. Consequently, a fixed magnitude threshold is not portable across systems, whereas a learned boundary remains effective within the reported scope.
- System dependence: A fixed |I0| threshold calibrated on one system can misclassify the three-phase fault pair on another system.The ordering reverses: on the second dataset, median |I0| is 1 × 10−3 for L-L-L-G versus 0.70 for L-L-L.
- Physical interpretation: Zero-sequence current depends on both a ground path and asymmetry among phase currents, so ground involvement alone does not guarantee substantial residual current.Balanced bolted three-phase-to-ground faults can leave phase currents symmetric and produce little residual.
- Physical interpretation: The second dataset appears dominated by balanced three-phase-to-ground events, while the TLFaultDataset has sufficient asymmetry to produce measurable residual current.A controlled sweep over fault-resistance asymmetry is identified as future work to establish this mechanism directly.
- Methodological consequence: A learned boundary using zero-sequence and remaining features attains F1 above 0.99 for the pair on both systems.The reported cross-system observation motivates learned classification rather than a fixed-threshold relay rule in this setting.
4.8 Computational Cost
The complete two-stage pipeline is evaluated for CPU inference without GPU acceleration. Its measured per-sample cost supports comparison with the federated deep-learning baseline on both accuracy and hardware requirements.
- Inference cost: 0.052 ms per sample is required by the complete pipeline on the second dataset during CPU inference.The measurement includes Branch A scoring, Branch B scoring, and Stage 2 classification.
- Inference cost: Approximately 1.9 × 104 samples per second are processed on a single machine.The comparison is framed against TLFed without GPU hardware or distributed training infrastructure.
5 Discussion and Future Work
The discussion finds that the architecture, zero-sequence features, and learned per-system boundaries drive cross-system performance, while false alarms and simulation-only evaluation limit deployment scope.
- Architecture: The allocation rule identifies structurally invisible or marginally weak fault classes automatically, rather than requiring advance specification.It identifies Line-fault on the primary dataset and L-L on the second, with detection rates of 0.9% and 45.3% at the reported operating points.
- Feature perspective: Zero-sequence components resolve the L-L-L versus L-L-L-G ambiguity on both systems, but the separating direction is system-dependent.A fixed rule such as “|I0| is larger for ground faults” is therefore unsuitable; the boundary must be learned.
- Cross-system transfer: The same architecture and allocation rule transfer across systems, whereas models fitted on one dataset perform near chance when applied directly to the other.The dominant cross-dataset failure is inversion of the three-phase class pair.
- Trade-off: Adding the Branch B detector increases the primary dataset’s false alarm rate, creating an operational burden in high-volume monitoring.Its decision threshold directly controls the trade-off, which depends on deployment priorities.
- Future work: Both datasets are simulated, so field validation remains necessary because operational measurements include noise, sensor drift, harmonics, and transient disturbances.The pipeline also lacks temporal context and does not localize faults along the line.
6 Conclusion
The paper addresses imbalanced data and fault signatures overlapping with normal operation by separating recall-oriented detection from precision-oriented classification. Across two independent systems, the framework improves difficult fault recognition, achieves strong end-to-end accuracy, and shows that learned, system-specific boundaries matter.
- Conclusion: Separating detection from classification addresses datasets dominated by normal samples and fault signatures nearly indistinguishable from normal operation.Detection is optimized for recall, while classification is optimized for per-class precision.
- Conclusion: 95.8% Line-fault end-to-end accuracy was achieved on the primary dataset, up from 31.3%.The improvement combines an Isolation Forest with an automatically allocated supervised binary detector for unresolved classes.
- Conclusion: 97.25% end-to-end accuracy across all classes was achieved on the second dataset, exceeding TLFed’s 94.84% without federated or GPU infrastructure.Inference took 0.05 ms per sample on CPU.
- Conclusion: A learned boundary is necessary because zero-sequence signature direction varies with fault balance and grounding impedance across networks.A fixed threshold calibrated on one network would fail on the other, whereas a learned boundary succeeds on both.