⚡ electrical engineering

A Hybrid Two-Stage Machine Learning Pipeline for Fault Detection and Classification in Power Transmission Systems

This paper proposes a hybrid two-stage machine learning pipeline that decouples anomaly detection from fault classification using an Isolation Forest-OR-fusion strategy and zero-sequence feature engineering, achieving over 95% accuracy in detecting and classifying transmission faults on imbalanced datasets while outperforming existing federated benchmarks without requiring GPU infrastructure.

Original authors: Sahil Manikshete, Atharva Gujarathi, Thanh Long Vu, Akhtar Hussain, Van-Hai Bui

Published 2026-08-26

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Sahil Manikshete, Atharva Gujarathi, Thanh Long Vu, Akhtar Hussain, Van-Hai Bui

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The electric grid is the invisible nervous system of modern society, carrying power from distant generators to the homes and factories that depend on it. This vast network relies on high-voltage transmission lines to move electricity efficiently over long distances. However, these lines are vulnerable. When a fault occurs—caused by a lightning strike, a broken wire, or equipment failure—the resulting surge of energy can damage expensive machinery or trigger widespread blackouts. To prevent this, the grid uses protection systems that act like automatic circuit breakers, designed to spot a problem and cut the power in a fraction of a second. For decades, these systems have relied on fixed rules: if the current rises above a certain number, trip the switch. But as the grid becomes more complex, with power flowing in new directions and coming from diverse sources, those simple rules sometimes fail. They struggle to tell the difference between a dangerous fault and a normal, albeit unusual, fluctuation in the system, leaving critical gaps in safety.

Researchers have turned to machine learning to fill these gaps, hoping to teach computers to recognize the subtle signatures of a failing line. Yet, a new study reveals that simply handing a computer a massive list of data and asking it to find the bad spots is not enough. The problem is twofold: first, the data is heavily skewed, with normal operation appearing millions of times more often than a fault, causing the computer to ignore the rare events it is supposed to find. Second, some specific types of faults look so much like normal operation that a single computer model cannot tell them apart. To solve this, a team of engineers and computer scientists has developed a two-step approach that separates the job of "noticing something is wrong" from the job of "figuring out exactly what is wrong."

The researchers tested their method on two different simulated power networks. The first was a large, complex system with seven measurement points and nearly 580,000 data samples, while the second was a smaller, single-point system with about 12,000 samples. In the large network, they found that a standard computer model trying to do everything at once failed miserably at spotting a specific type of trouble called a "line fault," which occurs when a transmission line itself is damaged. In this scenario, the standard model only caught about 31 percent of these faults. The researchers realized that the computer was confused because the electrical signals from a line fault were nearly identical to the signals from a healthy grid.

To fix this, they built a pipeline with two distinct stages. The first stage acts as a broad net, designed to catch anything that looks even slightly out of the ordinary. They used a technique called an Isolation Forest, which learns what "normal" looks like by studying only healthy data. If a new reading deviates from this pattern, the system flags it. However, the researchers knew this net would miss the line faults because they look so normal. So, they added a second, specialized detector that only activates for the specific types of faults the first net misses. This specialized detector is trained to recognize the exact pattern of a line fault. The two detectors work together: if either one raises an alarm, the system moves to the second stage.

The second stage is where the system identifies the specific type of problem. Once a sample is flagged as a fault, a different computer model sorts through the possibilities to determine if it is a lightning strike, a broken wire, or a ground connection. Crucially, the researchers added a specific type of mathematical calculation to the data before feeding it to the computer. They calculated what is known as zero-sequence components, which measure the imbalance in the electrical current that flows to the ground. This calculation is based on a fundamental principle of electricity that states currents in a healthy, balanced system should sum to zero. When a fault involves the ground, this balance is broken. By including this specific calculation, the computer could finally distinguish between a fault that touches the ground and one that does not, a task it previously could not do.

The results were striking. On the large network, the new two-stage system improved the detection of line faults from 31 percent to nearly 96 percent. On the smaller network, the system achieved an overall accuracy of 97.25 percent, correctly identifying faults and normal operation alike. This performance was better than a previous, more complex system that required expensive graphics hardware and a distributed network of computers to run. The new method ran on a standard computer processor in just 0.05 milliseconds per sample, making it fast enough for real-world use without needing a supercomputer.

Perhaps the most important discovery came from comparing the two different networks. The researchers found that while the zero-sequence calculation was essential for solving the problem, the way the signal behaved was not the same in both systems. In one network, a ground fault produced a larger signal than a non-ground fault, but in the other, the opposite was true. This means that a simple, fixed rule—like "if the signal is bigger than X, it is a ground fault"—would work on one system but fail on the other. The researchers concluded that the computer must learn the specific rule for each new grid it is installed in, rather than relying on a universal rule. This finding suggests that the best protection systems will not be rigid, pre-programmed devices, but adaptable tools that learn the unique electrical personality of the grid they protect.

The study also highlighted a trade-off. By making the system more sensitive to catch every possible fault, the researchers increased the number of false alarms, where the system flags a problem that isn't there. In the context of power grid safety, this is an acceptable cost. It is far better to investigate a false alarm than to miss a real fault that could cause a blackout. The researchers noted that the system is flexible enough to adjust this balance depending on the needs of the specific grid. While the current results are based on simulated data, the approach offers a clear path forward for real-world deployment. The next step will be to test these methods on actual power lines, where the noise and unpredictability of the real world will provide the ultimate test of the system's reliability.

Technical Summary: A Hybrid Two-Stage Machine Learning Pipeline for Fault Detection and Classification in Power Transmission Systems

Problem Statement
Rapid and accurate fault detection in high-voltage transmission networks is critical for grid reliability. However, existing single-model machine learning approaches face two primary challenges in this domain:

Severe Class Imbalance: Transmission fault datasets are dominated by normal operation samples (e.g., 88.1% in the primary dataset), causing classifiers to bias toward the majority class and fail to detect rare but critical faults.
Signature Overlap: Certain fault types, particularly "Line faults," produce electrical signatures nearly indistinguishable from normal operation. In the primary evaluation, a single-model approach achieved only 31.3% end-to-end accuracy for these faults, rendering them effectively invisible to standard anomaly detectors trained on normal data.

Traditional protection methods relying on fixed thresholds struggle with modern networks characterized by distributed energy resources and bidirectional power flow. While deep learning and federated learning (e.g., TLFed) have shown promise, they often treat all fault classes uniformly and rely on complex infrastructure (GPUs, distributed clients) without specifically addressing the structural blind spots of near-normal fault signatures.

Methodology
The paper proposes a hybrid two-stage machine learning pipeline that decouples fault detection from fault classification, optimizing each stage for distinct objectives.

Stage 1: Hybrid Fault Detection (Recall-Oriented)
This stage determines whether a fault is present, prioritizing recall over precision to minimize missed faults. It employs an OR-fusion rule combining two branches:
- Branch A (Unsupervised): An Isolation Forest trained exclusively on normal data to detect general anomalies.
- Branch B (Conditional Supervised): A dedicated supervised binary Random Forest detector allocated automatically during training for any fault class that Branch A fails to resolve (defined as a detection rate below a threshold $\tau$ ).
- Fusion: A sample is flagged as a fault if either Branch A or Branch B raises an alarm. This ensures that classes invisible to the anomaly detector are captured by the supervised branch.
Stage 2: Fault Type Classification (Precision-Oriented)
This stage operates only on samples flagged by Stage 1. It uses a Random Forest multiclass classifier trained exclusively on true fault samples. This separation allows the classifier to specialize in distinguishing between fault types without the noise of normal operation data.
Feature Engineering
The framework utilizes a network-size-independent feature operator $\phi$ applied per measurement point. For $L$ measurement points, the pipeline generates $18L$ features from six raw channels (three-phase currents and voltages). The 18 features per point include:
- 6 raw channels.
- 8 statistical descriptors (RMS, max/mean absolute values, imbalance ratios).
- 4 zero-sequence quantities (signed and absolute $I_0$ and $V_0$ derived from Fortescue's symmetrical components).
  The inclusion of zero-sequence components is critical for resolving ambiguities between faults with and without ground involvement.

Key Contributions

Conditional Two-Stage Architecture: The paper introduces a mechanism where a supervised detector is dynamically allocated only for "blind-spot" classes that an unsupervised anomaly detector cannot resolve. This adapts to the specific characteristics of the system rather than assuming a fixed set of difficult classes.
Network-Size Independent Feature Construction: The feature engineering is defined as a per-point operator, allowing the same pipeline to function on networks with varying numbers of measurement points ( $L=7$ and $L=1$ ) without architectural modification.
Learned Zero-Sequence Boundaries: The study demonstrates that zero-sequence features are essential for distinguishing three-phase faults from three-phase-to-ground faults. Crucially, it shows that the direction of the zero-sequence signature is system-dependent (influenced by fault balance and grounding impedance), arguing for a learned decision boundary over fixed relay thresholds.

Results
The framework was evaluated on two independent datasets: the TLFaultDataset (7 measurement points, 578,923 samples, severe imbalance) and a Second Dataset (1 measurement point, 12,001 samples, balanced).

Primary Dataset (TLFaultDataset):
- The hybrid pipeline raised the end-to-end accuracy for "Line faults" from 31.3% (single-model baseline) to 95.8%.
- Branch B successfully identified "Line faults" as the blind-spot class, increasing their detection rate from 0.9% (Isolation Forest alone) to 98.5%.
- Stage 2 achieved a macro-averaged F1-score of 0.99 on detected faults.
- Zero-sequence features resolved the confusion between L-L-L and L-L-L-G faults, significantly improving classification performance on this dataset.
Second Dataset (Independent Validation):
- The pipeline achieved 97.25% end-to-end accuracy across all classes (including normal operation), exceeding the TLFed federated benchmark of 94.84%.
- Unlike the primary dataset, the allocation rule identified "L-L" (Line-to-Line) faults as the blind-spot class, demonstrating the mechanism's adaptability.
- Inference time was 0.05 ms per sample on a CPU, requiring no GPU or federated infrastructure.
- Ablation Studies: Removing zero-sequence features caused the classifier to fail at distinguishing three-phase from three-phase-to-ground faults on both datasets. Specifically, on the second dataset, the F1-score for the three-phase versus three-phase-to-ground class pair improved from 0.39 to 0.997 upon adding zero-sequence features.
- The direction of the zero-sequence signature ( $|I_0|$ magnitude) was found to be anti-predictive on the second dataset compared to the primary dataset, confirming that fixed thresholds are not portable across systems.

Significance and Claims
The paper claims that decomposing the fault diagnosis task into detection and classification stages, combined with physics-informed feature engineering, offers a more robust solution than increasing model complexity or relying on fixed thresholds.

Adaptability: The automatic allocation of supervised detectors allows the system to adapt to the specific "blind spots" of different transmission networks without prior specification.
Physical Grounding: The results validate that zero-sequence components provide necessary discriminative information for ground-involved faults, but the decision boundary must be learned rather than fixed due to system-dependent variations in fault balance and grounding impedance.
Efficiency: The pipeline achieves state-of-the-art accuracy on independent datasets without the computational overhead of deep learning or federated learning, making it suitable for deployment in resource-constrained substation environments.

The authors note limitations, including the use of simulated data (lacking real-world noise and harmonics) and the lack of temporal context in the current pipeline, identifying field validation and the incorporation of temporal dynamics as necessary future steps.

More like this