Leakage-Safe Cross-Domain Partial-Discharge Void Diagnosis: A Benchmark on the PURE Dataset

High-voltage power cables are the arteries of modern civilization, carrying electricity across cities and landscapes. To keep them running safely, engineers must detect a subtle warning sign called a partial discharge. This is not a full electrical short that causes a blackout, but a tiny, localized spark that jumps across a small gap inside the insulation. Over time, these sparks eat away at the material, eventually leading to failure. For decades, experts have tried to build computer programs that can listen to these sparks and identify exactly what kind of defect is causing them. The hope has been that a smart algorithm could look at the shape of an electrical pulse and tell the difference between a single air bubble trapped in the plastic, two bubbles side-by-side, or two bubbles stacked on top of each other. If a computer could do this reliably, it could predict cable failures before they happen.

However, a new study reveals a significant hurdle in this quest. While computer models can become incredibly accurate when tested on data from the exact same conditions they were trained on, they often fail completely when faced with a slightly different real-world situation. Researchers from the Anhui University of Science and Technology set out to test this problem using a public collection of electrical measurements. They wanted to know if a model trained to recognize defects in one type of plastic under one type of voltage could still recognize those same defects when the plastic changed or the voltage changed. Their findings suggest that the current generation of these diagnostic tools is far less robust than previously thought, and that the way researchers have been testing them in the past may have been hiding this weakness.

The researchers focused on a specific dataset containing thousands of electrical pulses recorded from thin sheets of plastic. These sheets had artificial holes drilled into them to simulate three different types of defects: a single hole, two holes side-by-side, and two holes in a line. The data was collected under four different conditions: two types of plastic (low-density polyethylene and polypropylene) and two types of electrical pressure (alternating current and direct current). In total, the study analyzed 5,800 individual electrical pulses grouped into 58 separate recording files. The crucial step the team took was to treat each entire recording file as a single unit. In previous studies, researchers often mixed individual pulses from the same file into both the training and testing groups. This is like giving a student a practice test and then testing them on the exact same questions; it inflates the score. By keeping all pulses from one file together, the researchers ensured that when they tested the model, it was truly encountering a new, unseen measurement session, not just another pulse from a session it had already memorized.

When the researchers tested their models using this strict, leakage-safe method, the results were stark. A standard computer model that performed nearly perfectly when tested on data from the same material and voltage condition saw its accuracy collapse when the conditions changed. When the model was asked to identify defects in a different type of plastic, its ability to correctly classify the defects dropped to a level barely better than random guessing. Similarly, when the voltage type changed from alternating to direct current, the model's performance plummeted. The most troubling finding was that the model completely failed to recognize the single-hole defect in these new conditions, often misidentifying it as a different type of defect or missing it entirely. This failure happened regardless of whether the researchers used simple statistical tools or complex, modern deep-learning networks. Even when the researchers tried to teach the models to be "domain-invariant"—meaning they tried to force the models to ignore the differences between the plastic types and voltage conditions—the models still could not learn to recognize the defects in the new environments.

The study also investigated why these failures occurred. They found that the problem was not that the models were too simple or lacked enough computing power. Instead, the models had learned to rely on subtle clues that were specific to the exact conditions of the training data, such as the specific way the plastic reacted to the voltage or the unique signature of the recording equipment. When those conditions changed, the clues the models relied on disappeared, and the models were left blind. The researchers discovered that the most critical factor for failure was how the different types of defects shifted relative to each other in the data. If the model could not keep the identity of the single-hole defect distinct from the others when the environment changed, the entire system failed. They also tested a technique called data augmentation, where the training data is artificially tweaked with small changes like noise or shifting the signal in time. This technique helped the model perform better when it was trained on a mix of all four conditions and then tested on a new one, but it did not help when the model was trained on just one condition and tested on another.

Perhaps the most sobering part of the research came from a small set of data involving actual power cables, rather than the thin plastic sheets used for the main tests. The models, which had been trained on the plastic sheets, were completely unable to recognize the defects in the real cables. They misclassified every single cable record, consistently guessing the wrong type of defect. This suggests that the models had learned a representation of the defects that was specific to the thin plastic sheets and did not translate to the complex geometry of a real cable. The researchers concluded that while these models are excellent at describing what they have already seen, they are not yet ready to be trusted with the job of diagnosing new, unseen power lines.

The study does not claim that artificial intelligence cannot solve this problem, but it argues that the path forward requires a fundamental change in how these tools are evaluated. The current practice of reporting high accuracy scores based on data that is not truly independent is misleading. To build a reliable diagnostic tool, engineers must test their models on data that comes from different materials, different voltages, and different recording setups, just as a real-world system would encounter. The researchers propose a new standard for testing that separates the training data from the testing data at the level of the entire recording file, ensuring that no information leaks between the two. They also emphasize that performance must be reported separately for each type of defect and each type of shift, because a model might work well for one type of defect but fail completely for another. Until these rigorous tests are passed, and until models are proven to work across different physical conditions, the promise of automated cable diagnosis remains out of reach. The work serves as a necessary reality check, reminding the scientific community that high accuracy in a controlled lab does not guarantee safety in the field.