A lightweight multi-scale deep learning approach for robust photovoltaic fault detection under noisy conditions

Solar panels are the quiet workhorses of the modern energy transition, turning sunlight into electricity without moving parts or smoke. Yet, like any complex machine, they are prone to breaking down. A single cracked wire, a patch of shade from a nearby tree, or a failing component can drag down the performance of an entire system, wasting energy and money. The challenge for engineers has long been figuring out exactly what is wrong, and how quickly, especially when the data coming from the panels is messy. Sensors often pick up random static, calibration errors, or sudden electrical spikes that look like faults but are just noise. Distinguishing a genuine problem from a glitch in the measurement has been a difficult puzzle, often requiring expensive equipment or slow, manual inspections that leave systems running inefficiently for too long.

A team of researchers at Kookmin University in South Korea has developed a new way to solve this problem, creating a digital system that can spot these issues with remarkable speed and accuracy, even when the data is imperfect. They built a specialized computer program designed to act like a highly trained diagnostician. Instead of relying on a single method to look at the data, this program uses a hybrid approach that mimics how a human expert might examine a problem from multiple angles at once. It looks for small, local patterns in the electrical signals while also understanding the broader context of how those signals relate to one another over time. This dual perspective allows it to ignore the background static and focus on the true signs of trouble.

The system, which the researchers named BLGSNet, was tested on a dataset representing six different states of a solar array: normal operation, a broken string of panels, a ground fault, a short circuit between strings, a failing component, and partial shading. In a perfect, noise-free environment, the system correctly identified the condition of the panels 99.36% of the time. This is a high bar, but the real test came when the researchers introduced realistic interference. They simulated the kind of sensor errors and electrical noise that happen in the real world, adding random static and sudden spikes to the data. Even when the data was corrupted by noise at a level of 10%, the system maintained an accuracy of 98.08%. When the noise was increased to a challenging 15%, it still managed to diagnose the faults correctly 94.87% of the time. This stability is crucial because it means the system does not need to be paired with complex, slow filters to clean the data first; it can handle the messiness directly.

What makes this approach particularly effective is its design. The researchers combined two different types of digital processing. One part of the system scans the data using three different "lenses" simultaneously, looking for patterns that are very short, medium, and slightly longer. This allows it to catch faults that might look different depending on how closely you examine them. The other part of the system then takes these findings and looks at them in both forward and backward directions, understanding how the electrical features connect to one another in a sequence. This combination allows the system to learn the unique "fingerprint" of a fault without being confused by the surrounding noise. Unlike older methods that often struggle when the data gets dirty or require massive amounts of computing power, this new design is lightweight. It is small enough to run on a compact, low-power computer the size of a credit card, known as an NVIDIA Jetson Nano, which was used in the team's laboratory tests.

To prove the system works outside of a computer simulation, the team built a small-scale solar setup in their lab. They connected two solar panels to a battery and an inverter, then ran the system through various fault conditions, including disconnected wires and artificial shading. The system processed the live data in real time, making a diagnosis in less than 19 milliseconds. This speed is fast enough to allow for immediate action. When the system detects a problem, it does not just say "something is wrong." It translates that finding into a clear, actionable plan. If it identifies a string fault, it recommends checking the wiring and fuses. If it sees partial shading, it suggests looking for debris or obstructions on the panel surface. This decision-support layer turns a raw data point into a practical instruction for a maintenance worker.

The researchers also compared their new system against several existing methods, including traditional machine learning models and other deep learning networks. The results showed that while other models could perform well in clean conditions, their accuracy dropped significantly when noise was introduced. The new system, however, remained robust. It achieved this high performance while using far fewer computing resources than its competitors, requiring less than half the memory and a fraction of the processing power of other advanced models. This efficiency suggests that such technology could be deployed widely, even on small or remote solar installations where powerful computers are not available.

While the results are promising, the researchers are careful to note the boundaries of their current work. The testing was done on a relatively small dataset and a limited two-panel setup in a controlled lab environment. The "real-world" data they used for further validation was generated using rule-based labels rather than verified maintenance records from actual field failures. This means the system has not yet been tested over years of operation in diverse weather conditions across large solar farms. However, the consistency of the results across different noise levels and the successful real-time demonstration on physical hardware provide a strong foundation. The work demonstrates that it is possible to build a fault detection system that is not only accurate but also resilient to the imperfect, noisy reality of the physical world, offering a path toward more reliable and self-sustaining solar energy networks.