TRIBRID-Optimized MobileNetV3 Framework for Explainable Printed Circuit Board Defect Classification

Imagine you are a detective trying to solve a mystery, but instead of looking for fingerprints, you are looking for tiny, invisible cracks in the circuit boards that power your phone, your car, and even the satellites in space. These circuit boards, called Printed Circuit Boards (PCBs), are the nervous systems of modern electronics. If even a microscopic hole is missing or a tiny wire is touching where it shouldn't, the whole device can fail. For a long time, humans had to squint at these boards under microscopes to find the trouble, which was slow and tiring. Then, computers got smart enough to look at pictures of these boards and spot the errors automatically. This is where "Deep Learning" comes in—a type of computer brain that learns to recognize patterns just like we do, but much faster. However, there's a catch: these computer brains are often like black boxes. They might say, "This board is broken," but they can't explain why or point to the exact spot. In a factory, you need to trust the machine, so you need to know what it's looking at.

This paper introduces a clever new way to fix that problem. The researchers built a system that doesn't just spot defects on circuit boards; it also explains its reasoning, like a detective showing you the evidence. They used a lightweight computer brain called MobileNetV3, which is designed to be fast and efficient, perfect for running on factory machines that don't have super-computer power. But here's the twist: they didn't just let the computer learn on its own. They used a special "coach" called TRIBRID to automatically tune the computer's settings, finding the perfect recipe for learning without a human having to guess. Finally, they added a tool called Grad-CAM, which acts like a high-tech highlighter, drawing a glowing map over the exact part of the image that made the computer decide a defect was there. The result? A system that is not only incredibly accurate but also transparent, showing exactly where it found the problem.

The Detective's New Toolkit

The story begins with the challenge of inspecting Printed Circuit Boards (PCBs). These are the green (or sometimes blue or black) boards you see inside almost every electronic device. They hold all the tiny components together and let electricity flow between them. If a hole is missing, a wire is broken, or two wires touch by accident, the device stops working. Traditionally, humans checked these boards, but it was slow and prone to mistakes because people get tired. Later, machines called Automated Optical Inspection (AOI) systems took over, using simple rules to find defects. But these machines struggled when the boards got complicated or the lighting changed.

Enter Deep Learning. Think of a Deep Learning model as a student who has looked at millions of pictures and learned to recognize a "dog" or a "cat" without anyone telling them what features to look for. In this paper, the researchers wanted to teach a computer to recognize six specific types of PCB defects: missing holes, mouse bites (tiny nicks on the edge), open circuits (broken wires), short circuits (wires touching), spurs (unwanted metal spikes), and spurious copper (extra copper where it shouldn't be).

The researchers chose a specific type of computer brain called MobileNetV3. Imagine this as a "lightweight" detective. Unlike the heavy, slow detectives (like ResNet or EfficientNet) that need massive computers to think, MobileNetV3 is designed to be fast and efficient, making it perfect for real-time factory use. However, even a smart detective needs the right settings to work perfectly. If the learning rate is too fast, it misses details; too slow, and it takes forever. Usually, humans have to guess these settings, which is like trying to tune a radio by turning the knob randomly.

To solve this, the team used a method called TRIBRID. Think of TRIBRID as a super-smart coach that automatically adjusts the detective's settings. It combines three different strategies to find the perfect combination of learning speed, how much information to ignore (dropout), how many images to look at at once (batch size), and how many layers of the brain to retrain. It doesn't just guess; it explores different options, learns from the best ones, and even uses a bit of randomness to avoid getting stuck in a bad spot.

Finally, to make sure the factory workers trust the system, the researchers added Grad-CAM. This is the "highlighter" tool. When the computer says, "This is a short circuit," Grad-CAM draws a heat map over the image, glowing red on the exact spot that convinced the computer. It proves the computer isn't just guessing based on the background; it's actually looking at the defect.

The Big Test

The researchers tested their new system on a public dataset containing 2,953 images of these six defect types. They split the data so the computer could learn on 70% of the images, practice on 15%, and then take a final test on the remaining 15%.

The results were striking. When they compared their TRIBRID-optimized MobileNetV3 against other models, the new system won every time.

• A standard, unoptimized computer brain (a custom CNN) only got about 84.42% of the defects right.
• The standard MobileNetV3, without the TRIBRID coach, got 99.32% right.
• But the TRIBRID-Optimized MobileNetV3 achieved a perfect 100.00% accuracy, precision, recall, and F1 score.

This means the system correctly identified every single defect in the test set without making a single mistake. It also performed just as well as much heavier, slower models like ResNet18 and EfficientNet-B0, which also scored around 99.77%, but the new system did it while staying lightweight and fast.

To prove that the TRIBRID coach was the secret sauce, the team ran an experiment called an "ablation study." They compared TRIBRID against two other ways of tuning the settings: Random Search (guessing randomly) and PSO (a different optimization method).

• All three methods managed to reach 100% accuracy.
• However, the TRIBRID method was the most efficient. It only needed to retrain 9 layers of the computer brain to get the perfect result.
• The Random Search method had to retrain 12 layers, and the PSO method had to retrain 15 layers.

This suggests that TRIBRID is smarter at finding the right settings, allowing the system to keep more of its original "knowledge" from its initial training while still adapting perfectly to the new task.

Seeing is Believing

The most exciting part of the paper is the "Explainability" section. The researchers didn't just want the computer to be right; they wanted to know why it was right. They used Grad-CAM to visualize the decision-making process.

When they showed the computer an image of a "Missing Hole," the heat map glowed brightly right over the empty hole. When they showed a "Mouse Bite," the glow highlighted the jagged edge of the board. For "Open Circuits" and "Short Circuits," the system correctly focused on the broken lines or the accidental connections. Even for tricky defects like "Spurs" and "Spurious Copper," the heat maps accurately pointed to the tiny metal spikes and extra copper blobs.

This is crucial for industrial use. If a factory manager sees a red glow on a specific spot, they can trust the machine. If the glow was on the background or a random spot, they would know the machine was confused. The paper shows that the system is not just a "black box" guessing; it is a transparent tool that highlights exactly what it sees.

What This Means

The paper concludes that this TRIBRID-optimized framework is a powerful tool for the future of manufacturing. It combines the speed of a lightweight model with the intelligence of automatic tuning and the trustworthiness of visual explanations. The researchers note that while their system worked perfectly on the cropped images they tested, the next step is to see how it handles full, uncut images in real-world factory settings. They also hope to eventually put this system on edge devices (small computers right on the factory floor) to catch defects in real-time.

In short, the authors have built a system that is not only a perfect detective but also a very honest one, showing its work and proving that with the right coaching, even a lightweight computer brain can solve complex industrial mysteries.