Introduction to Dissolved Gas Analysis based Hybrid Confidence Fusion Model for improvement of Maintenance of Insulating Oil based Power Equipment

Imagine the massive, humming transformers that keep our lights on and our phones charged. Inside these giants, a special liquid called insulating oil acts like a silent guardian, cooling the equipment and stopping electricity from jumping where it shouldn't. But when things go wrong inside—like a tiny spark or a hot spot forming—the oil starts to break down, releasing invisible gases like hydrogen or methane. Think of these gases as the "smoke" from a fire you can't see yet. For decades, engineers have used a technique called Dissolved Gas Analysis (DGA) to sniff out these gases and guess what's wrong. It's a bit like a doctor listening to a heartbeat; the pattern of the gases tells a story about the machine's health. However, reading this story has always been tricky. Some methods are like rigid checklists that get confused when the numbers are small, while others are like complex maps that can overlap and leave you lost. And while smart computer programs (Artificial Intelligence) can learn to spot patterns, they sometimes act like black boxes that don't explain why they made a guess, or they get confused if they haven't seen that exact problem before.

This paper introduces a new way to solve that confusion, called the "Hybrid Confidence Fusion Model." Instead of relying on just one detective, the authors suggest bringing a whole team together. They take the best parts of the old, rule-based checklists, the visual maps (like the Duval Triangle), and the smart AI computers, and make them vote on what the problem is. But here's the clever part: not every detective gets an equal vote. The model assigns a "confidence score" to each method based on how reliable it is for the specific situation. If the AI is usually right about electrical sparks but bad about heat, and the old checklist is great at heat but bad at sparks, the model weighs their opinions accordingly. The result is a single, more trustworthy diagnosis that combines the strengths of everyone. In their tests, this team-up approach reached an accuracy of 89%, beating most of the individual methods working alone. The authors suggest this makes maintenance safer and more reliable, especially when the equipment is acting weird or the fault is a mix of different problems.

The Story of the Silent Guardians

Deep inside the power grid, transformers and other heavy-duty machines are covered in a bath of insulating oil. This oil is the lifeblood of the equipment, keeping it cool and preventing electricity from arcing where it shouldn't. But when a machine starts to sicken—maybe due to a loose wire causing a spark or a coil overheating—the oil begins to decompose. It's like a piece of toast burning; the heat or electricity breaks the chemical bonds in the oil, releasing gases like hydrogen, methane, and ethylene. These gases dissolve in the oil, waiting to be discovered.

For a long time, engineers have used Dissolved Gas Analysis (DGA) to find these gases. It's a bit like a doctor taking a blood sample. By measuring how much of each gas is present, they can tell if the machine has a thermal fault (overheating) or an electrical fault (sparking). The paper explains that these gases form through specific chemical reactions. For instance, when the oil breaks down, it can produce methane (CH4) or acetylene (C2H2), and the type of gas tells you the "temperature" or "energy" of the fault.

The Detective's Dilemma

The problem is that figuring out exactly what's wrong isn't always easy. The paper highlights three main ways engineers have tried to solve this puzzle, and each has its own flaws.

First, there are the Rule-Based Methods. These are like strict instruction manuals. They use fixed ratios (comparing the amount of one gas to another) or look at a chart to decide the fault. The paper mentions standards like IEC 60599 and IEEE C57.104, which are the official rulebooks. While these are simple and standardized, they can be rigid. If the gas levels are very low, or if the numbers fall right on the edge of a category, these methods might give conflicting answers or miss the fault entirely. It's like trying to sort marbles by size using a ruler that only has markings for "small" and "large," leaving the medium ones in a confusing gray area.

Second, there are Graphical Methods, like the famous Duval Triangle and Duval Pentagon. Imagine plotting the gas levels on a colorful map divided into zones. Each zone represents a different type of fault. This is much more visual and helps with low gas concentrations, but the paper notes that these maps have fixed boundaries. Real-world faults don't always fit neatly into a triangle; they can be messy, overlapping, or non-linear. It's like trying to draw a perfect circle around a cloud; the shape just doesn't fit.

Third, there are Artificial Intelligence (AI) methods. These are the smart computers that learn from thousands of past examples to recognize patterns. They are great at handling complex, messy data that the rulebooks can't figure out. However, the paper points out that AI has its own issues. It relies heavily on the data it was trained on; if it sees a new type of fault it hasn't learned, it might guess wrong. Also, AI can be a "black box," meaning it gives an answer but doesn't explain the physical reason why, which makes engineers hesitant to trust it blindly.

The Team-Up Solution

This is where the authors, Joydip Dhar, Sovan Dalai, and Saibal Chatterjee, step in with their Hybrid Confidence Fusion Model. Instead of picking one detective, they built a system where all the detectives work together and vote.

Here is how the "fusion" works:

1. The Voting Booth: The model takes the results from the best rule-based methods (like the Duval Triangle) and the best AI methods.
2. The Confidence Score: Not every vote counts the same. During a training phase, the model looks at how accurate each method is. If a specific AI model is great at spotting electrical faults but bad at thermal ones, it gets a high "confidence score" for electrical faults and a lower one for thermal ones.
3. The Weighted Decision: When a new sample comes in, the model calculates a weighted score for each possible fault. It uses a mathematical formula to combine these scores, prioritizing the methods that are most reliable for that specific situation.

Think of it like a group of friends trying to guess the answer to a trivia question. One friend is a history buff, another is a science whiz, and a third is good at guessing. If the question is about history, the history buff's guess counts for more. If it's about science, the science whiz gets the loudest voice. The Hybrid Confidence Fusion Model does exactly this, but for transformer faults. It dynamically adjusts who gets to speak up based on who is usually right.

What They Found

The authors tested their new model against the old methods using real-world data. They compared the accuracy of their "Hybrid Confidence Fusion" (HCF) model against individual methods like Machine Learning (ML) models, the Characteristic Gas Ensemble (CGE), and the Duval Triangle 1 (DT1).

The results were promising. The individual methods had varying levels of success:

• The Duval Triangle 1 (a popular graphical method) had an accuracy of about 72%.
• The Machine Learning models (ML1 and ML2) ranged from 76% to 80%.
• The Hybrid Confidence Fusion Model achieved an accuracy of 89%.

The paper also looked at other metrics like Precision and Recall (which measure how often the model is right when it says "fault" and how many actual faults it catches). The HCF model scored a precision of 0.78 and a recall of 0.78, with an F1-Score of 0.77. This suggests that by combining the methods, the model became more robust, especially when dealing with "mixed fault scenarios" where the machine might have both heat and electrical issues happening at once.

Why It Matters

The paper concludes that while rule-based methods are good for being easy to understand, and AI is good for spotting complex patterns, neither is perfect on its own. The Hybrid Confidence Fusion Model suggests a middle path. It doesn't throw away the old, trusted rules, nor does it blindly trust the new AI. Instead, it creates a system where they complement each other.

The authors emphasize that this approach helps resolve conflicts when different methods give different answers. By using "adaptive confidence scores," the model can handle sparse data (when there isn't much gas to measure) and mixed faults better than the isolated techniques. While the paper notes that this is a simulation and testing on specific datasets, the 89% accuracy suggests that this "team-up" strategy is a strong direction for the future of maintaining our power equipment. It promises a future where we can detect problems earlier and keep the lights on with greater confidence.