False-Alarm-Controlled Evaluation of Gas Indices and Deep Learning for Early Detection of Coal Spontaneous Combustion

Deep underground, where coal seams lie hidden beneath tons of rock, a slow and invisible danger can begin. Coal does not just sit there; it breathes. Over time, it reacts with the oxygen in the air, a process called oxidation that generates heat. Usually, this heat dissipates into the surrounding rock and airflow. But sometimes, the heat builds up faster than it can escape, causing the coal to warm itself until it ignites without a spark. This phenomenon, known as spontaneous combustion, is one of the oldest and most persistent hazards in mining. It can turn a productive mine into a sealed-off disaster zone, releasing toxic gases and greenhouse emissions that linger for decades. Because the fire starts slowly and without smoke or visible flame, miners rely entirely on the air itself to tell them something is wrong. For over a century, the standard way to listen to this air has been to measure specific gases and compare them using simple math ratios. The most famous of these is a calculation called Graham's index, which compares the amount of carbon monoxide in the air to the amount of oxygen that has been used up. If the ratio gets too high, it is supposed to signal a fire.

However, a new study by Ulas Cinar from Canakkale Onsekiz Mart University suggests that this long-standing method may be fundamentally flawed when used as an automatic alarm system. The research does not just look at whether these methods can spot a fire when one is already burning; it asks a more difficult question: how often do they scream "fire" when there is nothing but quiet air? In the world of safety monitoring, an alarm that rings constantly is worse than no alarm at all, because workers eventually learn to ignore it. Cinar developed a new way to test these warning systems that forces them to prove they can distinguish between a real threat and a false one. By running thousands of computer simulations of mine air and then testing the methods against real records from three different coal mines, the study found that the traditional Graham's index fails completely under these strict conditions. In fact, the study reveals that the index often behaves in the exact opposite way it should: it sounds the loudest when there is no fire, and stays silent when a fire is actually burning.

The core of the problem lies in the math behind the index. Graham's index divides the amount of carbon monoxide by the "oxygen deficiency," which is simply the difference between the normal amount of oxygen in fresh air and the amount currently in the mine. In well-ventilated tunnels where fresh air flows freely, the oxygen deficiency is tiny, often close to zero. When you divide a number by a number that is nearly zero, even the slightest error in measurement causes the result to explode into a huge, meaningless value. It is like trying to measure a tiny change in a vast ocean by looking at a single drop; the math becomes unstable. In the simulations, this instability meant that the index would frequently jump to alarm levels during perfectly normal, fire-free periods. When the researchers applied this test to real mine data, the failure was even more dramatic. In seven out of thirteen fire-free episodes in ventilated areas, the index soared to values as high as 44, far above the standard alarm threshold of 1.0. Meanwhile, in two confirmed fires that were happening behind sealed-off walls, the index never rose above 0.50. The system was screaming at the wrong times and staying quiet at the right times.

To solve this, the study introduced a new protocol that treats every warning method, whether a simple math formula or a complex computer program, as a continuous score rather than a fixed switch. The researchers then asked: "If we set the sensitivity of this system so that it only raises a false alarm 10 percent of the time, how many real fires does it catch?" This approach levels the playing field, forcing every method to prove its worth under the same strict rules. When they ran this test on the simulated data, Graham's index failed to detect a single fire at any setting that kept false alarms low. It was simply not a reliable detector. The study also tested advanced computer models known as deep learning networks, which are designed to learn patterns from vast amounts of data. These models were trained on the simulated air data and then tested on the real mine records without any further adjustment. The results showed that while these complex models could find some fires, they did not perform significantly better than a much simpler approach: watching the level of carbon monoxide alone.

In the real-world test, a simple threshold on the carbon monoxide sensor detected all three confirmed fires, providing a warning an average of 345 hours before the fire was officially confirmed by temperature sensors. The complex computer models detected two of the three fires, with similar lead times. The study concludes that the added complexity of the deep learning models did not translate into better safety performance in this specific context. The signal of a developing fire is so strong in the carbon monoxide levels that a simple check is often sufficient, provided the system is tuned to avoid false alarms. The failure of the traditional index was not due to a lack of chemical understanding, but rather a numerical weakness that makes it useless as a standalone alarm in ventilated areas. The researchers suggest that the industry should move away from relying on these unstable ratios for automatic warnings. Instead, they recommend using the raw concentration of carbon monoxide, perhaps with a simple rule to ignore the data if the oxygen levels are too high to make the calculation meaningful.

The study also highlights a limitation in how safety research is often reported. Many previous papers claim high accuracy for their models by measuring how well they predict gas levels or by testing them only on days when a fire was known to be happening. This approach hides the fact that the system might be ringing the alarm bell every single day when no fire exists. By insisting on measuring the false-alarm rate alongside the detection rate, this paper shows that a method can look perfect on paper while being useless in practice. The findings are based on a combination of computer simulations and a small set of real mine records, which means the results are specific to the conditions tested. The study does not claim that deep learning is useless for mining safety, but rather that in this specific task, a simple, well-tuned sensor check outperforms a complex ratio that is prone to mathematical errors. The ultimate goal is to ensure that when an alarm sounds in a mine, it is a signal that demands immediate attention, not a noise that has been ignored for years.