Operating-Condition-Decoupled Sparse Order-Domain Gear Fault Monitoring for Range-Extended Hybrid Agricultural Machinery in Hilly and Mountainous Areas

In the rugged, uneven terrain of hills and mountains, heavy machinery faces a unique challenge: the ground is never flat, and the work is never steady. When a tractor or harvester climbs a slope or plows through thick soil, its engine and transmission must constantly adjust, causing the speed of its rotating parts to fluctuate wildly. For engineers who need to listen to these machines to detect early signs of wear or damage, this variability is a major obstacle. Traditional methods of listening for trouble rely on finding specific, steady sounds that indicate a broken gear or a worn bearing. However, when the machine's speed changes, those sounds shift pitch and blur, making it difficult to tell if a strange noise is a sign of a serious fault or just a normal reaction to a steep hill. This uncertainty leads to false alarms, where a healthy machine is flagged as broken, or missed detections, where a real problem goes unnoticed until it causes a failure in the field.

To solve this, researchers at Henan University of Science and Technology and the Longmen Laboratory developed a new way to monitor the health of gears in hybrid agricultural machines designed for these difficult landscapes. Instead of trying to listen to the machine's sounds as they happen in real time, they created a method that translates the vibration data into a format that stays steady even when the machine speeds up or slows down. By simulating thousands of hours of operation on virtual hills, they trained a system to ignore the noise caused by changing slopes and loads, focusing only on the specific patterns that indicate a gear is damaged. Their approach successfully distinguished between healthy gears and faulty ones, even when the machine was operating under conditions it had never seen before, offering a promising path toward keeping hybrid farm equipment running safely and efficiently in the world's most challenging fields.

The core of this work addresses a specific type of vehicle: the range-extended hybrid agricultural machine. These are powerful tractors and harvesters that use a small engine to generate electricity, which then charges a battery or directly powers an electric motor to turn the wheels. This design allows them to operate with high power and low emissions, but it also introduces complex vibrations. The engine starts and stops, the battery charges and discharges, and the electric motor switches modes, all while the machine climbs slopes and pulls heavy loads. In a traditional diesel tractor, the vibration patterns are relatively predictable. In these hybrid machines, the vibration is a chaotic mix of the engine's rhythm, the electric motor's hum, the grinding of the gears, and the jolts from the uneven ground. When a gear inside the transmission begins to fail, its warning signs are often buried under this chaotic background noise.

The researchers began by building a detailed computer simulation of how these machines behave in hilly terrain. They did not rely on real-world data alone, because collecting enough examples of actual broken gears on expensive farm equipment is difficult and dangerous. Instead, they created a virtual environment that mimicked the real world with high precision. They programmed the simulation to generate long driving cycles that included plowing, transporting goods, and navigating mixed hilly routes. The simulation accounted for every factor that changes the machine's vibration: the steepness of the slope, the resistance of the soil, the speed of the vehicle, the charge level of the battery, and the state of the range-extender engine. Crucially, they used a "paired" strategy. For every single simulated drive, they generated two versions of the vibration data: one representing a perfectly healthy machine and one representing the same machine with a specific gear fault. Both versions experienced the exact same hill, the same load, and the same speed changes. The only difference was the presence of the fault. This ensured that when the computer learned to tell them apart, it was learning to recognize the fault itself, not just the difference between a steep hill and a flat road.

Once they had generated 2,000 of these paired samples, the researchers faced the task of extracting the useful information from the raw vibration signals. In a standard analysis, engineers look at the frequency of the vibrations, which is like listening to the pitch of a sound. However, because the machine's speed was constantly changing, the pitch of the gear fault would drift up and down, spreading its energy across a wide range and making it hard to spot. The team used a technique called computed order tracking to fix this. Instead of measuring the vibration at equal time intervals, they measured it at equal angles of the drive shaft's rotation. Imagine a spinning wheel: if you take a photo every time the wheel turns a specific amount, the picture of a crack on the rim will always be in the same spot, even if the wheel is spinning faster or slower. By applying this logic to the vibration data, the researchers transformed the shifting, drifting signals into a stable map where the fault always appeared in the same place, regardless of how fast the machine was moving.

With the data now organized into this stable map, the next step was to find the specific features that mattered. The researchers extracted dozens of measurements from different parts of this map, looking for peaks in energy, changes in average levels, and the relative strength of signals compared to their surroundings. This created a massive list of potential clues, many of which were just noise or irrelevant details caused by the battery or the engine. To cut through this clutter, they used a mathematical tool that automatically selected only the most important clues and discarded the rest. This process, known as sparse feature selection, acted like a filter that removed the background noise of the machine's operation, leaving behind a small, precise set of indicators that were strongly linked to the gear fault and completely unrelated to the battery charge or the slope of the hill.

Finally, the researchers tested their system to see if it could actually identify the faults. They split their simulated data into training sets and test sets, asking the computer to learn from one group and then predict the health of the other. They compared their new method against several older, standard techniques. The results were clear: the new approach, which combined the stable angle-based map with the smart feature filter, significantly outperformed the traditional methods. In repeated tests, the system correctly identified the gear condition with an accuracy of about 94 percent. This was notably better than the standard frequency-based methods, which struggled to keep up with the variable speeds. Perhaps more importantly, the system proved to be robust when faced with conditions it had never seen before. When the researchers tested it on data involving speeds, loads, or slopes that were outside the range of the training data, the new method maintained high accuracy, while the older methods saw their performance drop sharply.

The study also revealed that the specific features the system learned to rely on were indeed related to the gears and not to the machine's other systems. The selected indicators were concentrated in specific areas of the vibration map that correspond to the rotation of the drive shaft, and they showed almost no connection to the battery's state of charge or the engine's on-off cycles. This confirmed that the method was successfully isolating the mechanical fault from the complex hybrid powertrain. The researchers found that even a simpler type of computer classifier could achieve high accuracy once the data was properly prepared, suggesting that the real breakthrough was in how the data was organized and cleaned, rather than in the complexity of the final decision-making tool.

While the results are promising, the authors are careful to note that these findings come from a simulation. The virtual environment, though detailed, cannot perfectly replicate the messy reality of a real farm, where sensors might be mounted differently, or where the transmission path might be affected by rust or dirt in ways the computer model did not anticipate. The study serves as a proof of concept, demonstrating that the approach is theoretically sound and capable of handling the extreme variability of hillside farming. The next step, as the researchers outline, is to build a real machine equipped with the necessary sensors and test the system in the field. If the method holds up in the real world, it could provide farmers with a reliable way to monitor their expensive hybrid equipment, preventing unexpected breakdowns and ensuring that these advanced machines can continue to work the difficult, sloping fields that define modern agriculture.