Research on thermal error modeling of high-speed motorized spindles based on SC-ANFIS

Precision machining is the art of carving metal with such exactness that the difference between a perfect part and a flawed one is measured in micrometers, a scale so small it is invisible to the naked eye. For the machines that perform this work, the greatest enemy is not a dull tool or a shaky hand, but heat. As a high-speed motor spins inside a machine tool, friction and electrical resistance generate warmth that causes the metal components to expand. This expansion is not uniform; the machine bends and shifts in ways that are difficult to predict, pushing the cutting tool slightly off its intended path. Even a tiny shift of a few micrometers can ruin a part meant for an aircraft engine or a medical device. For decades, engineers have tried to compensate for this by measuring the temperature at various spots on the machine and using those numbers to guess how much the machine has expanded. However, the relationship between heat and movement is complex and changes depending on how fast the machine is spinning, making simple guesses unreliable.

A team of researchers at Hunan University and industry partners has tackled this problem by developing a new way to predict these thermal shifts, specifically for high-speed motorized spindles that use ceramic ball bearings. Instead of relying on traditional mathematical formulas that often fail under complex conditions, they turned to a method that mimics how humans learn from patterns. Their approach involves two main steps: first, figuring out which temperature sensors actually matter, and second, using a smart computer model to connect those temperatures to the machine's movement. The researchers found that by carefully selecting just three critical temperature points out of ten possible locations and feeding that data into a specialized learning system, they could predict the machine's thermal drift with significantly higher accuracy than existing methods, even when the machine was running at different speeds.

To understand the challenge, one must look at the machine itself. The researchers studied a specific type of motorized spindle, a compact unit where the motor and the cutting tool are integrated into a single rotating shaft. Inside, ceramic ball bearings and a built-in motor generate intense heat. When the machine runs, this heat does not spread evenly; it creates hot spots near the bearings and the motor, causing the metal to expand in uneven patterns. To map this invisible heat, the team first used a computer simulation to visualize the temperature distribution, confirming that the front and rear bearings, along with the motor housing, were the primary sources of thermal stress. They then built a physical test rig equipped with ten temperature sensors placed at strategic locations around the spindle, from the front housing to the rear bearing chamber. They also installed a high-precision sensor to measure the actual movement of the spindle tip as it heated up.

The team ran experiments where the spindle spun at five different speeds, ranging from 3,000 to 15,000 revolutions per minute, for two hours each time. As the machine warmed up, the sensors recorded the temperature changes every five seconds, while the displacement sensor tracked how far the spindle tip moved. The data revealed a clear pattern: at lower speeds, the temperature rose gently and the machine expanded slowly. But as the speed increased, the heat built up faster, and the temperature differences between various parts of the machine grew more extreme. At the highest speeds, the temperature difference between the front and rear of the spindle exceeded ten degrees Celsius, creating a complex thermal environment that was difficult to model with standard equations.

The first hurdle the researchers faced was deciding which of the ten temperature sensors to use for their prediction model. Using all ten would introduce too much noise and confusion, as many sensors were measuring temperatures that were too similar to each other. To solve this, they employed a two-step filtering process. First, they grouped the sensors based on how similarly their temperatures changed over time, effectively clustering the ten points into three distinct zones. Then, they analyzed which sensor in each zone was most closely linked to the actual movement of the spindle. This process allowed them to discard the redundant sensors and select just three "sensitive" points: one at the front housing, one near the motor, and one at the rear bearing. These three points captured the essential thermal behavior of the entire system without the clutter of extra data.

With the input data streamlined, the researchers turned to the prediction model. Traditional methods often struggle because they try to force the data into rigid, pre-defined categories, which can lead to errors when the machine behaves in unexpected ways. The team instead used a system called a subtractive clustering adaptive neuro-fuzzy inference system. In plain terms, this is a computer program that learns to recognize patterns in the data without being told exactly what to look for. It scans the relationship between the three selected temperatures and the resulting thermal drift, automatically identifying the most important "rules" that govern the machine's behavior. Unlike older models that create a massive, unwieldy set of rules that can confuse the system, this new method finds the most efficient set of rules by focusing on the densest clusters of data. This allows the model to adapt quickly to different operating speeds and remain accurate even when the thermal conditions are highly nonlinear.

The results of the study were striking. The researchers tested their new model against several other common prediction methods, including standard neural networks and linear regression, across all the different spindle speeds. At lower speeds, all the models performed reasonably well, but as the speed increased and the thermal behavior became more complex, the older models began to fail. Their predictions drifted away from the actual measurements, with errors growing significantly. In contrast, the new model maintained its precision. Even at the most demanding speed of 7,500 revolutions per minute, where the machine's thermal behavior was most chaotic, the new model predicted the thermal drift with an error of less than 1.6 micrometers. This was a substantial improvement over the other methods, which produced errors more than 30 percent larger. The model's ability to stay accurate across such a wide range of speeds suggests it has successfully learned the underlying physical laws of the machine's thermal expansion, rather than just memorizing specific data points.

The significance of this work lies in its potential to make high-precision manufacturing more reliable. By accurately predicting how much a machine will expand due to heat, engineers can program the machine to automatically adjust its movements in real-time, effectively canceling out the thermal error before it affects the final product. The study demonstrates that a smart combination of sensor selection and advanced pattern recognition can solve a problem that has long plagued the industry. The researchers showed that you do not need a massive array of sensors or a complex physical simulation to get the job done; instead, a focused approach that identifies the right data and uses a flexible learning system can provide the robustness needed for the next generation of precision tools. This approach offers a practical path forward for maintaining the extreme accuracy required in aerospace and automotive manufacturing, where the cost of a single thermal error can be immense.