Kalman filter-based error prediction with forgetting factor iterative learning control scheme

In the world of modern manufacturing, machines are often tasked with performing the same motion over and over again, like a robotic arm welding a car frame or a printer laying down layers of plastic. To make these machines move with extreme precision, engineers use a strategy called iterative learning control. The idea is simple yet powerful: the machine tries a movement, measures how far off it was from the target path, and then uses that mistake to adjust its next attempt. With each repetition, the machine learns, refining its motion until it follows the desired path perfectly. However, this learning process faces a significant hurdle in the real world: noise. In a factory, machines are constantly bombarded by random, unpredictable jolts and vibrations—things like a sudden draft of air, a fluctuation in power, or a loose component. These random disturbances are different every time, meaning the machine cannot simply "remember" them to cancel them out later. When a machine tries to learn from a movement that was ruined by a random jolt, it often learns the wrong lesson, causing its performance to degrade or get stuck with a permanent error.

To solve this, researchers at Lanzhou University of Technology have developed a new way to help these machines learn more effectively. They combined the standard learning method with a mathematical tool known as a Kalman filter, which acts like a smart noise-canceling ear. While the learning controller tries to correct the machine's path based on past attempts, the Kalman filter looks at the current data and estimates what part of the error is a genuine mistake the machine needs to learn from, and what part is just random noise that should be ignored. By separating the signal from the static in real-time, the system can make much sharper corrections. The researchers tested this new approach, which they call a composite control scheme, using computer simulations that mimic the behavior of industrial machinery. They found that by integrating this noise-filtering capability, the machine not only learned faster but also achieved a much higher level of accuracy, reducing the average error by more than half compared to traditional methods.

The core challenge the researchers addressed is that standard learning systems often struggle when the environment is not perfectly clean. In a typical setup, if a machine encounters a random vibration during one cycle, the control system might mistake that vibration for a flaw in its own movement and try to compensate for it in the next cycle. Since the vibration was random and won't happen again in the same way, the machine ends up over-correcting, introducing a new error. This cycle of confusion prevents the machine from ever reaching its full potential. The new method proposed in this paper introduces a "forgetting factor," a mechanism that allows the system to weigh recent information more heavily than older data, ensuring it doesn't get stuck on outdated mistakes. More importantly, it layers a Kalman filter on top of this learning process. This filter acts as a predictor, constantly estimating the true state of the machine by looking at the pattern of errors and filtering out the random spikes. It essentially asks, "Is this deviation a real problem with my path, or just a momentary glitch?" and adjusts the learning accordingly.

To verify their idea, the team created a digital model of a machine moving along a specific path, similar to a robotic arm tracing a curve. They programmed the simulation to include random noise, mimicking the chaotic conditions of a real factory floor. They then ran two different scenarios: one where the machine used the standard learning method, and another where it used their new combined method with the Kalman filter. The results were striking. In the standard scenario, the machine's path wobbled significantly, and the error fluctuated wildly even after many attempts. The machine was constantly reacting to the random noise, unable to settle into a smooth, precise motion. In contrast, the machine using the new method quickly found its rhythm. The random noise was effectively smoothed out, allowing the machine to focus on the actual shape of the path it needed to follow.

The data from these simulations showed a clear improvement in performance. When the system reached a steady state, the average error for the standard method was roughly 20.22 units, while the new method reduced this to about 10.08 units. This represents a 50 percent improvement in accuracy. Furthermore, the stability of the new system was far superior. The standard method showed a high degree of variation, with the error jumping up and down by large amounts, whereas the new method kept the error tightly controlled. The researchers also looked at the root mean square error, a measure of how far off the machine was on average, and found a reduction of over 52 percent. These numbers indicate that the new approach does not just make small tweaks; it fundamentally changes how the machine handles uncertainty, allowing it to maintain high precision even when the environment is noisy.

The researchers also examined how the system behaved over many different runs to ensure the results were not just a lucky fluke. They ran the simulation 50 times, each time with slightly different random noise patterns, to see if the improvement held up consistently. The results were robust. In every run, the new method outperformed the standard one, with the control inputs becoming smoother and the tracking path staying much closer to the target. The standard method often produced jerky movements as it fought against the noise, but the new method produced a fluid, gradual adjustment. This suggests that the combination of the learning algorithm and the noise filter creates a system that is not only more accurate but also more stable and reliable. The machine learns to ignore the distractions and focus on the task, much like a skilled musician who can play a perfect melody even when there is background chatter in the room.

This work suggests a practical path forward for industrial automation. By adding a layer of intelligent prediction to existing learning systems, engineers can build machines that are more resilient to the messy reality of factory environments. The method does not require expensive new hardware; it is a software-based upgrade that makes better use of the data the machine is already collecting. The researchers demonstrated that this approach works well for linear systems, which are common in many manufacturing processes, and they provided mathematical proof that the system will remain stable and converge to a solution under these conditions. While the study was conducted through simulation rather than on a physical robot, the rigorous testing and the clear mathematical foundation suggest that the benefits would translate to real-world applications. The ability to cut errors in half and stabilize performance against random interference offers a compelling solution for industries where precision is paramount, from assembling delicate electronics to machining heavy components. The study concludes that by letting the machine distinguish between what it needs to learn and what it should simply ignore, we can achieve a level of control that was previously difficult to reach.