In this study, visual recognition with a charge-coupled device (CCD) image feedback control system was used to record the movement of a coplanar XXY stage. The position of the stage is fedback through the image positioning method, and the positioning compensation of the stage is performed by the image compensation control parameter. The image resolution was constrained and resulted in an average positioning error of the optimized control parameter of 6.712 µm, with the root mean square error being 2.802 µm, and the settling time being approximately 7 s. The merit of a long short-term memory (LSTM) deep learning model is that it can identify long-term dependencies and sequential state data to determine the next control signal. As for improving the positioning performance, LSTM was used to develop a training model for stage motion with an additional dial indicator with an accuracy of 1 μm being used to record the XXY position information. After removing the assisting dial indicator, a new LSTM-based XXY feedback control system was subsequently constructed to reduce the positioning error. In other words, the morphing control signals are dependent not only on time, but also on the iterations of the LSTM learning process. Point-to-point commanded forward, backward and repeated back-and-forth repetitive motions were conducted. Experimental results revealed that the average positioning error achieved after using the LSTM model was 2.085 µm, with the root mean square error being 2.681 µm, and a settling time of 2.02 s. With the assistance of LSTM, the stage exhibited a higher control accuracy and less settling time than did the CCD imaging system according to three positioning indices.
Keywords: XXY stage; feedback control; image recognition; long short-term memory.