Kalman filter tracking of sequence spot centroid ablated by femtosecond laser

In the process of processing monocrystalline silicon by femtosecond laser, there is a phenomenon of derived plasma luminescence, which formed the time series spot. The laser spot image contains some effective information related to process parameters, such as ablation power of femtosecond laser, processing speed and so on. Through the comprehensive analysis of the geometric features of the sequential spot image, it can be used in the closed-loop feedback system of micromachining based on the features of the plasma spot image. Because the spot centroid can be used to locate the spot position accurately, the sequential spot centroid obtained by gray centroid method was used as the sampling data, which was tracked by Kalman filter, and then the movement law of the spot was observed and analyzed. In this paper, the discrete linear Kalman filter algorithm is used to establish motion model of the spot centroid, and the next movement state of spot is predicted, and then the trajectory of the spot is obtained. Generally, the traditional Kalman filter has some shortcomings, such as inaccurate model, too much noise, and it is difficult to achieve the optimal estimation, therefore, the application of Kalman filter in spot centroid tracking is optimized. Firstly, the process noise is updated in real time to correct the direction of acceleration accurately and reduce the uncontrollability of process noise by introducing the factorl; Secondly, combined with Wiener filter, the observation value is optimized to minimize the influence of observation noise; Thirdly, starting from the Kalman filter equation, the actual value of the previous moment is connected with the optimal estimated value by weighted addition based on the "Identity of A-OE" operator, so as to improve the prediction accuracy of the current time. Finally, the EF-BP neural network is introduced on the basis of the "Identity of A-OE" operator, and a stable training system is used to optimize the weight repeatedly, so that the error can reach the expected value. The results show that the error of the observation value is reduced by 8.2503%, the error of the optimal filter estimation value is reduced to 0.3490, and the error reduction range is 6.7662. The centroid tracking accuracy of the ablated spot image is improved obviously.