Abstract:The meters in the natural gas station are the windows for the interaction between workers and equipment, which can reflect the operation status of the plant. However, many old-fashioned instruments in the station yard cannot read the readings remotely, and manual reading is a waste of manpower, and it is necessary to carry out intelligent reading research on them. Aiming at the above problems, a new method based on quadruped robot as carrier motion control, and target tracking task and image processing through deep reinforcement learning (DQN) to read the instrument representation number is adopted. Firstly, through the deep network model of the improved DQN algorithm, according to the robot learning effect in the simulated environment, the action reward function is designed and adjusted, and the top-level decision control system of the robot is designed. The instrument target tracking task under the input of one-dimensional and two-dimensional state parameters is realized. Secondly, on the basis of meter positioning and meter registration, K-means clustering binarization is used to obtain a dial with clear scale; the image is inscribed circle, and then a pointer is added in the middle of the image to rotate, during the rotation process Accurately calculate the angle with the highest coincidence between the pointer and the dial to obtain the corresponding scale. Experiments show that this algorithm can achieve accurate tracking of instrument targets and reduce calculation time during the movement process, and greatly improve the accuracy and efficiency of instrument tracking and identification, providing an effective guarantee for instrument safety monitoring in natural gas stations.
