Design And Implementation Of Industrial Instrument Panel Identification System In Unattended Mode

In the field of industrial manufacturing,it is often necessary to collect visual information from instrument panels and recognize their readings.Traditional manual detection and identification methods require a great deal of work,resulting in low detection efficiency and accuracy.Therefore,designing an industrial instrument panel detection and identification system is of great significance.This paper combines unmanned aerial vehicle(UAV)inspection with deep learning technology to design and implement an industrial instrument panel recognition system.To achieve an unmanned industrial instrument panel recognition system,the target detection model needs to be deployed on embedded devices.However,the detection speed of existing target detection models is significantly affected when running on embedded devices.To improve the detection speed of the target detection model on embedded devices,researchers have proposed model compression methods.This paper uses the YOLOv7 target detection model for embedded model deployment.However,the existing model compression methods are not suitable for YOLOv7 model because it has a special network structure.In addition,when UAVs use line patrol to inspect instrument panels,factors such as lighting,shadows,and UAV movement may cause visual blurring,affecting the completion of the inspection task.To solve this problem,researchers have proposed various sensor fusion and positioning inspection methods suitable for UAVs.However,most studies focus on outdoor UAV patrol tasks,which is different from the application scenario of the system designed in this paper.The system designed in this paper needs to complete indoor patrol tasks.Based on the mentioned issues and the design and implementation of the system,the main research work and innovation of this paper are as follows:(1)In response to the speed reduction issue that may occur when deploying the YOLOv7 target detection model on embedded devices,this paper proposes a model compression method.The method is based on an improved pruning and quantization combined compression algorithm for YOLOv7.By considering the RepConv module used by YOLOv7,the pruning algorithm has been improved to ensure pruning correctness.Then,a quantization algorithm is introduced to further compress the YOLOv7 model,significantly reducing its computational complexity,and thereby improving the computational speed on mobile terminals.(2)This article proposes a visual and Ultra Wide Band(UWB)fusion inspection algorithm to address lighting,shadow,and visual blur issues that UAVs may encounter during dashboard inspections.The method includes two stages: line inspection and positioning.In the line inspection stage,an upward-facing camera is used as a visual sensor for image segmentation.When an exception occurs in the transmitted image from the visual sensor,the algorithm automatically switches to UWB navigation mode while continuously reading the feedback information from the visual sensor until the transmitted image returns to normal.In the positioning stage,AprilTag and UWB are used for fusion positioning,which achieves more precise UAV positioning by fusing input results from both sources.(3)A detection and recognition system for industrial dashboards has been designed and implemented.The system includes a mobile unmanned aerial vehicle(UAV)side and a ground server side.The UAV side conducts dashboard inspections using the algorithm proposed in(2),detects and collects dashboard images using the compressed YOLOv7 model proposed in(1),and then transmits the collected original images to the ground server.At the server side,a quality evaluation model is used to select the best image,which is then filtered for excess background information using the YOLOv7 model again.Next,the U-Net semantic segmentation model is used to extract the dial and pointer information of the dashboard,and traditional image processing techniques are used to process it.Finally,the angle method is used to identify the dashboard.In addition,the system can present uploaded dashboard images and their recognition results through a web client designed using Flask technology.