Thermal Fault Identification Of Power Equipment Based On Convolutional Neural Network

Power equipment has the characteristics of different target sizes and difficult vision extraction.In view of this characteristic,the characteristic gold tower structure model is generally used to identify it.General Single Shot MultiBox Detector(SSD),Feature Pyramid Networks for Object Detection(FPN),scale-transferrable Object Each feature map of Detection(STDN)and other models only comes from a single layer in the trunk network,which is not comprehensive enough.Therefore,this paper adopts a multi-scale and multi-level multi-level Feature Pyramid Network(M2DET)model to identify thermal faults of power equipment.This paper first introduces the research background and development status of thermal fault detection and convolutional neural network of power equipment,summarizes the relevant theory and algorithm principle of convolutional neural network,and analyzes the model structure of M2DET in detail.Firstly,the influence of the choice of confidence degree on the recognition accuracy is compared.The confidence levels were set as 0.3,0.4 and 0.5 for comparative analysis.When the confidence is set to 0.5,the average accuracy increases by 0.18%and 0.57%compared with 0.3 and 0.4,respectively.In addition,the batch size was optimized and set as 2,4 and 8 respectively.The confidence degree is uniformly set as 0.5 for selection.Experimental results show that the detection effect is optimal when the confidence degree is 0.5 and the batch size is 4.Finally,the M2DET model is compared with the traditional SSD model.Compared with the traditional SSD model,the average detection accuracy of M2DET model is improved by 3.3%,indicating a great improvement in detection accuracy.Based on Jetson Nano platform,thermal fault identification of power equipment is studied in this paper.First,the paper introduces the reason why Jetson Nano B01 is selected.Jetson Nano B01 has a CUDA core,which makes it more convenient to transplant GPU algorithm.In order to run the target detection model on the device,the backbone network VGG-16 used in this paper is pruned.Finally,the process and steps of TensorRT acceleration algorithm are analyzed and the detection model is accelerated by TensorRT acceleration algorithm.The test results show that the average detection time of each infrared image is 0.34s on the embedded platform,which meets the expected detection requirements.In summary,combined with the embedded platform,this paper mainly studies and optimizes the target recognition method for thermal fault of power equipment.The research results have important reference significance for thermal fault identification system of power equipment.Figure[43]Table[9]Reference[66]...