Electrical Equipment Thermal Fault Detection Based On Embedded Deep Learning

Statistically,more than 90% of electric system accidents are caused by electrical equipment failure,while more than 50% of the fault equipment appear abnormal thermal symptoms in the early stage.It is essential to remove hidden problems and equipment fault in time to ensure the safe and stable operation of electric system.Infrared thermal imaging is to detect and receive the infrared radiation energy emitted from the targeted surface and obtain the temperature distribution state of the targeted surface through electrical signal processing.Therefore,infrared technology is quite applicable for thermal fault detection of electrical equipment.Deep learning can automatically learn sample characteristics from large amounts of data.However,higher computational complexity and higher performance equipment becomes its limits.Thus,the restrict of limited resources on embedded devices become a challenging issue in deploying deep learning models.In this context,this thesis combines deep learning with infrared technology to propose a thermal fault detection method for electrical equipment based on embedded deep learning,which makes intelligent detection of electric equipment possible.To meet the need for speed,this thesis designs an efficient electrical equipment detection algorithm EEED(Embedded Electrical Equipment Detection)based on the YOLOv3 object detection algorithms and depthwise separable convolutions,which can greatly improve the speed within the acceptable range of accuracy loss.To meet the need for high accuracy,this thesis build the Storm parallel processing platform to parallelize deep learning,which is used to deploy high-accuracy and large-scale deep learning model.The scheduling algorithm are improved to enhance Storm parallel processing platform fault tolerance.Since infrared radiation will fade in the air,BP neural network is used in this thesis to correct the temperature measurement and make it close to the actual temperature value.Experimental results show that the proposed method can meet the requirements of real-time detection accuracy,speed,fault tolerance and low power consumption.