Fault Diagnosis Of Oil Immersed Transformer Based On Convolution Neural Network

In the power system,whether the transformer can work normally and stably will directly affect the safe operation of the power grid.A transformer failure could lead to the collapse of a large area or the entire grid,causing a huge impact on human life and huge social and economic losses.Therefore,it is vital to monitor transformers in real time to detect,diagnose and solve potential transformer faults as early as possible.In recent years,with the rapid development of artificial intelligence technology and the implementation of the "smart grid" policy,pattern recognition based on artificial neural networks has been widely used in the field of fault diagnosis in power systems.In the case of transformers,the traditional diagnostic methods can no longer meet the requirements of fast,accurate and intelligent diagnosis of transformer faults.Therefore,this paper proposes an improved AlexNet network model to identify the temperature of oil-immersed transformers and determine whether they are operating normally or whether a fault has occurred in a region where the temperature is too high.Firstly,infrared thermal images of oil-immersed transformers in different operating states are collected using an infrared thermal camera,and a large amount of Gaussian noise and pepper noise is present in the operating environment of oil-immersed transformers.The Gaussian filtering algorithm combined with the median filtering algorithm is finally used to remove the Gaussian and pepper noise from the thermal images of oil-immersed transformers.Then,an image enhancement algorithm based on wavelet transform is used to enhance the infrared thermal image of the oil-immersed transformer to make the image clearer and more distinctive,laying the foundation for the subsequent neural network to extract features.The image pre-processed data set is then expanded using the data expansion method to increase the diversity of the data samples.The expanded dataset was then trained with fine-tuned LeNet,AlexNet and VGG-16 neural networks,and the recognition accuracy and loss values of each model were analysed.Matlab experiments revealed that the AlexNet network had higher accuracy and better recognition than the other networks,so this paper improves the original AlexNet network model on the basis of it.Next,the AlexNet network model is improved.The number of convolutional kernels and the number of layers in the convolutional layer were firstly discussed in this paper to determine the impact on the recognition accuracy of the neural network model.Next,the original ReLU activation function was replaced by the Leaky ReLU activation function,which solved the problem that the neurons could not extract features without learning when the input was negative.Finally,the recognition accuracy of the improved AlexNet network is 99.2%,which is 2.4%better than that of the original AlexNet network.