Recognition Method Of Engine Carbon Deposit Based On Parallel Convolutional Neural Network

The operation of the car engine will inevitably produce carbon deposits.The current main method of judgment is to rely on the intuitive experience of the driver,such as unstable idling of the vehicle,acceleration and frustration,etc.,and the method of removing carbon deposits is only through regular maintenance and cleaning of the throttle Way.However,due to the different operating conditions of the vehicle and the quality of fuel,the severity of carbon deposits is also different.Therefore,the regular maintenance method cannot be targeted to remove the carbon deposits of the engine.The method has important practical value.In this paper,the endoscopic image of the engine piston crown is taken as the research object.For different degrees of carbon deposit images,a recognition method based on convolution neural network is designed to achieve five types of piston crown product: mild,mild,moderate,severe and severe.For the determination of carbon level,the accuracy of identification reached 82.92%.In addition,GPU parallel mode is adopted for the designed convolutional neural network to accelerate the training speed of the network.The specific work is as follows:(1)By constructing a convolutional neural network,the image of the engine piston crown is identified and classified.First,the two methods of median filtering and piecewise linear transformation are used to preprocess the image,which lays the foundation for subsequent carbon image recognition and classification.At the same time,different convolutional layers,different convolution kernel sizes,and different iteration times are analyzed.The impact on the accuracy of the image recognition of the engine piston carbon deposits.(2)Analyze the specific processes of forward propagation and back propagation in different network layers of convolutional neural networks.The forward and backward propagation processes of convolutional neural networks are implemented in parallel under a single GPU.Perform the same number of iterations as the CPU platform to calculate the time required for training.Experiments show that GPU has more advantages in the training of convolutional neural networks,and the maximum acceleration ratio of a single GPU is 12.26.(3)On the multi-GPU platform,the synchronous update method and the asynchronous stochastic gradient descent method are used to train the convolutional neural network in parallel.The asynchronous stochastic gradient descent method overcomes the need for the parameter server to wait for all GPU feedback parameter changes in the synchronous update method The shortcomings caused by the amount of training time,can send the latest model parameters to each GPU and start the next round of training.Experiments show that under the same number of iterations,the asynchronous stochastic gradient descent method is better for parallel training of convolutional neural networks.