| Centralized heating is an important public utility.At present,centralized heat supply is developing rapidly and the scale of heating supply is expanding in our country.The centralized heating system mainly includes three parts: heat source,heating pipeline network and heat users.The heating pipeline network is the weak link in the heating system,and the fault rate of the pipeline network is increasing as the service life of the network increases.Pipe network leakage is one of the main faults of the heating pipeline network.Leakage of the heating pipeline network will cause energy waste and economic loss,reduce the heating quality of customers and even affect the operation of the whole heating system.Therefore,scientific and accurate diagnosis of heating pipeline network leakage faults would be of great economic advantage and practical significance.In this thesis,deep convolutional neural network is used to diagnose the leakage of heating pipeline network.Firstly,convolutional neural network models for the diagnosis of the degree of leakage and the location of leakage in the heating pipeline network are established under the Tensor Flow deep learning framework of Python platform.After calculation and comparison,for the network model input is one-dimensional data with low data complexity.And I designed the network structure containing convolution-pooling and specifically set up the parameters and optimization methods of the network structure.Meanwhile,in order to further improve the model training speed and to solve the problem of internal covariate translation during training,a batch normalization layer is introduced into the heating network leakage location diagnosis model.Secondly,the experimental heating pipeline network in the laboratory is used as the research object.The normal and leakage conditions data of the pipeline network are obtained through experiments and a simulation model of the experimental network is established.The model can simulate the normal and leakage conditions of the pipeline network and obtain the corresponding simulation data of the network.The dataset is classified into normal conditions（no leakage）,leakage rate of 1.1%,2.5%,4% and 5.5% according to the leakage volume.The dataset is classified into normal conditions（no leakage）,leakage rate of 1.1%,2.5%,4% and5.5% according to the leakage volume,and I add the label classification of 0～4 to each of these five categories for training and testing the heating network leakage degree diagnosis model.I combine the above 5 leakages and 13 different leak locations into 53 leak conditions,and add label classifications from 0 to 52 respectively,for training and testing the heating network leakage location diagnosis model.In practical applications,the leakage data are usually unlabeled,and it is difficult to classify the leakage data accurately.To solve this situation,the experimental data and simulation data are divided into training sets according to a certain ratio to train the model,and a model with low dependence on the experimental data is obtained,the experimental data is used as a test set to test and evaluate the training of the model.Finally,I trained the convolutional neural network model for heating Pipeline network leakage degree diagnosis,the optimal number of convolution-pooling layer layers,number of convolution kernels,convolution kernel size,learning rate and other important parameters of the model were defined.The test results show that the accuracy of the model for heating pipeline network leakage degree diagnosis is 98.82%,and the macro-F1 is 0.988.The diagnosis of leakage degree is basically completed.Using the same parameter settings of the heating pipeline network leakage degree diagnosis model,testing the convolutional neural network model without batch normalization layer for heating Pipeline network leakage location diagnosis,the recognition accuracy of the model was 96.78%.Due to the increase of data labels,the recognition accuracy of the location diagnosis model with the same parameter settings becomes lower.Adding a batch normalization layer,by adjusting the layer positions,it was determined that the model with the convolution layer-batch normalization layer-activation function-pooling layer-full connection layer structure had the highest recognition effect and iteration speed,the accuracy of leakage location diagnosis for the experimental heating pipeline network was97.75%,and the macro-F1 is 0.977. |
