This application claims priority to Chinese Patent Application No. 202210638253.6 with a filing date of Jun. 7, 2022. The content of the aforementioned application, including any intervening amendments thereto, is incorporated herein by reference.
The present disclosure relates to the technical field of fault diagnosis of power transformers, and in particular, to a method for diagnosing transformer fault based on a deep coupled dense convolutional neural network.
A power transformer is a key device in a power transmission and transformation network, and its fault endangers safe and stable operation of an entire power system and causes a huge economic loss. Most of power transformers in service have exceeded an expected rated service life, because most of them were installed before 1980. In order to ensure reliable and effective operation of a transformer, real-time fault diagnosis of the transformer is particularly important. Dissolved gas-in-oil analysis (DGA) is an online monitoring technology. When a thermal fault or an electrical fault occurs on an oil immersed transformer, a content of dissolved gas in oil changes accordingly. A mutual relation between the content of the dissolved gas in the oil and a fault type of the transformer is studied to determine a health state of the transformer.
Traditional method for diagnosing transformer faults includes a ratio method and an image method. When a fault point falls on a boundary line, diagnosis of a fault type is very uncertain. Different diagnosis methods have different results, resulting in a low fault diagnosis accuracy rate. A series of artificial intelligence methods such as K-nearest neighbor, support vector machine, fuzzy theory, artificial neural network, deep belief network have been used to process DGA data. Some achievements have been made, but there are still shortcomings in a learning ability, processing efficiency, feature extraction, and the like. In recent years, different from a previous alternative method, a deep learning method has been applied more robustly and effectively. A convolutional neural network has been relatively mature in visual recognition, image processing, and fault diagnosis.
The dissolved gas in the oil uses five kinds of characteristic gas (H2, CH4, C2H6, C2H4, and C2H2) as characteristic quantities. If only a few characteristic quantities are input, and data is unbalanced, a deep convolutional neural network is prone to overfitting. There are still some difficulties in applying a deep learning technology to fault diagnosis of a transformer with dissolved gas in oil: 1) There are only a few and unbalanced transformer datasets. 2) There are only a little characteristic gas, in other words, there are only a few characteristics at an input end. 3) A network training process is prone to overfitting, and oscillation occurs in the training process.
The technical issue to be resolved in the present disclosure is to provide a method for diagnosing transformer fault based on a deep coupled dense convolutional neural network to overcome the defects in the prior art.
The present disclosure resolves the technical issue with the following technical solution:
The present disclosure provides a method for diagnosing transformer fault based on a deep coupled dense convolutional neural network, including the following steps:
Further, step 1 in the present disclosure includes:
samplei={xi,1,xi,2, . . . ,xi,j,yi}i∈[1,N]
where samplei represents data of the dissolved gas in the oil in an ith sample, and there are N data samples in total; xi,j represents a content of jth characteristic gas in the ith sample; and yi represents a state of the transformer in the ith sample;
Further, the adaptive synthetic oversampling method in step 2 in the present disclosure includes:
where xAj represents a content of jth characteristic gas of sample point A, xBj represents a content of jth characteristic gas of sample point B, and d represents a Euclidean distance between sample point A and sample point B;
r
i=Δi/K
where K represents a quantity of adjacent samples, Δi represents a quantity of majority-class samples adjacent to an ith sample, and ri represents a proportion of the majority-class samples adjacent to the ith sample;
where G represents the total quantity of samples to be synthesized for the minority-class sample, and gi represents a quantity of samples to be synthesized for the ith sample;
s
i
=x
i+(xzi−xi)×λ
where xi represents the ith data sample, λ represents a random number, and si represents a new data sample generated for the ith sample; and
Further, step 3 in the present disclosure includes:
Further, step 4 in the present disclosure includes:
x
m
=F
m([xm-2,xm-1])
where xm represents an input value of a network at an mth layer, namely, an output value of a network at an (m−1)th layer, and Fm represents a calculation function of the mth layer;
y=ΣWX+b
where x represents an input value, w represents a weight, b represents an offset, and y represents an output;
where x represents an input value of the layer, and f(x) represents an output value of the layer; and
Further, the transformer fault diagnosis model in step 4 in the present disclosure is specifically as follows:
Further, step 5 in the present disclosure includes:
The present disclosure achieves the following beneficial effects: The method for diagnosing transformer fault based on a deep coupled dense convolutional neural network in the present disclosure can resolve a problem that a fault diagnosis accuracy rate of the transformer based on the deep convolutional neural network is low due to insufficient and unbalanced fault samples in the dissolved gas in the oil and few characteristic quantities. Considering a mutual relation between the characteristic gas and the fault state comprehensively, the present disclosure provides a method for building the deep dense convolutional neural network to effectively alleviate oscillation and overfitting in a training process.
The present disclosure is described in further detail with reference to the accompanying drawings and embodiments.
To make the objectives, technical solutions, and advantages of the present disclosure clearer, the present disclosure is further described below in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely intended to explain the present disclosure, rather than to limit the present disclosure. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present disclosure without creative efforts shall fall within the protection scope of the present disclosure.
Datasets of dissolved gas in oil of each substation in the normal and fault states are collected from relevant literature over the years and actual test data of electric power companies. Each group of data includes five kinds of characteristic gas: hydrogen (H2), methane (CH4), ethane (C2H6), ethylene (C2H4), and acetylene (C2H2), and a state of the transformer.
The dataset of the dissolved gas in the oil in step 1 is as follows:
samplei={xi,1,xi,2, . . . ,xi,j,yi}i∈[1,N]
where samplei represents data of the dissolved gas in the oil in an ith sample, and there are N data samples in total; xi,j represents a content of jth characteristic gas in the ith sample; and yi represents a state of the transformer in the ith sample.
In step 1, normalization is performed according to the following formula:
In step 1, the label is set for the state of the transformer in a form of a numerical serial number. The transformer has a total of seven states: normal state, low-temperature overheating, mid-temperature overheating, high-temperature overheating, partial discharge, low-energy discharge, and high-energy discharge. Labels for the normal state, low-temperature overheating, mid-temperature overheating, high-temperature overheating, partial discharge, low-energy discharge, and high-energy discharge are defined as 0, 1, 2, 3, 4, 5, and 6 respectively. A quantity of samples collected in each state is shown in Table 1.
Step 2: Expand the obtained datasets in step 1 by using an adaptive synthetic oversampling technology, to form a new dataset.
A calculation process of the adaptive synthetic oversampling technology in step 2 may include:
where xAj represents a content of jth characteristic gas of sample point A, xBj represents a content of jth characteristic gas of sample point B, and d represents a Euclidean distance between sample point A and sample point B;
r
i=Δi/K
where K represents a quantity of adjacent samples, Δi represents a quantity of majority-class samples adjacent to an ith sample, and ri represents a proportion of the majority-class samples adjacent to the ith sample;
where G represents the total quantity of samples to be synthesized for the minority-class sample, and gi represents a quantity of samples to be synthesized for the ith sample;
s
i
=x
i+(xzi−xi)×λ
where xi represents the ith data sample, λ represents a random number, and si represents a new data sample generated for the ith sample; and
Step 3: Perform, in a form of a two-dimensional matrix, feature reconstruction on characteristic gas dissolved in the oil.
The feature reconstruction in step 3 is to construct a content of the characteristic gas in a form of one-dimensional matrix into the two-dimensional matrix through permutation and combination, so as to increase characteristic quantities, and comprehensively considering a relationship between each gas content change and the fault state.
The one-dimensional matrix composed of the five kinds of characteristic gas is constructed into the two-dimensional matrix through permutation and combination. When two kinds of characteristic gas are used, there are a total of 10 combination modes, such that a 2*10 two-dimensional matrix is formed, and so on. A 3*10 two-dimensional matrix is constructed for three kinds of characteristic gas, and a 4*5 two-dimensional matrix is constructed for four kinds of characteristic gas, as shown in
Step 4: Build a transformer fault diagnosis model based on a deep coupled dense convolutional neural network.
The deep coupled dense convolutional neural network in step 4 fuses values calculated by two previous convolutional layers in a depth direction, to serve as an input value of a next convolutional layer as shown in
x
m
=F
m([xm-2,xm-1])
In the above formula, xm represents an input value of a network at an mth layer, namely, an output value of a network at an (m−1)th layer, and Fm represents a calculation function of the mth layer.
The calculation function mainly includes five basic calculation processes: convolution calculation, standardization, activation functions, pooling, and discarding.
A simplified formula of the convolution calculation is as follows:
y=ΣWX+b
where x represents an input value, w represents a weight, b represents an offset, and y represents an output.
The standardization is capable of making data conform to a standard normal distribution with an average value of 0 and a standard deviation of 1. The activation functions mainly include a relu function, a tanh function, and a softmax function. The convolutional layer uses the relu function, a fully connected layer uses the tanh function, and an output layer uses the softmax function.
In the above formula, x represents an input value of the layer, and f(x) represents an output value of the layer.
The pooling can reduce an amount of characteristic data in a convolutional neural network, and mainly includes maximum pooling and average pooling. A discarding layer is mainly used to discard some neurons, so as to effectively prevent overfitting of the convolutional neural network.
The building of the transformer fault diagnosis model in step 4 is a design process of the deep coupled dense convolutional neural network, mainly including two parts: setting a quantity of convolutional layers in a coupled dense module, and setting a quantity of coupled dense modules in the deep coupled dense convolutional neural network.
The dataset in step 5 is divided into the training set and the test set in a proportion of 4:1.
As described in step 5, three two-dimensional matrices in step 3 are used as inputs of the deep coupled dense convolutional neural network, and are input into their corresponding deep coupled dense convolutional neural networks, and feature data obtained is flattened into one-dimensional data by using a flatten layer, and the one-dimensional data is input into the subsequent fully connected layer. The output layer uses the softmax function. The label set in step 1 is used as the output to train the network. An overall network architecture is shown in
An Adam algorithm is used in a training process. 512 pieces of data are put for iteration each time, such that the deep coupled dense convolutional neural network performs iteration and modifies a parameter value continuously, to obtain the transformer fault diagnosis model in which the training set and the test set have a small loss function and a high accuracy rate. According to step 5, data of dissolved gas in the oil of the transformer in the test set is input, and a final diagnosis accuracy rate is 94.05%.
In conclusion, the present disclosure realizes diagnosis of the fault state of the transformer. The adaptive synthetic oversampling technology makes fault sample data of the transformer more sufficient and balanced. The reconstruction of the characteristic gas increases the characteristic quantities, and comprehensively considers a mutual relation between each kind of gas and the transformer state. The transformer fault diagnosis model based on the deep coupled dense convolutional neural network can achieve a high fault diagnosis accuracy rate.
It should be understood that those of ordinary skill in the art can make improvements or transformations based on the above description, and all these improvements and transformations should fall within the protection scope of the appended claims of the present disclosure.
Number | Date | Country | Kind |
---|---|---|---|
202210638253.6C | Jun 2022 | CN | national |