Method for diagnosing transformer fault based on deep coupled dense convolutional neural network
A method for diagnosing transformer fault based on a deep coupled dense convolutional neural network, includes: obtaining datasets of dissolved gas in oil of a transformer in normal and fault states; expanding the datasets by using an adaptive synthetic oversampling method; performing, in a form of a two-dimensional matrix, feature reconstruction on characteristic gas dissolved in the oil; building a transformer fault diagnosis model based on a deep coupled dense convolutional neural network; and dividing an expanded dataset into a training set and a test set, and taking the two-dimensional matrix as an input of the deep coupled dense convolutional neural network and a set label as an output to train the network to obtain a fault diagnosis model. The present disclosure can resolve a problem that a fault diagnosis accuracy rate of the transformer is low due to insufficient and unbalanced fault samples in the dissolved gas in the oil.
1. A method for diagnosing a transformer fault based on a deep coupled dense convolutional neural network, comprising the following steps:
step 1: obtaining datasets of dissolved gas in oil of a transformer in normal and fault states, normalizing the datasets of the dissolved gas in the oil, and setting a label;
step 2: expanding the obtained datasets of the dissolved gas in the oil in step 1 by using an adaptive synthetic oversampling method, to form a new dataset;
step 3: performing, in a form of a two-dimensional matrix, feature reconstruction on characteristic gas dissolved in the oil;
step 4: building a transformer fault diagnosis model based on a deep coupled dense convolutional neural network; and
step 5: dividing the new expanded dataset in step 2 into a training set and a test set, taking the two-dimensional matrix in step 3 as an input of the deep coupled dense convolutional neural network and the set label in step 1 as an output to train the deep coupled dense convolutional neural network, and calculating an accuracy rate based on the test set to obtain a trained transformer fault diagnosis model;
wherein: the transformer fault diagnosis model in step 4 comprises a quantity of coupled dense modules; the coupled dense modules comprise a quantity of convolutional layers; the number of coupled dense modules is 1 to 4; the number of convolutional layers is 2 to 6;
the coupled dense modules are configured to train a sample and measurement accuracy rate of test sets; the convolutional layers are configured to train the dataset and measurement accuracy rate of test sets;
the transformer fault diagnosis model is configured to fuse values calculated by two previous convolutional layers in a depth direction as the input value of a next convolutional layer:
x m =F m ([ x m−2 ,x m−1 ])
wherein x m represents an input value of a network at an m th layer, namely, an output value of a network at an (m−1) th layer, and F m represents a calculation function of the m th layer;
the calculation function mainly comprises 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
wherein x represents an input value, w represents a weight, b represents an offset, and y represents an output;
the standardization is configured to making data conform to a standard normal distribution with an average value of 0 and a standard deviation of 1; the activation functions mainly comprise a Relu function, a Tanh function, and a softmax function; and the convolutional layer uses the relu function, a fully connected layer uses the tanh function, and an output layer uses the softmax function;
Relu
f
(
x
)
=
max
(
0
,
x
)
Than
f
(
x
)
=
e
x
-
e
-
x
e
x
+
e
-
x
soft
max
f
(
x
)
=
e
x
i
∑
i
=
0
n
e
x
i
wherein x represents an input value of the layer, and f(x) represents an output value of the layer; and
the pooling is configured to reducing an amount of characteristic data in a convolutional neural network, and mainly comprises maximum pooling and average pooling; and a discarding layer is mainly used to discard some neurons, so as to effectively prevent overfitting of the convolutional neural network.
2. The method according to claim 1 , wherein step 1 comprises:
obtaining the dataset of the dissolved gas in the oil according to the following formula:
sample i={x i,1 ,x i,2 , . . . ,x i,j ,y i }i∈[ 1 ,N]
wherein samplei represents data of the dissolved gas in the oil in an i th sample, and there are N data samples in total; x i,j represents a content of j th characteristic gas in the i th sample; and y i represents a state of the transformer in the i th sample;
performing normalization according to the following formula:
X
i
,
j
′
=
X
i
,
j
∑
j
=
0
X
i
,
j
setting the label for the state of the transformer in a form of a numerical serial number.
3. The method according to claim 1 , wherein the adaptive synthetic oversampling method in step 2 comprises:
(1) assuming that a quantity of samples to be synthesized for a minority-class sample is G, and for each minority-class sample samplei, finding K adjacent samples by using a Euclidean distance formula in n-dimensional space:
d
=
∑
j
=
1
n
(
x
Aj
-
x
Bj
)
2
wherein x Aj represents a content of j th characteristic gas of sample point A, x Bj represents a content of j th characteristic gas of sample point B, and d represents a Euclidean distance between sample point A and sample point B;
(2) calculating a proportion of majority-class samples among the found K adjacent samples;
r i =Δ i /K
wherein K represents a quantity of adjacent samples, Δ i represents a quantity of majority-class samples adjacent to an i th sample, and r i represents a proportion of the majority-class samples adjacent to the i th sample;
(3) calculating a quantity of samples to be synthesized for each minority-class sample samplei:
g
i
=
r
i
∑
i
r
i
×
G
wherein G represents the total quantity of samples to be synthesized for the minority-class sample, and g i represents a quantity of samples to be synthesized for the i th sample;
(4) selecting a minority-class sample x zi randomly from the K adjacent samples of each minority-class sample samplei to generate a synthetic data sample:
s i =x i +( x zi −x i )×λ
wherein x i represents the i th sample, λ represents a random number, and s i represents a new data sample generated for the i th sample; and
performing the above steps to expand the dataset to make datasets in various fault states more balanced.
4. The method according to claim 1 , wherein step 3 comprises:
constructing a content of the characteristic gas in a form of a 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.
5. The method according to claim 1 , wherein step 5 comprises:
dividing the new dataset into the training set and the test set in a proportion of 4:1;
taking the two-dimensional matrix in step 3 as the input of the deep coupled dense convolutional neural network and the set label in step 1 as the output to train the network; and taking a loss function and a fault diagnosis accuracy rate as reference standards; and
performing training by using an Adam algorithm, such that the deep coupled dense convolutional neural network performs iteration and modifies a parameter value continuously, and calculating the loss function and the accuracy rate based on the test set to obtain the trained transformer fault diagnosis model.