CN113392773A - Transformer fault diagnosis system of convolutional neural network based on map information fusion - Google Patents

Transformer fault diagnosis system of convolutional neural network based on map information fusion Download PDF

Info

Publication number
CN113392773A
CN113392773A CN202110670329.9A CN202110670329A CN113392773A CN 113392773 A CN113392773 A CN 113392773A CN 202110670329 A CN202110670329 A CN 202110670329A CN 113392773 A CN113392773 A CN 113392773A
Authority
CN
China
Prior art keywords
information
map
spectrum
model
information fusion
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202110670329.9A
Other languages
Chinese (zh)
Inventor
戴金林
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Qifeng Technology Co ltd
Original Assignee
Qifeng Technology Co ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Qifeng Technology Co ltd filed Critical Qifeng Technology Co ltd
Priority to CN202110670329.9A priority Critical patent/CN113392773A/en
Publication of CN113392773A publication Critical patent/CN113392773A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/08—Feature extraction
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/25—Fusion techniques
    • G06F18/253—Fusion techniques of extracted features
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/12—Classification; Matching

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • General Physics & Mathematics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Evolutionary Computation (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Evolutionary Biology (AREA)
  • Biophysics (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Molecular Biology (AREA)
  • General Health & Medical Sciences (AREA)
  • Computational Linguistics (AREA)
  • Biomedical Technology (AREA)
  • Health & Medical Sciences (AREA)
  • Signal Processing (AREA)
  • Complex Calculations (AREA)

Abstract

A transformer fault diagnosis system of a convolutional neural network based on map information fusion. The method comprises the following steps: step 1, collecting vibration signals of different fault states of a transformer, and calculating a corresponding fast spectrum kurtogram, a square envelope information spectrum and a square envelope spectrum information spectrum; step 2, extracting the characteristics of the information map by using a convolution algorithm to obtain a characteristic map; step 3, carrying out information fusion on the feature map in the step 2 by using a proposed square neighbor weighting algorithm to obtain an information fusion map; step 4, determining a label corresponding to the obtained information fusion map, and then training an improved CNN model as a sample and converging the CNN model; and 5, installing the diagnosis model obtained in the steps 1-4 in an upper computer program, processing a vibration signal uploaded by a signal acquisition system in real time, and sending an alarm through an information interaction system if a fault is detected. The model can accurately and effectively detect the fault of the transformer and has good practical application value.

Description

Transformer fault diagnosis system of convolutional neural network based on map information fusion
Technical Field
The invention relates to the field of transformer fault diagnosis, in particular to a transformer fault diagnosis system based on a convolutional neural network of map information fusion.
Background
Power transformers are critical devices in power systems, and their reliability directly affects the operational safety of the power system. The faults of the power transformer mainly include winding deformation, iron core looseness, tap joint faults, oil paper insulation aging and the like. Wherein the winding deformation accounts for about 15% of the total fault and the core looseness accounts for about 45% of the total fault. The deformation of the winding can further aggravate the vibration and cause the deterioration of turn-to-turn insulation performance, if the winding cannot be overhauled in time, the deformation of the winding is further increased under the next short-circuit current impact, and finally the winding can be seriously damaged under the action of small short-circuit current and overvoltage; when the iron core is loosened, no-load loss is increased, the iron core is overheated, and an iron core insulating layer is aged and falls off or even the iron core is burnt. Therefore, early identification of power transformer winding and core faults is of great significance to ensure safe operation of the power system.
From the current research, in the field of transformer equipment fault diagnosis, the fault diagnosis by using a signal processing and deep learning model has wide application. For the signal processing technology, factors such as noise interference in the environment, weak characteristics of fault signals and the like can bring great interference; for the deep learning technology, the application of the model is further limited by the problems of training sample feature difference, insufficient generalization performance of the model and the like.
The patent related to transformer fault diagnosis in China is 'a transformer fault diagnosis method based on a deep forest model' (202010621042.2), the patent firstly takes a non-coding ratio of analysis data of dissolved gas in transformer oil as a characteristic parameter of the deep forest model, then a deep forest model DF is constructed, the deep forest model DF extracts more characteristic information from multi-dimensional data of transformer faults through multi-granularity scanning, and the effect of diagnosing and identifying the fault type of a transformer is optimal through training of cascade forests, but the patent does not consider the generalization problem of the model. The invention discloses a transformer core loosening defect diagnosis method based on a vibration frequency spectrum matrix (201910468382.3), which comprises the steps of firstly converting vibration signals into frequency domain signals, then forming a frequency row matrix through the frequency signals, then forming an amplitude row matrix by using amplitude signals in a frequency spectrum diagram, multiplying the two matrixes to obtain a second core coefficient, then calculating the change rate of the second core coefficient relative to the first core coefficient as a judgment factor for judging the loosening degree of the core, and finally comparing the judgment factor with a preset defect threshold value to judge whether the core is loosened.
Disclosure of Invention
In order to solve the problems, the invention provides a transformer fault diagnosis system of a convolutional neural network based on map information fusion on the basis of an information map, a fusion algorithm thereof and a CNN model. In view of the fact that the information map contains rich frequency band information, the frequency domain information contained in the fault signal is extracted by using a fast spectral kurtosis map, a square envelope information spectrum and a square envelope spectrum information spectrum, and then a square neighbor weighting algorithm is proposed to fuse the three information maps to obtain an information fusion map, so that the image characteristics input by a model are enhanced; in addition, aiming at the problems that the traditional CNN model is easy to be over-fitted and insufficient in generalization, the patent provides an improved CNN model, wherein an LEU activation function is provided, a dynamic dropout layer is provided, a loss function for solving the imbalance and sparsity of samples is provided, and the generalization of the model is improved. In order to achieve the purpose, the invention provides a transformer fault diagnosis system of a convolutional neural network based on atlas information fusion, which comprises the following specific steps:
step 1, obtaining an information map: collecting vibration signals of the transformer in different fault states, and calculating a corresponding fast spectrum kurtosis diagram, a square envelope information spectrum and a square envelope spectrum information spectrum, wherein a PCB acceleration sensor selects 608A11, and a data acquisition card selects NI 9234;
step 2, extracting convolution characteristics: extracting the characteristics of the information map by using a convolution algorithm to obtain a characteristic map;
and step 3, fusing the characteristic information: carrying out information fusion on the feature map in the step 2 by using a proposed square neighbor weighting algorithm to obtain an information fusion map;
step 4, training an improved CNN model: determining a label corresponding to the obtained information fusion map, and then training an improved CNN model as a sample and converging the CNN model;
step 5, model on-line monitoring: and (4) installing the diagnosis model obtained in the steps 1-4 in an upper computer program, processing a vibration signal uploaded by a signal acquisition system in real time, and sending an alarm through an information interaction system if a fault is detected.
Further, the process of calculating the fast spectral kurtogram, the squared envelope information spectrum, and the squared envelope spectrum information spectrum in step 1 can be represented as follows:
assuming that a discrete time-domain vibration signal x (n) (1, 2.. multidot.l) of length L is acquired, the corresponding spectral kurtosis K in the fast spectral kurtosis mapY(f) Is defined as:
Figure BDA0003118453260000021
where {. denotes the mathematical expectation, | - | denotes the absolute value, and H (n, f) denotes the complex envelope of the signal x (n) at the frequency f.
For a squared envelope information spectrum, the negative entropy of its spectrum in the time domain is defined as:
Figure BDA0003118453260000022
in the formula, SEx(n; f, Δ f) denotes that x (n) is within a frequency band range [ f- Δ f/2, f + Δ f/2]The inner squared envelope.
For the square envelope spectrum information spectrum, the definition of the spectrum negative entropy in the frequency domain is as follows:
Figure BDA0003118453260000031
in the formula, SESx(alpha; f, delta f) represents x (n) in the frequency band range [ f-delta f/2, f + delta f/2]Inner square envelope spectrum, alpha is the cycle frequency.
Further, the specific description of the feature extraction of the information map by using the convolution algorithm in the step 2 is as follows:
processing the information map by using a convolution kernel in a Fullpadding mode, wherein the specification of the convolution kernel is 3 multiplied by 3, and the moving step length is set to be 1;
further, the specific steps of performing information fusion on the feature map by using the proposed square neighbor weighting algorithm in the step 3 are as follows:
step 3.1, determining values of a fast spectral kurtogram, a squared envelope information spectrum and a squared envelope information spectrum at the ith row and jth column positions
Figure BDA0003118453260000032
And
Figure BDA0003118453260000033
and calculating the average value
Figure BDA0003118453260000034
Namely, it is
Figure BDA0003118453260000035
Step 3.2, calculate
Figure BDA0003118453260000036
And
Figure BDA0003118453260000037
to
Figure BDA0003118453260000038
Distance d squaredp(p is 1,2,3) and determining corresponding neighbor weight coefficients
Figure BDA0003118453260000039
The calculation formula is as follows:
Figure BDA00031184532600000310
step 3.3, calculating the value If of the corresponding pixel point after information fusionijThe expression is as follows:
Figure BDA00031184532600000311
further, the specific steps of training the improved CNN model in step 4 are as follows:
step 4.1, sequentially performing feature extraction on the information fusion map by using the convolution layer and an LEU activation function to obtain a feature map, wherein the LEU activation function is defined as:
Figure BDA00031184532600000312
step 4.2, performing Pooling treatment on the feature map by using Max Pooling Pooling to realize dimension reduction of the feature map;
and 4.3, performing feature map sparsification by using the proposed dynamic dropout layer, wherein the dynamic dropout layer sets some pixel points in the feature map to be 0 by random probability in each model training process, and the expression of the random probability is as follows:
rp＝e-m
wherein m is a thinning coefficient and satisfies a uniform distribution, i.e., m to U [1,10 ].
Step 4.4, continuing the convolution-LEU activation function-pooling-dropout algorithm flow from the step 4.1 to the step 4.3 to further process the feature map to obtain a final feature map;
step 4.5, unfolding the characteristic diagram obtained in the step 4.4 in a Flatten form, connecting two full connecting layers, and then realizing classification and identification of the image through Softmax logistic regression, wherein an activation function still selects an LEU;
step 4.6, repeating the steps 4.1 to 4.5 until the convergence of the loss function or the iteration reaches the set times 100, wherein the loss function LICNNThe design is as follows:
LICNN＝Lfl+Lre+Lsp
in the formula, LflIs the Focal loss function, LreFor regular penalty loss, LspFor dropout layer sparse term loss, specific expressions are respectively as follows:
Figure BDA0003118453260000041
Figure BDA0003118453260000042
where μ denotes the equilibrium coefficient, γ is the adjustment coefficient, ptIs the predicted probability of the model, p is the output probability of the model, and y is the logical output of the model.
Figure BDA0003118453260000043
In the formula, λ represents a regularization coefficient, wiThe number of the convolution kernel weight coefficients is N.
Figure BDA0003118453260000044
In the formula, s represents the number of neurons in a dropout layer, rho is a sparsity parameter,
Figure BDA0003118453260000045
is the average activity of the j dropout layer neurons.
The transformer fault diagnosis system of the convolutional neural network based on the map information fusion has the beneficial effects that: the invention has the technical effects that:
1. the method utilizes the fast spectral kurtosis map, the square envelope information spectrum and the square envelope spectrum information spectrum to extract frequency domain information contained in the fault signal, and then proposes a square neighbor weighting algorithm to fuse the three information maps to obtain an information fusion map, thereby enhancing the image characteristics input by the model;
2. aiming at the problems that a traditional CNN model is easy to be over-fitted and insufficient in generalization, the patent provides an improved CNN model, wherein an LEU (Lee activating unit) activating function is provided to solve the problem that the gradient of the traditional sigmoid activating function is easy to disappear during back propagation, and meanwhile, a dynamic dropout layer and a loss function for solving the imbalance and sparsity of a sample are provided, so that the generalization of the model is improved, and the accurate and effective diagnosis of the transformer fault is realized.
Drawings
FIG. 1 is a flow chart of the present invention;
FIG. 2 is a network architecture diagram of a graph information fused convolutional neural network used in the present invention;
FIG. 3 is a schematic diagram of an online diagnostic system.
Detailed Description
The invention is described in further detail below with reference to the following detailed description and accompanying drawings:
the invention provides a transformer fault diagnosis system based on a convolutional neural network of map information fusion, and aims to realize accurate and effective diagnosis of transformer faults and improve power transmission capacity. FIG. 1 is a flow chart of the present invention, and the steps of the present invention will be described in detail in conjunction with the flow chart.
Step 1, obtaining an information map: collecting vibration signals of the transformer in different fault states, and calculating a corresponding fast spectrum kurtosis diagram, a square envelope information spectrum and a square envelope spectrum information spectrum, wherein a PCB acceleration sensor selects 608A11, and a data acquisition card selects NI 9234;
the process of calculating the fast spectral kurtogram, the squared envelope information spectrum, and the squared envelope spectrum information spectrum in step 1 can be represented as follows:
assuming that a discrete time-domain vibration signal x (n) (1, 2.. multidot.l) of length L is acquired, the corresponding spectral kurtosis K in the fast spectral kurtosis mapY(f) Is defined as:
Figure BDA0003118453260000051
where {. denotes the mathematical expectation, | - | denotes the absolute value, and H (n, f) denotes the complex envelope of the signal x (n) at the frequency f.
For a squared envelope information spectrum, the negative entropy of its spectrum in the time domain is defined as:
Figure BDA0003118453260000052
in the formula, SEx(n; f, Δ f) denotes that x (n) is within a frequency band range [ f- Δ f/2, f + Δ f/2]The inner squared envelope.
For the square envelope spectrum information spectrum, the definition of the spectrum negative entropy in the frequency domain is as follows:
Figure BDA0003118453260000053
in the formula, SESx(alpha; f, delta f) represents x (n) in the frequency band range [ f-delta f/2, f + delta f/2]Inner square envelope spectrum, alpha is the cycle frequency.
Step 2, extracting convolution characteristics: extracting the characteristics of the information map by using a convolution algorithm to obtain a characteristic map;
the specific description of the information map feature extraction by using the convolution algorithm in the step 2 is as follows:
processing the information map by using a convolution kernel in a Full padding mode, wherein the specification of the convolution kernel is 3 multiplied by 3, and the moving step length is set to be 1;
and step 3, fusing the characteristic information: carrying out information fusion on the feature map in the step 2 by using a proposed square neighbor weighting algorithm to obtain an information fusion map;
the specific steps of utilizing the proposed square nearest neighbor weighting algorithm to perform information fusion on the feature map in the step 3 are as follows:
step 3.1, determining values of a fast spectral kurtogram, a squared envelope information spectrum and a squared envelope information spectrum at the ith row and jth column positions
Figure BDA0003118453260000061
And
Figure BDA0003118453260000062
and calculating the average value
Figure BDA0003118453260000063
Namely, it is
Figure BDA0003118453260000064
Step 3.2, calculate
Figure BDA0003118453260000065
And
Figure BDA0003118453260000066
to
Figure BDA00031184532600000611
Distance d squaredp(p is 1,2,3) and determining corresponding neighbor weight coefficients
Figure BDA0003118453260000067
The calculation formula is as follows:
Figure BDA0003118453260000068
step 3.3, calculating the value If of the corresponding pixel point after information fusionijThe expression is as follows:
Figure BDA0003118453260000069
step 4, training an improved CNN model: determining a label corresponding to the obtained information fusion map, and then training an improved CNN model as a sample and converging the CNN model;
the specific steps of training the improved CNN model in step 4 are as follows:
step 4.1, sequentially performing feature extraction on the information fusion map by using the convolution layer and an LEU activation function to obtain a feature map, wherein the LEU activation function is defined as:
Figure BDA00031184532600000610
step 4.2, performing Pooling treatment on the feature map by using Max Pooling Pooling to realize dimension reduction of the feature map;
and 4.3, performing feature map sparsification by using the proposed dynamic dropout layer, wherein the dynamic dropout layer sets some pixel points in the feature map to be 0 by random probability in each model training process, and the expression of the random probability is as follows:
rp＝e-m
wherein m is a thinning coefficient and satisfies a uniform distribution, i.e., m to U [1,10 ].
Step 4.4, continuing the convolution-LEU activation function-pooling-dropout algorithm flow from the step 4.1 to the step 4.3 to further process the feature map to obtain a final feature map;
step 4.5, unfolding the characteristic diagram obtained in the step 4.4 in a Flatten form, connecting two full connecting layers, and then realizing classification and identification of the image through Softmax logistic regression, wherein an activation function still selects an LEU;
step 4.6, repeating the steps 4.1 to 4.5 until the convergence of the loss function or the iteration reaches the set times 100, wherein the loss function LICNNThe design is as follows:
LICNN＝Lfl+Lre+Lsp
in the formula, LflIs the Focal loss function, LreFor regular penalty loss, LspFor dropout layer sparse term loss, specific expressions are respectively as follows:
Figure BDA0003118453260000071
Figure BDA0003118453260000072
where μ denotes the equilibrium coefficient, γ is the adjustment coefficient, ptIs the predicted probability of the model, p is the output probability of the model, and y is the logical output of the model.
Figure BDA0003118453260000073
In the formula, λ represents a regularization coefficient, wiThe number of the convolution kernel weight coefficients is N.
Figure BDA0003118453260000074
In the formula, s represents the number of neurons in a dropout layer, rho is a sparsity parameter,
Figure BDA0003118453260000075
is the average activity of the j dropout layer neurons.
Step 5, model on-line monitoring: and (4) installing the diagnosis model obtained in the steps 1-4 in an upper computer program, processing a vibration signal uploaded by a signal acquisition system in real time, and sending an alarm through an information interaction system if a fault is detected.
FIG. 2 is a network architecture diagram of a graph information fused convolutional neural network used in the invention. The structure diagram can be divided into two major modules, which are respectively: an information fusion module and an improved CNN module. For the information fusion module, frequency domain information contained in the fault signal is extracted by using a fast spectral kurtosis diagram, a square envelope information spectrum and a square envelope spectrum information spectrum, then a square neighbor weighting algorithm is proposed to fuse the three information diagrams to obtain an information fusion diagram, and then the information fusion diagram is used as a training sample to train an improved CNN model, so that the image characteristics input by the CNN model are enhanced. For an improved CNN module, the problem that a traditional CNN model is easy to over-fit and insufficient in generalization is considered, the improved CNN model is provided, an LEU activation function is provided to solve the problem that the gradient of the traditional sigmoid activation function is easy to disappear during back propagation, a dynamic dropout layer and a loss function combining unbalanced and sparse samples are provided, the generalization of the model is improved, and accurate and effective diagnosis of transformer faults is achieved. In general, the sensor and the acquisition system thereof acquire the vibration signals of the transformer, and finally judge whether the signals contain fault information or not through algorithm fusion and characteristic extraction of CNN at different levels, so that the safety of the transformer is guaranteed and the power transmission capacity is improved.
Fig. 3 is a schematic diagram of an online diagnostic system, and it can be seen that: the 608A11-PCB acceleration sensor is installed on each transformer, and then the NI9234 data acquisition card converts the signals acquired by the sensor through a signal conditioning circuit and uploads the signals to a data acquisition system; and then the server calls the convolutional neural network with map information fusion to classify and identify the uploaded data, if a fault is found, an alarm message is sent out through an alarm, and otherwise, the monitoring is continued. The whole system guarantees the working safety of the transformer and improves the power transmission capacity of the transformer.
The above description is only a preferred embodiment of the present invention, and is not intended to limit the present invention in any way, but any modifications or equivalent variations made according to the technical spirit of the present invention are within the scope of the present invention as claimed.

Claims (5)

1. The transformer fault diagnosis system of the convolutional neural network based on map information fusion specifically comprises the following steps:
step 1, obtaining an information map: acquiring vibration signals of the transformer in different fault states, and calculating a corresponding fast spectrum kurtogram, a square envelope information spectrum and a square envelope spectrum information spectrum;
the process of calculating the fast spectral kurtogram, the squared envelope information spectrum, and the squared envelope spectrum information spectrum in step 1 can be represented as follows:
assuming that a discrete time-domain vibration signal x (n) (1, 2.. multidot.l) of length L is acquired, the corresponding spectral kurtosis K in the fast spectral kurtosis mapY(f) Is defined as:
Figure FDA0003118453250000011
where {. denotes the mathematical expectation, | - | denotes the absolute value, and H (n, f) denotes the complex envelope of the signal x (n) at the frequency f.
For a squared envelope information spectrum, the negative entropy of its spectrum in the time domain is defined as:
Figure FDA0003118453250000012
in the formula, SEx(n; f, Δ f) denotes that x (n) is within a frequency band range [ f- Δ f/2, f + Δ f/2]The inner squared envelope.
For the square envelope spectrum information spectrum, the definition of the spectrum negative entropy in the frequency domain is as follows:
Figure FDA0003118453250000013
in the formula, SESx(alpha; f, delta f) represents x (n) in the frequency band range [ f-delta f/2, f + delta f/2]Inner square envelope spectrum, alpha is cycle frequency;
step 2, extracting convolution characteristics: extracting the characteristics of the information map by using a convolution algorithm to obtain a characteristic map;
and step 3, fusing the characteristic information: carrying out information fusion on the feature map in the step 2 by using a proposed square neighbor weighting algorithm to obtain an information fusion map;
the specific steps of utilizing the proposed square nearest neighbor weighting algorithm to perform information fusion on the feature map in the step 3 are as follows:
step 3.1, determining values of a fast spectral kurtogram, a squared envelope information spectrum and a squared envelope information spectrum at the ith row and jth column positions
Figure FDA0003118453250000014
And
Figure FDA0003118453250000015
and calculating the average value
Figure FDA0003118453250000016
Namely, it is
Figure FDA0003118453250000017
Step 3.2, calculate
Figure FDA0003118453250000018
And
Figure FDA0003118453250000019
to
Figure FDA00031184532500000110
Distance d squaredp(p is 1,2,3) and determining corresponding neighbor weight coefficients
Figure FDA00031184532500000111
The calculation formula is as follows:
Figure FDA0003118453250000021
step 3.3, calculating the value If of the corresponding pixel point after information fusionijThe expression is as follows:
Figure FDA0003118453250000022
step 4, training an improved CNN model: determining a label corresponding to the obtained information fusion map, and then training an improved CNN model as a sample and converging the CNN model;
step 5, model on-line monitoring: and (4) installing the diagnosis model obtained in the steps 1-4 in an upper computer program, processing a vibration signal uploaded by a signal acquisition system in real time, and sending an alarm through an information interaction system if a fault is detected.
2. The transformer fault diagnosis system based on the convolutional neural network of map information fusion of claim 1, wherein: in the step 1, 608A11 is selected as the PCB acceleration sensor, and NI9234 is selected as the data acquisition card.
3. The transformer fault diagnosis system based on the convolutional neural network of map information fusion of claim 1, wherein: the specific description of the information map feature extraction by using the convolution algorithm in the step 2 is as follows:
processing the information map by using a convolution kernel in a Full padding mode, wherein the specification of the convolution kernel is 3 multiplied by 3, and the moving step length is set to be 1;
4. the transformer fault diagnosis system based on the convolutional neural network of map information fusion of claim 1, wherein: the specific steps of training the improved CNN model in step 4 are as follows:
step 4.1, sequentially performing feature extraction on the information fusion map by using the convolution layer and an LEU activation function to obtain a feature map, wherein the LEU activation function is defined as:
Figure FDA0003118453250000023
step 4.2, performing Pooling treatment on the feature map by using Max Pooling Pooling to realize dimension reduction of the feature map;
and 4.3, performing feature map sparsification by using the proposed dynamic dropout layer, wherein the dynamic dropout layer sets some pixel points in the feature map to be 0 by random probability in each model training process, and the expression of the random probability is as follows:
rp＝e-m
wherein m is a thinning coefficient and satisfies a uniform distribution, i.e., m to U [1,10 ].
Step 4.4, continuing the convolution-LEU activation function-pooling-dropout algorithm flow from the step 4.1 to the step 4.3 to further process the feature map to obtain a final feature map;
step 4.5, unfolding the characteristic diagram obtained in the step 4.4 in a Flatten form, connecting two full connecting layers, and then realizing classification and identification of the image through Softmax logistic regression, wherein an activation function still selects an LEU;
step 4.6, repeating the steps 4.1 to 4.5 until the convergence of the loss function or the iteration reaches the set times 100, wherein the loss function LICNNThe design is as follows:
LICNN＝Lfl+Lre+Lsp
in the formula, LflIs the Focal loss function, LreFor regular penalty loss, LspFor dropout layer sparse term loss, specific expressions are respectively as follows:
Figure FDA0003118453250000031
Figure FDA0003118453250000032
where μ denotes the equilibrium coefficient, γ is the adjustment coefficient, ptIs the predicted probability of the model, p is the output probability of the model, and y is the logical output of the model.
Figure FDA0003118453250000033
In the formula, λ represents a regularization coefficient, wiThe number of the convolution kernel weight coefficients is N.
Figure FDA0003118453250000034
In the formula, s represents the number of neurons in a dropout layer, rho is a sparsity parameter,
Figure FDA0003118453250000035
is the average activity of the j dropout layer neurons.
5. The transformer fault diagnosis system based on the convolutional neural network of map information fusion of claim 1, wherein: the online monitoring of the model in step 5 is specifically described as follows:
and (4) installing the diagnosis model obtained in the steps 1-4 in an upper computer program, processing the vibration signal uploaded by the signal acquisition system in real time, if a fault is detected, sending an alarm through the information interaction system, and otherwise, continuing to monitor in real time.
CN202110670329.9A 2021-06-17 2021-06-17 Transformer fault diagnosis system of convolutional neural network based on map information fusion Pending CN113392773A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110670329.9A CN113392773A (en) 2021-06-17 2021-06-17 Transformer fault diagnosis system of convolutional neural network based on map information fusion

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110670329.9A CN113392773A (en) 2021-06-17 2021-06-17 Transformer fault diagnosis system of convolutional neural network based on map information fusion

Publications (1)

Publication Number Publication Date
CN113392773A true CN113392773A (en) 2021-09-14

Family

ID=77621833

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110670329.9A Pending CN113392773A (en) 2021-06-17 2021-06-17 Transformer fault diagnosis system of convolutional neural network based on map information fusion

Country Status (1)

Country Link
CN (1) CN113392773A (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113782113A (en) * 2021-09-17 2021-12-10 黄河水利职业技术学院 Method for identifying gas fault in transformer oil based on deep residual error network

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113782113A (en) * 2021-09-17 2021-12-10 黄河水利职业技术学院 Method for identifying gas fault in transformer oil based on deep residual error network

Similar Documents

Publication Publication Date Title
CN109612708B (en) Power transformer on-line detection system and method based on improved convolutional neural network
US11474163B2 (en) Power transformer winding fault positioning method based on deep convolutional neural network integrated with visual identification
Wang et al. A method for rapidly evaluating reliability and predicting remaining useful life using two-dimensional convolutional neural network with signal conversion
CN106197999B (en) A kind of planetary gear method for diagnosing faults
CN112257530B (en) Rolling bearing fault diagnosis method based on blind signal separation and support vector machine
CN107449994A (en) Partial discharge method for diagnosing faults based on CNN DBN networks
CN105841961A (en) Bearing fault diagnosis method based on Morlet wavelet transformation and convolutional neural network
CN108398268A (en) A kind of bearing performance degradation assessment method based on stacking denoising self-encoding encoder and Self-organizing Maps
CN112200244A (en) Intelligent detection method for anomaly of aerospace engine based on hierarchical countermeasure training
CN114417926A (en) Power equipment partial discharge pattern recognition method and system based on deep convolution generation countermeasure network
Wang et al. Prediction model of natural gas pipeline crack evolution based on optimized DCNN-LSTM
CN107037327A (en) Partial discharges fault judges feature extracting method and decision method
CN112146882B (en) Bearing fault diagnosis method based on transfer learning vibration signal image recognition
CN110020637A (en) A kind of analog circuit intermittent fault diagnostic method based on more granularities cascade forest
CN102034111A (en) Method for identifying and detecting aircraft structural damage conditions in diversified way
CN113203914A (en) Underground cable early fault detection and identification method based on DAE-CNN
CN113392773A (en) Transformer fault diagnosis system of convolutional neural network based on map information fusion
CN112160877A (en) Fan bearing fault diagnosis method based on SFA and CNN
CN117347796B (en) Intelligent gateway-based switching equipment partial discharge diagnosis system and method
CN111079647A (en) Circuit breaker defect identification method
CN117110744A (en) Transformer fault diagnosis method and system based on voiceprint analysis
CN116559728A (en) Power transmission line fault diagnosis method based on wavelet transformation-multilayer perceptron
CN115840116A (en) Switch cabinet partial discharge fault mode identification method based on GAN-CNN
CN114781427A (en) Wind generating set rolling bearing fault diagnosis system based on information fusion technology
CN114936947A (en) High-voltage direct-current transmission line fault diagnosis method based on GADF-VGG16

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination