A method and system for identifying dissolved gas content in transformer oil

Abstract

The present invention discloses a method and system for identifying the dissolved gas content in transformer oil, and belongs to the technical field of transformer operation status analysis. The method of the present invention comprises: preprocessing multiple groups of gas content data, generating sample data, and setting labels for the sample data; initializing deep migration network parameters, and inputting the sample data with the labels set and the data of the target transformer into the deep migration network to obtain the predicted labels of the target transformer data; determining the classification data of the gas in the sample data and the target transformer data according to the labels and predicted labels set for the sample data; inputting the classified data into the deep migration network to obtain the dissolved gas content in the target transformer oil. The present invention can extract the corresponding gas content characteristics in the transformer oil, identify the gas state in the oil according to different gas contents, accurately identify transformer faults, and provide excellent guarantees for the reliable operation and maintenance of power transformers.

Description

Technical Field

[0001] The present invention relates to the technical field of transformer operating state analysis, and more specifically, to a method and system for identifying dissolved gas content in transformer oil. Background Art

[0002] The transformer insulating oil will gradually deteriorate and decompose into low molecular gas due to various factors such as electricity and heat during operation. By analyzing the gas content in the oil, the insulating oil state can be identified, which can provide an effective basis for transformer equipment fault analysis and diagnosis. However, due to the different types and severity of internal transformer faults, the components and content of dissolved gas in the oil are different. Summary of the invention

[0003] In view of the above problems, the present invention proposes a method for identifying the dissolved gas content in transformer oil, comprising:

[0004] Acquire multiple sets of gas content data in transformer oil under normal operating conditions and abnormal operating conditions of the transformer, pre-process the multiple sets of gas content data, generate sample data, and set labels for the sample data;

[0005] Initialize the deep migration network parameters, and input the sample data with set labels and the data of the target transformer into the deep migration network to obtain the predicted label of the target transformer data;

[0006] According to the labels set for the sample data and the predicted labels, the classification data of the gas in the sample data and the target transformer data are determined;

[0007] The classified data is input into the deep migration network to obtain the dissolved gas content in the target transformer oil.

[0008] Optionally, the sample data is no less than 20,000 copies, and the sample data is the content data of H2, CH4, C2H2, C2H4 and C2H6 in transformer oil.

[0009] Optional, abnormal operating states include: transformer low-temperature overheating state, medium-temperature overheating state, high-temperature overheating state, partial discharge state, low-energy discharge state and high-energy discharge state.

[0010] Optionally, the preprocessing is normalization processing.

[0011] Optionally, deeply migrate network parameters, including: number of hidden layers, number of hidden layer nodes, initialized network weights and thresholds.

[0012] The present invention also proposes a system for identifying the dissolved gas content in transformer oil, comprising:

[0013] The acquisition module obtains multiple sets of gas content data in the transformer oil under normal operating conditions and abnormal operating conditions, pre-processes the multiple sets of gas content data, generates sample data, and sets labels for the sample data;

[0014] Initialization module, initializes the deep migration network parameters, and inputs the sample data with set labels and the data of the target transformer into the deep migration network to obtain the predicted label of the target transformer data;

[0015] A processing module determines classification data of gases in the sample data and target transformer data according to labels set for the sample data and predicted labels;

[0016] The output module inputs the classified data into the deep migration network to obtain the dissolved gas content in the target transformer oil.

[0017] Optionally, the sample data is no less than 20,000 copies, and the sample data is the content data of H2, CH4, C2H2, C2H4 and C2H6 in transformer oil.

[0018] Optional, abnormal operating states include: transformer low temperature overheating state, medium temperature overheating state, high temperature overheating state, partial discharge state, low energy discharge state and high energy discharge state.

[0019] Optionally, the preprocessing is normalization processing.

[0020] Optionally, deeply migrate network parameters, including: number of hidden layers, number of hidden layer nodes, initialized network weights and thresholds.

[0021] The present invention can extract the corresponding gas content characteristics in transformer oil, identify the gas state in the oil according to different gas contents, accurately identify transformer faults, and provide excellent guarantee for the reliable operation and maintenance of power transformers. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 A flow chart of a method for identifying dissolved gas content in transformer oil according to the present invention;

[0023] Figure 2 A method for identifying the dissolved gas content in transformer oil according to the present invention;

[0024] Figure 3 A method for identifying the dissolved gas content in transformer oil according to the present invention;

[0025] Figure 4 A method for identifying the dissolved gas content in transformer oil

[0026] Figure 5The present invention is a system structure diagram for identifying the dissolved gas content in transformer oil. DETAILED DESCRIPTION

[0027] Now, exemplary embodiments of the present invention are described with reference to the accompanying drawings. However, the present invention can be implemented in many different forms and is not limited to the embodiments described herein. These embodiments are provided to disclose the present invention in detail and completely and to fully convey the scope of the present invention to those skilled in the art. The terms used in the exemplary embodiments shown in the accompanying drawings are not intended to limit the present invention. In the accompanying drawings, the same units / elements are marked with the same reference numerals.

[0028] Unless otherwise specified, the terms (including technical terms) used herein have the commonly understood meanings to those skilled in the art. In addition, it is understood that the terms defined in commonly used dictionaries should be understood to have the same meanings as those in the context of the relevant fields, and should not be understood as idealized or overly formal meanings.

[0029] The present invention proposes a method for identifying the dissolved gas content in transformer oil, such as Figure 1 As shown, including:

[0030] Acquire multiple sets of gas content data in transformer oil under normal operating conditions and abnormal operating conditions of the transformer, pre-process the multiple sets of gas content data, generate sample data, and set labels for the sample data;

[0031] Initialize the deep migration network parameters, and input the sample data with set labels and the data of the target transformer into the deep migration network to obtain the predicted label of the target transformer data;

[0032] According to the labels set for the sample data and the predicted labels, the classification data of the gas in the sample data and the target transformer data are determined;

[0033] The classified data is input into the deep migration network to obtain the dissolved gas content in the target transformer oil.

[0034] Among them, the sample data is no less than 20,000 pieces, and the sample data are the content data of H2, CH4, C2H2, C2H4 and C2H6 in transformer oil.

[0035] Abnormal operating conditions include: transformer low-temperature overheating condition, medium-temperature overheating condition, high-temperature overheating condition, partial discharge condition, low-energy discharge condition and high-energy discharge condition.

[0036] The preprocessing is normalization.

[0037] Deep migration network parameters, including: number of hidden layers, number of hidden layer nodes, initialized network weights and thresholds.

[0038] The present invention will be further described below in conjunction with embodiments:

[0039] Sample data selection: When the transformer is operating normally, the contents of H2, CH4, C2H2, C2H4 and C2H6 are very small. When it is overheated at low temperature, the contents of CH4 and C2H6 are relatively high, and C2H4 is less than or even absent in comparison with C2H6. As the temperature rises, the content of C2H4 increases significantly. When the transformer is locally discharged, H2 and CH4 are mainly produced. When the transformer is discharged at low energy, the growth of C2H4 content is relatively fast compared with other components. When the transformer is discharged at high energy, a large amount of H2 and C2H2 are produced. Therefore, the contents of H2, CH4, C2H2, C2H4 and C2H6 are used as the standards for state identification. A total of 20,000 sets of gas data in transformer oil from some engineering examples are selected as sample source domain data.

[0040] Initialize network parameters: Before performing overall class center alignment, the network parameters need to be initialized. The parameters of the deep migration network mainly include the number of hidden layers, the number of nodes in each hidden layer, the network weight and threshold. For the gas data in transformer oil, the number of hidden layers is set to 2, the number of nodes in each hidden layer is 20, and the network weight and threshold are randomly initialized.

[0041] Deep neural network with overall class center alignment: The input is labeled source domain data and unlabeled target domain data, and the predicted label of the target domain data is obtained through overall class center alignment.

[0042] Deep neural network with various class center alignment: The data of each category in the source domain and the target domain are obtained through the real label of the source domain data and the predicted label of the target domain data. After obtaining the data of each category in the source domain and the target domain, the deep neural network with various class center alignment is used to classify the target domain data to obtain the classification result of the target domain data.

[0043] The key to the transfer learning network is the transfer strategy, that is, how to describe the relationship between the source domain data and the target domain data to achieve effective knowledge transfer. The class center alignment transfer strategy for each category of data is constrained to the last hidden layer of the network to align the features of each category of data in the source domain and the target domain. In the classification process, it is necessary to estimate the center of each category of data in the target domain. Because the target domain data does not contain the true label, the predicted label of the target domain data is needed to estimate the class center.

[0044] In order to improve the accuracy of target domain data prediction labels and class centroid estimation, the overall class centroid alignment migration strategy is first used to reduce the overall distribution drift of source domain data and target domain data. The prediction label is obtained by using the overall class centroid alignment. Then, each class centroid alignment is used to align the features of each category of data in the source domain and the target domain. Finally, the classification method is called the Centroid Alignment Deep Transfer Learning algorithm (CADTL). Its four-layer network structure is as follows: Figure 2 shown.

[0045] like Figure 2 As shown in the figure, the overall centroid alignment can achieve the overall distribution alignment of the source domain data and the target domain data. It is stipulated that the first layer is the input layer, the 2nd to Lth layers are hidden layers, and the last layer (L+1 layer) is the softmax output layer. W(L) and b(L) represent the connection weights and thresholds between the Lth layer to the L+1th layer, respectively, and C represents the number of categories.

[0046] Therefore, the forward propagation formula of the overall class center alignment network is as follows:

[0047] z (k) =W (k-1) a (k-1) +b (k-1) ,k=2,...,L

[0048] a (k) =f(z (k) ),k=2,...,L

[0049] Among them, the first layer is the input layer, and let x = a (1) Represents the original input features of the data, a (k) represents the output features of the kth layer of the network. f(·) represents the activation function, and the ReLU function is used here.

[0050] The L+1th layer of the network is the softmax output layer, and its formula is as follows:

[0051]

[0052] In the overall class center alignment, the last hidden layer of the network provides the features of the source domain and target domain data, so the overall class center alignment (OCA) strategy acts on the last hidden layer (layer L), and its cost function is shown in the following formula.

[0053]

[0054] Among them, a t(L) and a s(L)Represent the output features of the Lth layer of the target domain and the Lth layer of the source domain, respectively. t and N s They represent the number of samples in the target domain and the number of samples in the source domain respectively.

[0055] Combining the overall class center alignment strategy with the softmax classification error, the overall objective function of the overall class center alignment network can be obtained, as shown in the following formula.

[0056] J=J (softmax) +βJ (OCA)

[0057] Among them, β represents the weight of the overall class center alignment in the loss function.

[0058] After obtaining the predicted labels of the target domain data through overall class center alignment, we can use the true labels of the source domain data and the predicted labels of the target domain data to obtain the data of each category in the source domain and the target domain. After obtaining the data of each category in the source domain and the target domain, we can perform various class center alignment constraints on the data of each category in the source domain and the target domain in RDNN.

[0059] Various types of class center alignment strategies can achieve the distribution alignment of each category of data in the source domain and the target domain. The symbol specification and forward propagation process of this network are the same as those of the overall class center alignment network. In various types of class center alignment networks, the last hidden layer of the network provides the features of the source domain and target domain data, so various types of class center alignment strategies act on the Lth layer of the network, and its cost function is shown in the following formula.

[0060]

[0061] Among them, ajt(L) and ajs(L) represent the L-th layer output features of the j-th category data in the target domain and the L-th layer output features of the j-th category data in the source domain, respectively. Ntj and Nsj represent the number of samples of the j-th category in the target domain and the number of samples of the j-th category in the source domain, respectively.

[0062] By combining various class center alignment strategies with the softmax classification error, the overall objective function of various class center alignment networks can be obtained, as shown in the following formula.

[0063] J=J (softmax) +βJ (CCA)

[0064] Among them, β represents the weight of the alignment of each class center in the loss function.

[0065] The contents of five characteristic gases, H2, CH4, C2H2, C2H4, and C2H6, are selected as the basis for analyzing and identifying the gas state in the oil. The fault type of the transformer is determined based on the identified gas state in the oil. Since the difference in the contents of the five characteristic gases may be large, in order to make the characteristic variable values ​​more in line with the input needs of the network, all input data are normalized according to the following formula:

[0066]

[0067] In the formula, x scaled is the normalized gas content data, x is the original value of the gas content, and x max is the maximum gas content, x min is the minimum gas content.

[0068] The transformer status is divided into 7 categories: low temperature overheating, medium temperature overheating, high temperature overheating, partial discharge, low energy discharge, high energy discharge and normal. The network output is the encoding of these 7 types of status. The encoding of the transformer status type is shown in Table 1. The network structure for identifying the gas status in transformer oil is as follows: Figure 3 As shown, there are 5 nodes in the input layer, 7 nodes in the output layer, and two hidden layers.

[0069] Table 1

[0070] Serial number Transformer status type coding 1 Low temperature overheating [1 0 0 0 0 0 0] 2 Medium temperature overheating [0 1 0 0 0 0 0] 3 High temperature overheating [0 0 1 0 0 0 0] 4 Partial Discharge [0 0 0 1 0 0 0] 5 Low energy discharge [0 0 0 0 1 0 0] 6 High Energy Discharge [0 0 0 0 0 1 0] 7 normal [0 0 0 0 0 0 1]

[0071] Table 2

[0072] Serial number Fault type Training samples Test samples 1 normal 1552 1093 2 Low temperature overheating 1728 1127 3 Medium temperature overheating 1755 1209 4 High temperature overheating 1705 1121 5 Partial Discharge 1773 1245 6 Low energy discharge 1785 1117 7 High Energy Discharge 1702 1088 total 12000 8000

[0073] In order to verify the recognition effect of the deep migration network on the gas state in transformer oil, 20,000 groups of gas analysis data in transformer oil with actual fault results are selected. The corresponding 7 states are shown in Table 2, of which 12,000 groups are used as training data and 8,000 groups are used as test data. Because the class center can represent the average characteristic attribute of the data, reducing the difference between the class centers of the source domain data and the target domain data can obtain similar data distributions of the source domain and the target domain. The class center alignment effect and classification effect of CADTL are verified respectively below.

[0074] The purpose of the CADTL algorithm is to align the center of each type of data in the source domain and the target domain. The center alignment effect of some test data in four states: low temperature overheating, medium temperature overheating, high temperature overheating and high energy discharge is shown in the figure. Figure 4 As shown;

[0075] We select the 1st, 2nd, 3rd, and 4th categories of gas analysis data in transformer oil as examples. The 1st, 2nd, 3rd, and 4th categories are low-temperature overheating, medium-temperature overheating, high-temperature overheating, and high-energy discharge, respectively. For the original data before alignment, you can select any one, and display the center difference between the source domain and the target domain data, such as Figure 4 As shown in (a), (c), (e), and (g) of ; for the alignment features obtained through the network, it is necessary to select the features of the source domain and target domain data of any two nodes with non-zero feature values, and display the alignment effects of these features, as shown in Figure 4 (b), (d), (f), (h). Figure 4 (a), (c), (e), and (g) show that the center difference between the source domain data and the target domain data is large, and the alignment effect is poor. Figure 4 (b), (d), (f), and (h) show that the centers of the data features of the source and target domains after conversion are very close, achieving the purpose of aligning the centers of each category of data in the source and target domains. This is because the CADTL algorithm performs nonlinear transformation on the source and target domain data and constrains the features of the source and target domains to minimize the class center, thereby reducing the center difference between the source and target domain data.

[0076] In order to compare the performance of the deep migration network model in this paper, three methods, namely the three-ratio method, BP neural network and SVM, were used for experiments. The diagnostic accuracy of each fault type is shown in Table 3. Table 4 shows some sample values ​​in the test samples.

[0077] Table 3 shows the accuracy of the three-ratio method, BP neural network, SVM, and deep migration network in diagnosing the fault type of the test samples. It can be seen from Table 3 that the accuracy of the three-ratio method is 71.2%, the accuracy of BP is 78.3%, the accuracy of SVM is 86.8%, and the accuracy of the deep migration network is 96.5%. The deep migration network has the highest accuracy and has a higher accuracy, which can more accurately identify the gas state in the transformer insulating oil.

[0078] Table 3

[0079]

[0080]

[0081] Table 4

[0082]

[0083] The present invention also provides a system 200 for identifying the dissolved gas content in transformer oil. Figure 5 As shown, including:

[0084] The acquisition module 201 obtains multiple sets of gas content data in transformer oil under normal operating conditions and abnormal operating conditions of the transformer, pre-processes the multiple sets of gas content data, generates sample data, and sets labels for the sample data;

[0085] Initialization module 202, initializing the deep migration network parameters, and inputting the sample data with the set label and the data of the target transformer into the deep migration network to obtain the predicted label of the target transformer data;

[0086] Processing module 203, determining classification data of gas in sample data and target transformer data according to labels and predicted labels set for sample data;

[0087] The output module 204 inputs the classified data into the deep migration network to obtain the dissolved gas content in the target transformer oil.

[0088] Among them, the sample data is no less than 20,000 pieces, and the sample data are the content data of H2, CH4, C2H2, C2H4 and C2H6 in transformer oil.

[0089] Abnormal operating conditions include: transformer low-temperature overheating condition, medium-temperature overheating condition, high-temperature overheating condition, partial discharge condition, low-energy discharge condition and high-energy discharge condition.

[0090] The preprocessing is normalization.

[0091] Deep migration network parameters, including: number of hidden layers, number of hidden layer nodes, initialized network weights and thresholds.

[0092] The present invention can extract the corresponding gas content characteristics in transformer oil, identify the gas state in the oil according to different gas contents, accurately identify transformer faults, and provide excellent guarantee for the reliable operation and maintenance of power transformers.

[0093] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can adopt the form of complete hardware embodiments, complete software embodiments, or embodiments in combination with software and hardware. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code. The scheme in the embodiments of the present application can be implemented in various computer languages, for example, object-oriented programming language Java and literal scripting language JavaScript, etc.

[0094] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0095] These computer program instructions may also be stored in a computer readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture including an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0096] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0097] Although the preferred embodiments of the present application have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications falling within the scope of the present application.

[0098] Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalents, the present application is also intended to include these modifications and variations.

Claims

1. A method for identifying dissolved gas content in transformer oil, the method comprising: Acquire multiple sets of gas content data in transformer oil under normal operating conditions and abnormal operating conditions of the transformer, pre-process the multiple sets of gas content data, generate sample data, and set labels for the sample data; Initialize the deep migration network parameters, and input the sample data with set labels and the data of the target transformer into the deep migration network to obtain the predicted label of the target transformer data; According to the labels set for the sample data and the predicted labels, the classification data of the gas in the sample data and the target transformer data are determined; Input the classified data into the deep migration network to obtain the dissolved gas content in the target transformer oil; The initializing the deep migration network parameters includes: Before performing overall centroid alignment, the network parameters need to be initialized. The parameters of the deep migration network mainly include the number of hidden layers, the number of nodes in each hidden layer, the network weight and threshold. For the gas data in transformer oil, the number of hidden layers is set to 2, the number of nodes in each hidden layer is 20, and the network weight and threshold are randomly initialized. Deep Migration Network, including: Deep neural network with overall class center alignment: The input is labeled source domain data and unlabeled target domain data, and the predicted label of the target domain data is obtained through overall class center alignment; Deep neural network with various class center alignment: The data of each category in the source domain and the target domain are obtained through the real label of the source domain data and the predicted label of the target domain data. After obtaining the data of each category in the source domain and the target domain, the deep neural network with various class center alignment is used to classify the target domain data to obtain the classification result of the target domain data; The overall centroid alignment realizes the overall distribution alignment of the source domain data and the target domain data. It stipulates that the first layer is the input layer, the 2nd to Lth layers are hidden layers, and the last layer, i.e., the L+1th layer, is the softmax output layer. W(L) and b(L) represent the connection weight and threshold between the Lth layer and the L+1th layer, respectively, and C represents the number of categories. The forward propagation formula of the overall class center alignment network is as follows: z (k) =W (k-1) a (k-1) +b (k-1) ,k=2,...,L a (k) =f(z (k) ),k=2,...,L Among them, the first layer is the input layer, and let x = a (1) Represents the original input features of the data, a (k) represents the output feature of the kth layer of the network, f() represents the activation function, and the ReLU function is used here; The L+1th layer of the network is the softmax output layer, and its formula is as follows: In the overall class center alignment, the last hidden layer of the network provides the features of the source domain and the target domain data, so the overall class center alignment OCA strategy acts on the last hidden layer and the Lth layer, and its cost function is shown in the following formula; Among them, a t(L) and a s(L) Represent the output features of the Lth layer of the target domain and the Lth layer of the source domain, respectively. t and N s Represent the number of samples in the target domain and the number of samples in the source domain respectively; Combining the overall class center alignment strategy with the softmax classification error, the overall objective function of the overall class center alignment network can be obtained as shown in the following formula; J=J (softmax) +βJ (OCA) Among them, β represents the weight of the overall class center alignment in the loss function; After obtaining the predicted labels of the target domain data through overall class center alignment, the real labels of the source domain data and the predicted labels of the target domain data are used to obtain the data of each category in the source domain and the target domain. After obtaining the data of each category in the source domain and the target domain, various class center alignment constraints are applied to the data of each category in the source domain and the target domain in the RDNN. Various types of class center alignment strategies realize the distribution alignment of each category of data in the source domain and the target domain. The symbol specification and forward propagation process of the network are the same as those of the overall class center alignment network. In various types of class center alignment networks, the last hidden layer of the network provides the features of the source domain and target domain data. Various types of class center alignment strategies act on the Lth layer of the network, and their cost functions are shown in the following formula; Among them, ajt(L) and ajs(L) represent the L-th layer output features of the j-th category data in the target domain and the L-th layer output features of the j-th category data in the source domain, respectively. Ntj and Nsj represent the number of samples of the j-th category in the target domain and the number of samples of the j-th category in the source domain, respectively. By combining various class center alignment strategies with the softmax classification error, the overall objective function of various class center alignment networks can be obtained, as shown in the following formula; J=J (softmax) +βJ (CCA) Among them, β represents the weight of the alignment of each class center in the loss function.

2. According to the method of claim 1, the sample data is not less than 20,000 copies, and the sample data is the content data of H2, CH4, C2H2, C2H4 and C2H6 in transformer oil.

3. The method according to claim 1, wherein the abnormal operating state comprises: Transformer low temperature overheating state, medium temperature overheating state, high temperature overheating state, partial discharge state, low energy discharge state and high energy discharge state. The method according to claim 1 , wherein the preprocessing is a normalization process.

5. A system for identifying dissolved gas content in transformer oil, the system comprising: The acquisition module obtains multiple sets of gas content data in the transformer oil under normal operating conditions and abnormal operating conditions, pre-processes the multiple sets of gas content data, generates sample data, and sets labels for the sample data; Initialization module, initializes the deep migration network parameters, and inputs the sample data with set labels and the data of the target transformer into the deep migration network to obtain the predicted label of the target transformer data; A processing module determines classification data of gases in the sample data and target transformer data according to labels set for the sample data and predicted labels; The output module inputs the classified data into the deep migration network to obtain the dissolved gas content in the target transformer oil; The initializing the deep migration network parameters includes: Before performing overall centroid alignment, the network parameters need to be initialized. The parameters of the deep migration network mainly include the number of hidden layers, the number of nodes in each hidden layer, the network weight and threshold. For the gas data in transformer oil, the number of hidden layers is set to 2, the number of nodes in each hidden layer is 20, and the network weight and threshold are randomly initialized. Deep Migration Network, including: Deep neural network with overall class center alignment: The input is labeled source domain data and unlabeled target domain data, and the predicted label of the target domain data is obtained through overall class center alignment; Deep neural network with various class center alignment: The data of each category in the source domain and the target domain are obtained through the real label of the source domain data and the predicted label of the target domain data. After obtaining the data of each category in the source domain and the target domain, the deep neural network with various class center alignment is used to classify the target domain data to obtain the classification result of the target domain data; The overall centroid alignment realizes the overall distribution alignment of the source domain data and the target domain data. It stipulates that the first layer is the input layer, the 2nd to Lth layers are hidden layers, and the last layer, i.e., the L+1th layer, is the softmax output layer. W(L) and b(L) represent the connection weight and threshold between the Lth layer and the L+1th layer, respectively, and C represents the number of categories. The forward propagation formula of the overall class center alignment network is as follows: z (k) =W (k-1) a (k-1) +b (k-1) ,k=2,...,L a (k) =f(z (k) ),k=2,...,L Among them, the first layer is the input layer, and let x = a (1) Represents the original input features of the data, a (k) represents the output feature of the kth layer of the network, f() represents the activation function, and the ReLU function is used here; The L+1th layer of the network is the softmax output layer, and its formula is as follows: In the overall class center alignment, the last hidden layer of the network provides the features of the source domain and the target domain data, so the overall class center alignment OCA strategy acts on the last hidden layer and the Lth layer, and its cost function is shown in the following formula; Among them, a t(L) and a s(L) Represent the output features of the Lth layer of the target domain and the Lth layer of the source domain, respectively. t and N s Represent the number of samples in the target domain and the number of samples in the source domain respectively; Combining the overall class center alignment strategy with the softmax classification error, the overall objective function of the overall class center alignment network can be obtained as shown in the following formula; J=J (softmax) +βJ (OCA) Among them, β represents the weight of the overall class center alignment in the loss function; After obtaining the predicted labels of the target domain data through overall class center alignment, the real labels of the source domain data and the predicted labels of the target domain data are used to obtain the data of each category in the source domain and the target domain. After obtaining the data of each category in the source domain and the target domain, various class center alignment constraints are applied to the data of each category in the source domain and the target domain in the RDNN. Various types of class center alignment strategies realize the distribution alignment of each category of data in the source domain and the target domain. The symbol specification and forward propagation process of the network are the same as those of the overall class center alignment network. In various types of class center alignment networks, the last hidden layer of the network provides the features of the source domain and target domain data. Various types of class center alignment strategies act on the Lth layer of the network, and their cost functions are shown in the following formula; Among them, ajt(L) and ajs(L) represent the L-th layer output features of the j-th category data in the target domain and the L-th layer output features of the j-th category data in the source domain, respectively. Ntj and Nsj represent the number of samples of the j-th category in the target domain and the number of samples of the j-th category in the source domain, respectively. By combining various class center alignment strategies with the softmax classification error, the overall objective function of various class center alignment networks can be obtained, as shown in the following formula; J=J (softmax) +βJ (CCA) Among them, β represents the weight of the alignment of each class center in the loss function.

6. The system according to claim 5, wherein the sample data is no less than 20,000 copies, and the sample data is the content data of H2, CH4, C2H2, C2H4 and C2H6 in transformer oil.

7. The system according to claim 5, wherein the abnormal operating state comprises: Transformer low temperature overheating state, medium temperature overheating state, high temperature overheating state, partial discharge state, low energy discharge state and high energy discharge state. The system according to claim 5 , wherein the preprocessing is a normalization process.

Images