Direct-current converter transformer implicit fault diagnosis method and system based on digital twinning

Abstract

The application discloses a DC converter transformer implicit fault diagnosis method and system based on digital twinning, and the method comprises the following steps: collecting the operation state data of an extra-high voltage DC converter transformer; using the operation state data to establish a converter transformer digital twinning model in a virtual space and performing real-time updating on the model; constructing an information system S through the operation state data and the digital twinning model data, and obtaining a characteristic parameter set N of the information system S by using a data processing method of a rough set theory; constructing a deep optimization network through local sparse design and training the network; inputting the real-time collected characteristic parameter set N into the trained deep optimization network to obtain a fault diagnosis result; and the application has the advantages that the efficiency and accuracy of fault diagnosis are improved.

Description

Technical Field

[0001] This invention relates to the field of transformer fault diagnosis, specifically to a method and system for diagnosing latent faults in DC converter transformers based on digital twins. Background Technology

[0002] In ultra-high voltage direct current (UHVDC) transmission systems, converter transformers play a crucial role, serving not only as a "bridge" connecting the DC and AC systems but also as the safety guardians of the converter units within the DC converter station. Due to their vital position and high cost in power transmission, it is difficult to conduct relevant tests on UHVDC converter transformers. To ensure their normal operation and timely and accurate fault diagnosis, companies have equipped them with multiple protection systems. However, due to inherent limitations, the failure rate remains high.

[0003] Latent faults are a critical type of fault in power companies. Their most significant characteristic is their difficulty in detection, leading to secondary losses and potentially causing serious personal injury and equipment damage. Therefore, their identification and prevention are extremely important. As latent functions are invisible to operators, even highly skilled personnel cannot detect their proper operation when they are not needed. Latent faults remain undetected unless an accident occurs; this is their most obvious characteristic and the key to identifying them. Based on digital twin technology, the physical body of the converter transformer can be mapped to a digital twin. By combining this with deployed sensors, real-time data is collected and updated to ensure consistency between the physical entity and the digital twin. The digital twin reflects the operating status of the converter transformer, facilitating the diagnosis of latent faults. Therefore, a method and system for diagnosing latent faults in UHVDC converter transformers based on digital twins is needed.

[0004] Chinese Patent Publication No. CN112684379A discloses a transformer fault diagnosis system and method based on digital twins, comprising a physical system module, a digital twin module, and a fault diagnosis module. The method specifically includes the following steps: using the physical system module to create a three-dimensional model of the transformer and acquire transformer operating status data collected by sensors; using the digital twin module to create a digital twin model of the transformer, generate simulation data, and calibrate the digital twin model; using the fault diagnosis module to perform fault diagnosis on the transformer, combining the physical entity of the transformer with the virtual model, correcting the digital twin model based on the status monitoring data obtained from sensors and the simulation data of the digital twin model, extracting feature parameters, and using a BP neural network algorithm to diagnose the type of fault and analyze the possible causes of the fault. However, the BP neural network algorithm has a large number of parameters, resulting in long network fitting time and low fitting accuracy, thus leading to low network training efficiency and poor performance, affecting the efficiency and accuracy of fault diagnosis. Summary of the Invention

[0005] The technical problem to be solved by this invention is that the existing method for diagnosing latent faults in DC converter transformers based on digital twins uses a BP neural network algorithm for fault diagnosis. This algorithm has a large number of parameters, resulting in long network fitting time and low fitting accuracy, which affects the efficiency and accuracy of fault diagnosis.

[0006] This invention solves the above-mentioned technical problems through the following technical means: a method for diagnosing latent faults in DC converter transformers based on digital twins, comprising the following steps:

[0007] Step 1: Collect operating status data of the UHVDC converter transformer;

[0008] Step 2: Build a digital twin model of the converter transformer in virtual space using operational status data and update the model in real time;

[0009] Step 3: Construct information system S using operational status data and digital twin model data, and obtain the feature parameter set N of information system S using data processing methods based on rough set theory;

[0010] Step 4: Construct a deep optimization network through local sparse design and train the network;

[0011] Step 5: Input the real-time collected feature parameter set N into the trained deep optimization network to obtain the fault diagnosis results.

[0012] Furthermore, step two includes:

[0013] Step 2.1: Construct a three-dimensional model of the UHVDC converter transformer using 3ds Max software based on the operating status data;

[0014] Step 2.2: Load the constructed 3D model of the UHVDC converter transformer into Unity3D software to build a digital twin model of the converter transformer.

[0015] Furthermore, step two involves updating the model in real time, including:

[0016] Step 2.3: Add virtual sensors to the digital twin model of the converter transformer to obtain the operating status data of the digital twin model of the converter transformer in the virtual environment, i.e., simulation data;

[0017] Step 2.4: Compare the simulation data of the obtained converter transformer digital twin model with the operating status data of the physical entity to determine the similarity between the converter transformer digital twin model and the physical entity, and continuously adjust the parameters of the converter transformer digital twin model to meet the preset accuracy requirements.

[0018] Furthermore, the data processing method of rough set theory includes data cleaning, data integration, and data transformation. Data cleaning involves detecting and processing noise and outliers in the data; data integration involves merging and integrating data from different data sources and eliminating redundancy; and data transformation involves normalizing, discretizing, or reducing the dimensionality of the data to convert it into the required data format.

[0019] Furthermore, the deep optimization network includes a first convolutional layer, a first max pooling layer, a second convolutional layer, a second max pooling layer, a third convolutional layer, a third max pooling layer, a local sparse network A, a local sparse network B, an average pooling layer, and a classifier, all connected in sequence. The local sparse network A includes a fourth max pooling layer, a fourth convolutional layer, a fifth convolutional layer, and a sixth convolutional layer. The inputs of the fourth max pooling layer, the fifth convolutional layer, and the sixth convolutional layer are all connected to the third max pooling layer. The output of the fourth max pooling layer is connected to the input of the fourth convolutional layer. The outputs of the fourth to sixth convolutional layers are superimposed to output a first superposition result. The local sparse network B includes a seventh convolutional layer and an eighth convolutional layer. The inputs of the seventh and eighth convolutional layers both receive the first superposition result. The outputs of the seventh and eighth convolutional layers are superimposed to output a second superposition result. The second superposition result is input to the average pooling layer.

[0020] Furthermore, the kernel size of the first convolutional layer is 64×1, the kernel size of the second convolutional layer is 3×1, the kernel size of the third convolutional layer is 3×1, the kernel size of the fourth convolutional layer is 1×1, the kernel size of the fifth convolutional layer is 3×1, the kernel size of the sixth convolutional layer is 1×1, the kernel size of the seventh convolutional layer is 3×1, and the kernel size of the eighth convolutional layer is 1×1.

[0021] Furthermore, the process of training the deep optimization network in step four is as follows:

[0022] For the designed deep optimization network, based on the test data, the network parameters are initialized, the network output value is calculated through the backpropagation algorithm, and the error between the entire network output value and the true value is compared. The network parameters are then adjusted to gradually reduce the error. A fixed batch size of test data is continuously fed into the network for training until all test data is used up or the error is minimized, at which point training stops. Finally, the trained deep optimization network is validated using test samples.

[0023] This invention also provides a latent fault diagnosis system for DC converter transformers based on digital twins, comprising:

[0024] The data acquisition unit is used to collect operating status data of the UHVDC converter transformer;

[0025] The twin model construction unit is used to build a digital twin model of the converter transformer in virtual space using operating status data and to update the model in real time.

[0026] The data processing unit is used to construct the information system S using operational status data and digital twin model data, and to obtain the feature parameter set N of the information system S using data processing methods based on rough set theory.

[0027] Neural network building blocks are used to construct and train deep optimized networks through local sparse design.

[0028] The fault diagnosis unit is used to input the real-time acquired feature parameter set N into the trained deep optimization network to obtain the fault diagnosis result.

[0029] Furthermore, the twin model building unit is also used for:

[0030] Step 2.1: Construct a three-dimensional model of the UHVDC converter transformer using 3ds Max software based on the operating status data;

[0031] Step 2.2: Load the constructed 3D model of the UHVDC converter transformer into Unity3D software to build a digital twin model of the converter transformer.

[0032] Furthermore, the twin model construction unit updates the model in real time, including:

[0033] Step 2.3: Add virtual sensors to the digital twin model of the converter transformer to obtain the operating status data of the digital twin model of the converter transformer in the virtual environment, i.e., simulation data;

[0034] Step 2.4: Compare the simulation data of the obtained converter transformer digital twin model with the operating status data of the physical entity to determine the similarity between the converter transformer digital twin model and the physical entity, and continuously adjust the parameters of the converter transformer digital twin model to meet the preset accuracy requirements.

[0035] Furthermore, the data processing method of rough set theory includes data cleaning, data integration, and data transformation. Data cleaning involves detecting and processing noise and outliers in the data; data integration involves merging and integrating data from different data sources and eliminating redundancy; and data transformation involves normalizing, discretizing, or reducing the dimensionality of the data to convert it into the required data format.

[0036] Furthermore, the deep optimization network includes a first convolutional layer, a first max pooling layer, a second convolutional layer, a second max pooling layer, a third convolutional layer, a third max pooling layer, a local sparse network A, a local sparse network B, an average pooling layer, and a classifier, all connected in sequence. The local sparse network A includes a fourth max pooling layer, a fourth convolutional layer, a fifth convolutional layer, and a sixth convolutional layer. The inputs of the fourth max pooling layer, the fifth convolutional layer, and the sixth convolutional layer are all connected to the third max pooling layer. The output of the fourth max pooling layer is connected to the input of the fourth convolutional layer. The outputs of the fourth to sixth convolutional layers are superimposed to output a first superposition result. The local sparse network B includes a seventh convolutional layer and an eighth convolutional layer. The inputs of the seventh and eighth convolutional layers both receive the first superposition result. The outputs of the seventh and eighth convolutional layers are superimposed to output a second superposition result. The second superposition result is input to the average pooling layer.

[0037] Furthermore, the kernel size of the first convolutional layer is 64×1, the kernel size of the second convolutional layer is 3×1, the kernel size of the third convolutional layer is 3×1, the kernel size of the fourth convolutional layer is 1×1, the kernel size of the fifth convolutional layer is 3×1, the kernel size of the sixth convolutional layer is 1×1, the kernel size of the seventh convolutional layer is 3×1, and the kernel size of the eighth convolutional layer is 1×1.

[0038] Furthermore, the process of training the deep optimization network in the neural network building unit is as follows:

[0039] For the designed deep optimization network, based on the test data, the network parameters are initialized, the network output value is calculated through the backpropagation algorithm, and the error between the entire network output value and the true value is compared. The network parameters are then adjusted to gradually reduce the error. A fixed batch size of test data is continuously fed into the network for training until all test data is used up or the error is minimized, at which point training stops. Finally, the trained deep optimization network is validated using test samples.

[0040] The advantages of this invention are:

[0041] (1) This invention constructs a deep optimization network based on a deep convolutional neural network through local sparse design, which greatly reduces the number of parameters in the network, thereby reducing the risk of network overfitting, reducing fitting time, and improving fitting accuracy, thus improving the training efficiency and training effect of the network, and improving the efficiency and accuracy of fault diagnosis.

[0042] (2) This invention utilizes the digital twin method to effectively diagnose hidden faults in UHVDC converter transformers, reducing the personnel handling pressure required for fault diagnosis while achieving both economy and safety, thereby enabling rapid handling of faults in UHVDC converter stations. Attached Figure Description

[0043] Figure 1 This is a flowchart of the method for diagnosing latent faults in a DC converter transformer based on digital twins, as disclosed in an embodiment of the present invention.

[0044] Figure 2 This is a schematic diagram of the network architecture in the DC converter transformer latent fault diagnosis method based on digital twin disclosed in this embodiment of the invention, wherein... Figure 2 (a) is a convolutional neural network without local sparse design. Figure 2 (b) is a deep optimization network constructed through local sparse design;

[0045] Figure 3 This is a schematic diagram of a local sparse structure in the digital twin-based method for diagnosing latent faults in DC converter transformers disclosed in this invention. Figure 3 (a) Figure 3 (b) are schematic diagrams of the local sparse structures of local sparse network A and local sparse network B, respectively. Detailed Implementation

[0046] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0047] Example 1

[0048] like Figure 1 As shown, this invention provides a method for diagnosing latent faults in DC converter transformers based on digital twins, comprising the following steps:

[0049] S1. Collect operating status data of the UHVDC converter transformer by sensors deployed on the converter transformer body;

[0050] S2. A digital twin model of the converter transformer is established in virtual space using operational status data, and the digital twin is updated in real time by acquiring data from the physical entity to meet reliability requirements; the specific process is as follows:

[0051] Step 2.1: Construct a three-dimensional model of the UHVDC converter transformer using 3ds Max software based on the operating status data;

[0052] Step 2.2: Load the constructed 3D model of the UHVDC converter transformer into Unity3D software to build a digital twin model of the converter transformer.

[0053] Step 2.3: Add virtual sensors to the digital twin model of the converter transformer to obtain the operating status data of the digital twin model of the converter transformer in the virtual environment, i.e., simulation data;

[0054] Step 2.4: Compare the simulation data of the obtained converter transformer digital twin model with the operating status data of the physical entity to determine the similarity between the converter transformer digital twin model and the physical entity, and continuously adjust the parameters of the converter transformer digital twin model to meet the preset accuracy requirements.

[0055] S3. Construct an information system S using operational status data and digital twin model data. Use rough set theory to process data and obtain the feature parameter set N of the information system S. The main function of rough set is to retain the "useful" feature parameters and delete the "useless" feature parameters that affect the diagnostic accuracy by using attribute reduction methods while ensuring classification ability. Finally, obtain the feature parameter set that is most sensitive to decision-making and has the smallest dimension.

[0056] In practical applications, the data processing method of rough set theory includes data cleaning, data integration, and data transformation. Data cleaning involves detecting and processing noise and outliers in the data; data integration involves merging and integrating data from different data sources and eliminating redundancy; and data transformation involves normalizing, discretizing, or reducing the dimensionality of the data to convert it into the required data format.

[0057] It should be noted that rough set theory is a current technology, and the general process of rough set theory in processing data is as follows:

[0058] Step 3.1: Construct an information system, i.e. a knowledge representation method, using the data collected from the UHVDC converter transformer, as follows: S=(U,A,V,f)

[0059] Where U represents a non-empty finite set of objects, i.e., the object set; A represents a non-empty finite set of attributes; and f: U→A→V represents an information function of an object attribute value.

[0060] Step 3.2: In the rough set constructed from the converter transformer data, the information system contains redundancy, with these pieces of information having varying degrees of importance. This redundancy needs to be reduced. A feature parameter set N for the information system S is constructed using rough set reduction methods, and N is a non-empty set.

[0061] S4. Construct a deep optimization network through local sparse design and train the network; the specific process is as follows:

[0062] Firstly, as Figure 2 (a) is a convolutional neural network without local sparse design. Figure 2 (b) is a deep optimization network constructed through local sparse design. The input is the feature parameter set N, BN is the batch normalization layer, Conv is the convolutional layer, Max is the max pooling layer, ⊕ is the stacking operation on the feature dimension, Avg is the average pooling layer, the number in each convolutional layer represents the size of the convolutional kernel, and the number below each network layer represents the size of the output: length × width × depth. The output size of each batch normalization layer BN is the same as that of its corresponding convolutional layer.

[0063] The number of parameters in the first and second convolutional layers of a convolutional neural network without local sparse design is:

[0064] N Conυ1 =K m ×K n ×C in ×C out =64×1×1×16=1024

[0065] N Conυ2 =K m ×K n ×C in ×C out =3×1×16×32=1536

[0066] The parameter values ​​of the fourth and fifth convolutional layers and the fully connected layers are:

[0067] N Conυ4 =K m ×K n ×C in ×C out =3×1×64×64=12288

[0068] N conυ5 =K m ×K n ×C in×C out 3 × 1 × 64 × 64 = 12288

[0069] N FC =C in ×C out = (2 × 1 × 64) × 100 = 12800

[0070] Among them, K m K represents the height of the convolution kernel. n C represents the width of the convolution kernel. in C represents the number of input channels. out This represents the number of output channels. By comparing the calculated parameter values, it can be observed that in convolutional neural networks without local sparse design, the number of parameters in the network continuously increases with network depth, N... Conυ5 The number of parameters is N Conυ1 10 times.

[0071] After local sparse design, such as Figure 2 As shown in (b), the deep optimization network includes a first convolutional layer, a first max pooling layer, a second convolutional layer, a second max pooling layer, a third convolutional layer, a third max pooling layer, a local sparse network A, a local sparse network B, an average pooling layer, and a classifier, all connected in sequence. The local sparse network A includes a fourth max pooling layer, a fourth convolutional layer, a fifth convolutional layer, and a sixth convolutional layer. The inputs of the fourth max pooling layer, the fifth convolutional layer, and the sixth convolutional layer are all connected to the third max pooling layer. The output of the fourth max pooling layer is connected to the input of the fourth convolutional layer. The outputs of the fourth to sixth convolutional layers are superimposed to output a first superimposed result. The local sparse network B includes a seventh convolutional layer and an eighth convolutional layer. The inputs of the seventh and eighth convolutional layers both receive the first superimposed result. The outputs of the seventh and eighth convolutional layers are superimposed to output a second superimposed result. The second superimposed result is input to the average pooling layer.

[0072] The kernel size of the first convolutional layer is 64×1, the kernel size of the second convolutional layer is 3×1, the kernel size of the third convolutional layer is 3×1, the kernel size of the fourth convolutional layer is 1×1, the kernel size of the fifth convolutional layer is 3×1, the kernel size of the sixth convolutional layer is 1×1, the kernel size of the seventh convolutional layer is 3×1, and the kernel size of the eighth convolutional layer is 1×1.

[0073] Calculate the number of parameters for local sparse network A and local sparse network B after local sparse design:

[0074]

[0075]

[0076] By using local sparse design, a large convolutional layer is broken down into a combination of multiple smaller convolutional layers, compared to... Figure 2 (a) A convolutional neural network without local sparse design. This design significantly reduces the number of parameters in the network; and the number of parameters is further reduced by replacing the fully connected layers in the non-locally sparse convolutional neural network with parameterless pooling layers. Based on Hebbian theory, neural networks can be designed as locally sparse structures. To match the output feature map size, the stride of the convolutional layers in locally sparse network B is designed to be 2. Schematic diagrams of the locally sparse structures of locally sparse networks A and B are shown below. Figure 3 (a) and Figure 3 (b) Description, in which Figure 3 (a) represents the dimensionality reduction style, where the output feature map size is half that of the input. Figure 3 (b) is the general style, where the feature map size is the same after input and output.

[0077] For the designed deep optimization network, based on the test data, the network parameters are initialized, the network output value is calculated through the backpropagation algorithm, and the error between the entire network output value and the true value is compared. The network parameters are then adjusted to gradually reduce the error. A fixed batch size of test data is continuously fed into the network for training until all test data is used up or the error is minimized, at which point training stops. Finally, the trained deep optimization network is validated using test samples.

[0078] S5. Input the real-time collected feature parameter set N into the trained deep optimization network. After the feature parameter set N is input into the network, it first goes through three convolutional layers to further extract the features. The size of the convolutional kernel is represented by the numbers in the convolutional layer. Then, it is further processed and integrated through local sparse network A and local sparse network B with local sparse structure. Finally, the fault diagnosis results of UHVDC converter transformer are output by the average pooling layer and the Softmax layer.

[0079] Through the above technical solutions, this invention constructs a deep optimization network based on a deep convolutional neural network by using local sparse design, which significantly reduces the number of parameters in the network, thereby reducing the risk of overfitting, reducing fitting time, and improving fitting accuracy. This improves the training efficiency and effect of the network, and enhances the efficiency and accuracy of fault diagnosis.

[0080] Example 2

[0081] Based on Embodiment 1, Embodiment 2 of the present invention also provides a DC converter transformer latent fault diagnosis system based on digital twin, including:

[0082] The data acquisition unit is used to collect operating status data of the UHVDC converter transformer;

[0083] The twin model construction unit is used to build a digital twin model of the converter transformer in virtual space using operating status data and to update the model in real time.

[0084] The data processing unit is used to construct the information system S using operational status data and digital twin model data, and to obtain the feature parameter set N of the information system S using data processing methods based on rough set theory.

[0085] Neural network building blocks are used to construct and train deep optimized networks through local sparse design.

[0086] The fault diagnosis unit is used to input the real-time acquired feature parameter set N into the trained deep optimization network to obtain the fault diagnosis result.

[0087] Specifically, the twin model building unit is also used for:

[0088] Step 2.1: Construct a three-dimensional model of the UHVDC converter transformer using 3ds Max software based on the operating status data;

[0089] Step 2.2: Load the constructed 3D model of the UHVDC converter transformer into Unity3D software to build a digital twin model of the converter transformer.

[0090] More specifically, the twin model construction unit updates the model in real time, including:

[0091] Step 2.3: Add virtual sensors to the digital twin model of the converter transformer to obtain the operating status data of the digital twin model of the converter transformer in the virtual environment, i.e., simulation data;

[0092] Step 2.4: Compare the simulation data of the obtained converter transformer digital twin model with the operating status data of the physical entity to determine the similarity between the converter transformer digital twin model and the physical entity, and continuously adjust the parameters of the converter transformer digital twin model to meet the preset accuracy requirements.

[0093] Specifically, the data processing method of rough set theory includes data cleaning, data integration, and data transformation. Data cleaning involves detecting and processing noise and outliers in the data. Data integration involves merging and integrating data from different data sources and eliminating redundancy. Data transformation involves normalizing, discretizing, or reducing the dimensionality of the data to convert it into the required data format.

[0094] Specifically, the deep optimization network includes a first convolutional layer, a first max pooling layer, a second convolutional layer, a second max pooling layer, a third convolutional layer, a third max pooling layer, a local sparse network A, a local sparse network B, an average pooling layer, and a classifier, all connected in sequence. The local sparse network A includes a fourth max pooling layer, a fourth convolutional layer, a fifth convolutional layer, and a sixth convolutional layer. The inputs of the fourth max pooling layer, the fifth convolutional layer, and the sixth convolutional layer are all connected to the third max pooling layer. The output of the fourth max pooling layer is connected to the input of the fourth convolutional layer. The outputs of the fourth to sixth convolutional layers are stacked to output a first stacking result. The local sparse network B includes a seventh convolutional layer and an eighth convolutional layer. The inputs of the seventh and eighth convolutional layers both receive the first stacking result. The outputs of the seventh and eighth convolutional layers are stacked to output a second stacking result. The second stacking result is input to the average pooling layer.

[0095] More specifically, the kernel size of the first convolutional layer is 64×1, the kernel size of the second convolutional layer is 3×1, the kernel size of the third convolutional layer is 3×1, the kernel size of the fourth convolutional layer is 1×1, the kernel size of the fifth convolutional layer is 3×1, the kernel size of the sixth convolutional layer is 1×1, the kernel size of the seventh convolutional layer is 3×1, and the kernel size of the eighth convolutional layer is 1×1.

[0096] Specifically, the process of training the deep optimization network in the neural network construction unit is as follows:

[0097] For the designed deep optimization network, based on the test data, the network parameters are initialized, the network output value is calculated through the backpropagation algorithm, and the error between the entire network output value and the true value is compared. The network parameters are then adjusted to gradually reduce the error. A fixed batch size of test data is continuously fed into the network for training until all test data is used up or the error is minimized, at which point training stops. Finally, the trained deep optimization network is validated using test samples.

[0098] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for diagnosing latent faults in DC converter transformers based on digital twins, characterized in that, Includes the following steps: Step 1: Collect operating status data of the UHVDC converter transformer; Step 2: Build a digital twin model of the converter transformer in virtual space using operational status data and update the model in real time; Step 3: Construct information system S using operational status data and digital twin model data, and obtain the feature parameter set N of information system S using data processing methods based on rough set theory; Step 4: Construct a deep optimization network through local sparse design and train the network; Step 5: Input the real-time collected feature parameter set N into the trained deep optimization network to obtain the fault diagnosis results.

2. The method for diagnosing latent faults in DC converter transformers based on digital twins according to claim 1, characterized in that, Step two includes: Step 2.1: Construct a three-dimensional model of the UHVDC converter transformer using 3ds Max software based on the operating status data; Step 2.2: Load the constructed 3D model of the UHVDC converter transformer into Unity3D software to build a digital twin model of the converter transformer.

3. The method for diagnosing latent faults in DC converter transformers based on digital twins according to claim 2, characterized in that, Step two involves updating the model in real time, including: Step 2.3: Add virtual sensors to the digital twin model of the converter transformer to obtain the operating status data of the digital twin model of the converter transformer in the virtual environment, i.e., simulation data; Step 2.4: Compare the simulation data of the obtained converter transformer digital twin model with the operating status data of the physical entity to determine the similarity between the converter transformer digital twin model and the physical entity, and continuously adjust the parameters of the converter transformer digital twin model to meet the preset accuracy requirements.

4. The method for diagnosing latent faults in DC converter transformers based on digital twins according to claim 1, characterized in that, The data processing method of rough set theory includes data cleaning, data integration, and data transformation. Data cleaning involves detecting and processing noise and outliers in the data. Data integration involves merging and integrating data from different data sources and eliminating redundancy. Data transformation involves normalizing, discretizing, or reducing the dimensionality of the data to convert it into the required data format.

5. The method for diagnosing latent faults in DC converter transformers based on digital twins according to claim 1, characterized in that, The deep optimization network includes a first convolutional layer, a first max pooling layer, a second convolutional layer, a second max pooling layer, a third convolutional layer, a third max pooling layer, a local sparse network A, a local sparse network B, an average pooling layer, and a classifier, all connected in sequence. The local sparse network A includes a fourth max pooling layer, a fourth convolutional layer, a fifth convolutional layer, and a sixth convolutional layer. The inputs of the fourth max pooling layer, the fifth convolutional layer, and the sixth convolutional layer are all connected to the third max pooling layer. The output of the fourth max pooling layer is connected to the input of the fourth convolutional layer. The outputs of the fourth to sixth convolutional layers are stacked to output a first stacking result. The local sparse network B includes a seventh convolutional layer and an eighth convolutional layer. The inputs of the seventh and eighth convolutional layers both receive the first stacking result. The outputs of the seventh and eighth convolutional layers are stacked to output a second stacking result. The second stacking result is input to the average pooling layer.

6. The method for diagnosing latent faults in DC converter transformers based on digital twins according to claim 5, characterized in that, The kernel size of the first convolutional layer is 64×1, the kernel size of the second convolutional layer is 3×1, the kernel size of the third convolutional layer is 3×1, the kernel size of the fourth convolutional layer is 1×1, the kernel size of the fifth convolutional layer is 3×1, the kernel size of the sixth convolutional layer is 1×1, the kernel size of the seventh convolutional layer is 3×1, and the kernel size of the eighth convolutional layer is 1×1.

7. The method for diagnosing latent faults in DC converter transformers based on digital twins according to claim 1, characterized in that, The process of training the deep optimization network in step four is as follows: For the designed deep optimization network, based on the test data, the network parameters are initialized, the network output value is calculated through the backpropagation algorithm, and the error between the entire network output value and the true value is compared. The network parameters are then adjusted to gradually reduce the error. A fixed batch size of test data is continuously fed into the network for training until all test data is used up or the error is minimized, at which point training stops. Finally, the trained deep optimization network is validated using test samples.

8. A latent fault diagnosis system for DC converter transformers based on digital twins, characterized in that, include: The data acquisition unit is used to collect operating status data of the UHVDC converter transformer; The twin model construction unit is used to build a digital twin model of the converter transformer in virtual space using operating status data and to update the model in real time. The data processing unit is used to construct the information system S using operational status data and digital twin model data, and to obtain the feature parameter set N of the information system S using data processing methods based on rough set theory. Neural network building blocks are used to construct and train deep optimized networks through local sparse design. The fault diagnosis unit is used to input the real-time acquired feature parameter set N into the trained deep optimization network to obtain the fault diagnosis result.

9. The DC converter transformer latent fault diagnosis system based on digital twin as described in claim 8, characterized in that, The twin model building unit is also used for: Step 2.1: Construct a three-dimensional model of the UHVDC converter transformer using 3ds Max software based on the operating status data; Step 2.2: Load the constructed 3D model of the UHVDC converter transformer into Unity3D software to build a digital twin model of the converter transformer.

10. The DC converter transformer latent fault diagnosis system based on digital twin according to claim 9, characterized in that, The twin model construction unit updates the model in real time, including: Step 2.3: Add virtual sensors to the digital twin model of the converter transformer to obtain the operating status data of the digital twin model of the converter transformer in the virtual environment, i.e., simulation data; Step 2.4: Compare the simulation data of the obtained converter transformer digital twin model with the operating status data of the physical entity to determine the similarity between the converter transformer digital twin model and the physical entity, and continuously adjust the parameters of the converter transformer digital twin model to meet the preset accuracy requirements.

Images