Extra-high voltage converter area fault identification method based on improved MobileNetV2 network

Abstract

The invention discloses an extra-high voltage converter area fault identification method and system based on an improved MobileNetV2 network, and a medium, and the method comprises the steps: carrying out the modeling simulation of an extra-high voltage DC power transmission system, setting different fault conditions of a converter area, and obtaining an original fault signal; performing wavelet decomposition on the original fault signal, selecting a decomposed signal having the maximum correlation with the original fault signal to construct a feature data set, and dividing the feature data set into a training set and a test set; inputting the training set into the improved MobileNetV2 model to train a neural network model, and learning fault features; when the training frequency of the current neural network model reaches a set effect, the training of the improved MobileNetV2 model is completed; and inputting the test set into the trained improved MobileNetV2 model to carry out fault identification so as to obtain an extra-high voltage converter area fault identification result. The fault identification accuracy and identification efficiency of the converter area can be improved.

Description

Technical Field

[0001] This invention relates to the field of power system fault identification technology, specifically to a method, system, and medium for fault identification in ultra-high voltage converter areas based on an improved MobileNetV2 network. Background Technology

[0002] Current methods for fault identification in converter areas primarily rely on criterion-based approaches to determine fault conditions. These methods require constructing action criteria using electrical information collected after a fault occurs. While offering advantages for rapid fault identification, they also suffer from limitations such as incomplete consideration of fault types and numerous, complex criterion conditions. A better solution to address the issues of numerous, complex criterion conditions and comprehensive fault identification would significantly improve fault repair efficiency and reduce power outage losses. Therefore, researching more comprehensive and accurate fault identification methods for converter areas has significant theoretical and engineering implications.

[0003] Currently, due to the continuous optimization of power flow layout, a cross-provincial and cross-regional power transmission pattern of "West-to-East Power Transmission" has been formed, and ultra-high voltage direct current (UHVDC) transmission projects have also been vigorously developed. The interconnection of AC and DC at different voltage levels has continuously increased the complexity of the power system. With the rapid development of smart grids and communication technologies, the efficient transmission of massive amounts of data not only helps in the training and optimization of artificial intelligence models, but also lays the foundation for data-driven fault diagnosis technologies.

[0004] Existing technologies use SVM for fault feature extraction, but SVM is typically applied to binary classification problems. For classification problems involving multiple faults, manual threshold setting is required. Existing fault identification methods for converter areas suffer from complex criteria, excessive computation, and insufficient comprehensiveness in fault identification. Summary of the Invention

[0005] The purpose of this invention is to provide a method, system, and medium for fault identification in ultra-high voltage converter areas based on an improved MobileNetV2 network, which solves the problems of complex criteria, numerous calculations, and incomplete fault identification in existing converter area fault diagnosis methods.

[0006] This invention is achieved through the following technical solution: In a first aspect, the first embodiment of the present invention provides a method for fault identification in an ultra-high voltage converter area based on an improved MobileNetV2 network, comprising: Modeling and simulating the UHVDC transmission system, setting different fault conditions in the converter area in the simulation model, and obtaining the original fault signals of the three-phase current signals between the Y-bridge and D-bridge converters, as well as the DC current signals of the high-voltage outlet side and the neutral end in the simulation model. The original fault signal is decomposed by wavelet decomposition. The decomposed signal with the highest correlation to the original fault signal is selected to construct a feature dataset. The feature dataset is then divided into a training set and a test set. The training set is input into the improved MobileNetV2 model to train the neural network model and learn fault characteristics; Determine whether the current neural network model has reached the set effect after training. If the current neural network model has reached the set effect after training, the improved MobileNetV2 model training is completed. If the current neural network model has not reached the set effect after training, adjust the network parameters to build the optimal improved MobileNetV2 model. The improved MobileNetV2 model, after being trained, is used to identify faults by inputting the test set, and the fault identification results of the UHV converter area are obtained.

[0007] Furthermore, the fault conditions include: six types of converter valve short circuits occurring in both the Y-bridge and D-bridge; three types of single-phase ground faults occurring in both the Y-bridge and D-bridge; three types of two-phase short circuits occurring in both the Y-bridge and D-bridge; a ground fault at the high-voltage end of the converter on the DC side; a ground fault at the midpoint of the converter on the DC side; a ground fault at the neutral end of the converter on the DC side; a short circuit from the high-voltage end to the midpoint of the converter; a short circuit from the midpoint to the neutral end of the converter; and a short circuit from the high-voltage end to the neutral end of the converter.

[0008] Furthermore, the improved MobileNetV2 model includes two dilated convolutional modules, a Ghost module, six inverted residual modules, a ULSAM attention mechanism module, a global average pooling module, and a fully connected module connected in sequence. The two dilated convolutional modules are a first dilated convolutional module and a second dilated convolutional module. The first dilated convolutional module contains one convolutional layer with a convolutional size of 3×1, a stride of 1×1, a kernel size of 32, and an activation function of ReLU. The second dilated convolutional module contains one convolutional layer with a convolutional size of 3×1, a stride of 2×1, a kernel size of 32, a dilation factor of 2, and an activation function of ReLU. The Ghost module comprises two convolutional layers and one batch normalization layer connected in sequence. The first convolutional layer has a size of 5×1, a stride of 1×1, a kernel size of 8, and an activation function of ReLU. The second convolutional layer has a size of 1×1, a stride of 1×1, a kernel size of 8, and no activation function. The batch normalization layer has a dimension of 16.

[0009] Furthermore, after each convolution operation in the dilated convolution module and the inverted residual module, a batch normalization layer is set, and after the global average pooling module performs global average pooling, Dropout is used for random deletion.

[0010] Furthermore, the proportion of random deletions performed by Dropout is 0.3.

[0011] Furthermore, the learning fault features are adaptively estimated using an adaptive matrix estimation method with an initial learning rate of 0.01, and the cross-entropy loss function is selected as the objective function.

[0012] Furthermore, the cross-entropy loss function is calculated using the following formula: ; Where Loss is the cross-entropy loss function, and C is the class. For the true value, These are predicted values.

[0013] Secondly, another embodiment of the present invention provides a fault identification system for ultra-high voltage converter areas based on an improved MobileNetV2 network, comprising: a modeling and simulation module, a data sample construction module, a model training module, a training count determination module, and a fault identification module. The modeling and simulation module is used to model and simulate the UHVDC transmission system. Different fault conditions are set in the converter area in the simulation model. The original fault signals are obtained in the simulation model, including the three-phase current signals between the Y-bridge and D-bridge converters and the DC current signals of the high-voltage outlet side and the neutral end. The data sample construction module performs wavelet decomposition on the original fault signal, selects the decomposed signal with the highest correlation to the original fault signal to construct a feature dataset, and divides the feature dataset into a training set and a test set. The model training module is used to input the training set into the improved MobileNetV2 model to train the neural network model and learn fault characteristics; The training count judgment module is used to determine whether the current neural network model has reached the set effect in terms of training count. If the current neural network model has reached the set effect in terms of training count, the improved MobileNetV2 model training is completed. If the current neural network model has not reached the set effect in terms of training count, the network parameters are adjusted to build the optimal improved MobileNetV2 model. The fault identification module is used to input the test set into the trained improved MobileNetV2 model for fault identification, and obtain the fault identification results of the UHV converter area.

[0014] Furthermore, the improved MobileNetV2 model includes two dilated convolutional modules, a Ghost module, six inverted residual modules, a ULSAM attention mechanism module, a global average pooling module, and a fully connected module connected in sequence. The two dilated convolutional modules are a first dilated convolutional module and a second dilated convolutional module. The first dilated convolutional module contains one convolutional layer with a convolutional size of 3×1, a stride of 1×1, a kernel size of 32, and an activation function of ReLU. The second dilated convolutional module contains one convolutional layer with a convolutional size of 3×1, a stride of 2×1, a kernel size of 32, a dilation factor of 2, and an activation function of ReLU. The Ghost module comprises two convolutional layers and one batch normalization layer connected in sequence. The first convolutional layer has a size of 5×1, a stride of 1×1, a kernel size of 8, and an activation function of ReLU. The second convolutional layer has a size of 1×1, a stride of 1×1, a kernel size of 8, and no activation function. The batch normalization layer has a dimension of 16.

[0015] Thirdly, another embodiment of the present invention provides a computer-readable storage medium storing a computer program, the computer program including program instructions, which, when executed by a processor, cause the processor to perform the method described in the first embodiment above.

[0016] Compared with the prior art, the present invention has the following advantages and beneficial effects: This invention provides a method, system, and medium for fault identification in UHV converter areas based on an improved MobileNetV2 network. It introduces deep learning into UHV converter area fault identification, proposes an improved MobileNetV2 model, and uses fault data after wavelet decomposition as model input to remove redundant fault information, thereby improving the accuracy of fault identification in the converter area. The improved MobileNetV2 model is used to extract fault features and classify different fault conditions, enabling rapid learning and memorization of fault features. This fully utilizes the extraction and classification capabilities of the improved MobileNetV2 model, improving the efficiency of fault identification in UHV converter areas. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be considered as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort. In the drawings: Figure 1 A flowchart of a fault identification method for ultra-high voltage converter areas based on an improved MobileNetV2 network, provided for the first embodiment of the present invention; Figure 2 A fault type diagram for a specific example of an ultra-high voltage converter area; Figure 3 To improve the network structure diagram of the MobileNetV2 model; Figure 4 A fault simulation model diagram of the UHV converter area; Figure 5 For training accuracy convergence curve; Figure 6 For training loss value curves; Figure 7 To improve the confusion matrix of the MobileNetV2 test set; Figure 8 The diagram below shows a structural block diagram of a fault identification system for an ultra-high voltage converter area based on an improved MobileNetV2 network, provided as a second embodiment of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings. The illustrative embodiments and descriptions of the present invention are only used to explain the present invention and are not intended to limit the present invention.

[0019] like Figure 1 As shown, the first embodiment of the present invention provides a fault identification method for ultra-high voltage converter areas based on an improved MobileNetV2 network, comprising the following steps: Step 1: Model and simulate the UHVDC transmission system, setting up different fault scenarios in the converter area within the simulation model. Specific fault locations are as follows: Figure 2 As shown, the original fault signals are obtained in the simulation model by acquiring the three-phase current signals between the Y-bridge and D-bridge converters, as well as the DC current signals at the high-voltage outlet side and the neutral end.

[0020] There are 30 typical fault types within the converter area, including twelve types of converter valve short circuits, six types of single-phase ground faults at the AC input terminal of the converter, six types of two-phase short circuits at the AC input terminal of the converter, one type of DC-side high-voltage terminal ground fault, one type of DC-side midpoint ground fault, one type of DC-side neutral terminal ground fault, one type of high-voltage terminal to midpoint short circuit, one type of midpoint to neutral terminal short circuit, and one type of high-voltage terminal to neutral terminal short circuit. Specific fault types are shown in Table 1. Table 1. Specific Fault Types in the Converter Area

[0021] Specifically, different parameter combinations are used to simulate various faults using PSCAD for each fault condition, increasing both the number of fault samples and the quality of the fault data. PSCAD (Power Systems Computer Aided Design) is a widely used electromagnetic transient simulation software worldwide, with EMTDC as its core simulation calculation. PSCAD provides a graphical user interface for EMTDC (Electromagnetic Transients including DC).

[0022] Step 2: Perform wavelet decomposition on the original fault signal to remove redundancy, select the decomposed signal with the highest correlation to the original fault signal to construct a feature dataset, set labels for the constructed dataset, and divide the feature dataset into training set and test set.

[0023] The wavelet decomposition formula is as follows: , , in, These are the low-frequency approximation coefficients for the (j+1)th layer. For the high-frequency detail coefficients of the (j+1)th layer, It is a low-pass filter. For high-pass filter, j represents the number of decomposition layers (starting from 0, j=0 corresponds to the original signal). Denotes the length of the coefficients in layer j, which satisfies , where n represents the position index of the coefficient in the current layer (layer j), and the range is 0 ≤ n < . ;k represents the position index of the coefficient in the next layer (j+1th layer), in the range 0 ≤ k < By using the fault data after wavelet decomposition as input, redundant fault information can be eliminated, thereby improving the accuracy of fault identification in the converter area.

[0024] Step 3: Input the training set into the improved MobileNetV2 model to train the neural network model. Dilated convolutions capture a wider range of contextual information. After passing through the Ghost module, inverted residual module, and ULSAM attention mechanism module, the expressive power of features is improved, and features are extracted efficiently. The data after in-depth feature extraction is then subjected to global average pooling to train the neural network model and learn fault features. The Ghost module generates feature maps more efficiently, reducing computation and parameters while maintaining the model's feature extraction capabilities as much as possible. The inverted residual module uses an inverted residual structure, first increasing dimensionality and then decreasing it, utilizing operations such as depthwise separable convolutions to enhance feature expressive power while reducing computational costs. The ULSAM attention mechanism module uses an attention mechanism to focus the model on important features and suppress unimportant features, improving the model's ability to capture key information. The improved MobileNetV2 model utilizes dilated convolution modules to capture a wider range of contextual information, expanding the receptive field. Combined with the lightweight MobileNetV2 network, it better captures temporal fault information without increasing parameters compared to the original MobileNetV2 network. Furthermore, the introduction of the Ghost module and ULSAM attention mechanism into the MobileNetV2 network enhances the model's lightweightness and efficiency while strengthening its focus on key features. This improves robustness and expressiveness during feature extraction, thereby increasing the efficiency of fault identification in the converter area.

[0025] like Figure 3As shown, the fault identification model based on the improved MobileNetV2 model includes two dilated convolutional modules, a Ghost module, six inverted residual modules, a ULSAM attention mechanism module, a global average pooling module, and a fully connected module connected in sequence. The two dilated convolutional modules are the first and second dilated convolutional modules. The first dilated convolutional module contains one convolutional layer with a convolutional size of 3×1, a stride of 1×1, a kernel size of 32, and a ReLU activation function. The second dilated convolutional module contains one convolutional layer with a convolutional size of 3×1, a stride of 2×1, a kernel size of 32, a dilation factor of 2, and a ReLU activation function. The Ghost module contains two convolutional layers and one batch normalization layer connected in sequence. The first convolutional layer has a size of 5×1, a stride of 1×1, a kernel size of 8, and a ReLU activation function. The second convolutional layer has a size of 1×1, a stride of 1×1, and a kernel size of 8. The batch normalization layer has a dimension of 16 and no activation function. The six inverse residual modules are: the first inverse residual module, the second inverse residual module, the third inverse residual module, the fourth inverse residual module, the fifth inverse residual module, and the sixth inverse residual module. The first inverse residual module contains three convolutional layers: the first layer has a convolutional size of 1×1, a stride of 1×1, a kernel size of 96, a batch normalization of 96, and a ReLU6 activation function; the second layer has a convolutional size of 5×1, a stride of 2×1, a kernel size of 96, a batch normalization of 96, and a ReLU6 activation function; the third layer has a convolutional size of 1×1, a stride of 1×1, a kernel size of 24, and no activation function. The second inverse residual module... The first convolutional module contains three convolutional layers. The first layer has a convolutional size of 1×1, a stride of 1×1, a kernel size of 144, and a ReLU6 activation function. The second layer has a convolutional size of 5×1, a stride of 2×1, a kernel size of 144, and a ReLU6 activation function. The third layer has a convolutional size of 1×1, a stride of 1×1, a kernel size of 32, and no activation function. The third inverse residual module also contains three convolutional layers. The first layer has a convolutional size of 1×1, a stride of 1×1, a kernel size of 192, and a ReLU6 activation function. The second layer has a convolutional size of 5×1, a stride of 2×1, a kernel size of 192, and a ReLU6 activation function. The third layer has a convolutional size of 1×1, a stride of 1×1... The first convolutional layer has a kernel size of 64 and no activation function; the fourth inverse residual module contains three convolutional layers. The first layer has a convolutional size of 1×1, a stride of 1×1, a kernel size of 384, and a ReLU6 activation function; the second layer has a convolutional size of 5×1, a stride of 1×1, a kernel size of 384, and a ReLU6 activation function; the third layer has a convolutional size of 1×1, a stride of 1×1, a kernel size of 128, and no activation function; the fifth inverse residual module contains three convolutional layers. The first layer has a convolutional size of 1×1, a stride of 1×1, a kernel size of 768, and a ReLU6 activation function; the second layer has a convolutional size of 5×1, a stride of 1×1, a kernel size of 768, and a ReLU6 activation function.The third convolutional layer has a size of 1×1, a stride of 1×1, a kernel size of 128, and no activation function. The sixth inverse residual module contains three convolutional layers: the first layer has a size of 1×1, a stride of 1×1, a kernel size of 768, and a ReLU6 activation function; the second layer has a size of 5×1, a stride of 2×1, a kernel size of 768, and a ReLU6 activation function; the third layer has a size of 1×1, a stride of 1×1, a kernel size of 256, and no activation function. The ULSAM attention mechanism module contains one global average pooling layer of size 1, one fully connected layer with 16 neurons and a ReLU activation function, and one fully connected layer with 256 neurons and a Sigmoid activation function. The global average pooling module contains one global average pooling layer of size 1. The fully connected module has 30 neurons.

[0026] In both the dilated convolution and inverse residual modules, a batch normalization layer is added after each convolutional operation. The global average pooling module performs global average pooling followed by random deletion using Dropout at a ratio of 0.3. The fully connected module handles fault identification. Random deletion using Dropout after the global average pooling layer prevents overfitting. Cross-entropy loss is employed to reduce the difference between the output and the expected value, further preventing overfitting during training.

[0027] The learning fault features are learned using the adaptive moment estimation (Adam) method with an adaptive learning rate of 0.01, and the cross-entropy loss function is selected as the objective function.

[0028] The formula for calculating the cross-entropy loss function is as follows: , Where Loss is the cross-entropy loss function, and C is the number of classes; The actual value; These are predicted values.

[0029] Step 4: Determine whether the current neural network model has reached the set effect after training. If the current neural network model has reached the set effect after training, the training of the improved MobileNetV2 model is complete. If the current neural network model has not reached the set value after training, repeat steps 3 and 4 to repeatedly adjust the network parameters to build the optimal improved MobileNetV2 model. The network parameters include batch size, learning rate and number of iterations.

[0030] Step 5: Input the test set into the trained improved MobileNetV2 model for fault identification, completing the identification of faults in the UHV converter area and obtaining the fault identification results. Using the training set as input to the improved MobileNetV2 model allows it to learn fault information more efficiently over a wider range, while avoiding increasing model complexity.

[0031] This invention proposes a fault identification method for UHV converter areas based on an improved MobileNetV2 network. It introduces deep learning into UHV converter area fault identification, proposing an improved MobileNetV2 model. Using fault data after wavelet decomposition as model input, it eliminates redundant fault information, improving the accuracy of converter area fault identification. The improved MobileNetV2 model is used to extract fault features and classify different fault conditions, enabling rapid learning and memorization of fault features. This fully utilizes the extraction and classification capabilities of the improved MobileNetV2 model, thereby improving the efficiency of UHV converter area fault identification.

[0032] The implementation process of the present invention will be further described below with reference to embodiments: Fault conditions in the UHV converter area, such as Figure 2 As shown, fault simulation is performed in PSCAD, and the fault simulation model is as follows. Figure 4 As shown.

[0033] The fault conditions were set with different fault times and different transition resistances. For the fault occurrence time, six different fault times were selected: 1.001s, 1.004s, 1.007s, 1.010s, 1.013s, and 1.016s. For the transition resistance, values ​​ranging from 0Ω to 200Ω were selected, increasing in 5Ω increments, resulting in 40 different cases. All data were sampled at 20kHz, generating 240 samples for each fault type, and divided into training and test sets in a 7:3 ratio. The training accuracy convergence curve and the loss curve are shown below. Figure 5 , 6 As shown. Figure 7 As shown in the figure, the confusion matrix visualization demonstrates the correspondence between the prediction results of the improved MobileNetV2 model and the true labels.

[0034] By verifying 30 fault scenarios in the UHV converter area, the improved MobileNetV2 model achieved a fault identification accuracy of 98.43%, ensuring both comprehensive fault identification and good fault identification accuracy.

[0035] like Figure 8As shown, the second embodiment of the present invention provides a fault identification system for UHV converter areas based on an improved MobileNetV2 network, including: a modeling and simulation module, a data sample construction module, a model training module, a training count judgment module, and a fault identification module; The modeling and simulation module is used to model and simulate the UHVDC transmission system. Different fault conditions are set in the converter area in the simulation model. The original fault signals are obtained in the simulation model, including the three-phase current signals between the Y-bridge and D-bridge converters, as well as the DC current signals on the high-voltage outlet side and the neutral end. The data sample construction module performs wavelet decomposition on the original fault signal, selects the decomposed signal with the highest correlation to the original fault signal to construct a feature dataset, and divides the feature dataset into a training set and a test set. The model training module is used to input the training set into the improved MobileNetV2 model to train the neural network model and learn fault characteristics; The training count judgment module is used to determine whether the current neural network model has reached the set effect after training. If the current neural network model has reached the set effect after training, the improved MobileNetV2 model training is completed. If the current neural network model has not reached the set effect after training, the network parameters are adjusted to build the optimal improved MobileNetV2 model. The fault identification module is used to input the test set into the trained improved MobileNetV2 model for fault identification, and obtain the fault identification results of the UHV converter area.

[0036] The improved MobileNetV2 model includes two dilated convolutional modules, a Ghost module, six inverted residual modules, a ULSAM attention mechanism module, a global average pooling module, and a fully connected module connected in sequence. The two dilated convolutional modules are designated as the first dilated convolutional module and the second dilated convolutional module. The first dilated convolutional module contains one convolutional layer with a convolutional size of 3×1, a stride of 1×1, a kernel size of 32, and an activation function of ReLU. The second dilated convolutional module contains one convolutional layer with a convolutional size of 3×1, a stride of 2×1, a kernel size of 32, a dilation factor of 2, and an activation function of ReLU. The Ghost module contains two convolutional layers and one batch normalization layer connected in sequence. The first convolutional layer has a size of 5×1, a stride of 1×1, a kernel size of 8, and an activation function of ReLU. The second convolutional layer has a size of 1×1, a stride of 1×1, a kernel size of 8, and no activation function. The batch normalization layer has a dimension of 16.

[0037] The proposed fault identification system for UHV converter areas based on an improved MobileNetV2 network utilizes a modeling and simulation module to model and simulate the UHVDC transmission system. It performs wavelet decomposition on fault data collected from the PSCAD simulation model to remove redundant fault information, thus improving the accuracy of fault identification in the converter area. Furthermore, the improved MobileNetV2 model is used to extract fault features and classify different fault conditions, enabling rapid learning and memorization of fault features. This fully leverages the extraction and classification capabilities of the improved MobileNetV2 model, thereby enhancing the efficiency of fault identification in UHV converter areas.

[0038] The fourth embodiment of the present invention also provides a computer-readable storage medium storing a computer program, the computer program including program instructions, which, when executed by a processor, cause the processor to execute the intelligent processing method for process approval data based on a large model described in the first embodiment above.

[0039] The computer-readable storage medium can be an internal storage unit of the terminal described in the foregoing embodiments, such as the terminal's hard drive or memory. The computer-readable storage medium can also be an external storage device of the terminal, such as a plug-in hard drive, Smart Media Card (SMC), Secure Digital (SD) card, or Flash Card equipped on the terminal. The computer-readable storage medium may include both internal and external storage units of the terminal. The computer-readable storage medium is used to store the computer program and other programs and data required by the terminal. The computer-readable storage medium can also be used to temporarily store data that has been output or will be output.

[0040] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0041] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the terminals and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0042] In the several embodiments provided in this application, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. In addition, the mutual coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices or units, or may be electrical, mechanical or other forms of connection.

[0043] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for fault identification in ultra-high voltage converter areas based on an improved MobileNetV2 network, characterized in that, include: Modeling and simulating the UHVDC transmission system, setting different fault conditions in the converter area in the simulation model, and obtaining the original fault signals of the three-phase current signals between the Y-bridge and D-bridge converters, as well as the DC current signals of the high-voltage outlet side and the neutral end in the simulation model. The original fault signal is decomposed by wavelet decomposition. The decomposed signal with the highest correlation to the original fault signal is selected to construct a feature dataset. The feature dataset is then divided into a training set and a test set. The training set is input into the improved MobileNetV2 model to train the neural network model and learn fault characteristics; Determine whether the current neural network model has reached the set effect after training. If the current neural network model has reached the set effect after training, the improved MobileNetV2 model training is completed. If the current neural network model has not reached the set effect after training, adjust the network parameters to build the optimal improved MobileNetV2 model. The improved MobileNetV2 model, after being trained, is used to identify faults by inputting the test set, and the fault identification results of the UHV converter area are obtained.

2. The method for fault identification in UHV converter areas based on an improved MobileNetV2 network as described in claim 1, characterized in that, The fault conditions include: six types of converter valve short circuits occurring in both the Y-bridge and D-bridge; three types of single-phase ground faults occurring in both the Y-bridge and D-bridge; three types of two-phase short circuits occurring in both the Y-bridge and D-bridge; a ground fault at the high-voltage end of the converter on the DC side; a ground fault at the midpoint of the converter on the DC side; a ground fault at the neutral end of the converter on the DC side; a short circuit from the high-voltage end to the midpoint of the converter; a short circuit from the midpoint to the neutral end of the converter; and a short circuit from the high-voltage end to the neutral end of the converter.

3. The method for fault identification in UHV converter areas based on an improved MobileNetV2 network as described in claim 1, characterized in that, The improved MobileNetV2 model includes two dilated convolutional modules, a Ghost module, six inverted residual modules, a ULSAM attention mechanism module, a global average pooling module, and a fully connected module connected in sequence. The two dilated convolutional modules are a first dilated convolutional module and a second dilated convolutional module. The first dilated convolutional module contains one convolutional layer with a convolutional size of 3×1, a stride of 1×1, a kernel size of 32, and an activation function of ReLU. The second dilated convolutional module contains one convolutional layer with a convolutional size of 3×1, a stride of 2×1, a kernel size of 32, a dilation factor of 2, and an activation function of ReLU. The Ghost module comprises two convolutional layers and one batch normalization layer connected in sequence. The first convolutional layer has a size of 5×1, a stride of 1×1, a kernel size of 8, and an activation function of ReLU. The second convolutional layer has a size of 1×1, a stride of 1×1, a kernel size of 8, and no activation function. The batch normalization layer has a dimension of 16.

4. The method for fault identification in UHV converter areas based on an improved MobileNetV2 network as described in claim 3, characterized in that, Each convolutional operation in the dilated convolution module and the inverted residual module is followed by a batch normalization layer. After global average pooling, the global average pooling module performs global average pooling and then uses Dropout for random deletion.

5. The method for fault identification in UHV converter areas based on an improved MobileNetV2 network as described in claim 4, characterized in that, The Dropout random deletion ratio is 0.

3.

6. The method for fault identification in UHV converter areas based on an improved MobileNetV2 network as described in claim 1, characterized in that, The learning fault features are learned using an adaptive matrix estimation method with an adaptive learning rate of 0.01, and the cross-entropy loss function is selected as the objective function.

7. The method for fault identification in UHV converter areas based on an improved MobileNetV2 network as described in claim 6, characterized in that, The cross-entropy loss function is calculated using the following formula: ； Where Loss is the cross-entropy loss function, and C is the class. For the true value, These are predicted values.

8. A fault identification system for ultra-high voltage converter areas based on an improved MobileNetV2 network, characterized in that, include: The module includes a modeling and simulation module, a data sample construction module, a model training module, a training iteration determination module, and a fault identification module. The modeling and simulation module is used to model and simulate the UHVDC transmission system. Different fault conditions are set in the converter area in the simulation model. The original fault signals are obtained in the simulation model, including the three-phase current signals between the Y-bridge and D-bridge converters and the DC current signals of the high-voltage outlet side and the neutral end. The data sample construction module performs wavelet decomposition on the original fault signal, selects the decomposed signal with the highest correlation to the original fault signal to construct a feature dataset, and divides the feature dataset into a training set and a test set. The model training module is used to input the training set into the improved MobileNetV2 model to train the neural network model and learn fault characteristics; The training count judgment module is used to determine whether the current neural network model has reached the set effect in terms of training count. If the current neural network model has reached the set effect in terms of training count, the improved MobileNetV2 model training is completed. If the current neural network model has not reached the set effect in terms of training count, the network parameters are adjusted to build the optimal improved MobileNetV2 model. The fault identification module is used to input the test set into the trained improved MobileNetV2 model for fault identification, and obtain the fault identification results of the UHV converter area.

9. The UHV converter area fault identification system based on the improved MobileNetV2 network as described in claim 8, characterized in that, The improved MobileNetV2 model includes two dilated convolutional modules, a Ghost module, six inverted residual modules, a ULSAM attention mechanism module, a global average pooling module, and a fully connected module connected in sequence. The two dilated convolutional modules are a first dilated convolutional module and a second dilated convolutional module. The first dilated convolutional module contains one convolutional layer with a convolutional size of 3×1, a stride of 1×1, a kernel size of 32, and an activation function of ReLU. The second dilated convolutional module contains one convolutional layer with a convolutional size of 3×1, a stride of 2×1, a kernel size of 32, a dilation factor of 2, and an activation function of ReLU. The Ghost module comprises two convolutional layers and one batch normalization layer connected in sequence. The first convolutional layer has a size of 5×1, a stride of 1×1, a kernel size of 8, and an activation function of ReLU. The second convolutional layer has a size of 1×1, a stride of 1×1, a kernel size of 8, and no activation function. The batch normalization layer has a dimension of 16.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, the computer program including program instructions that, when executed by a processor, cause the processor to perform the method as described in any one of claims 1-7.

Images