Transformer substation grounding grid corrosion detection and positioning method based on improved optimization algorithm

Abstract

The invention relates to the field of electric power system detection, and provides a transformer substation grounding grid corrosion detection and positioning method based on an improved optimization algorithm, which comprises the following steps: detecting a grounding grid by using a ground penetrating radar device and an electromagnetic induction instrument, obtaining GPR and EMI data, inputting the data into an MATLAB / Simulink platform for analog simulation, respectively carrying out space-time convolution processing on the received GPR and EMI data, and carrying out optimization processing on the GPR and EMI data; extracting feature information of time and space dimensions; a self-attention algorithm is applied to a processing result of space-time convolution, a long-range dependency relationship is captured, meanwhile, dynamic weight distribution is carried out on different dimension features through self-adaptive gating fusion, and feature information of time and space dimensions of GPR and EMI is fused; comparing the fused feature information with pre-stored feature information of an original grounding grid to generate a detection result; according to the method, accurate detection and positioning of corrosion of the transformer substation grounding system are realized by improving a multi-modal data fusion framework of the space-time convolutional neural network and the self-attention mechanism.

Description

Technical Field

[0001] The present invention relates to the field of power system detection, and specifically to a substation grounding grid corrosion detection and positioning method based on multimodal data fusion improved spatiotemporal convolutional neural network and self-attention mechanism, which is used for accurate positioning of substation grounding system corrosion. Background Art

[0002] With the development of the times, my country's industrial production has entered a new stage. Driven by Industry 4.0, the demand for electricity, as the main energy source in production development, has grown rapidly, which has also brought new pressure and challenges to the power system. In the power grid, power plants / substations are an important part of power transmission, which integrates various key equipment such as generators, transformers, secondary equipment, communication equipment, etc. Many of them are connected to the grounding grid by grounding down conductors. The strong grounding effect and the earth's current dispersion capacity can prevent the voltage rise of electrical equipment. Therefore, the grounding grid is of great significance to ensure the safety of power equipment.

[0003] At present, the integrity and reliability of the substation grounding system, as a key infrastructure for the safe operation of the power system, directly affects the stability and safety of the entire power grid. However, due to long-term exposure to complex underground environments, the grounding system inevitably faces corrosion problems, which not only reduces the electrical performance of the system, but also increases the risk of equipment failure and safety accidents.

[0004] The corrosion of the grounding system is mainly electrochemical corrosion, and its occurrence principle is based on the chemical reaction between metals and ions in an electrolyte environment. When grounded metal materials (such as steel or copper) are exposed in moist soil, they will corrode due to oxygen, moisture and other electrolytes in the soil. Traditional grounding system detection methods, such as visual inspection and resistance measurement, often have difficulty in identifying hidden corrosion problems in a timely and accurate manner, especially in the complex grounding networks of large substations. At the same time, existing single-mode detection methods, such as ground penetrating radar (GPR), are a physical exploration method that uses ultra-high frequency pulsed electromagnetic waves to detect the characteristics of underground targets, but they often face problems such as insufficient accuracy and weak anti-interference ability in complex environments. Summary of the invention

[0005] In order to solve the above technical problems, the present invention provides a substation grounding grid corrosion detection and positioning method based on an improved optimization algorithm, which is used for the precise positioning of substation grounding system corrosion by improving the multimodal data fusion framework of spatiotemporal convolutional neural network (ST-CNN) and self-attention mechanism, so as to solve the problems in the prior art.

[0006] A method for detecting and locating corrosion of a substation grounding grid based on an improved optimization algorithm, comprising:

[0007] Step 1: Detect the grounding grid by using a ground penetrating radar (GPR) device and an electromagnetic induction (EMI) instrument to obtain GPR and EMI data of the grounding grid;

[0008] Step 2, the multimodal data input stage, receives GPR and EMI data, and performs simulation through the MATLAB / Simulink platform. The parameter setting of the simulation model refers to the technical specifications and operation data of the multimodal data detection and analysis method applied to the substation grounding system to ensure that the simulation results are of practical significance.

[0009] Step 3, feature extraction stage, perform spatiotemporal convolution processing on the received GPR and EMI data respectively, and extract feature information in time and space dimensions; extract features of GPR and EMI data through the improved spatiotemporal convolutional neural network, set the number of convolution layers to 5, the convolution kernel size to 3*3, and the hidden state dimension of the self-attention mechanism to 128;

[0010] In step 4, the self-attention mechanism stage, the self-attention algorithm is applied to the processing results of spatiotemporal convolution to capture long-range dependencies and enhance the model's perception of key features. At the same time, dynamic weights are assigned to features of different dimensions through adaptive gated fusion, and the feature information of the time and space dimensions of GPR and EMI are fused.

[0011] Step 5, the output stage, compares the fused feature information with the pre-stored feature information of the original grounding grid in a normal state to generate the final corrosion detection result.

[0012] Preferably, in step 3, the received GPR and EMI data are subjected to spatiotemporal convolution processing respectively, and feature information of time and space dimensions is extracted; wherein spatial features are extracted from GPR data, and time features are extracted from EMI data; the specific process of extracting spatial features from GPR data is as follows:

[0013] The multi-scale convolution algorithm is used to extract spatial features of GPR data. The formula is as follows:

[0014]

[0015] Among them, conv means extracting the features to be extracted by the convolution kernel through the convolution operation of the convolution layer, including spatial features, X is the input GPR data point, k is the convolution kernel, F is the extracted feature map, and the spatial features of different scales are captured by changing the size of the convolution kernel. i is the scale index of different scales, and k is the scale index of different scales. i are convolution kernels of different scales, It is the extracted i-scale GPR data feature map.

[0016] Preferably, the specific process of extracting time features from EMI data is as follows:

[0017] A multi-layer dilated convolution structure is used to extract the timing features of EMI data. The formula is as follows:

[0018]

[0019] Among them, l is the number of convolutional layers in the convolutional structure, and lD represents the multi-expansion convolution calculation identifier, which means that at the very beginning of this spatiotemporal convolution processing, that is, layer 0, no real convolution operation is performed to change the shape of the data;

[0020]

[0021] in, Represents the timing characteristic diagram of the EMI data at the i-th time scale; ω i express The corresponding fusion weight can be learned through the self-attention mechanism, F EMI This is the extracted EMS data feature map.

[0022] Preferably, in step 4, the specific process of applying the self-attention algorithm to the processing result of the spatiotemporal convolution is as follows:

[0023] First, the feature map F EMI or F i =F final As input, it is processed through the self-attention mechanism:

[0024] F global =Attention(F final ,F final ,F final )

[0025] This step generates a feature representation F containing global information global ;

[0026] Then, the local features and global features are fused to obtain the final feature representation:

[0027] F output =Concat(F final ,F global )W o

[0028] Among them, W o It is a weight matrix obtained through learning, which is used to dynamically adjust the importance of local and global features. It can be determined and adjusted through random initialization operations and subsequent training algorithms including back propagation algorithms.

[0029] In addition, in order to further enhance the model's ability to integrate multi-scale features, a multi-head attention mechanism is introduced. The multi-head attention mechanism learns multiple attention functions in parallel to capture information from different representation subspaces. Its formula can be expressed as:

[0030] MultiHead(Q,U,V)=Concat(head1,...,head h )W o

[0031]

[0032] Among them, Q, U and V represent the query vector, key vector and value vector respectively, which are obtained by inputting F output It is obtained through linear transformation and is used to measure the degree of association between different elements and extract corresponding value information in the self-attention mechanism. h is the total number of attention heads, and j is the index of the corresponding weight of the vector. and The learnable weight matrix is ​​a set of parameters used to linearly transform the input data, thereby affecting the calculation of the query vector, key vector, and value vector, as well as the integration of the final output. They determine which parts of the input data the model focuses on and how to combine information operations, and can be determined and adjusted through random initialization operations and subsequent training algorithms including backpropagation algorithms.

[0033] Get the final output head j After that, each head (head1,...,head h ) is concatenated with the results of the Attention operation to obtain a concatenated matrix, which is then combined with the learnable weight matrix Perform matrix multiplication to obtain the final output feature map F' after multi-head attention mechanism processing;

[0034] This F' is evaluated based on the feature quality including information entropy calculation and label data correlation. Information entropy reflects the amount of information contained in the feature. For the features extracted from GPR and EMI data, the lower the information entropy, the smaller the uncertainty of the feature, and there is insufficient information to effectively distinguish different grounding grid states. Evaluation and screening are performed according to the preset information entropy threshold;

[0035] The label data is the GPR and EMI data of historical ground grid corrosion conditions. Through comparison, it is evaluated and screened according to the thresholds set by the staff. Among them, those that pass the evaluation are used as input for subsequent adaptive gating fusion, and those that fail the evaluation are subjected to feature enhancement.

[0036] Preferably, the specific process of dynamically allocating weights to features of different dimensions through adaptive gated fusion is as follows:

[0037] The feature vector F of GPR and EMI data after being processed by improved spatiotemporal convolution and self-attention mechanism respectively GPR ' and F EMI ', adaptive gated fusion will calculate a gating vector G to control the fusion ratio of the two features. The calculation of the gating vector G is as follows:

[0038] G=σ(W g (F GPR :F EMI )+b g )

[0039] Among them, σ represents the sigmoid activation function, W g and b g are the learnable weight matrix and bias term respectively, and “:” represents the feature concatenation operation;

[0040] Preferably, in step 4, the specific process of fusing the feature information of the time and space dimensions of GPR and EMI is as follows:

[0041] After obtaining the gate vector, the fusion process of multimodal features can be expressed as:

[0042] F fused =G⊙F GPR +(1-G)⊙F EMI

[0043] Among them, ⊙ represents element-wise multiplication, F fused is the final fused feature information.

[0044] Preferably, in step 5, the fused feature information F fused Compare the characteristic information of the original grounding grid in a normal state stored in advance with the distance measurement method, including Euclidean distance or cosine similarity, calculate the degree of difference between the two, and judge whether the grounding grid is corroded and the location of the corrosion according to the preset threshold;

[0045] At the same time, after the corrosion position is calculated, the position accuracy is further confirmed. If the position accuracy meets the requirements of subsequent operations including repair processing, the result is output. If the position accuracy does not meet the requirements of subsequent operations including repair processing, the parameters of the EMI data after convolution processing are adjusted, including adjusting the convolution kernel size, step size or filling method, to meet the requirements of accurate spatial feature data.

[0046] Compared with the prior art, the present invention has the following beneficial effects:

[0047] 1. The present invention realizes accurate extraction of substation grounding system corrosion information by improving the spatial characteristics of GPR and the temporal characteristics of EMI through a multimodal data fusion framework based on an improved spatiotemporal convolutional neural network and a self-attention mechanism. At the same time, the introduction of a self-attention mechanism enhances the global correlation between features and improves the accuracy and efficiency of data fusion.

[0048] 2. The present invention dynamically integrates multi-modal feature technology through an adaptive gating mechanism. In the data fusion stage, the technology adopts an adaptive gating mechanism to dynamically integrate the data features of GPR and EMI modes. This mechanism can effectively deal with noise and interference in complex environments, improve positioning accuracy and robustness, and provide strong support for the precise monitoring and maintenance of substation grounding grids.

[0049] 3. The present invention improves the ability to capture local patterns by improving the spatiotemporal convolutional neural network, and uses the self-attention mechanism to model long-range dependencies, thereby achieving a comprehensive and in-depth understanding of the corrosion information of the substation grounding system in the feature extraction stage, which not only improves the utilization rate of the data, but also lays a solid foundation for subsequent data fusion and precise positioning. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 It is a schematic diagram of the stages and steps of the method of the present invention;

[0051] Figure 2 It is a schematic diagram of the implementation process of the method of the present invention. DETAILED DESCRIPTION

[0052] The following embodiments of the present invention are described in further detail in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0053] The present invention provides a method for detecting and locating corrosion of a substation grounding grid based on an improved optimization algorithm, comprising:

[0054] Step 1: Detect the grounding grid by using a ground penetrating radar (GPR) device and an electromagnetic induction (EMI) instrument to obtain GPR and EMI data of the grounding grid;

[0055] Step 2, the multimodal data input stage, receives GPR and EMI data, and performs simulation through the MATLAB / Simulink platform. The parameter setting of the simulation model refers to the technical specifications and operation data of the multimodal data detection and analysis method applied to the substation grounding system to ensure that the simulation results are of practical significance.

[0056] Step 3, feature extraction stage, perform spatiotemporal convolution processing on the received GPR and EMI data respectively, and extract feature information in time and space dimensions; extract features of GPR and EMI data through the improved spatiotemporal convolutional neural network, set the number of convolution layers to 5, the convolution kernel size to 3*3, and the hidden state dimension of the self-attention mechanism to 128;

[0057] In step 4, the self-attention mechanism stage, the self-attention algorithm is applied to the processing results of spatiotemporal convolution to capture long-range dependencies and enhance the model's perception of key features. At the same time, dynamic weights are assigned to features of different dimensions through adaptive gated fusion, and the feature information of the time and space dimensions of GPR and EMI are fused.

[0058] Step 5, the output stage, compares the fused feature information with the pre-stored feature information of the original grounding grid in a normal state to generate the final corrosion detection result.

[0059] Example:

[0060] like Figure 1-2 As shown, in this embodiment, in a certain large substation, the grounding grid plays a vital role in ensuring the safe and stable operation of the power system. Due to environmental influences, the grounding grid is easily affected by factors such as electrochemical corrosion, resulting in a decrease in grounding performance. Therefore, regular corrosion detection and positioning are required.

[0061] To detect the grounding grid, inspectors carry ground penetrating radar (GPR) devices and electromagnetic induction (EMI) instruments into the substation;

[0062] Using the GPR device, slowly move along the predetermined detection path above the grounding grid to emit electromagnetic waves into the ground. The electromagnetic waves will be reflected when encountering the grounding grid and the different media around it, such as normal metal parts, corroded parts, soil, etc. The GPR receives these reflected waves and records relevant data.

[0063] At the same time, turn on the electromagnetic induction (EMI) instrument, which obtains EMI data by sensing the changes in the magnetic field around the ground grid. During the entire detection process, ensure that the detection range covers the entire ground grid area to obtain comprehensive GPR and EMI data.

[0064] Multimodal data collection is achieved by combining ground penetrating radar (GPR) with electromagnetic induction (EMI) instruments, which comprehensively obtains grounding grid information from different physical mechanisms, overcomes the problems of insufficient accuracy and weak anti-interference ability faced by traditional single-modal detection methods in complex environments, and improves detection accuracy and reliability.

[0065] In the multimodal data input stage, the GPR and EMI data obtained from the substation grounding grid detection are transmitted to the computer and input on the MATLAB / Simulink platform. The parameters of the simulation model are set according to the technical specifications of the multimodal data detection and analysis method applied to the substation grounding system and the previous operation data of the substation grounding grid. For example, the electromagnetic parameters and propagation model parameters in the simulation model are determined by referring to the soil type, grounding grid material, previous corrosion conditions and other data of the substation to ensure that the simulation results can truly reflect the actual situation of the substation grounding grid and provide a reliable basis for subsequent feature extraction and analysis.

[0066] To extract spatial features from GPR data, a multi-scale convolution algorithm is used to extract spatial features from GPR data. The formula is as follows:

[0067]

[0068] Among them, conv means extracting the features to be extracted by the convolution kernel through the convolution operation of the convolution layer, including spatial features, X is the input GPR data point, k is the convolution kernel, F is the extracted feature map, and the spatial features of different scales are captured by changing the size of the convolution kernel. i is the scale index of different scales, and k is the scale index of different scales. i are convolution kernels of different scales, It is the extracted i-scale GPR data feature map.

[0069] To extract the time dimension features of EMI, a multi-layer dilated convolution structure is used to extract the timing features of EMI data. The formula is as follows:

[0070]

[0071] Among them, l is the number of convolutional layers in the convolutional structure, and lD represents the multi-expansion convolution calculation identifier, which means that at the very beginning of this spatiotemporal convolution processing, that is, layer 0, no real convolution operation is performed to change the shape of the data;

[0072]

[0073] in, Represents the timing characteristic diagram of the EMI data at the i-th time scale; ω i express The corresponding fusion weight can be learned through the self-attention mechanism, F EMI This is the extracted EMS data feature map.

[0074] In the self-attention mechanism stage, first, the feature map F EMI or F i =F final As input, it is processed through the self-attention mechanism:

[0075] F global =Attention(F final ,F final ,F final )

[0076] This step generates a feature representation F containing global information global In this process, the self-attention mechanism calculates the correlation between features, allowing the model to pay attention to the correlation between features at different locations. For example, in grounding grid detection, the relationship between corrosion features at different locations can be associated to better judge the overall corrosion situation. Subsequently, the local features and global features are fused to obtain the final feature representation:

[0077] F output =Concat(F final ,F global )W o

[0078] Among them, W o It is a weight matrix obtained through learning, which is used to dynamically adjust the importance of local and global features. It can be determined and adjusted through random initialization operations and subsequent training algorithms including back propagation algorithms.

[0079] In order to further enhance the model's ability to integrate multi-scale features, a multi-head attention mechanism is introduced. The multi-head attention mechanism learns multiple attention functions in parallel to capture information from different representation subspaces. Its formula can be expressed as:

[0080] MultiHead(Q,U,V)=Concat(head1,...,head h )W o

[0081]

[0082] Among them, Q, U and V represent the query vector, key vector and value vector respectively, which are obtained by inputting F output It is obtained through linear transformation and is used to measure the degree of association between different elements and extract corresponding value information in the self-attention mechanism. h is the total number of attention heads, and j is the index of the corresponding weight of the vector. and The learnable weight matrix is ​​a set of parameters used to linearly transform the input data, thereby affecting the calculation of the query vector, key vector, and value vector, as well as the integration of the final output. They determine which parts of the input data the model focuses on and how to combine information operations, and can be determined and adjusted through random initialization operations and subsequent training algorithms including backpropagation algorithms.

[0083] Get the final output head j After that, each head (head1,...,head h ) is concatenated with the results of the Attention operation to obtain a concatenated matrix, which is then combined with the learnable weight matrix Perform matrix multiplication to obtain the final output feature map F' after multi-head attention mechanism processing;

[0084] This F' is evaluated based on the feature quality including information entropy calculation and label data correlation. Information entropy reflects the amount of information contained in the feature. For the features extracted from GPR and EMI data, the lower the information entropy, the smaller the uncertainty of the feature, and there is insufficient information to effectively distinguish different grounding grid states. Evaluation and screening are performed according to the preset information entropy threshold;

[0085] The label data is the GPR and EMI data of historical ground grid corrosion conditions. Through comparison, it is evaluated and screened according to the thresholds set by the staff. Among them, those that pass the evaluation are used as input for subsequent adaptive gating fusion, and those that fail the evaluation are subjected to feature enhancement.

[0086] For example, for GPR features with low information entropy, the features can be enhanced by increasing the scale range of multi-scale convolution or adjusting the convolution kernel parameters; for EMI features with weak correlation with label data, the dilation rate or number of layers of dilated convolution can be adjusted to improve the features.

[0087] The specific process of dynamically allocating weights to features of different dimensions through adaptive gated fusion is as follows:

[0088] The feature vector F of GPR and EMI data after being processed by improved spatiotemporal convolution and self-attention mechanism respectively GPR ' and F EMI ', adaptive gated fusion will calculate a gating vector G to control the fusion ratio of the two features. The calculation of the gating vector G is as follows:

[0089] G=σ(W g (F GRP :F EMI )+b g )

[0090] Among them, σ represents the sigmoid activation function, W g and b g are the learnable weight matrix and bias term respectively, and “:” represents the feature concatenation operation;

[0091] The specific process of fusing the feature information of GPR and EMI in time and space dimensions is as follows:

[0092] After obtaining the gate vector, the fusion process of multimodal features can be expressed as:

[0093] F fused =G⊙F GPR +(1-G)⊙F EMI

[0094] Among them, ⊙ represents element-wise multiplication, F fused is the final fused feature information.

[0095] Dynamically integrate multi-modal feature technology through adaptive gating mechanism. In the data fusion stage, this technology adopts adaptive gating mechanism to dynamically integrate the data features of GPR and EMI modes. This mechanism can effectively deal with noise and interference in complex environments, improve positioning accuracy and robustness, and provide strong support for accurate monitoring and maintenance of substation grounding grids.

[0096] The fused feature information F fused Compare the characteristic information of the original grounding grid in a normal state stored in advance with the distance measurement method, including Euclidean distance or cosine similarity, calculate the degree of difference between the two, and judge whether the grounding grid is corroded and the location of the corrosion according to the preset threshold;

[0097] At the same time, after the corrosion position is calculated, the position accuracy is further confirmed. If the position accuracy meets the requirements of subsequent operations including repair processing, the result is output. If the position accuracy does not meet the requirements of subsequent operations including repair processing, the parameters of the EMI data after convolution processing are adjusted, including adjusting the convolution kernel size, step size or filling method, to meet the requirements of accurate spatial feature data.

[0098] In summary, the substation grounding grid corrosion detection and positioning method based on the improved optimization algorithm realizes multimodal data acquisition by integrating ground penetrating radar (GPR) and electromagnetic induction (EMI) instruments, mines and integrates data features with the help of improved spatiotemporal convolutional neural networks, self-attention mechanisms and multi-head attention mechanisms, and optimizes data fusion by relying on adaptive gated fusion mechanisms. It works synergistically from multiple dimensions to comprehensively improve the efficiency and quality of substation grounding grid corrosion detection and positioning.

[0099] It provides strong technical support for ensuring the safe and stable operation of substations, extending the service life of grounding grids, and promoting the advancement of power system operation and maintenance technology, and has extremely significant economic and social benefits.

[0100] The embodiments of the present invention are provided for the purpose of illustration and description. Although the embodiments of the present invention have been shown and described above, it can be understood that the above embodiments are exemplary and cannot be understood as limitations of the present invention. Ordinary technicians in this field can change, modify, replace and modify the above embodiments within the scope of the present invention.

Claims

1. A method for detecting and locating corrosion of a substation grounding grid based on an improved optimization algorithm, characterized in that: include: Step 1: Detect the grounding grid by using a ground penetrating radar (GPR) device and an electromagnetic induction (EMI) instrument to obtain GPR and EMI data of the grounding grid; Step 2, multi-modal data input stage, receiving GPR and EMI data, and performing simulation through MATLAB / Simulink platform. The parameter setting of the simulation model refers to the technical specifications and operation data of the multi-modal data detection and analysis method applied to the substation grounding system to ensure that the simulation results are of practical significance; Step 3, feature extraction stage, perform spatiotemporal convolution processing on the received GPR and EMI data respectively, and extract feature information in time and space dimensions; extract features of GPR and EMI data through the improved spatiotemporal convolutional neural network, set the number of convolution layers to 5, the convolution kernel size to 3*3, and the hidden state dimension of the self-attention mechanism to 128; Step 4, the self-attention mechanism stage, applies the self-attention algorithm to the processing results of spatiotemporal convolution to capture long-range dependencies and enhance the model's perception of key features. At the same time, adaptive gated fusion is used to dynamically assign weights to features of different dimensions, and the feature information of the time and space dimensions of GPR and EMI is fused; Step 5, the output stage, compares the fused feature information with the pre-stored feature information of the original grounding grid in a normal state to generate the final corrosion detection result.

2. The method for detecting and locating corrosion of a substation grounding grid based on an improved optimization algorithm as claimed in claim 1, characterized in that: In step 3, the received GPR and EMI data are subjected to spatiotemporal convolution processing respectively, and feature information of time and space dimensions is extracted; wherein spatial features are extracted from GPR data, and time features are extracted from EMI data; the specific process of extracting spatial features from GPR data is as follows: The multi-scale convolution algorithm is used to extract spatial features of GPR data. The formula is as follows: Among them, conv means extracting the features to be extracted by the convolution kernel through the convolution operation of the convolution layer, including spatial features, X is the input GPR data point, k is the convolution kernel, F is the extracted feature map, and the spatial features of different scales are captured by changing the size of the convolution kernel. i is the scale index of different scales, and k is the scale index of different scales. i are convolution kernels of different scales, It is the extracted i-scale GPR data feature map.

3. The method for detecting and locating corrosion of a substation grounding grid based on an improved optimization algorithm as claimed in claim 2, characterized in that: The specific process of extracting time characteristics from EMI data is as follows: A multi-layer dilated convolution structure is used to extract the timing features of EMI data. The formula is as follows: Among them, l is the number of convolutional layers in the convolutional structure, and lD represents the multi-expansion convolution calculation identifier, which means that at the very beginning of this spatiotemporal convolution processing, that is, layer 0, no real convolution operation is performed to change the shape of the data; in, Represents the timing characteristic diagram of the EMI data at the i-th time scale; ω i express The corresponding fusion weight can be learned through the self-attention mechanism, F EMI This is the extracted EMS data feature map.

4. The method for detecting and locating corrosion of a substation grounding grid based on an improved optimization algorithm as claimed in claim 1, characterized in that: In step 4, the specific process of applying the self-attention algorithm to the processing result of the spatiotemporal convolution is as follows: First, the feature map F EMI or F i =F final As input, it is processed through the self-attention mechanism: F global =Attention(F final ,F final ,F final This step generates a feature representation F containing global information global ; Then, the local features and global features are fused to obtain the final feature representation: F output =Concat(F final ,F global )W o Among them, W o It is a weight matrix obtained through learning, which is used to dynamically adjust the importance of local and global features. It can be determined and adjusted through random initialization operations and subsequent training algorithms including back propagation algorithms; In addition, in order to further enhance the model's ability to integrate multi-scale features, a multi-head attention mechanism is introduced. The multi-head attention mechanism learns multiple attention functions in parallel to capture information from different representation subspaces. Its formula can be expressed as: MultiHead(Q,U,V)=Concat(head1,...,head h )W o Among them, Q, U and V represent the query vector, key vector and value vector respectively, which are obtained by inputting F output It is obtained through linear transformation and is used to measure the degree of association between different elements and extract corresponding value information in the self-attention mechanism. h is the total number of attention heads, and j is the index of the corresponding weight of the vector. and A learnable weight matrix is ​​a set of parameters used to perform a linear transformation on the input data, thereby affecting the generation of query vectors, key vectors, and value vectors and the integration of the final output. They determine which parts of the input data the model focuses on and how to combine information operations. They can be determined and adjusted through random initialization operations and subsequent training algorithms including back-propagation algorithms. Get the final output head j After that, each head (head1,...,head h ) is concatenated with the Attention operation to obtain a concatenated matrix, which is then combined with the learnable weight matrix Perform matrix multiplication to obtain the final output feature map F' after multi-head attention mechanism processing; This F' is evaluated based on the feature quality including information entropy calculation and label data correlation. Information entropy reflects the amount of information contained in the feature. For the features extracted from GPR and EMI data, the lower the information entropy, the smaller the uncertainty of the feature, and there is insufficient information to effectively distinguish different grounding grid states. Evaluation and screening are performed according to the preset information entropy threshold; The label data is the GPR and EMI data of historical ground grid corrosion conditions. Through comparison, it is evaluated and screened according to the thresholds set by the staff. Among them, those that pass the evaluation are used as input for subsequent adaptive gating fusion, and those that fail the evaluation are subjected to feature enhancement.

5. The method for detecting and locating corrosion of a substation grounding grid based on an improved optimization algorithm as claimed in claim 1, characterized in that: The specific process of dynamically allocating weights to features of different dimensions through adaptive gated fusion is as follows: The feature vector F of GPR and EMI data after being processed by improved spatiotemporal convolution and self-attention mechanism respectively GPR ' and F EMI ', adaptive gated fusion will calculate a gating vector G to control the fusion ratio of the two features. The calculation of the gating vector G is as follows: G=σ(W g (F GPR :F EMI )+b g ) Among them, σ represents the sigmoid activation function, W g and b g are the learnable weight matrix and bias term respectively, and ":" indicates the feature concatenation operation.

6. The method for detecting and locating corrosion of a substation grounding grid based on an improved optimization algorithm as claimed in claim 5, characterized in that: In step 4, the specific process of fusing the feature information of the time and space dimensions of GPR and EMI is as follows: After obtaining the gate vector, the fusion process of multimodal features can be expressed as: F fused =G⊙F GPR +(1-G)⊙F EMI Among them, ⊙ represents element-wise multiplication, F fused is the final fused feature information.

7. The method for detecting and locating corrosion of a substation grounding grid based on an improved optimization algorithm as claimed in claim 6, characterized in that: In step 5, the fused feature information F fused Compare the characteristic information of the original grounding grid in a normal state stored in advance with the distance measurement method, including Euclidean distance or cosine similarity, calculate the degree of difference between the two, and judge whether the grounding grid is corroded and the location of the corrosion according to the preset threshold; At the same time, after the corrosion position is calculated, the position accuracy is further confirmed. If the position accuracy meets the requirements of subsequent operations including repair processing, the result is output. If the position accuracy does not meet the requirements of subsequent operations including repair processing, the parameters of the EMI data after convolution processing are adjusted, including adjusting the convolution kernel size, step size or filling method, to meet the requirements of accurate spatial feature data.

Images