GIS fault detection method and system based on multispectral image feature fusion

Abstract

The application provides a GIS fault detection method and system based on multispectral image feature fusion, which comprises the following steps: alternately arranging a plurality of optical filters at the output end of an optical fiber to obtain visible light images, infrared spectrum images and ultraviolet spectrum images inside a GIS; pre-training a FasterViT model using the visible light images; removing the last layer classifier of the FasterViT model and taking the remaining part as a feature extractor of the visible light images, the infrared spectrum images and the ultraviolet spectrum images; connecting the extracted features through a series operation; inputting the fused feature vector into an SCNs classifier to obtain a detection result. The application solves the technical problems that the internal fault condition of a GIS device is difficult to detect, the device state information obtained by a single monitoring method is not comprehensive enough, and the information acquisition accuracy and efficiency are relatively low.

Description

Technical Field

[0001] This invention relates to the field of GIS equipment fault detection, and specifically to a GIS fault detection method and system based on multispectral image feature fusion. Background Technology

[0002] Gas-Insulated Switchgear (GIS) is a gas-insulated, metal-enclosed switchgear using SF6 as the insulating medium, widely used in high-voltage, ultra-high-voltage, and extra-high-voltage fields. However, the enclosed nature of GIS makes it difficult to diagnose internal faults. Problems can lead to sudden, large-scale power outages, and even explosions and fires, resulting in direct and indirect economic losses in the hundreds of millions of yuan and serious negative social impacts. Furthermore, GIS suffers from long condition assessment cycles, limited monitoring methods, and weak real-time performance. The fully sealed structure of GIS makes fault location and repair difficult, complex, and results in longer average downtime repair times than conventional equipment, with larger outage areas often involving numerous non-faulty components. Therefore, analyzing and monitoring multispectral signals within GIS equipment is of great significance.

[0003] For example, the existing invention patent application document CN112417717A, entitled "A Method for Laser Focused Detection Imaging Analysis and a Readable Storage Medium for Internal Detection of GIS," includes the following methods: constructing a photo-acoustic conversion model based on a pre-established experimental platform; establishing a finite element model of laser-excited ultrasonic waves based on the photo-acoustic conversion model; extracting sound source characteristics based on the finite element model; and determining the ultrasonic information of laser focused detection imaging based on the sound source characteristics. Another example is the existing invention patent application document CN108646148A, entitled "A Precise Location Method for Fault Diagnosis of GIS Equipment Based on Photoelectric Detection Technology," which includes: gas detection: detecting the pressure of sulfur hexafluoride gas in the GIS equipment and recording the detection data; sensory detection: detecting the sound, rust, heat generation, and component damage during the operation of the GIS equipment and recording the detection data; data analysis: comprehensively analyzing the data from gas detection and sensory detection; when the data is normal, internal detection of the GIS equipment is required; and ultrasonic detection: during the internal detection process described in the data analysis, due to localized... During discharge, ultrasonic waves are generated on the outer wall of the GIS equipment cavity. Ultrasonic sensors are used to measure the partial discharge signal and record the data. Ultra-high frequency (UHF) detection: The operating GIS equipment is filled with high-pressure sulfur hexafluoride gas, which has high insulation strength and breakdown field strength. When the partial discharge is small, the gas breakdown process is rapid, generating a steep pulse current. The pulse radiates UHF electromagnetic wave signals in all directions. UHF sensors detect these signals and record the data. Comprehensive analysis: By integrating and analyzing the data from the ultrasonic and UHF detection methods, accurate fault location in the GIS equipment can be achieved. As can be seen from the specific implementation details of the aforementioned existing methods, the existing technologies for monitoring GIS equipment only use indirect detection methods to evaluate its operating status. There is no effective means to directly display its internal condition. Multi-spectral optical imaging observation using ultraviolet, visible, and infrared light inside the GIS becomes a more direct means of diagnosing defects such as partial discharge, poor contact, and overheating. Based on the above problems, how to use new monitoring methods to monitor multi-spectral imaging of internal ultraviolet, infrared, and visible light, supplement the shortcomings of existing monitoring methods, and improve the on-site status detection level of GIS equipment is a key issue in solving the causes of GIS equipment failures and abnormal states.

[0004] GIS systems have limitations in using single monitoring methods such as fiber optic image transmission, ultraviolet discharge measurement, and infrared thermometry. Ultraviolet, infrared, and visible light monitoring are only effective for certain external faults; ultraviolet discharge measurement is only effective for insulation defects caused by certain electric field distortions; and infrared thermometry monitors slowly developing faults. In practical engineering applications, due to the complex and ever-changing environmental interference and the complex internal structure of GIS equipment, single monitoring methods cannot accurately obtain comprehensive GIS equipment status information, easily leading to misjudgments and false alarms.

[0005] In summary, existing technologies suffer from technical problems such as difficulty in detecting internal faults in GIS equipment, insufficient comprehensiveness of equipment status information obtained through single monitoring methods, and low accuracy and efficiency in information acquisition. Summary of the Invention

[0006] The technical problem to be solved by this invention is: how to solve the technical problems in the prior art where it is difficult to detect internal faults in GIS equipment, the equipment status information obtained by a single monitoring method is not comprehensive enough, and the accuracy and efficiency of information acquisition are low.

[0007] This invention solves the above-mentioned technical problems by employing the following technical solution: A GIS fault detection method based on multispectral image feature fusion includes:

[0008] S1. At the output end of the image acquisition fiber optic cable, at least two filters are alternately arranged to acquire multispectral images inside the GIS equipment, including: visible light images, infrared spectral images and ultraviolet spectral images.

[0009] S2. Using visible light images, perform initialization and training operations on the FasterViT model to obtain the trained FasterViT model.

[0010] S3. Remove the last layer of the trained FasterViT model as a classifier to obtain the FasterViT feature extractor for multispectral images.

[0011] S4. Using the FasterViT feature extractor for multispectral images, extract FasterViT features from visible light images, infrared spectral images, and ultraviolet spectral images respectively.

[0012] S5. Perform a concatenation operation on the FasterViT features of the infrared and ultraviolet spectral images to obtain a fused feature vector;

[0013] S6. Input the fused feature vector into the preset SCNs classifier to process and obtain the GIS fault detection results.

[0014] This invention utilizes multispectral image feature fusion for fault detection in GIS equipment, enabling intelligent decision-making regarding fault mechanisms and states to support rapid handling of GIS faults. This invention addresses the challenges of detecting internal faults in GIS equipment and the inadequacy of single monitoring methods.

[0015] This invention fuses visible light images, infrared spectral images, and ultraviolet spectral images together and uses deep learning to detect GIS fault status, reducing the personnel processing burden required for alarm information and thus enabling rapid handling of GIS faults.

[0016] In a more specific technical solution, in step S1, a specific narrow wavelength of light transmission is designed for each filter.

[0017] This invention utilizes optical fibers and lenses to probe the interior of GIS (Gas Insulated Switchgear) to obtain images, effectively solving the problem of difficulty in detecting the internal conditions of enclosed GIS systems. By acquiring visible light, infrared, and ultraviolet spectral images, it is possible to detect problems such as partial discharge, poor contact, and overheating within the GIS, enabling a comprehensive diagnosis of its internal condition.

[0018] In a more specific technical solution, step S2 includes:

[0019] S21. Using no less than two consecutive convolutional layers, transform and process the visible light image to obtain overlapping patches. Project the overlapping patches into a D-dimensional embedding and use the embedding tokens of the D-dimensional embedding to perform batch normalization operations. Here, a ReLU activation function is set after each consecutive convolutional layer to obtain visible light embedding features.

[0020] S22. Using downsampling blocks, the spatial resolution of visible light embedding features is reduced by a preset factor, so that residual convolutional spatial features can be obtained through normalization operations and residual convolutional blocks.

[0021] S23. Using the window attention method, subglobal CTs are introduced to summarize local windows, so as to perform self-attention on connected tokens and perform local and global information exchange operations. Through the alternation of subglobal (CTs) and local window self-attention, hierarchical attention parameters are formed.

[0022] S24. This allows global information to propagate through the fully connected layer and the Softmax layer, completing the model training operation of the FasterViT model.

[0023] This invention utilizes the FastViT method when pre-training a model using visible light images, introducing a novel hierarchical attention method that can compute cross-window interactions at a lower computational cost, with better accuracy and faster speed.

[0024] In a more specific technical solution, step S22 uses the following logic to define the residual convolutional block:

[0025]

[0026] In the formula, BN represents batch normalization, and GELU represents the activation function based on the Gaussian error function.

[0027] In a more specific technical solution, step S23 includes:

[0028] S231. Using the following logic, the input feature map is divided into an n×n local window:

[0029]

[0030] Where k is the window size and H is the height of the input image, the input features are segmented using the following logic:

[0031]

[0032] in, denoted as a local window, and x represents the input feature map.

[0033] S232, Pooling with L=2 for each local window C Tokens are pooled to initialize subglobal CTs:

[0034]

[0035] In the formula, Conv 3×3 This indicates efficient location coding. AvgPool and AvgPool represent carrier tokens and feature pool operations, respectively.

[0036] S233. Using the attention logic of the following hierarchical attention block HAT, process subglobal CTs to obtain local tokens. and carrier token Calculate local tokens separately carrier token The interaction information between them is used to perform short-range spatial information modeling operations, thereby obtaining global information propagation data.

[0037] In a more specific technical solution, step S233 also includes:

[0038] S2331. Using the following logic, obtain the local token. and carrier token

[0039]

[0040] In the formula, LN represents layer normalization, MHSA represents multi-head self-attention, γ1 and γ2 are both learnable single-channel scale multipliers, and MLP d→4d→d It is a 2-layer MLP with GeLU activation function.

[0041] S2332. Connect local features and CTs to enable local windows to access corresponding subglobal CTs:

[0042]

[0043] S2333. Using the following logic, for local tokens carrier token Perform the attention process to obtain the attention token.

[0044]

[0045] S2334. Split the attention token to set up a layered attention layer:

[0046]

[0047] S2335. Iteratively execute steps S2331 to S2334 to propagate global information.

[0048] In a more specific technical solution, step S2335 utilizes the following logic for global information propagation:

[0049]

[0050] Upsample(.) is the downsampling operation, and Merge(.) is the merging operation.

[0051] In a more specific technical solution, step S6 includes:

[0052] S61. Using the following logic, suppose there exists a single hidden layer feedforward network with L-1 hidden nodes:

[0053]

[0054] In the formula, β j =[β j,1 ,...,β j,r ] T To output the weight vector, p j (·) is a random basis function. and b j These are the input weights and the bias, respectively. p j ∈[-θ,θ];

[0055] S62. Use the following logic to represent the current residual:

[0056] e L-1 (F q )=f(F q )-f L-1 (F q )=[e L-1,1 (F q ),…,e L-1,r (F q )],

[0057] e L-1,z (F q )=[e L-1,z (F q 1),…,e L-1,z (F q N )] T , z = 1, ..., r.

[0058] Where e(.) represents the residual;

[0059] S63 Order:

[0060]

[0061] Among them, a L In the Lth hidden layer output matrix, p L (·) is the Lth random basis function. and b L These are the Lth input weights and biases, respectively.

[0062] Let F represent the fused feature vector. q The Lth th The activation of each new hidden node is used to express the current hidden layer output matrix using the following logic:

[0063] A L =[a1,...,a L ]

[0064] S64. Using the preset variable w L,z During the construction process, inequality constraints are used to assign values ​​to the hidden parameters in the current hidden layer output matrix, making the residual zero, thereby generating new hidden nodes:

[0065]

[0066] In the formula, 0 < ξ < 1, u L It is an increasing sequence of non-negative real numbers, and

[0067] S65. For L hidden nodes, use the following logic to obtain the GIS fault detection results.

[0068] In a more specific technical solution, step S65 uses the following logic to obtain the GIS fault detection result:

[0069]

[0070] In the formula, For matrix G L Moore-Penrose generalized inverse, ||·|| F Given the Frobenius norm, the output labels of the video are U = {u1, ..., u}. L}

[0071] In more specific technical solutions, GIS fault detection systems based on multispectral image feature fusion include:

[0072] The multispectral image acquisition module is used to alternately arrange no less than two filters at the output end of the image acquisition fiber optic cable to acquire multispectral images inside the GIS equipment, including: visible light images, infrared spectral images and ultraviolet spectral images.

[0073] The model training module is used to initialize and train the FasterViT model using visible light images, and obtain the trained FasterViT model. The model training module is connected to the multispectral image acquisition module.

[0074] The feature extractor building module is used to remove the last layer of the trained FasterViT model as a classifier to obtain the multispectral image FasterViT feature extractor. The feature extractor building module is connected to the model training module.

[0075] The feature extraction module is used to extract FasterViT features from visible light images, infrared spectral images, and ultraviolet spectral images using the FasterViT feature extractor. The feature extraction module is connected to the feature extractor construction module.

[0076] The fusion feature acquisition module is used to perform a concatenation operation on the FasterViT features of infrared and ultraviolet spectral images to obtain a fused feature vector. The fusion feature acquisition module is connected to the feature extraction module.

[0077] The fault detection result acquisition module is used to input the fused feature vector into a preset SCNs classifier to process and obtain GIS fault detection results. The fault detection result acquisition module is connected to the fused feature acquisition module.

[0078] The present invention has the following advantages over the prior art:

[0079] This invention utilizes multispectral image feature fusion for fault detection in GIS equipment, enabling intelligent decision-making regarding fault mechanisms and states to support rapid handling of GIS faults. This invention addresses the challenges of detecting internal faults in GIS equipment and the inadequacy of single monitoring methods.

[0080] This invention fuses visible light images, infrared spectral images, and ultraviolet spectral images together and uses deep learning to detect GIS fault status, reducing the personnel processing burden required for alarm information and thus enabling rapid handling of GIS faults.

[0081] This invention utilizes optical fibers and lenses to probe the interior of GIS (Gas Insulated Switchgear) to obtain images, effectively solving the problem of difficulty in detecting the internal conditions of enclosed GIS systems. By acquiring visible light, infrared, and ultraviolet spectral images, it is possible to detect problems such as partial discharge, poor contact, and overheating within the GIS, enabling a comprehensive diagnosis of its internal condition.

[0082] This invention utilizes the FastViT method when pre-training a model using visible light images, introducing a novel hierarchical attention method that can compute cross-window interactions at a lower computational cost, with better accuracy and faster speed.

[0083] This invention solves the technical problems in the prior art, such as the difficulty in detecting internal faults in GIS equipment, the incompleteness of equipment status information obtained by a single monitoring method, and the low accuracy and efficiency of information acquisition. Attached Figure Description

[0084] Figure 1 This is a schematic diagram of the model architecture and data flow processing of the GIS fault detection method based on multispectral image feature fusion according to Embodiment 1 of the present invention;

[0085] Figure 2 This is a schematic diagram of the GIS fault detection method based on multispectral image feature fusion according to Embodiment 1 of the present invention;

[0086] Figure 3 This is a schematic diagram illustrating the specific steps of training the FasterViT model in Embodiment 1 of the present invention;

[0087] Figure 4 This is a schematic diagram illustrating the visualization principle of hierarchical attention in feature space according to Embodiment 1 of the present invention. Detailed Implementation

[0088] 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.

[0089] Example 1

[0090] like Figure 1 and Figure 2 As shown, the GIS fault detection method based on multispectral image feature fusion provided by this invention includes the following basic steps:

[0091] Step S1: Alternately arrange multiple filters at the fiber optic output end to acquire visible light images, infrared spectral images and ultraviolet spectral images inside the GIS respectively;

[0092] In this embodiment, multiple filters are alternately arranged at the fiber optic output end. Each filter is designed to allow only light of a specific narrow wavelength to pass through and block all other light. Three optical fibers are inserted into the GIS device, and each optical fiber outputs a visible light image, an infrared spectrum image, and an ultraviolet spectrum image in sequence.

[0093] In this embodiment, the image size is:

[0094] H×W×3

[0095] Where H is the height of the video image, W is the width of the video image, and 3 represents that the image is an RGB image;

[0096] Step S2: Train the initial FasterViT model using visible light images;

[0097] like Figure 3 As shown, in this embodiment, step S2 of training the FasterViT model further includes the following specific steps:

[0098] Step S21: The input visible light image is transformed into overlapping patches through two consecutive 3×3 convolutional layers, and the overlapping patches are projected into a D-dimensional embedding; the embedding tokens are further batch normalized, and the ReLU activation function is used after each convolution.

[0099] In this embodiment, the stride of each layer of a consecutive 3×3 convolutional layer can be set to, for example, 2;

[0100] Step S22: Using downsampling blocks, the spatial resolution is reduced by a factor of 2 between the two stages; then the spatial features are normalized using a 2D layer, followed by a convolutional layer with a kernel of 3×3 and a stride of 2. This is then followed by a residual convolutional block.

[0101] In this embodiment, the aforementioned residual convolution block is defined as follows:

[0102]

[0103] Among them, BN (Batch Normalization) is batch normalization, and GELU (Gaussian Error LinearUnits) is an activation function based on the Gaussian error function;

[0104] In this embodiment, the aforementioned steps S21 and S22 are repeated at least twice. See [link to relevant documentation]. Figure 1 Phase 1 and Phase 2 in the process;

[0105] Step S23: Using a novel window attention method, CTs (carrier tokens) are introduced. By performing self-attention on connected tokens, local and global information exchange is promoted. Through the alternation of subglobal (CTs) and local (window) self-attention, hierarchical attention is formed.

[0106] like Figure 4 As shown, in this embodiment, CTs (carrier tokens) are introduced based on the Swin Transformer, which serve as a summary for the entire local window. The first Conv Block is applied to the CTs to summarize and transmit global information. Then, the local window tokens and CTs are concatenated, so that each local window can only access its own set of CTs.

[0107] In this embodiment, the input feature map is divided into an n×n local window, where:

[0108]

[0109] Where k is the window size, and the division formula is:

[0110]

[0111] The key strategy employed in this embodiment is to formulate CTs, which helps to obtain a much larger attention footprint than local windows at low cost. First, pooling is performed with L=2 for each window. C Use tokens to initialize CTs:

[0112]

[0113] Among them, Conv 3×3 This indicates efficient location coding. AvgPool and AvgPool represent carrier tokens and feature pool operations, respectively.

[0114] In this embodiment, c is set to 1, which can be changed to control the delay. CTs initialization is performed only once during the parsing phase.

[0115] In this embodiment, after passing through the HAT (Hierarchical Attention) block, CTs undergo an attention procedure within each HAT block:

[0116]

[0117] Where LN represents layer normalization, MHSA (represents multi-head self attention), γ is a learnable single-channel scale multiplier, and MLP... d→4d→d It is a 2-layer MLP with GeLU activation function.

[0118] In this embodiment, in order to model short-range spatial information, local tokens were calculated respectively. and carrier token The interaction between them. First, local features and CTs are connected. Each local window can only access its corresponding CTs:

[0119]

[0120] These tokens undergo another set of attention processes:

[0121]

[0122] Finally, the token is further split and used in subsequent hierarchical attention layers:

[0123]

[0124] In this embodiment, the process described in formulas (4) to (7) is iteratively applied to many layers of this stage. Global information propagation is performed to further facilitate remote interaction. Finally, the output of this stage is calculated as follows:

[0125]

[0126] Step S24: After passing through the fully connected layer, input to the Softmax layer;

[0127] Step S3: Remove the last layer of the FasterViT model as a classifier and use the remaining part as a feature extractor for the multispectral image.

[0128] In this embodiment, multispectral images include, but are not limited to: visible light images, external spectral images, and ultraviolet spectral images;

[0129] Step S4: Using the FasterViT feature extractor from the previous step, extract features from the visible light image, infrared spectral image, and ultraviolet spectral image respectively;

[0130] Step S5: Concatenate the features extracted from the infrared spectral image and the ultraviolet spectral image;

[0131] Step S6: Input the fused feature vector into the SCNs classifier to obtain the GIS fault detection results;

[0132] In this embodiment, the fused feature vectors are:

[0133]

[0134] The input is fed into SCNs, where q is a specific convolutional layer, i = 1, ..., N;

[0135] In this embodiment, it is assumed that there exists a single hidden layer feedforward network with L-1 hidden nodes, denoted as:

[0136]

[0137] In the formula, β j =[β j,1 ,...,β j,r ] T The output weight vector; p j (·) represents a random basis function; and b j These are the input weights and the bias, respectively. p j ∈[-θ,θ];

[0138] The current residual can be expressed as:

[0139] e L-1 (F q )=f(F q )-f L-1 (F q )=[e L-1,1 (F q ),…,e L-1,r (F q )],

[0140] e L-1,z (Fq )=[e L-1,z (F q 1),…,e L-1,z (F q N )] T , z = 1, ..., r.

[0141] In this embodiment, let:

[0142]

[0143] F represents q The Lth th The activation of a new hidden node. Therefore, the current hidden layer output matrix can be represented as:

[0144] A L =[a1,...,a L ]

[0145] Using a set of variables w L,z During the construction process, inequality constraints are used to assign values ​​to the hidden parameters, and the residuals are forced to zero to generate a new hidden node:

[0146]

[0147] Among them, 0<ξ<1,u L It is an increasing sequence of non-negative real numbers, and For L hidden nodes, a standard least squares solution is available. All can be found:

[0148]

[0149] in, For matrix G L Moore-Penrose generalized inverse, ||·|| F It is a Frobenius norm. The output labels of the video are U = {u1,...,u...} L This is the result of fault diagnosis.

[0150] In summary, this invention utilizes multispectral image feature fusion for fault detection in GIS equipment, enabling intelligent decision-making regarding fault mechanisms and states to support rapid handling of GIS faults. This invention addresses the challenges of detecting internal faults in GIS equipment and the inadequacy of single monitoring methods.

[0151] This invention fuses visible light images, infrared spectral images, and ultraviolet spectral images together and uses deep learning to detect GIS fault status, reducing the personnel processing burden required for alarm information and thus enabling rapid handling of GIS faults.

[0152] This invention utilizes optical fibers and lenses to probe the interior of GIS (Gas Insulated Switchgear) to obtain images, effectively solving the problem of difficulty in detecting the internal conditions of enclosed GIS systems. By acquiring visible light, infrared, and ultraviolet spectral images, it is possible to detect problems such as partial discharge, poor contact, and overheating within the GIS, enabling a comprehensive diagnosis of its internal condition.

[0153] This invention utilizes the FastViT method when pre-training a model using visible light images, introducing a novel hierarchical attention method that can compute cross-window interactions at a lower computational cost, with better accuracy and faster speed.

[0154] This invention solves the technical problems in the prior art, such as the difficulty in detecting internal faults in GIS equipment, the incompleteness of equipment status information obtained by a single monitoring method, and the low accuracy and efficiency of information acquisition.

[0155] 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 GIS fault detection method based on multispectral image feature fusion, characterized in that, The method includes: S1. At the output end of the image acquisition fiber optic cable, at least two filters are alternately arranged to acquire multispectral images inside the GIS equipment, including: visible light images, infrared spectral images and ultraviolet spectral images. S2. Using visible light images, perform initialization and training operations on the FasterViT model to obtain the trained FasterViT model. S3. Remove the last layer of the trained FasterViT model as a classifier to obtain the FasterViT feature extractor for multispectral images. S4. Using the FasterViT feature extractor for multispectral images, extract FasterViT features from visible light images, infrared spectral images, and ultraviolet spectral images respectively. S5. Perform a concatenation operation on the FasterViT features of the infrared and ultraviolet spectral images to obtain a fused feature vector; S6. Input the fused feature vector into the preset SCNs classifier to process and obtain the GIS fault detection results; S6 includes: S61. Using the following logic, suppose there exists a... A single-hidden-layer feedforward network with 1 hidden node: （9） In the formula, To output the weight vector, For random basis functions, and These are the input weights and the bias, respectively. ; S62. Use the following logic to represent the current residual: ， ， 。 In the formula, Represents the residual; S63 Order: In the formula, No. In the hidden layer output matrix, For the first Item random basis function, and The first Input the weights and biases; To represent the fused feature vector The The activation of each new hidden node is used to express the current hidden layer output matrix using the following logic: S64. Using preset variables During the construction process, inequality constraints are used to assign values ​​to the hidden parameters in the current hidden layer output matrix, making the residual zero, thereby generating new hidden nodes: （10） In the formula, It is an increasing sequence of non-negative real numbers, and , ; S65, Regarding Given a hidden node, the following logic is used to obtain the GIS fault detection results.

2. The GIS fault detection method based on multispectral image feature fusion according to claim 1, characterized in that, In step S1, a specific narrow wavelength of light transmission is designed for each of the filters.

3. The GIS fault detection method based on multispectral image feature fusion according to claim 1, characterized in that, Step S2 includes: S21. Using no fewer than two consecutive convolutional layers, the visible light image is transformed and processed to obtain overlapping patches, and the overlapping patches are projected onto... In dimensional embedding, utilizing the Embedding of dimensional embedding Perform batch normalization, wherein after each of the consecutive convolutional layers, set Activation functions are used to obtain visible light embedding features; S22. Using a downsampling block, the spatial resolution of the visible light embedding feature is reduced by a preset factor, so as to obtain residual convolutional spatial features through normalization operation and residual convolution block. S23. Using the window attention method, subglobal CTs are introduced to summarize local windows, so as to perform self-attention on connected tokens and perform local and global information exchange operations. Through the alternation of the subglobal (CTs) and the local window self-attention, a hierarchical attention parameter is formed. S24. The global information propagation data passes through the fully connected layer and the Softmax layer to complete the model training operation of the FasterViT model.

4. The GIS fault detection method based on multispectral image feature fusion according to claim 3, characterized in that, In step S22, the residual convolutional block is defined using the following logic: （1） In the formula, BN represents batch normalization, and GELU represents the activation function based on the Gaussian error function.

5. The GIS fault detection method based on multispectral image feature fusion according to claim 3, characterized in that, Step S23 includes: S231. Using the following logic, the input feature map is divided into... The local window mentioned above: in, Let H be the window size and H be the height of the input image. The input features are then segmented using the following logic: （2） S232, Pooling for each of the local windows The tokens are pooled to initialize the subglobal CTs. （3） In the formula, This indicates efficient location coding. and These represent carrier tokens and feature pool operations, respectively. S233. Using the attention logic of the following hierarchical attention block HAT, process the subglobal CTs to obtain local tokens. and carrier token Calculate the local tokens respectively The carrier token The interaction information between them is used to perform short-range spatial information modeling operations, thereby obtaining global information propagation data.

6. The GIS fault detection method based on multispectral image feature fusion according to claim 5, characterized in that, Step S233 further includes: S2331. Obtain the local token using the following logic. and the carrier token : （4） In the formula, LN represents layer normalization, and MHSA represents bullish self-attention. Both are learnable single-channel scale multipliers. It is a 2-layer MLP with GeLU activation function. S2332. Connect local features and CTs so that the local window can access the corresponding subglobal CTs: （5） S2333. Using the following logic, the local token... The carrier token Perform attention process operations to obtain attention tokens: （6） S2334. Perform a splitting operation on the attention token to set up a layered attention layer: （7） S2335. Iteratively execute steps S2331 to S2334 to propagate global information.

7. The GIS fault detection method based on multispectral image feature fusion according to claim 6, characterized in that, In step S2335, the global information propagation is performed using the following logic: （8） In the formula, For downsampling operation, This is a merge operation.

8. The GIS fault detection method based on multispectral image feature fusion according to claim 1, characterized in that, In step S65, the GIS fault detection result is obtained using the following logic: （11） In the formula, For matrix Moore Penrose in a generalized inverse sense, The output labels for the video are Frobenius norm and are... .

9. A GIS fault detection system based on multispectral image feature fusion, used to execute the GIS fault detection method based on multispectral image feature fusion as described in any one of claims 1 to 8, characterized in that, The system includes: A multispectral image acquisition module is used to alternately arrange no less than two filters at the output end of the image acquisition optical fiber to acquire multispectral images inside the GIS equipment, wherein the multispectral images include: visible light images, infrared spectral images and ultraviolet spectral images; The model training module is used to perform initialization and training operations on the FasterViT model using visible light images, and to obtain the trained FasterViT model. The model training module is connected to the multispectral image acquisition module. A feature extractor building module is used to remove the last layer of the trained FasterViT model as a classifier to obtain a multispectral image FasterViT feature extractor. The feature extractor building module is connected to the model training module. The feature extraction module is used to extract FasterViT features from the visible light image, the infrared spectral image, and the ultraviolet spectral image using the multispectral image FasterViT feature extractor, respectively. The feature extraction module is connected to the feature extractor construction module. The fusion feature acquisition module is used to perform a concatenation operation on the FasterViT features of the infrared spectral image and the ultraviolet spectral image to obtain a fusion feature vector. The fusion feature acquisition module is connected to the feature extraction module. The fault detection result acquisition module is used to input the fused feature vector into a preset SCNs classifier to process and obtain GIS fault detection results. The fault detection result acquisition module is connected to the fused feature acquisition module.

Images