Power transformation equipment monitoring system and device based on edge calculation
Through the substation equipment monitoring system based on edge computing and the heterogeneous knowledge distillation super-resolution network model, combined with multi-source data for monitoring and image fusion, the real-time and security problems of monitoring data transmission in the existing technology are solved, and efficient, accurate and safe monitoring of substation equipment is achieved.
Patent Information
- Application Number
- CN202411410013.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-10
- Publication Date
- 2025-03-11
AI Technical Summary
When facing centralized multi-source data monitoring of substation equipment, existing substation equipment monitoring technology has problems such as large network transmission burden, poor real-time and accuracy, and noise and ionosphere interference in industrial environments affect the security of data transmission.
A substation equipment monitoring system based on edge computing is adopted, combined with multi-source data for monitoring, and a heterogeneous knowledge distillation super-resolution network model is designed for infrared-visible images fusion and super-resolution, and the data interaction is encrypted using codec methods to ensure security.
Real-time monitoring of the single operating status of the substation equipment is realized, taking into account the overall operating status of the system, improving the accuracy and safety of monitoring, reducing the system's computing burden, and improving the response speed.
Smart Images
Figure CN119599878A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of substation equipment status monitoring, and more specifically to a substation equipment monitoring system and device based on edge computing. Background Art
[0002] With the rapid development of smart grid technology, integrated substation networking is increasingly being used in urban distribution networks due to its efficient use of ground space, low sensitivity to climate and environmental changes, high reliability and strong scalability. However, frequent failures of substation equipment due to manufacturing process defects and external force damage pose a severe challenge to the stable operation of the distribution system. Therefore, in-depth research on the monitoring technology of hidden faults of substation equipment is of vital importance to ensure the stable and reliable operation of the distribution network.
[0003] At present, the technical means for substation equipment monitoring mainly include traditional methods based on data analysis and intelligent methods based on deep learning. The former evaluates data by building a regular model of normal state and fault state, while the latter uses a deep learning architecture to fit complex distribution relationships to achieve accurate analysis of operating parameters, such as partial discharge and inductance. However, in the face of the demand for centralized multi-source data monitoring of substation equipment, the complex operating environment not only increases the transmission burden of the network communication system, but also affects the real-time and accuracy of the back-end monitoring platform due to the uncertainty of the communication environment. At the same time, factors such as external noise and ionospheric interference in the industrial environment also seriously threaten the security of data transmission.
[0004] In order to improve the monitoring effect, infrared-visible light image fusion technology has become a research hotspot because it can combine the complementary advantages of two modal images (spatial texture details of visible light images and thermal target highlight information of infrared images). Under the condition of high-resolution source images, this technology has shown good fusion effect, significantly enhancing the comprehensiveness and accuracy of the image description of the scene, and providing strong support for subsequent analysis. However, when processing low-resolution input images, the performance of traditional fusion methods is often greatly reduced, and means of improving resolution such as step processing have problems such as low efficiency and accumulation of artificial effects. Summary of the invention
[0005] In response to the above problems, the purpose of the present invention is to provide a substation equipment monitoring system and device based on edge computing, which realizes the monitoring of a single operating state of the substation equipment based on edge computing and multi-source data while taking into account the overall operating status of the system. At the same time, an infrared-visible light fusion super-resolution network model with a knowledge distillation architecture is designed to improve the visualization interface, and a coding and decoding method is given to encrypt and protect data interaction.
[0006] In order to achieve the above object, the present invention is implemented through the following technical solutions: The present invention discloses a substation equipment monitoring system based on edge computing, comprising: an edge perception layer, an edge intelligent terminal and a local service platform application layer, wherein the edge intelligent terminal is respectively connected to the edge perception layer and the local service platform application layer in data; The edge perception layer is used to detect the photoelectric, infrared and environmental data of the substation equipment, encode the data, and generate the encoded data to send to the edge intelligent terminal; The edge intelligent terminal is used to collect the coded data, decode the coded data to obtain various original parameter data, store them locally, and send them to the edge intelligent terminal; The application layer of the local service platform is used to visualize infrared images, visible light images, fusion visualization, and query related data of substation equipment images based on various types of original parameter data; train the score evaluation model and the heterogeneous knowledge distillation super-resolution network based on the historical operation data of the substation equipment, and use the score evaluation model to comprehensively analyze the operating characteristics of various types of original parameter data to perform score evaluation and visualize the evaluation results.
[0007] Furthermore, the edge sensing layer includes: a photoelectric sensing unit, an inductive sensing unit, an operating environment sensing unit and a coding module; Photoelectric sensing unit, used to collect visible light images and infrared images of substation equipment as photoelectric parameters; An inductance sensing unit is used to collect static parameters and dynamic parameters of the substation equipment and calculate loss parameters based on the static parameters and dynamic parameters; the static parameters include resistance and inductance, and the dynamic parameters include harmonic current, voltage and sound pressure; Operation environment sensing unit, used to collect environmental parameters of substation equipment; The encoding module is used to quantify and encode the photoelectric parameters, static parameters, dynamic parameters, loss parameters and environmental parameters, convert them into continuous value data, and send them to the edge intelligent terminal.
[0008] Further, the calculation of the loss parameters according to the static parameters and the dynamic parameters includes: The loss parameters are calculated according to the following formula: : ; in, is the loss of the substation equipment, R is the resistance of the substation equipment winding, is the eddy current loss, is the stray loss, ILN is the sum of the square roots of the currents under different harmonics, and They are eddy current loss coefficient and stray loss coefficient, ILN, and Satisfies the following formula: ; in, and are the current and frequency of each harmonic, f1 and I1 are the first harmonic and the first current.
[0009] Furthermore, the encoding module is specifically used for: The collected photoelectric parameters, static parameters, dynamic parameters, loss parameters and environmental parameters are normalized to the range of [0,1], and the values are quantized to an integer using the following formula as quantized data: ; Among them, x is the normalized parameter data, It represents the floor operation to obtain the largest integer less than x, Q=12, Q is the quantization bit; Transpose the effective bit of the quantized data to an integer value of the power of 2 (1*k), then multiply it by the generator matrix G(k*n) to obtain the coded codeword. Through inverse quantization, the codeword is converted into continuous value data and sent to the edge intelligent terminal; where G is , k is the parameter data length, are the elements of the right half of the matrix G.
[0010] Furthermore, the edge intelligent terminal includes: an acquisition interface, a decoding module, a storage module, a peripheral interface and a communication module; The acquisition interface is used to collect continuous value data and send it to the decoding module; The decoding module is used to receive continuous value data and calculate the syndrome vector Z by multiplying the corresponding code element with the matrix GT; if Z is not equal to 0, the decoding algorithm is started, using the formula , the code data is modified and inverse quantized through the offline trained GMM model to obtain parameter data; where, is the corrected parameter value, a is the value of the i-th non-zero element of the syndrome vector Z; A storage module, used for storing parameter data; Peripheral interface, used for connecting with peripherals; The communication module is used to send parameter data to the local service platform application layer.
[0011] Furthermore, the fusion visualization uses a heterogeneous knowledge distillation super-resolution network to output the fusion result; The heterogeneous knowledge distillation super-resolution network includes: a teacher network and a student network; The teacher network includes an encoder, a fusion layer and a decoder, which are used to obtain a source image and output a fused label image; The student network includes a feature extraction module, a feature fusion layer and a super-resolution module; The loss function of the teacher network is: ; Among them, Lt, , , and They represent total loss, pixel loss, structural similarity loss, gradient loss and perceptual loss respectively, and μ is the weight parameter; The loss function of the student network is: ; in, denote the high-resolution output fusion results of the teacher network and the student network, φ is the balance parameter, T m is the feature of the mth layer in the teacher network, S n is the feature of the nth layer in the student network, and H is a set of candidate feature pairs.
[0012] Further, the encoder includes a 3x3 convolutional layer and a dense connection block, the convolutional layer is used to extract shallow features, and the dense connection block includes three 3×3 convolutional layers; The fusion layer is used to merge the features of the infrared image and the visible light image into a fusion feature in a linear weighted form; The decoder consists of four convolutional layers, and the dimensions of the four convolutional layers are 64×64, 64×32, 32×16, and 16×1. The high-resolution image output of the decoder is: ; Where W2 is the weight of the decoder, α is a balance parameter that controls the input ratio of infrared and visible light image features, and the value of α ranges from 0 to 1. , , They respectively represent the infrared source image features, visible light source image features and fusion features extracted by the encoder.
[0013] Further, the feature extraction module includes an infrared image feature extraction branch and a visible light image feature extraction branch; Infrared image feature extraction branch, specifically used for: Use a 3×3 convolutional layer to extract the low-resolution infrared source image Shallow features ; Obtaining deeper features through a residual group ; Enhance the features through the DFABs module and finally get the DFA features ; Visible light image feature extraction branch, specifically used for: Use a 3×3 convolutional layer to extract a low-resolution visible light source image Shallow features ; Obtain deeper features through two cascaded residual groups ; Extract CEA features through CEA module .
[0014] Furthermore, the DFABs module includes four DFA modules, and the DFA modules are specifically used for: Will As input, the output features are processed by two cascaded residual channel attention modules ; Where C is the channel number, H and W are the length and width of the feature map respectively; Will As input, the output is processed by three convolutional layers one by one , , ; Will and After the Hadamard dot product, the feature map is output through the softmax layer ;in, ; The output of a single DFA module, i.e., the weighted feature E, is obtained by the following formula: ; Among them, δ is a learnable weight parameter, is the dot product operation; The weighted features E output by four cascaded DFA modules and the Linear weighting obtains the DFA feature ; The CEA module includes two convolutional layers and a corner attention mechanism submodule; The corner attention mechanism submodule is specifically used for: Will As the input of the corner attention mechanism submodule, the corner attention map is obtained by the following formula : ; Where, c=1,2,...,C, i=1,2,...,W, j=1,2,...,H, Indicated The feature map of the cth channel, Harris (*) indicates Harris algorithm processing, and γ is a coefficient between 0 and 1; The The features are obtained by outputting the first convolutional layer ; The output of the corner attention mechanism submodule, namely the corner feature map, is obtained by the following formula : ; Among them, β is initialized to 0; The The CEA features are obtained through the output of the second convolutional layer .
[0015] Furthermore, the feature extraction module is specifically used for: definition and The features of infrared images extracted by edge photoelectric sensing units are and visible light image features ; Where C is the channel number, H and W are the length and width of the feature map respectively; Calculated by the following formula and The L1 norm graph of and : ; Where, i=1,2,...,W, j=1,2,...,H; The fusion weight coefficient is calculated by the following formula : ; in, is the mean of the L1 norm graph; The fusion feature is calculated by the following formula : ; in, Represents the feature map of the cth channel of the fusion feature, c=1,2,...,C, is the dot product operation.
[0016] Correspondingly, the present invention also discloses a substation equipment monitoring device based on edge computing, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the function of the substation equipment monitoring system based on edge computing as described in any one of the above items is realized.
[0017] Compared with the prior art, the present invention has the following beneficial effects: 1. The present invention integrates the data interface of the existing single-source monitoring system and adjusts the interface and data format at the edge terminal to meet the data transmission and analysis and computing requirements of the edge computing device.
[0018] 2. The present invention adopts a coding and decoding method that can effectively defend against adversarial attacks on data-driven models of industrial monitoring systems. The coding and decoding scheme is universal and independent of specific machine learning models. Therefore, it can be applied to substation equipment monitoring systems in various scenarios. This enables the method to adapt to different application scenarios without relying on the details of a specific model.
[0019] 3. The heterogeneous knowledge distillation super-resolution network of the present invention can simultaneously complete the fusion and super-resolution of infrared and visible light images, solving the problem of poor fusion quality caused by the lack of high-resolution fusion image labels in previous methods. By introducing the teacher-student network structure, the teacher network generates high-resolution fusion image labels and uses them to guide the learning of the student network, realizing the joint optimization of fusion and super-resolution. It further realizes the visualization of digital twins, enabling operators to monitor, calculate, analyze and manage the substation equipment monitoring system.
[0020] It can be seen that compared with the prior art, the present invention has outstanding substantive features and significant progress, and the beneficial effects of its implementation are also obvious. BRIEF DESCRIPTION OF THE DRAWINGS
[0021] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.
[0022] Figure 1 It is a system structure diagram of a specific implementation mode of the present invention.
[0023] Figure 2 It is a functional schematic diagram of the encoding module and the decoding module of a specific implementation mode of the present invention.
[0024] Figure 3 It is a structural principle diagram of a heterogeneous knowledge distillation super-resolution network according to a specific implementation mode of the present invention.
[0025] Figure 4 It is a structural principle diagram of a DFABs module according to a specific embodiment of the present invention.
[0026] Figure 5 It is a structural principle diagram of a CEA module according to a specific implementation mode of the present invention.
[0027] Figure 6 It is a schematic diagram of the display interface of the local service platform application layer of a specific implementation mode of the present invention.
[0028] Figure 7 It is a schematic diagram of a heterogeneous knowledge distillation super-resolution network training curve according to a specific implementation mode of the present invention.
[0029] Figure 8 It is a schematic diagram of the fusion result of the heterogeneous knowledge distillation super-resolution network according to a specific implementation mode of the present invention. DETAILED DESCRIPTION
[0030] In order to enable those skilled in the art to better understand the scheme of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific implementation methods. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.
[0031] See also Figure 1 As shown, this embodiment provides a substation equipment monitoring system based on edge computing, including: an edge perception layer, an edge intelligent terminal and a local service platform application layer, and the edge intelligent terminal is respectively connected to the edge perception layer and the local service platform application layer.
[0032] The edge sensing layer is used to detect the photoelectric, infrared and environmental data of the substation equipment, encode the data, and generate the encoded data to send to the edge intelligent terminal. The edge sensing layer contains several photoelectric sensing units, inductive sensing units, and operating environment sensing units to collect relevant data information, and each sensing unit is equipped with a coding module.
[0033] Among them, the photoelectric sensing unit is used to collect visible light images and infrared images of the substation equipment as photoelectric parameters; the inductance sensing unit is used to collect static parameters and dynamic parameters of the substation equipment, and calculate the loss parameters based on the static parameters and dynamic parameters; the static parameters include resistance and inductance, and the dynamic parameters include harmonic current, voltage and sound pressure; the operating environment sensing unit is used to collect environmental parameters of the substation equipment; the encoding module is used to quantize and encode the photoelectric parameters, static parameters, dynamic parameters, loss parameters and environmental parameters, convert them into continuous value data, and send them to the edge intelligent terminal.
[0034] The edge intelligent terminal is used to collect coded data, decode the coded data to obtain various original parameter data, store them locally, and send them to the edge intelligent terminal. The edge intelligent terminal includes a storage module, a peripheral interface, a collection interface, a communication module, and a decoding module.
[0035] The application layer of the local service platform is used to visualize infrared images, visible light images, fusion visualization, and query related data of substation equipment images based on various types of original parameter data; train the score evaluation model and the heterogeneous knowledge distillation super-resolution network based on the historical operation data of the substation equipment, and use the score evaluation model to comprehensively analyze the operating characteristics of various types of original parameter data to perform score evaluation and visualize the evaluation results.
[0036] First, various sensor devices at the edge perception layer detect the relevant operating data of the substation equipment, and collect, preliminarily process and store the data through the edge intelligent terminal. Then, various data are sent to the local service platform through the communication module of the edge intelligent terminal, and various functions such as real-time monitoring, visualization, query, analysis and statistics of the operating data are realized on the platform. Two important functions are to train the score evaluation model based on the historical operating data of the substation equipment and to train the heterogeneous knowledge fusion super-resolution network. After obtaining the model, the relevant parameters are distributed to the computing unit of the edge intelligent terminal through the downlink data channel. The edge computing unit uses the intelligent algorithm issued by the local service platform to comprehensively analyze the operating characteristics of the substation equipment using the real-time operating data of the substation equipment, and finally realizes the functions of visualization and score evaluation. The data interaction is improved through the codec module to improve security.
[0037] This system can divide the training and actual operation of the visualization fusion and evaluation algorithm of substation equipment into local service platforms and edge intelligent units. Compared with traditional centralized monitoring methods, it reduces the system computing burden and improves monitoring reliability and response speed.
[0038] In a specific implementation, the photoelectric sensing unit is mainly responsible for collecting visible light images and infrared images; the inductive sensing unit is mainly responsible for collecting static parameters such as resistance and inductance of the transformer winding, dynamic parameters such as harmonic current, voltage, sound pressure and loss parameters; the operating environment sensing unit is mainly responsible for collecting environmental parameters such as air pressure and ambient water level.
[0039] The loss parameters specifically include: ; in, is the loss of the substation equipment, R is the resistance of the substation equipment winding, is the eddy current loss, is the stray loss, ILN is the sum of the square roots of the currents under different harmonics, and They are eddy current loss coefficient and stray loss coefficient respectively. The latter three parameters satisfy the following formula: ; Among them, and are the current and frequency of each harmonic, f1 and I1 refer to the first harmonic and current.
[0040] In the specific implementation, the security of industrial data interaction also needs to be taken seriously, because the attacker's goal is to achieve the purpose of destruction by tampering with industrial data, rather than directly attacking physical facilities. This type of adversarial data attack is easier to implement and more difficult to be discovered by regulators, such as disturbing noise or tampering with fake information. Therefore, this system sets up encoding modules and decoding modules at the edge perception layer and edge intelligent terminal respectively to ensure the security of data interaction.
[0041] like Figure 2 As shown, the specific functions of the encoding module and the decoding module are as follows: The encoding module is specifically used for: Quantization: First, normalize the collected parameter data to the range of [0,1], and quantize the value to an integer using the following formula: ; Among them, x is the normalized parameter data, represents the floor operation performed to obtain the largest integer less than x, Q=12, which is the quantization bit.
[0042] Coding: The first half of the significant bits (MSB) of the quantized data are transposed to an integer value of a power of 2 (1*k), and then multiplied by the generator matrix G (k*n) to obtain the encoded codeword. Through inverse quantization, the codeword is converted into continuous value data and sent to the edge intelligent terminal.
[0043] Among them, G is , k is the parameter data length, are the elements of the right half of the matrix G.
[0044] The decoding module is specifically used for: The edge intelligent terminal receives the instruction from the edge sensing device, and the code element is multiplied by GT to obtain the syndrome vector Z. If Z is not equal to 0, the decoding algorithm is started. The following formula is used to modify the code element data through the offline trained GMM model and then inverse quantize to obtain the parameter data.
[0045] ; in, is the corrected parameter value, and a is the value of the i-th non-zero element of the syndrome vector Z.
[0046] In a specific implementation, the image visualization specifically includes infrared image visualization, visible light image visualization, and fusion visualization. The fusion visualization method uses a fusion result with clear resolution output by a heterogeneous knowledge distillation super-resolution network.
[0047] It should be noted that some existing methods attempt to achieve image fusion and super-resolution at the same time, but their fusion performance is limited due to the lack of high-resolution fusion images as guidance. Infrared images and visible light images can capture complementary information under different lighting conditions. Integrating this information can improve the comprehensiveness and accuracy of scene descriptions and help subsequent visualization tasks. To overcome the above problems, this system designs a heterogeneous knowledge distillation super-resolution network, which contains multiple layers of attention embedding and can simultaneously achieve fusion and super-resolution of infrared and visible light images of substation equipment.
[0048] like Figure 3 As shown, the heterogeneous knowledge distillation super-resolution network specifically includes a teacher network and a student network.
[0049] The teacher network consists of an encoder, a fusion layer and a decoder. A high-definition fused label image is obtained by inputting a high-resolution source image into a pre-trained teacher network. The student network consists of a feature extraction module, a feature fusion layer and a super-resolution module. A high-resolution fused image can be obtained by feeding a low-resolution source image into the student network, and the fusion features and fusion results of the high-resolution source image in the teacher network are used to guide the training of the student network. Unlike conventional networks, the parameters of the teacher network of the present invention are less than those of the student network.
[0050] The teacher network mentioned above includes: The encoder in the teacher network contains a 3x3 convolutional layer and a densely-connected-block, where the convolutional layer is used to extract shallow features and the densely-connected block contains three 3×3 convolutional layers, and the output of each layer is cascaded to the next layer as input; The fusion layer in the teacher network combines the features of infrared and visible light images into a fused feature in a linear weighted form.
[0051] The decoder in the teacher network consists of four convolutional layers with dimensions of 64×64, 64×32, 32×16, and 16×1. The high-resolution image output of the decoder is: ; Where W2 is the weight of the decoder, α is a balancing parameter that controls the input ratio of infrared and visible light image features, and the value of α ranges from 0 to 1. , , They respectively represent the infrared source image features, visible light source image features and fusion features extracted by the encoder.
[0052] The student network mentioned above includes: The feature extraction module in the student network includes an infrared image feature extraction branch and a visible light image feature extraction branch, which are designed for infrared images and visible light images respectively.
[0053] Among them, the infrared image feature extraction branch includes: Initial stage: Use a 3×3 convolutional layer to extract a low-resolution infrared source image Shallow features .
[0054] Deep feature extraction: Obtaining deeper features through a residual group .
[0055] DFA features: Enhance the features through the DFABs module and finally get the DFA features .
[0056] Among them, the visible light image feature extraction branch includes: Initial stage: Also use a 3×3 convolution layer to extract a low-resolution visible light source image Shallow features .
[0057] Deep feature extraction: Obtain deeper features through two cascaded residual groups .
[0058] CEA feature extraction: Extract CEA features through the CEA module .
[0059] Since the frequency of image features changes with the depth of the network. If the features of adjacent network layers are combined and the correlation between features of different frequencies is extracted, the correlation can enhance the image features. In infrared images, intensity represents thermal radiation information. Inspired by the knowledge that features of different frequencies have complementary information, the features of adjacent network layers are fused together by the Hadamard product. Then, the DFA map is obtained by softmax. As the network is trained, the correlation of features from adjacent layers will be adaptively encoded into the DFA map. By weighting the features of the infrared image using the DFA map, more significant weights will be given to areas with higher intensity to enhance the thermal radiation information.
[0060] Based on the above ideas, the present invention designs DFAG, such as Figure 4 As shown, the DFABs module includes four DFA modules, wherein the DFA module includes: 1) As input, the output features are processed by two cascaded residual channel attention modules (RCAB) modules , C is the channel number, H and W are the length and width of the feature map respectively.
[0061] 2) As input, the output is processed by three convolutional layers one by one , , .
[0062] 3) and After the Hadamard dot product, it is output through the softmax layer Feature map.
[0063]
[0064] 4) The output of a single DFA module, i.e., the weighted feature E, is obtained through the following formula.
[0065]
[0066] Among them, δ is a learnable weight parameter, is the dot product operation.
[0067] 5) The weighted features E output by four cascaded DFA modules and the Linear weighting obtains the DFA feature .
[0068] For visible light images, if more attention is paid to features corresponding to textures and edges, it will be beneficial to retain the saliency information of the source image, and better feature representation can be obtained by capturing the distribution relationship of corner points between objects in the image. Therefore, the present invention designs a CEA module to predict the weights of visible image features at different spatial corner point positions.
[0069] like Figure 5 As shown, the CEA module specifically includes two convolutional layers and a corner attention mechanism sub-module.
[0070] The As the input of the CEA module, it passes through the first 1×1 convolutional layer output Then it is fed into three sub-modules.
[0071] The corner attention mechanism submodules include: 1) As The input of the corner attention mechanism submodule is obtained by the following formula to obtain the corner attention map ; ; Where, c=1,2,...,C, i=1,2,...,W, j=1,2,...,H, Indicated The feature map of the c-th channel, Harris (*) indicates Harris algorithm processing, and γ is a coefficient between 0 and 1.
[0072] 2) The features are obtained through the 1×1 convolution layer output ; 3) The output of the corner attention mechanism submodule, i.e. the corner feature map, is obtained by the following formula: .
[0073] ; Among them, β is initialized to 0, and as the network's ability to extract features improves, the value will be assigned a larger value.
[0074] 4) The CEA feature is obtained through the output of the second 1×1 convolutional layer .
[0075] After obtaining two sets of features, they need to be fully combined. The feature fusion module is: 1) Definition and are the features of infrared image and visible light image extracted by edge photoelectric sensing unit, namely and , C is the channel number, H and W are the length and width of the feature map respectively; 2) Calculate using the following formula and The L1 norm graph of and ; ; Among them, i=1,2,...,W, j=1,2,...,H.
[0076] 3) Calculate the fusion weight coefficient using the following formula ; ; in, is the mean of the L1 norm graph.
[0077] 4) The fusion feature is calculated by the following formula .
[0078] ; in, Represents the feature map of the cth channel of the fusion feature, c=1,2,...,C, is the dot product operation.
[0079] In a specific implementation, the loss function of the teacher network is: ; Among them, Lt, , , and They represent total loss, pixel loss, structural similarity loss, gradient loss and perceptual loss respectively, and μ is the weight parameter.
[0080] The loss function of the student network is: ; in, They represent the high-resolution output fusion results of the teacher network and the student network respectively, φ is the balance parameter, Tm is the feature of the mth layer in the teacher network, Sn is the feature of the nth layer in the student network, and H is a set of candidate feature pairs.
[0081] In a specific implementation, the local service platform application layer may adopt a server-client (BS) architecture, wherein the server is responsible for model delivery and training and interacting with the client and edge terminals, and the client is responsible for interface display and interaction with the server.
[0082] like Figure 6 As shown, the client display interface includes: model training progress and setting interface, comprehensive monitoring and evaluation score display interface, infrared-visible light-fusion three-category switchable visualization display interface for substation equipment, real-time data display interface for multi-source operating parameters, and operation and maintenance status display interface.
[0083] As an example, the score evaluation model of this system can select the SVM model, and the measurement data set is collected from the physical system, totaling 100000 × 43. Among them, there are 14 static parameters and 22 non-static device parameters, which are sent to the edge intelligent terminal after being trained by the application layer of the local service platform.
[0084] As an example, the training curve of the heterogeneous knowledge distillation super-resolution network of this system is as follows: Figure 7 As shown, Figure 7 (a) shows the convergence curve of the training loss during the total student network training process under 4 times magnification, from Figure 7 (a) It can be seen that as the number of training times increases, the training loss gradually decreases. When it reaches 200 epochs, the loss curve converges from Figure 7(a) It can be seen that as the number of training times increases, the training loss gradually decreases. When it reaches 200 epochs, the loss curve converges, and the curve enlarged on the validation data set is as follows Figure 7 (b). Figure 7 As can be seen in (b), the verification accuracy improves with the increase in the number of training times, which indicates that with the increase in training, the difference between the fused images of the student network and the teacher network becomes smaller and smaller, as shown in Figure 8 As shown, the fusion results can clearly show signs of fever.
[0085] The present invention also provides a substation equipment monitoring device based on edge computing, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the function of the substation equipment monitoring system based on edge computing as described in any one of the above items is realized.
[0086] To sum up, the present invention divides the training and actual operation of the intelligent identification algorithm for the operating status of substation equipment groups into the local service platform and the edge intelligent unit. Compared with the traditional centralized monitoring method, it reduces the system computing burden, improves the monitoring reliability and the response speed to loss perception, and can effectively ensure the safe and stable operation of substation equipment.
[0087] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. As for the method disclosed in the embodiment, since it corresponds to the system disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.
[0088] The above is a detailed introduction to the substation equipment monitoring system and device based on edge computing provided by the present invention. This article uses specific examples to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea. It should be pointed out that for ordinary technicians in this technical field, without departing from the principles of the present invention, several improvements and modifications can be made to the present invention, and these improvements and modifications also fall within the scope of protection of the present invention.
Claims
1. A substation equipment monitoring system based on edge computing, characterized in that: include: Edge perception layer, edge intelligent terminal and local service platform application layer. The edge intelligent terminal is connected with the edge perception layer and the local service platform application layer respectively. The edge perception layer is used to detect the photoelectric, infrared and environmental data of the substation equipment, encode the data, and generate the encoded data to send to the edge intelligent terminal; The edge intelligent terminal is used to collect the coded data, decode the coded data to obtain various original parameter data, store them locally, and send them to the edge intelligent terminal; The application layer of the local service platform is used to visualize infrared images, visible light images, fusion visualization, and query related data of substation equipment images based on various types of original parameter data; train the score evaluation model and the heterogeneous knowledge distillation super-resolution network based on the historical operation data of the substation equipment, and use the score evaluation model to comprehensively analyze the operating characteristics of various types of original parameter data to perform score evaluation and visualize the evaluation results.
2. The substation equipment monitoring system based on edge computing according to claim 1 is characterized in that: The edge sensing layer includes: a photoelectric sensing unit, an inductive sensing unit, an operating environment sensing unit and a coding module; Photoelectric sensing unit, used to collect visible light images and infrared images of substation equipment as photoelectric parameters; An inductance sensing unit is used to collect static parameters and dynamic parameters of the substation equipment and calculate loss parameters based on the static parameters and dynamic parameters; the static parameters include resistance and inductance, and the dynamic parameters include harmonic current, voltage and sound pressure; The operating environment sensing unit is used to collect environmental parameters of the substation equipment; The encoding module is used to quantify and encode the photoelectric parameters, static parameters, dynamic parameters, loss parameters and environmental parameters, convert them into continuous value data, and send them to the edge intelligent terminal.
3. The substation equipment monitoring system based on edge computing according to claim 2 is characterized in that: The encoding module is specifically used for: The collected photoelectric parameters, static parameters, dynamic parameters, loss parameters and environmental parameters are normalized to the range of [0,1], and the values are quantized to an integer using the following formula as quantized data: ； Among them, x is the normalized parameter data, It represents the floor operation to obtain the largest integer less than x, Q=12, Q is the quantization bit; Transpose the effective bit of the quantized data to an integer value of the power of 2 (1*k), then multiply it by the generator matrix G(k*n) to obtain the coded codeword. Through inverse quantization, the codeword is converted into continuous value data and sent to the edge intelligent terminal; where G is , k is the parameter data length, are the elements of the right half of the matrix G.
4. The substation equipment monitoring system based on edge computing according to claim 3 is characterized in that: The edge intelligent terminal includes: an acquisition interface, a decoding module, a storage module, a peripheral interface and a communication module; The acquisition interface is used to collect continuous value data and send it to the decoding module; The decoding module is used to receive continuous value data and calculate the syndrome vector Z by multiplying the corresponding code element with the matrix GT; if Z is not equal to 0, the decoding algorithm is started, using the formula , the code data is modified and inverse quantized through the offline trained GMM model to obtain parameter data; where, is the corrected parameter value, a is the value of the i-th non-zero element of the syndrome vector Z; A storage module, used for storing parameter data; Peripheral interface, used for connecting with peripherals; The communication module is used to send parameter data to the local service platform application layer.
5. The substation equipment monitoring system based on edge computing according to claim 1 is characterized in that: The fusion visualization uses a heterogeneous knowledge distillation super-resolution network to output the fusion result; The heterogeneous knowledge distillation super-resolution network includes: a teacher network and a student network; The teacher network includes an encoder, a fusion layer and a decoder, which are used to obtain a source image and output a fused label image; The student network includes a feature extraction module, a feature fusion layer and a super-resolution module; The loss function of the teacher network is: ; Among them, Lt, , , and They represent total loss, pixel loss, structural similarity loss, gradient loss and perceptual loss respectively, and μ is the weight parameter; The loss function of the student network is: ; in, , They represent the high-resolution output fusion results of the teacher network and the student network, φ is the balance parameter, T m is the feature of the mth layer in the teacher network, S n is the feature of the nth layer in the student network, and H is a set of candidate feature pairs.
6. The substation equipment monitoring system based on edge computing according to claim 5 is characterized in that: The encoder includes a 3x3 convolutional layer and a dense connection block, the convolutional layer is used to extract shallow features, and the dense connection block includes three 3×3 convolutional layers; The fusion layer is used to merge the features of the infrared image and the visible light image into a fusion feature in a linear weighted form; The decoder consists of four convolutional layers, and the dimensions of the four convolutional layers are 64×64, 64×32, 32×16, and 16×1. The high-resolution image output of the decoder is: ； Where W2 is the weight of the decoder, α is a balance parameter that controls the input ratio of infrared and visible light image features, and the value of α ranges from 0 to 1. , , They respectively represent the infrared source image features, visible light source image features and fusion features extracted by the encoder.
7. The substation equipment monitoring system based on edge computing according to claim 6 is characterized in that: The feature extraction module includes an infrared image feature extraction branch and a visible light image feature extraction branch; Infrared image feature extraction branch, specifically used for: Use a 3×3 convolutional layer to extract the low-resolution infrared source image Shallow features ; Obtaining deeper features through a residual group ; Enhance the features through the DFABs module and finally get the DFA features ; Visible light image feature extraction branch, specifically used for: Use a 3×3 convolutional layer to extract a low-resolution visible light source image Shallow features ; Obtain deeper features through two cascaded residual groups ; Extract CEA features through CEA module .
8. The substation equipment monitoring system based on edge computing according to claim 7 is characterized in that: The DFABs module includes four DFA modules, and the DFA modules are specifically used for: Will As input, the output features are processed by two cascaded residual channel attention modules ; Where C is the channel number, H and W are the length and width of the feature map respectively; Will As input, the output is processed by three convolutional layers one by one , , ; Will and After the Hadamard dot product, the feature map is output through the softmax layer ;in, ; The output of a single DFA module, i.e., the weighted feature E, is obtained by the following formula: ； Among them, δ is a learnable weight parameter, is the dot product operation; The weighted features E output by four cascaded DFA modules and the Linear weighting obtains the DFA feature ; The CEA module includes two convolutional layers and a corner attention mechanism submodule; The corner attention mechanism submodule is specifically used for: Will As the input of the corner attention mechanism submodule, the corner attention map is obtained by the following formula : ； Where, c=1,2,...,C, i=1,2,...,W, j=1,2,...,H, Indicated The feature map of the cth channel, Harris (*) indicates Harris algorithm processing, and γ is a coefficient between 0 and 1; The The features are obtained by outputting the first convolutional layer ; The output of the corner attention mechanism submodule, namely the corner feature map, is obtained by the following formula : ； Among them, β is initialized to 0; The The CEA features are obtained through the output of the second convolutional layer .
9. The substation equipment monitoring system based on edge computing according to claim 8 is characterized in that: The feature extraction module is specifically used for: definition and The features of infrared images extracted by edge photoelectric sensing units are and visible light image features ; Where C is the channel number, H and W are the length and width of the feature map respectively; Calculated by the following formula and The L1 norm graph of and : ； Where, i=1,2,...,W, j=1,2,...,H; The fusion weight coefficient is calculated by the following formula : ； in, is the mean of the L1 norm graph; The fusion feature is calculated by the following formula : ； in, Represents the feature map of the cth channel of the fusion feature, c=1,2,...,C, is the dot product operation.
10. A substation monitoring device based on edge computing, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the functions of the substation equipment monitoring system based on edge computing as described in any one of claims 1 to 9 are realized.