A power equipment identification method based on deep learning

Abstract

The application belongs to the technical field of power system operation and maintenance, and specifically discloses a power equipment identification method based on deep learning. The method generates a pre-training lightweight model based on a constructed power equipment knowledge graph and a teacher-student model, and simultaneously uses a semi-supervised pseudo-label iteration strategy combined with manual review and an adaptive threshold to solve the problem of new equipment data scarcity, uses unlabeled data in a data-driven manner to continuously improve the identification accuracy of the model for new equipment, and reduces the dependence on high-cost manual labeling data. By constructing a power equipment spatial relationship graph, the isolated equipment detection task is converted into an inference task for structured graph data, so that the model explicitly learns and uses the topological layout and functional dependency relationship between equipment, reduces the misidentification rate caused by the independent detection of a single equipment, and improves the accuracy of operation and maintenance decision-making.

Description

Technical Field

[0001] This invention belongs to the field of power system operation and maintenance technology, specifically relating to a deep learning-based method for identifying power equipment. Background Technology

[0002] As the core carrier of the safe and stable operation of the power grid, the real-time monitoring and accurate identification of the operating status of power equipment is a key link in ensuring the reliability of the power system and reducing operation and maintenance costs. Traditional power equipment inspection relies on manual on-site inspection, which suffers from low efficiency, high labor intensity, great susceptibility to environmental influences, and high rates of missed inspections and misjudgments, making it difficult to meet the operation and maintenance needs of large-scale power grids. With the breakthroughs in deep learning technology in the field of computer vision, deep learning-based power equipment identification methods are gradually becoming the core technology direction to replace manual inspection. By analyzing power equipment inspection images captured by drones and fixed cameras, it can automatically locate the equipment position, identify the equipment type, and determine the operating status, significantly improving inspection efficiency and reducing labor costs. At the same time, it can achieve early defect warning through quantitative visual feature analysis, avoiding power grid accidents caused by equipment failures, which has important practical significance for ensuring the safe operation of the power system.

[0003] Currently, the mainstream solution for power equipment identification based on deep learning usually involves training the model on a large-scale labeled dataset, and power equipment often exists in clusters in inspection images.

[0004] However, the existing technical solutions still have some inherent defects: (1) New types of power equipment that are constantly emerging in the power system have a short application time, very few labeled samples, and are not included in the prior knowledge system of traditional power equipment. As a result, the model is prone to insufficient learning of the features of new equipment, reduced recognition accuracy, and inability to meet the operation and maintenance requirements.

[0005] (2) Existing identification methods only detect individual devices independently without considering the spatial relationship between devices. This makes it easy to misclassify adjacent similar devices as the same category, resulting in a high misidentification rate and affecting the accuracy of operation and maintenance decisions. Summary of the Invention

[0006] In view of this, in order to solve the problems mentioned in the background technology, a deep learning-based power equipment identification method is proposed.

[0007] The objective of this invention can be achieved through the following technical solution: This invention provides a deep learning-based method for identifying power equipment, comprising: constructing a power equipment knowledge graph and a teacher-student model, training the teacher model by combining the power equipment knowledge graph with a power equipment dataset, and transferring the learned common features to the student model through knowledge distillation to obtain a pre-trained lightweight model.

[0008] We acquire labeled and unlabeled datasets of new power equipment. Based on the labeled dataset, we fine-tune the pre-trained lightweight model to obtain the basic recognition model. Based on this model, we infer high-confidence pseudo-label samples from the unlabeled dataset. After verification, we merge the samples with the labeled dataset to iteratively fine-tune the basic recognition model.

[0009] A differential feature attention module is embedded in the basic recognition model to form a lightweight device basic recognition model.

[0010] The inspection image of the power equipment to be identified is input into the lightweight equipment basic recognition model, and the candidate regions of the power equipment are output.

[0011] Based on candidate regions of power equipment, a spatial relationship map of power equipment is constructed, and iterative relationship reasoning is performed on it, updating the node feature vectors in real time.

[0012] The final node feature vector is input into the classifier, which outputs the final accurate category. The initial category confidence and the final accurate category are combined to complete the type identification and status analysis of all power equipment in the power equipment inspection image.

[0013] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. The present invention generates a pre-trained lightweight model based on the constructed knowledge graph of power equipment and the teacher-student model. At the same time, the semi-supervised pseudo-label iteration strategy combined with manual review and adaptive threshold solves the problem of data scarcity for new equipment, provides a solid knowledge foundation for the model, and effectively utilizes unlabeled data. It continuously improves the model's recognition accuracy for new equipment in a data-driven manner and reduces the dependence on high-cost manual labeling data.

[0014] 2. This invention transforms isolated equipment detection tasks into reasoning tasks based on structured graph data by constructing a spatial relationship map of power equipment. This enables the model to explicitly learn and utilize the topological layout and functional dependencies between equipment, reducing the misidentification rate caused by independent detection of individual equipment, which easily leads to misclassifying adjacent similar equipment as the same category, and improving the accuracy of operation and maintenance decisions. Attached Figure Description

[0015] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a schematic diagram illustrating the implementation steps of the method of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and 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.

[0018] Please see Figure 1 As shown, this invention provides a deep learning-based method for identifying power equipment. The specific steps are as follows: construct a power equipment knowledge graph and a teacher-student model; train the teacher model by combining the power equipment knowledge graph with the power equipment dataset; and transfer the learned common features to the student model through knowledge distillation to obtain a pre-trained lightweight model.

[0019] In one feasible embodiment of the present invention, the specific process of the power equipment knowledge graph and teacher-student model includes: extracting physical parameters of several traditional power equipment from power industry standards, simultaneously extracting structural features of several traditional power equipment from power equipment operation and maintenance ledgers, fusing physical parameters and structural features to form an initial knowledge graph, and obtaining the power equipment knowledge graph after redundancy removal.

[0020] In one specific example, the physical parameters include device size proportions and material properties, which are stored in a standardized data format; the structural features include component locations and component quantities, which are stored in a structured data format.

[0021] The teacher model was built using a ResNet50 convolutional neural network, and the student model was built using a MobileNetV3 lightweight convolutional neural network. This model is referred to as the teacher-student model.

[0022] In one feasible embodiment of the present invention, the specific process of obtaining the pre-trained lightweight model includes: converting the power equipment knowledge graph into a knowledge constraint vector and inputting it into the teacher model.

[0023] Specifically, the process of converting the power equipment knowledge graph into a knowledge constraint vector includes: independently encoding discrete attributes such as material (cast iron, stainless steel), normalizing continuous parameters such as size ratio, encoding component relationships such as position (right side) through spatial vectors such as relative coordinate mapping, and finally merging the above into a fixed-dimensional knowledge constraint vector, such as a 512-dimensional vector.

[0024] The teacher model is trained using a dataset of power equipment stored in the database. The teacher model learns visual features (extracted from images) and knowledge constraints (extracted from knowledge vectors) simultaneously. The classification head is optimized using a cross-entropy loss function, enabling the teacher model to accurately identify power equipment categories. The feature output head generates common features of power equipment containing knowledge constraints, such as a transformer = specific size + cast iron material + terminals on top.

[0025] It should be noted that the power equipment dataset stored in the database is specifically a public dataset of traditional power equipment, such as a power equipment inspection image dataset or a substation equipment annotation dataset. The dataset must cover all equipment types in the knowledge graph and include samples from different angles, lighting conditions, and operating conditions.

[0026] The common features learned by the teacher model are transferred to the student model by using the KL divergence distillation loss function, resulting in a pre-trained lightweight model.

[0027] It should be noted that the KL divergence distillation loss function is used as the core loss to measure the difference in output distribution between the teacher model and the student model. The teacher output is a soft label, which contains more class probability information. At the same time, the auxiliary cross-entropy loss is used to constrain the classification accuracy of the student model with the true label.

[0028] Specifically, the process of transferring the common features learned by the teacher model to the student model includes the following: fixing the parameters of the teacher model and using the feature output head of the teacher model as a guidance signal; while learning image features, the student model aligns its common feature distribution with that of the teacher model through the KL divergence distillation loss function, such as forcing the student model's feature extraction result for the transformer to be close to the output of the teacher model; and finally training to obtain a pre-trained lightweight model with lightweight parameters but retaining the ability to extract common features.

[0029] It should be added that the training objective of the student model is to minimize the KL divergence distillation loss function and the possible cross-entropy loss with the real hard labels, so as to transfer the common features of knowledge enhancement learned by the teacher model to itself, and finally obtain a pre-trained lightweight model that is both lightweight and rich in prior knowledge.

[0030] We acquire labeled and unlabeled datasets of new power equipment. Based on the labeled dataset, we fine-tune the pre-trained lightweight model to obtain the basic recognition model. Based on this model, we infer high-confidence pseudo-label samples from the unlabeled dataset. After verification, we merge the samples with the labeled dataset to iteratively fine-tune the basic recognition model.

[0031] It should be noted that the aforementioned new power equipment specifically refers to power equipment that is not included in the power equipment knowledge graph and has been put into application, such as intelligent inspection robots, new energy grid-connected inverters, solid-state circuit breakers, and digital instrument transformers.

[0032] The labeled dataset specifically refers to a new type of power equipment dataset obtained by manually annotating the power company's new equipment grid connection acceptance records and pilot substation inspection images, such as by two or more senior operation and maintenance personnel annotating the equipment categories and key components. Its sample size is usually 50 to 200 images, i.e., small sample scenes, and it needs to cover different shooting angles such as front, side and top views, lighting conditions such as sunny days, cloudy days and night supplemental lighting, and operating conditions such as normal operation and minor defects.

[0033] The unlabeled dataset specifically refers to images extracted from inspection image libraries of similar substations and drone aerial video frames. The sample size is 10-20 times that of the labeled dataset, such as 500-4000 images. It needs to include various scenarios of new equipment, such as different installation locations and combinations with traditional equipment, but no manual labeling of categories is required.

[0034] In one feasible embodiment of the present invention, the specific process of iteratively fine-tuning the basic recognition model includes: using the pre-trained lightweight model as the initial weights, and fine-tuning the pre-trained lightweight model using the labeled dataset of new power equipment, so that the general features it has learned can be adapted to the new power equipment, thereby obtaining the basic recognition model.

[0035] It should be noted that the initial basic recognition model has the ability to initially identify new devices, but its performance is limited due to insufficient labeled data.

[0036] The basic recognition model is used to infer the bounding box, category and confidence of each image in the labeled dataset of new power equipment. Several images with confidence greater than a preset value are retained and recorded as high-confidence pseudo-label samples.

[0037] For example, the preset value can be 0.95, that is, prediction results with a confidence level higher than 0.95 are considered to be judgments that the model is very confident in, and their reliability is high. These prediction results, namely bounding boxes and categories, are directly used as high-confidence pseudo-label samples, thereby transforming unlabeled images into labeled training samples.

[0038] After the high-confidence pseudo-label samples are manually reviewed, they are merged with the labeled dataset to form a comprehensive enhanced training set.

[0039] It should be noted that the manual review aims to correct potential errors in high-confidence pseudo-label samples and prevent these errors from being amplified in subsequent iterations. Furthermore, the review can be a full inspection or a sampling inspection, depending on the quality requirements and labor costs.

[0040] The basic recognition model after adjusting the learning rate is fine-tuned multiple times using the comprehensive enhanced training set until the number of iterations reaches the preset upper limit. Then, the fine-tuning is stopped and the final basic recognition model is saved.

[0041] For example, the preset upper limit can be 3 times.

[0042] It should be noted that the process of generating pseudo-labels, manual review, data merging, and fine-tuning can be repeated multiple times. After each iteration, the basic recognition model will use the knowledge learned in the previous iteration to mine and label more and more difficult unlabeled samples, thereby achieving self-evolution and continuous performance improvement until convergence.

[0043] A differential feature attention module is embedded in the basic recognition model to form a lightweight device basic recognition model.

[0044] It should be noted that the embedded differential feature attention module is specifically embedded into the backbone network of the iteratively fine-tuned basic recognition model in a pluggable manner, usually added after the deep feature map of the basic recognition model and before the classifier.

[0045] The core of the differential feature attention module is an attention mechanism, but its goal is not to focus on all important features, but to focus on and amplify the differential features that can distinguish new power equipment from traditional equipment.

[0046] The working principle of the differential feature attention module specifically includes the following: (1) Feature input: The module receives feature maps from the backbone network. (2) Differential weight generation: The module learns an attention weight map through a small network such as a fully connected layer or a convolutional layer, and the learning process of the attention weight map is guided by the contrastive learning loss, which makes it tend to assign high weights to unique features. (3) Feature reconstruction: The learned attention weight map is multiplied point by point with the original feature map to generate a differentially enhanced feature map. In the feature map, features related to the unique structure of new power equipment, such as the special skirt design of a certain new type of insulator and the unique heat sink layout of transformers, are significantly enhanced, while features common to traditional power equipment, such as background and general shape, are relatively suppressed.

[0047] In one feasible embodiment of the present invention, the structure of the differential feature attention module includes a feature mapping layer, a contrastive learning layer, and a feature weighting layer.

[0048] The specific formation process of the lightweight equipment basic identification model includes: inputting the image features of the new power equipment into the feature mapping layer and extracting feature vectors.

[0049] In the contrastive learning layer, the weight distribution of the differential structural features is determined by calculating the similarity of the feature vectors between the new power equipment and the traditional power equipment.

[0050] Specifically, a triplet consisting of anchor samples, positive samples, and negative samples is constructed, and a triplet loss function is used to optimize the differential feature attention module so that the distance between the anchor sample and the positive sample in the feature space is less than the distance between the anchor sample and the negative sample.

[0051] In the feature weighting layer, the differential structural features are enhanced according to the weight distribution to obtain enhanced feature vectors, which are then input into the model classification layer.

[0052] This forms the basic identification model for lightweight equipment.

[0053] This invention generates a pre-trained lightweight model based on a constructed knowledge graph of power equipment and a teacher-student model. At the same time, a semi-supervised pseudo-label iteration strategy combined with manual review and adaptive thresholds solves the problem of data scarcity for new equipment, provides a solid knowledge foundation for the model, and effectively utilizes unlabeled data. It continuously improves the model's recognition accuracy for new equipment in a data-driven manner, reducing the dependence on high-cost manually labeled data.

[0054] The inspection image of the power equipment to be identified is input into the lightweight equipment basic recognition model, and the candidate regions of the power equipment are output.

[0055] In one feasible embodiment of the present invention, the candidate region for power equipment includes bounding box coordinates, preliminary category confidence, and initial visual semantic feature vector.

[0056] The specific process of outputting candidate regions for power equipment includes: the lightweight equipment basic recognition model extracts multi-scale features from the input power equipment inspection image to be identified, and generates a hierarchical feature map of the power equipment inspection image to be identified.

[0057] It should be noted that the specific generation process of the hierarchical feature map includes: (1) Backbone network: a deep residual network is used to extract basic features such as equipment outline and texture from the input power equipment inspection image to be identified; (2) Feature pyramid network: in conjunction with the backbone network, the basic features are fused and enhanced at multiple scales to generate a hierarchical feature map covering different receptive fields, which can be adapted to power equipment of different sizes in the inspection image, such as small insulators and large transformers.

[0058] The hierarchical feature map is output after generating candidate regions for power equipment through a region proposal network.

[0059] Based on candidate regions of power equipment, a spatial relationship map of power equipment is constructed, and iterative relationship reasoning is performed on it, updating the node feature vectors in real time.

[0060] In one feasible embodiment of the present invention, the specific construction process of the spatial relationship map of the power equipment includes: defining each candidate region of the power equipment as a node.

[0061] For any two nodes, if the Euclidean distance between the center points of their bounding boxes is less than a preset distance threshold or the intersection-union ratio of their bounding boxes is greater than zero, then an edge is established between the two nodes to represent their spatial proximity relationship.

[0062] The weight of an edge is determined by the reciprocal of the Euclidean distance between the center points of the two node bounding boxes, with the smaller the distance, the greater the weight. The weighted undirected graph constructed in this way is denoted as the spatial relationship graph of power equipment.

[0063] It should be noted that a fusion feature vector is generated for each node in the spatial relationship graph. The fusion feature vector is formed by concatenating the initial visual semantic feature vector of the node with a spatial location encoding vector. The spatial location encoding vector is a five-dimensional vector, and its elements are, in order, the normalized x-coordinate, normalized y-coordinate, normalized width, normalized height, and normalized area of ​​the center point of the bounding box.

[0064] In one feasible embodiment of the present invention, the specific process of real-time updating node feature vectors includes: using a multi-head graph attention network to perform iterative relationship reasoning on the spatial relationship graph of the power equipment.

[0065] It should be noted that the multi-head graph attention network consists of three stacked graph attention layers. Each graph attention layer contains eight parallel attention heads. In each graph attention layer, the input node features first pass through an independent linear transformation layer, and then the attention coefficients are calculated.

[0066] In each iteration, each attention head in the multi-head graph attention network independently calculates the attention coefficient between each node and its neighboring nodes, and the attention coefficient is weighted and summed over the feature vectors of the neighboring nodes.

[0067] It should be noted that the attention coefficients are obtained by mapping the transformed features of the two connected nodes through a single-layer feedforward network and applying a modified linear unit activation function to obtain the attention coefficients.

[0068] The output feature vectors of each attention head in the multi-head graph attention network are concatenated to form the output of the multi-head graph attention network, which generates the updated node feature vector after aggregation of context information.

[0069] This enables real-time updates of node feature vectors.

[0070] The final node feature vector is input into the classifier, which outputs the final accurate category. The initial category confidence and the final accurate category are combined to complete the type identification and status analysis of all power equipment in the power equipment inspection image.

[0071] In one feasible embodiment of the present invention, the specific process of outputting the final accurate category of the power equipment candidate region includes: the classifier is specifically a three-layer multilayer perceptron.

[0072] After receiving the final node feature vector, the multilayer perceptron passes it through two hidden layers containing batch normalization and modified linear unit activation functions, and finally outputs a probability distribution vector with a dimension equal to the preset total number of power equipment categories.

[0073] The category corresponding to the maximum probability value in the probability distribution vector is denoted as the final precise category of the final node feature vector.

[0074] It should be noted that after the classifier outputs the final recognition result, a graph pruning step is required. The specific steps include: traversing all nodes, and if the highest class confidence of a node is lower than a preset confidence threshold, then the node and all its associated edges are removed from the spatial relationship graph.

[0075] Traverse all edges. If an edge connects two nodes whose category combination does not conform to the predefined knowledge base of power equipment assembly rules, then remove that edge.

[0076] In one feasible embodiment of the present invention, the specific process of completing the type identification and status analysis of all power equipment in the power equipment inspection image includes: when the preliminary category confidence of a certain power equipment in the power equipment inspection image exceeds the preset preliminary category confidence limit and the final category confidence exceeds the preset final category confidence limit, the type of the power equipment can be identified, and redundant regions of the candidate regions belonging to the same category after screening are merged, thereby completing the type identification of all power equipment in the power equipment inspection image.

[0077] It should be noted that the specific method for merging redundant regions among candidate regions belonging to the same category after screening is as follows: if the bounding box of a certain power equipment in the power equipment inspection image is greater than or equal to a preset bounding box threshold, such as 0.5, it indicates that the same equipment is being identified repeatedly. In this case, the candidate region with the highest confidence in the final category is retained, and the remaining redundant regions are removed.

[0078] It should also be noted that if there is a conflict in the power equipment inspection image where the initial category confidence of a certain power equipment exceeds the preset initial category confidence limit but the final category confidence is lower than the preset final category confidence limit, if the spatial distance between the power equipment and another high-confidence power equipment is less than the preset distance, then the type of the other high-confidence power equipment will be used as the type of the power equipment; otherwise, the power equipment will be marked as a power equipment to be manually reviewed.

[0079] Visual detail features such as color distribution, edge integrity, and abnormal area area of ​​each type of power equipment in the power equipment inspection images are extracted, and they are matched with the corresponding category features in the state feature library, and the similarity is calculated according to the cosine similarity calculation formula.

[0080] It should be noted that the state feature library predefines the visual features of the normal state and typical defect state for each type of power equipment, thereby forming the state feature library. A specific example is shown in Table 1 below.

[0081] Table 1 Examples of State Feature Library

[0082] transformer The casing is free of oil stains and the heat sink is intact. The casing has brown oil stains and the heat sink is deformed. insulator The umbrella skirt has no cracks and the color is even. The umbrella skirt is damaged and has black burn marks on the surface. Smart Inverter The indicator light is green and there are no obstructions. The indicator light is red and there is dust accumulation on the casing.

[0083] The operating status corresponding to the category feature with the highest similarity is taken as the operating status of each power device of the identified type, and the matching similarity is output as the status confidence, thereby completing the status analysis of all power devices in the power equipment inspection image.

[0084] It should be noted that the operating status includes, but is not limited to, normal, oil leakage, and damage.

[0085] For example, the state-type association verification in the above steps ensures that the state and type match. For example, oil leakage can correspond to a transformer but not to an insulator. If a mismatch occurs, such as marking an insulator as oil leakage, it is automatically corrected to an abnormal state identification and awaits review.

[0086] It should also be noted that the final result integrates the information of each type of power equipment identified in the power equipment inspection images and outputs a structured result, which includes the equipment location, i.e., the bounding box coordinates (x1, y1, x2, y2), the equipment type, i.e., the final precise category, such as 110kV transformer, silicone rubber insulator, the operating status, i.e., normal or defect type, such as normal, oil leakage defect, and the confidence information, i.e., the final category confidence and status confidence.

[0087] This invention transforms isolated equipment detection tasks into reasoning tasks based on structured graph data by constructing a spatial relationship map of power equipment. This enables the model to explicitly learn and utilize the topological layout and functional dependencies between equipment, reducing the misidentification rate caused by independent detection of individual equipment, which easily leads to misclassifying adjacent similar equipment as belonging to the same category, and improving the accuracy of operation and maintenance decisions.

[0088] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0089] Those skilled in the art will recognize that the algorithmic steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this application.

[0090] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0091] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0092] Finally, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for identifying power equipment based on deep learning, characterized in that: include: A knowledge graph of power equipment and a teacher-student model are constructed. The knowledge graph of power equipment is combined with the power equipment dataset to train the teacher model. The learned common features are transferred to the student model through knowledge distillation to obtain a pre-trained lightweight model. We acquire labeled and unlabeled datasets of new power equipment, fine-tune the pre-trained lightweight model based on the labeled dataset to obtain the basic recognition model, and use this model to infer high-confidence pseudo-label samples from the unlabeled dataset. After verification, we merge the samples with the labeled dataset to iteratively fine-tune the basic recognition model. A differential feature attention module is embedded in the basic recognition model to form a lightweight device basic recognition model. Input the inspection image of the power equipment to be identified into the lightweight equipment basic recognition model, and output the candidate regions of the power equipment; A spatial relationship map of power equipment is constructed based on candidate regions of power equipment, and iterative relationship reasoning is performed on it, with node feature vectors updated in real time. The specific construction process of the spatial relationship map of the power equipment includes: Each of the aforementioned power equipment candidate regions is defined as a node; For any two nodes, if the Euclidean distance between the center points of their bounding boxes is less than a preset distance threshold or the intersection-union ratio of their bounding boxes is greater than zero, then an edge is established between the two nodes. The weight of the edge is determined by the reciprocal of the Euclidean distance between the center points of the two node bounding boxes, and the smaller the distance, the greater the weight. The weighted undirected graph constructed in this way is called the spatial relationship graph of power equipment. The candidate regions for power equipment include bounding box coordinates, preliminary category confidence, and initial visual semantic feature vectors. The final node feature vector is input into the classifier, which outputs the final accurate category. The initial category confidence and the final accurate category are combined to complete the type identification and status analysis of all power equipment in the power equipment inspection image. The specific process of the power equipment knowledge graph and teacher-student model includes: Physical parameters of several traditional power equipment are extracted from power industry standards, and structural features of several traditional power equipment are extracted from power equipment operation and maintenance ledgers. The physical parameters and structural features are integrated to form an initial knowledge graph. After redundancy removal, the power equipment knowledge graph is obtained. A teacher model was built using a ResNet50 convolutional neural network, and a student model was built using a MobileNetV3 lightweight convolutional neural network. This model is referred to as the teacher-student model. The specific process of obtaining the pre-trained lightweight model includes: The power equipment knowledge graph is converted into a knowledge constraint vector and input into the teacher model; The teacher model is trained using a dataset of power equipment stored in the database. The teacher model learns visual features and knowledge constraints simultaneously. The classification head is optimized using the cross-entropy loss function, enabling the teacher model to accurately identify power equipment categories. The feature output head generates common features of power equipment containing knowledge constraints. The common features learned by the teacher model are transferred to the student model by using the KL divergence distillation loss function to obtain a pre-trained lightweight model. The specific process of iteratively fine-tuning the basic recognition model includes: The pre-trained lightweight model is used as the initial weights, and the pre-trained lightweight model is fine-tuned using the labeled dataset of new power equipment to adapt the learned general features to the new power equipment, thus obtaining the basic recognition model. The basic recognition model is used to infer the bounding box, category and confidence of each image in the unlabeled dataset of new power equipment. Several images with confidence greater than a preset value are retained and recorded as high-confidence pseudo-label samples. After the high-confidence pseudo-label samples are manually reviewed, they are merged with the labeled dataset to form a comprehensive enhanced training set; The basic recognition model after adjusting the learning rate is fine-tuned multiple times using the comprehensive enhanced training set until the number of iterations reaches the preset upper limit. Then, the fine-tuning is stopped and the final basic recognition model is saved.

2. The method for identifying power equipment based on deep learning according to claim 1, characterized in that: The structure of the differential feature attention module includes a feature mapping layer, a contrastive learning layer, and a feature weighting layer; The specific formation process of the lightweight equipment basic identification model includes: inputting the image features of the new power equipment into the feature mapping layer and extracting feature vectors; In the contrastive learning layer, the weight distribution of the differential structural features is determined by calculating the similarity of the feature vectors between the new power equipment and the traditional power equipment. In the feature weighting layer, the differential structural features are enhanced according to the weight distribution to obtain enhanced feature vectors, which are then input into the model classification layer. This forms the basic identification model for lightweight equipment.

3. The method for identifying power equipment based on deep learning according to claim 1, characterized in that: The specific process of outputting the candidate region of power equipment includes: the lightweight equipment basic recognition model extracts multi-scale features from the input power equipment inspection image to be identified, and generates a hierarchical feature map of the power equipment inspection image to be identified; The hierarchical feature map is output after generating candidate regions for power equipment through a region proposal network.

4. The deep learning-based power equipment identification method according to claim 3, characterized in that: The specific process of real-time updating of node feature vectors includes: A multi-head graph attention network is used to perform iterative relation reasoning on the spatial relation graph of the power equipment. In each iteration, each attention head in the multi-head graph attention network independently calculates the attention coefficient between each node and its neighboring nodes, and the attention coefficient is weighted and summed over the feature vectors of the neighboring nodes. The output feature vectors of each attention head in the multi-head graph attention network are concatenated to form the output of the multi-head graph attention network, which generates the updated node feature vector after aggregation of context information. This enables real-time updates of node feature vectors.

5. The method for identifying power equipment based on deep learning according to claim 1, characterized in that: The specific process for finalizing the precise category of the candidate region for output power equipment includes: The classifier is specifically a three-layer multilayer perceptron; After receiving the final node feature vector, the multilayer perceptron passes through two hidden layers containing batch normalization and modified linear unit activation functions, and finally outputs a probability distribution vector with a dimension equal to the preset total number of power equipment categories. The category corresponding to the maximum probability value in the probability distribution vector is denoted as the final precise category of the final node feature vector.

6. The deep learning-based power equipment identification method according to claim 5, characterized in that: The specific process for identifying the type and analyzing the status of all power equipment in the power equipment inspection images includes: When the initial category confidence of a power device in a power equipment inspection image exceeds the preset initial category confidence limit and the final category confidence exceeds the preset final category confidence limit, the type of the power device can be identified, and redundant regions belonging to the same category after screening can be merged to complete the type identification of all power devices in the power equipment inspection image. Visual detail features of each type of power equipment in the power equipment inspection images are extracted, and they are matched with the corresponding category features in the status feature library, and the similarity is calculated. The operating status corresponding to the category feature with the highest similarity is taken as the operating status of each power device of the identified type, and the matching similarity is output as the status confidence, thereby completing the status analysis of all power devices in the power equipment inspection image.

Images