A defect intelligent identification system for power transmission line inspection

Abstract

The application relates to the technical field of intelligent operation and maintenance of power systems, and particularly discloses a defect intelligent identification system for power transmission line inspection. The system comprises a dynamic scene modeling unit and an embedded adaptive reasoning unit. The dynamic scene modeling unit generates a scene state vector representing the coupling state of the environment and equipment by analyzing an image sequence and metadata. The embedded adaptive reasoning unit comprises a basic feature extraction network, a scene adaptation controller, a reconfigurable identification network and a feature modulator. The scene adaptation controller outputs control parameters according to the scene state vector, dynamically reconfigures the topology of the identification network and generates a feature modulation field to correct the extracted features in real time, and finally completes defect identification. The application enables the system to dynamically adjust the internal processing mechanism according to real-time inspection conditions, and improves the identification accuracy and adaptive ability in complex scenes.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance technology for power systems, and more specifically, to an intelligent defect identification system for transmission line inspection. Background Technology

[0002] As the core artery of the power system, the safe and stable operation of transmission lines is of paramount importance. Regular inspections are the primary means of detecting potential defects such as insulator damage and loose hardware, and preventing faults. With the expansion of the power grid and the trend towards unmanned inspections, traditional manual inspection methods face challenges in terms of efficiency, coverage, and objectivity. Therefore, developing intelligent and automated defect identification technologies is of great significance.

[0003] Currently, automatic recognition based on computer vision is the main technological direction, but it still has significant shortcomings in practical applications. Traditional image processing methods rely on manually designed features, which have poor robustness in complex and ever-changing natural scenes and struggle to cope with issues such as changes in lighting, background interference, and perspective differences. While deep learning-based methods have improved recognition performance, their performance heavily depends on large-scale, high-quality, and balanced labeled training data. Transmission line defect samples themselves have long-tailed distribution characteristics, and new defect types and environmental conditions are constantly emerging, leading to data scarcity challenges in actual model deployment. In addition, most existing models are static models; once trained, their structure and parameters remain fixed, lacking the ability to perceive and adaptively adjust to the real-time environment of specific inspection tasks. When faced with weather conditions, shooting angles, or new equipment not fully covered by training data, the model's recognition accuracy will significantly decrease, with increased false negative and false positive rates, making it difficult to meet the requirements of high-reliability operation and maintenance.

[0004] In summary, the main problems with existing technologies are: static models are difficult to adapt to dynamically changing inspection environments and diverse defect forms; their strong dependence on large-scale labeled data restricts their rapid deployment and updates in data-scarce scenarios; and the lack of online self-optimization capabilities leads to insufficient generalization and practicality under complex real-world conditions. Summary of the Invention

[0005] The present invention aims to overcome the above-mentioned shortcomings of the prior art and provide an intelligent identification system for transmission line inspection defects that can sense the environment, dynamically adapt, and continuously learn.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: A defect intelligent identification system for power transmission line inspection includes a dynamic scenario modeling unit and an embedded adaptive inference unit.

[0007] The dynamic scenario modeling unit is used to perform online analysis on the input inspection image sequence and associated metadata to generate a dynamically updated scenario state vector. ).

[0008] This unit decouples environmental interference from inherent device attributes and fuses time-series prediction information, enabling the scenario state vector ( It can quantitatively characterize the overall context state that currently affects defect identification.

[0009] The embedded adaptive inference unit is connected to the dynamic scenario modeling unit. This unit receives the scenario state vector ( It compares the current inspection image with the actual image and performs adaptive recognition. This includes: A basic feature extraction network is used to extract multi-scale feature maps from the input image. ); Context adaptation controller, used to adapt to the context state vector ( Generate structural control parameters ( ) and characteristic modulation parameters ( ); A reconfigurable recognition network comprising a set of predefined feature processing subnetworks, the reconfigurable recognition network being configured to operate according to the structural control parameters ( The sub-networks are dynamically selected and combined to form a temporary network topology that adapts to the current scenario. Feature modulator, used according to the feature modulation parameters ( ) generates a modulation field, and modulates the multi-scale feature map ( Correction is performed to obtain the modulated features ( ); The embedded adaptive inference unit is configured to utilize the temporary network topology to modify the modulated features ( The system processes the defect and outputs the type and location information of the defect.

[0010] Furthermore, the dynamic scenario modeling unit specifically includes: The time series analysis module is used to perform sliding window processing on image sequences and calculate the apparent consistency measure and feature statistics drift of local regions to quantify the dynamic changes of environmental disturbances. The attribute decoupling module is used to construct a cross-frame device component feature association graph and separate the device's stable inherent attribute features through iterative message passing of a graph neural network. The state encoding module is used to fuse the dynamic change information, the stable attribute features, and the device static description vector, and output the scenario state vector through a gated recurrent unit network. ).

[0011] Furthermore, the structural control parameters ( ) is a sparse vector whose non-zero weights are used to activate a subset of the predefined feature processing subnetwork and to perform weighted fusion on the output of that subset. The modulation field includes a spatial transformation field and a channel weight field, which are used to perform spatial geometric correction and position-by-position channel importance recalibration on the feature map, respectively.

[0012] In a preferred embodiment, the parameters of the embedded adaptive inference unit are obtained through a hierarchical optimization mechanism. This mechanism includes an inner-layer optimization process and an outer-layer optimization process: the inner-layer process simulates online adaptation and optimizes control parameters using a small number of samples; the outer-layer process, based on the performance after inner-layer adaptation, jointly optimizes all network parameters, enabling the system to achieve rapid adaptation capabilities.

[0013] In another preferred embodiment, the system further includes a scenario knowledge base for storing snapshots of typical scenario patterns and their corresponding optimization parameters. The scenario adaptation controller can match the current scenario with the knowledge base to quickly recall historical optimization parameters, accelerating the adaptation process. The system can also use a knowledge update unit to compress, evaluate, and store newly generated effective parameters in the knowledge base, enabling continuous learning.

[0014] As another preferred embodiment, the system further includes a closed-loop feedback link. This link converts the defect location information in the identification results into a spatial attention mask (…). The data is fed back to the dynamic scene modeling unit to modulate its analysis process of subsequent images, thereby forming a closed-loop optimization in which perception and recognition mutually enhance each other.

[0015] The technical effects and advantages of this invention are as follows: Compared to existing technologies, this invention constructs a dynamic scenario modeling unit and an embedded adaptive inference unit. First, it performs online analysis on the inspection image sequence, extracting and encoding the coupling state of environmental interference and device attributes as a scenario state vector. This vector then drives the scenario adaptation controller to generate control parameters in real time. These parameters, on the one hand, guide the reconfigurable recognition network to dynamically select and combine different feature processing sub-networks from its resource pool to form a temporary network topology adapted to the current scenario; on the other hand, they control the feature modulator to generate targeted modulation fields, correcting the spatial and channel dimensions of the basic features. This enables the recognition system to proactively adjust its internal processing structure and feature expression strategy according to the specific inspection conditions at every moment, thereby improving the model's recognition accuracy and stability in non-steady-state environments such as changes in illumination and complex backgrounds, and reducing performance fluctuations caused by environmental differences.

[0016] Compared to existing technologies, this invention employs a hierarchical optimization mechanism to train system parameters. This mechanism simulates an online adaptive process in the inner layer, forcing the system to learn how to quickly adjust its state using a small number of samples. In the outer layer, it optimizes all basic components based on the performance gained from this rapid adaptation. When faced with new defect types or unfamiliar inspection environments, the system can quickly and effectively self-adjust using a very small number of samples collected on-site (such as a few images manually reviewed at the beginning of the inspection). This significantly reduces reliance on pre-labeled, massive, and complete datasets, enhancing the system's practicality and deployability in data-scarce scenarios.

[0017] Compared to existing technologies, this invention introduces a continuously evolving architecture consisting of a contextual knowledge base and a closed-loop feedback loop. The contextual knowledge base stores optimized parameters under typical patterns and provides a warm start during inference to accelerate the adaptive process. The closed-loop feedback loop guides the recognition results back to the contextual modeling unit, making its subsequent analysis more focused on relevant areas. After the system runs, it can also compress effective adaptive results and selectively update the knowledge base. The system has the ability to accumulate experience and optimize strategies in each inspection task, achieving continuous iteration and improvement of performance. Attached Figure Description

[0018] Figure 1 This is a diagram showing the overall architecture of the intelligent identification system for defects in power transmission line inspection according to the present invention.

[0019] Figure 2 This is a flowchart of the internal processing of the dynamic scenario modeling unit of the present invention.

[0020] Figure 3 This is a flowchart of the dual-branch processing of the embedded adaptive inference unit of the present invention. Detailed Implementation

[0021] 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.

[0022] Example 1 As attached Figures 1 to 3 The system illustrates an intelligent defect identification system for power transmission line inspection. During operation, the system receives images of power transmission line equipment and associated metadata collected by the inspection equipment via a data interface. The images constitute a time series, and the metadata includes equipment identification, spatial location, and sensor attitude information. The system performs dynamic scenario modeling and embedded adaptive identification, and outputs the defect identification results.

[0023] S100. Generation of Dynamic Scenario State Vectors This step is used to perform dynamic scene modeling, corresponding to the function of the dynamic scene modeling unit, which converts the input image sequence and metadata into a fixed-dimensional scene state vector. .

[0024] Furthermore, this step S100 includes the following sub-steps: S110. Timing Interference Analysis.

[0025] The system performs block segmentation and temporal analysis on the input continuous image frames, dividing each image frame into segments. Non-overlapping grid regions (e.g., dividing an image into...) The grid, then ).

[0026] For each region Extract its color histogram and gradient direction histogram ,in Indicates time window Frame number within, For window length (e.g., This indicates that the current frame and the previous 4 frames are used.

[0027] Calculate the feature differences between corresponding regions in adjacent frames: and .

[0028] area At any moment Apparent consistency measure Defined as the average difference of this region across all adjacent frame pairs: ; in This is a preset positive scalar coefficient used to adjust the relative weights of color features and gradient features in the consistency measure (e.g., ).

[0029] Simultaneously, the mean of each regional feature within the time window is calculated. and standard deviation The characteristic statistic drift is obtained by comparing it with the corresponding statistic from the previous time window. and .

[0030] The output is a dynamic disturbance description tensor. , where each position Corresponding to a three-dimensional vector .

[0031] S120. Device attribute decoupling. The system processes the current frame image. Based on the device identifier in the metadata, retrieve the prior knowledge of the device's component topology from the pre-built ledger database.

[0032] Using a trained semantic segmentation network The process is performed to obtain pixel-level component classification masks. In the mask Within each connected region of a defined component, the feature maps extracted from the intermediate layers of the backbone network are subjected to average pooling to obtain a set of component feature vectors. .

[0033] Construct an undirected graph , where nodes correspond ,side Defined by the spatial adjacency relationship or prior topology between components.

[0034] The graph is processed using a two-layer graph attention network. The first layer of attention mechanism calculates the association weights between the features of the current node and its neighboring nodes. The second layer aggregates the weighted neighboring features and combines them with the node's own features.

[0035] This message passing process enables node feature encoding to contain more information related to the device's global structure. Global average pooling is then performed on all processed node feature vectors to obtain a vector representing the device's stable inherent properties. .

[0036] S130. State fusion coding and prediction. The system will use the output of S110... Flattened and compressed into a vector through a fully connected layer. .Will S120 output and the embedding vector of the device identifier The features are concatenated to form a comprehensive feature vector. .Will Input a gated recurrent unit (GRU). The internal calculations of the GRU are as follows: ; ; ; ; in, It is the hidden state of the previous time step (the initial state can be set as the zero vector). It is the Sigmoid function. This represents element-wise multiplication. These are the trainable parameters of the GRU.

[0037] Current hidden state Defined as a situational state vector ,Right now The GRU's circular connections enable it to fuse timing information, thereby improving the output... It can encode continuous temporal context information.

[0038] S200. Adaptive Recognition Based on Context Status This step is used to perform context-state-based adaptive recognition, which corresponds to the function of the embedded adaptive inference unit.

[0039] S210. Basic Feature Extraction. Input single-frame image. It is fed into a pre-trained convolutional neural network backbone. This network outputs feature maps at multiple scales, denoted as... ,in Indicates the first The output of each stage. The dimension is , For the number of channels, and For height and width.

[0040] S220. Generate adaptive control parameters. Scenario state vector. The input scenario adaptation controller. This controller contains two multilayer perceptrons.

[0041] The first MLP outputs the raw structural control parameters. , , This represents the number of sub-networks in the resource pool.

[0042] To achieve sparsity, for Applying Top-K operation: Selecting the top values ​​with the largest values A set of indices ,in Let be a preset positive integer, and (For example, , ).

[0043] Then, the structural control parameters for sparsification are calculated. : ; The second MLP generates characteristic modulation parameters , , The dimensions are determined by the modulator design. The parameters of the different feature processing subnetworks in the resource pool are configured during the pre-training phase to have specific response characteristics for different typical scenario interference patterns or device feature patterns.

[0044] S230. Feature map modulation. For the first... Layer feature map Feature modulator utilizes The corresponding part in Adjust it accordingly.

[0045] The modulator comprises two sub-networks. The spatial modulation sub-network is... and downsampling Input: Spatial transformation field, output: offset field and scaling field .

[0046] Channel modulation subnetwork with and The global average pooling vector is used as input, and the channel weight field is used as output. .

[0047] Modulated feature map The calculation is performed using the following steps: First, based on the offset field... right Perform deformable sampling based on bilinear interpolation (this operation can be achieved through deformable convolution) to obtain Then, element-wise calculations are performed using the scaling field and channel weight field: ,in This indicates element-wise multiplication.

[0048] S240. Dynamic Reconstruction and Identification of Network Structure

[0049] The system maintains a collection of A pre-trained feature processing subnetwork The resource pool. During inference, based on sparse vectors. Only the index belongs to the set. The subnetwork is activated.

[0050] The outputs of these activated subnetworks are weighted and summed, thereby weighting and fusing the outputs of the selected feature processing subnetworks through the sparse path activation weights to form the backbone output of a temporary network topology adapted to the current scenario: .

[0051] in, Represents the modulated multi-scale feature map Selected for input subnetwork Specific layer feature maps (e.g., the fourth-stage output feature map of the backbone network).

[0052] The data is then fed into a detection head. This head includes a region proposal network, a region of interest alignment layer, and a classification and bounding box regression branch, ultimately outputting a set of bounding boxes representing the defects. Category tag collection and confidence set .

[0053] As some preferred implementations, the system also includes the following processes to improve efficiency and performance: S300. Accelerated Reasoning and Continuous Learning Based on Contextual Knowledge Base S310. Knowledge Base-Assisted Rapid Parameter Initialization. The system maintains a contextual knowledge base, which stores... There are 1 record, and each record contains a scenario template vector. and corresponding parameter snapshots In step S220, the controller calculates... At the same time, calculate the current With all templates cosine similarity .

[0054] If the maximum similarity Exceeding a preset matching threshold (For example, ), then retrieve the corresponding .

[0055] The final parameters used are obtained through linear interpolation: , The calculation is similar. Among them, the weighting factor... It can be calculated from similarity, for example, defined as .

[0056] S320. Progressive updates to the knowledge base. After an inspection task is completed, the system collects data on those with a confidence level higher than a threshold. (For example, And for samples that have been confirmed as correct, record their corresponding scenario vectors. and final stability parameters .

[0057] To compress storage, parameter sharing and functional approximation methods are used to reconstruct the effective parameter set into an equivalent parameter set with a smaller parameter size.

[0058] One approach is to train a lightweight "student" multilayer perceptron. Its goal is to approximate the original modulation parameter generation function: the loss function is .

[0059] After training, store weight Then, calculations were performed. With all templates in the knowledge base The similarity.

[0060] If the maximum similarity is below a novelty threshold (For example, ), and use The performance loss on the validation set is less than the threshold. (For example, Then the new template-snapshot pair Add to the knowledge base.

[0061] S400 Closed-Loop Feedback Optimization Process S410. Generate semantic guidance feedback signal. Based on the defect detection results output by S240. Generate a spatial attention mask ,in This represents the original input image size.

[0062] initialization This is a zero matrix. For each detection box... Calculate its center Draw a two-dimensional Gaussian kernel at the center: ,in Proportional to the width and height of the frame, for example , For a preset scaling factor (e.g., ).

[0063] All Overlay At the corresponding position, and for Perform maximum value normalization.

[0064] S420. Feedback signals are integrated into scene modeling. The system processes the current frame image. At that time, the spatial attention mask generated and stored from the previously processed frames is used. Downsampled to the resolution required by each module to obtain .

[0065] In the time series analysis of S110, when calculating the apparent consistency measure of local regions, the following method is used: As a weight: .

[0066] In the device attribute decoupling of S120, The component feature map is multiplied element-wise before pooling.

[0067] S500. System Training and Initialization Before deploying this system, offline training is required to determine its learnable parameters. The semantic segmentation network is pre-trained using images of power transmission equipment with component annotations. The feature processing subnetwork in the resource pool... It is obtained by supervising defect detection training on data subsets that are clustered according to scenarios, with each subnetwork focusing on a specific scenario pattern.

[0068] The core parameters of the basic feature extraction network, scenario adaptation controller, feature modulator, and GRU network are jointly trained through a hierarchical optimization mechanism. The scenario knowledge base can be empty in the initial stage, or it can be initialized after training using representative scenarios and their corresponding parameters from the training data.

[0069] In summary, this system can repeatedly execute steps S100 to S400 on a continuous stream of input inspection images. Through the synergistic effect of dynamic scenario modeling, adaptive recognition, knowledge base utilization, and closed-loop feedback, it achieves continuous, adaptive, and intelligent recognition of transmission line defects. The recognition results can be used to generate inspection reports or trigger early warnings.

[0070] Finally, the following points should be noted: First, in the description of this application, it should be noted that, unless otherwise specified and limited, the terms "installation", "connection", and "linkage" should be interpreted broadly, and can be mechanical or electrical connections, or internal connections between two components, or direct connections. "Up", "down", "left", "right", etc. are only used to indicate relative positional relationships. When the absolute position of the described object changes, the relative positional relationship may change. Secondly: The accompanying drawings of the embodiments disclosed in this invention only involve the structures involved in the embodiments disclosed in this invention. Other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of this invention can be combined with each other. In conclusion, 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 defect intelligent identification system for power transmission line inspection, characterized in that, include: The dynamic scenario modeling unit is used to perform online analysis on the input inspection image sequence and associated metadata to generate dynamically updated scenario state vectors. An embedded adaptive inference unit is connected to the dynamic scenario modeling unit, the inference unit comprising: A basic feature extraction network is used to extract multi-scale feature maps from the input image; A scenario adaptation controller is used to receive the scenario state vector and responsively output structural control parameters and feature modulation parameters. A reconfigurable identification network has a set of predefined feature processing subnetworks, which is configured to dynamically select and combine specific subnetworks from the set of predefined feature processing subnetworks according to the structure control parameters to form a temporary network topology adapted to the current scenario. A feature modulator, connected to the output of the basic feature extraction network, is used to generate a modulation field according to the feature modulation parameters and perform a correction transformation on the multi-scale feature map to obtain modulated features; The embedded adaptive inference unit is configured to process the modulated features using the temporary network topology to output defect identification results.

2. The intelligent defect identification system for transmission line inspection according to claim 1, characterized in that, The dynamic scenario modeling unit includes: The time-series analysis module is used to perform sliding window processing on the inspection image sequence, calculate the apparent consistency measure and feature statistics drift of local image regions, so as to quantify the dynamic changes of environmental interference. The attribute decoupling module is used to construct a feature association graph of device components across image frames, and separate stable attribute features that characterize the inherent structure of the device through iterative message passing and aggregation operations on the graph. The state encoding module is used to fuse the quantized dynamic changes of environmental disturbances, the stable attribute features, and the static description vector from the equipment ledger, and output the scenario state vector containing the current state and prediction information through a sequence encoding network with memory units.

3. The intelligent defect identification system for transmission line inspection according to claim 1, characterized in that, The structural control parameters output by the scenario adaptation controller are a sparse vector, which is used to activate a subset from the set of predefined feature processing subnetworks. The non-zero weight values ​​in the sparse vector are used as path activation weights to perform weighted fusion on the outputs of the feature processing subnetworks in the activated subset.

4. The intelligent defect identification system for transmission line inspection according to claim 1, characterized in that, The modulation field generated by the feature modulator includes a spatial transformation field, which is used to generate a two-dimensional offset vector and a scaling factor for each spatial location in the multi-scale feature map, so as to apply geometric and intensity corrections to the pixel positions of the feature map.

5. The intelligent defect identification system for transmission line inspection according to claim 1 or 4, characterized in that, The modulation field generated by the feature modulator includes a channel weight field, which is used to assign an independent channel dimension weight vector to each spatial location in the multi-scale feature map in order to perform position-by-position channel importance recalibration on the feature map.

6. The intelligent defect identification system for transmission line inspection according to claim 1, characterized in that, In the set of predefined feature processing subnetworks, different subnetworks are pre-configured to have differentiated response characteristics to different typical scenario interference modes or device feature modes.

7. The intelligent defect identification system for transmission line inspection according to claim 1, characterized in that, The parameters of the embedded adaptive inference unit are obtained through a hierarchical optimization mechanism, which includes: The inner optimization process simulates online adaptation and rapidly optimizes the parameters generated by the scenario adaptation controller based on a support set consisting of a small number of samples for a specific scenario. The outer layer optimization process, based on the recognition performance of the temporary network topology after inner layer optimization on a query set containing different interference scenarios, jointly optimizes the parameters of the basic feature extraction network, the scenario adaptation controller, and the set of predefined feature processing sub-networks.

8. The intelligent defect identification system for transmission line inspection according to claim 1, characterized in that, It also includes a scenario knowledge base, which is connected to the scenario adaptation controller; The scenario knowledge base is used to store multiple typical scenario pattern templates, as well as pre-computed network configuration parameters associated with each template; When the scenario adaptation controller is running, it is also used to match the currently generated scenario state vector with the template in the scenario knowledge base, and load the corresponding pre-calculated network configuration parameters according to the matching result to assist in generating the structural control parameters and feature modulation parameters.

9. The intelligent defect identification system for transmission line inspection according to claim 1, characterized in that, It also includes closed-loop feedback links; The closed-loop feedback link is configured to convert the defect location information output by the embedded adaptive inference unit into a spatial attention mask, and feed the spatial attention mask back to the dynamic scenario modeling unit to modulate its internal calculation process when analyzing subsequent input inspection image sequences.

Images