Power transmission line defect image intelligent recognition and grading method

Abstract

The application relates to the fields of image processing and smart grid maintenance technology, and discloses a power transmission line defect image intelligent identification and grading method, which acquires high-dimensional visual features, preliminary classification confidence and a space mask matrix of an original image; calculates the topological risk weight of a target defect; splices the visual features and the topological risk weight to perform database retrieval, acquires a search prior grade and an evolution anchor point image; when the preliminary classification confidence is lower than a threshold value, decoupling interpolation and cross-level judgment are performed on the original image and the anchor point image in a latent space based on the mask matrix, and a gradient jump point is output; finally, the search prior grade, the topological risk weight and the gradient jump point are fused to calculate a comprehensive risk score, and a defect severity grade is output by comparing the threshold value. The application fuses historical experience, spatial mechanics attributes and evolution trend multi-dimensional parameters, solves the problem that a critical state defect is easily misjudged by single feature grading, and improves the reliability of the grading result.

Description

Technical Field

[0001] This invention relates to the fields of image processing and smart grid maintenance technology, specifically to a method for intelligent identification and classification of transmission line defect images. Background Technology

[0002] Currently, routine inspections of power transmission lines often utilize drones equipped with cameras to acquire images of the lines, which are then combined with deep learning models to automatically identify defects such as insulator damage and hardware corrosion. This conventional method primarily relies on extracting static pixel features from defective areas to directly output classification and grading results. However, surface defects in power transmission components often exhibit slow, gradual evolution. In real-world scenarios, the visual appearance captured by images frequently falls within the critical range of different severity levels. Relying solely on the visual features of a single two-dimensional image makes it difficult to provide a definitive severity assessment, causing the algorithm model to be highly susceptible to environmental lighting and grading biases.

[0003] Meanwhile, existing computer vision-based defect classification methods generally deviate from the actual physical operating environment of transmission lines. Visual defects of the same size and appearance pose significantly different actual threats to the safety of the entire line depending on whether they occur at critical connection nodes subjected to high mechanical tension or in non-stressed areas. Conventional image classification networks lack consideration of the spatial stress attributes of defects, leading to classification results that are often out of sync with the actual physical maintenance needs on site.

[0004] Furthermore, existing visual inspection methods are static, post-event assessments that only reflect the surface state of defects at the moment of capture, and cannot calculate the physical evolution of a target defect to the next severity level. Faced with complex operational and inspection requirements, current technologies lack a quantitative mechanism that can organically integrate historical evolution experience, spatial mechanical risks, and deterioration trend calculations. This lack of multi-dimensional decision-making and evaluation capabilities directly limits the reliability and practicality of intelligent inspection systems in classifying complex defects. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an intelligent identification and classification method for transmission line defect images. This method solves the problem that existing image recognition methods often rely on single visual features, lack the correlation between visual appearance and the physical stress structure of components, and cannot accurately quantify the actual failure degree of gradual defects, resulting in a high misjudgment rate of critical state defects.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for intelligent identification and classification of transmission line defect images, comprising the following steps: Acquire raw images of transmission lines, extract high-dimensional visual features, preliminary classification confidence scores, and binary space mask matrices; Extract the coordinates of the center point and the topological reference point of the target defect region, and calculate the topological risk weight by combining the nonlinear attenuation mechanism and engineering mechanical properties; A joint feature vector is obtained by performing feature concatenation on high-dimensional visual features and topological risk weights, and then searching in a preset feature vector database to obtain the prior retrieval level and evolution anchor point image. When the initial classification confidence is lower than the confidence threshold, based on the original image of the transmission line, the evolution anchor point image and the binary space mask matrix, the generative model performs decoupled interpolation and cross-level judgment in the latent space and outputs the gradient jump point. The prior retrieval level, topological risk weight, and gradient transition point are aligned and weighted to calculate a comprehensive risk score. The comprehensive risk score is then compared with a preset level threshold to output the final defect severity level.

[0007] In a preferred embodiment of the present invention, the step of calculating the topology risk weight by combining the nonlinear attenuation mechanism and engineering mechanical properties includes: Calculate the spatial geometric Euclidean distance between the coordinates of the center point and the coordinates of the topological reference point on a two-dimensional pixel plane; Retrieve the engineering stress coefficient of the component to which the current target defect belongs, as well as the preset distance attenuation scale factor, from the preset configuration database; Using a nonlinear mapping equation that includes a negative natural constant exponent, spatial geometric Euclidean distance, engineering stress coefficient, and distance attenuation scale factor are mapped into topological risk weights. Among them, the smaller the spatial geometric Euclidean distance, the closer the topological risk weight is to the theoretical upper limit determined by the engineering stress coefficient.

[0008] Furthermore, the step of performing feature concatenation on the high-dimensional visual features and topological risk weights to obtain a joint feature vector includes: L2 norm operations are used to numerically constrain high-dimensional visual features to obtain normalized visual features. The topological risk weights are mapped to a set interval using the extreme value normalization method to obtain normalized weights; Normalized weights are directly appended to the end of the normalized visual feature dimension, and tensor concatenation is performed to construct a one-dimensional joint feature vector.

[0009] Furthermore, the step of searching in the preset feature vector database includes: Based on the current defect topology risk weight and the preset high-risk topology threshold, a dynamic adjustment factor is calculated through a nonlinear smoothing function; The dynamic distance metric algorithm is applied to calculate the Euclidean distance between the joint feature vector of the current defect and each historical joint feature vector in the feature vector database at the visual feature level, and the absolute difference at the topological risk feature level, using the dynamic adjustment factor, so as to calculate the comprehensive similarity distance.

[0010] Further, the steps for obtaining the prior retrieval level and evolutionary anchor image include: The historical samples in the feature vector database are sorted in ascending order based on the similarity distance value, and the historical label information of the top K positions is extracted. The reciprocal of the similarity distance values ​​of each extracted historical sample is used as the statistical weight. The frequency of occurrence of each physical classification category is weighted and accumulated, and the classification level with the largest weighted accumulated value is selected as the retrieval prior level. Lock historical samples whose similarity distance reaches the minimum value in ascending order, and retrieve their corresponding original image data as evolution anchor images.

[0011] Furthermore, decoupled interpolation is performed in the latent space using a generative model, and its preceding data alignment process includes: The historical bounding box coordinates of the evolution anchor point image are retrieved, and the spatial position is aligned to the two-dimensional bounding box range of the original transmission line image using affine transformation operation to generate a spatially aligned anchor point image. The original image of the transmission line and the spatially aligned anchor point image are input into the encoder network to perform feature extraction and spatial dimensionality reduction operations, which are respectively mapped into independent high-order and low-dimensional latent feature tensors. An adaptive max-pooling downsampling operation is performed on the binary spatial mask matrix to align its spatial resolution with the latent feature tensor, thereby generating a latent mask matrix.

[0012] Furthermore, the specific process of performing decoupling interpolation is as follows: Set a discretely increasing interpolation step size variable; Constructing spatially decoupled interpolation equations based on the latent mask matrix; In the region where the latent mask matrix is ​​numerically represented as a defect active area, the feature data is linearly interpolated along the latent feature tensor of the original image of the transmission line to the latent feature tensor of the evolved anchor image. In the region where the potential mask matrix is ​​numerically represented as a frozen background area, the initial input features of the original image of the transmission line are locked. This generates a transition state tensor sequence with fixed background features and only the defect region evolves.

[0013] Furthermore, the steps for determining and outputting gradient transition points across levels include: The generated transition state tensor sequence is sequentially input into the decoder network in ascending order of step size to perform inverse mapping operation, and the output is a transition state synthetic image sequence composed of synthetic images; The transition state synthesized image sequence is fed into the detection network frame by frame for classification. When the classification category output by the detection network first jumps from a low level to a high level, the interpolation step size ratio of the corresponding image is recorded as the gradient jump point. If no level jump is detected, the gradient jump point will be assigned the default upper limit value; For cases where the initial classification confidence level is greater than or equal to the confidence threshold, the gradient jump point is directly assigned based on the initial classification category.

[0014] Further steps in calculating the comprehensive risk score include: Divide the prior retrieval level by the preset maximum level value to obtain the normalized prior retrieval level. Invert the gradient jump points to construct the risk gain term; The normalized prior level, topological risk weight, and risk gain term are multiplied by their respective preset weight coefficients and then summed to calculate the comprehensive risk score.

[0015] Further steps, including outputting the final defect severity level, include: The built-in logic comparator unit is invoked to compare the comprehensive risk score with the preset severity threshold and critical threshold in parallel. When the overall risk score is less than the severity threshold, a general defect is output. When the comprehensive risk score is greater than or equal to the severity threshold but less than the critical threshold, a severe defect is output. When the comprehensive risk score is greater than or equal to the critical threshold, a critical defect is identified and output.

[0016] This invention provides a method for intelligent identification and classification of defects in transmission line images. It has the following beneficial effects: 1. This invention constructs a joint feature vector by extracting high-dimensional visual features and topological risk weights, and introduces a dynamic adjustment factor based on a nonlinear smoothing function in database retrieval. When measuring the similarity of historical samples, it can adaptively adjust the distance ratio of visual parameters and positional parameters according to the different physical force positions of defects, thereby effectively reducing single visual misjudgments caused by lighting interference or limited shooting angles and improving the accuracy of output retrieval prior levels.

[0017] 2. This invention uses a binary spatial mask matrix to separate feature tensors into regions, performing linear interpolation only in the active defect areas identified by the mask towards the evolution anchor points, while keeping the initial feature input unchanged in the background regions. This method avoids background distortion during image generation and, by recording the step size ratio when classification levels change, outputs gradient jump points, enabling quantitative measurement of the evolution potential of defects in a gradual state.

[0018] 3. This invention performs polarity conversion and normalization alignment on the prior level of retrieval, topological risk weight, and gradient jump point, and obtains a comprehensive risk score by weighting with preset weight coefficients. This rating mechanism mathematically integrates empirical judgment based on historical matching, mechanical judgment based on spatial location, and trend judgment based on generation interpolation, overcoming the limitation of traditional methods that rely solely on static features of a single image for direct rating, and improving the consistency between the final output defect severity level result and the actual operating risk of the transmission line. Attached Figure Description

[0019] Figure 1 This is a system architecture diagram of the present invention; Figure 2 This is a flowchart of the method of the present invention; Figure 3 A schematic diagram illustrating the principle of topological risk weight calculation provided in this embodiment of the invention; Figure 4 This is a schematic diagram of the joint feature retrieval and prior classification process provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of a generative decoupled interpolation network architecture provided in an embodiment of the present invention; Figure 6 This is a schematic diagram illustrating the principle of potential space verification under spatial decoupling constraints provided in an embodiment of the present invention. Figure 7 A schematic diagram illustrating the principle of multidimensional matrix decision-making and comprehensive hierarchical classification provided in this embodiment of the invention.

[0020] Among them, 101 is an image sensing device; 102 is a computing and processing device; 201 is a feature extraction and topology analysis module; 202 is a risk weight calculation module; 203 is a feature database retrieval module; 204 is a decoupling interpolation generation and verification module; 205 is a comprehensive decision rating module; and 301 is a feature vector database. 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: Please see the appendix Figure 1 This invention provides an intelligent identification and grading system for transmission line defect images, which may include: Image sensing device 101 and computing processing device 102 communicatively connected to image sensing device 101.

[0023] The computing and processing device 102 has a built-in feature extraction and topology parsing module 201, a risk weight calculation module 202, a feature database retrieval module 203, a decoupling interpolation generation and verification module 204, and a comprehensive decision rating module 205.

[0024] The computing processing device 102 has an external or internal feature vector database 301.

[0025] The feature vector database 301 pre-stores verified historical transmission line defect image features and their corresponding physical classification labels.

[0026] The feature extraction and topology parsing module 201 is used to acquire the original image of the transmission line and perform network inference, output the two-dimensional bounding box of the target defect, high-dimensional visual features, preliminary classification confidence and topology benchmark, and generate a binary space mask matrix. The risk weight calculation module 202 is used to calculate the spatial geometric Euclidean distance between the center location of the target defect and the topological reference point, and to calculate the topological risk weight in combination with the engineering stress coefficient. The feature database retrieval module 203 is used to concatenate high-dimensional visual features and topological risk weights into a joint feature vector, perform dynamic metric comparison retrieval in the feature vector database 301, and output the retrieval prior level and evolution anchor point image. The decoupled interpolation generation verification module 204 is used to map the original image of the transmission line and the evolution anchor point image to the latent space, combine the binary space mask matrix to perform calculations on the latent space features to generate interpolation trajectory, generate a transition state synthetic image sequence through inverse mapping and feedback detection, and output the gradient jump point corresponding to the jump. The comprehensive decision rating module 205 is used to integrate the prior retrieval level, topological risk weight and gradient jump point to calculate the comprehensive risk score, and output the severity level of the defect in combination with the judgment threshold.

[0027] Please see the appendix Figure 2 This invention provides a method for intelligent identification and classification of transmission line defect images, comprising the following steps: S110: Obtain the original image of the transmission line and input it into the feature extraction network to extract the two-dimensional bounding box, high-dimensional visual features, preliminary classification confidence and topological reference points of the target defect, and generate a binary space mask matrix based on the coordinate range of the two-dimensional bounding box. S120 calculates the spatial geometric Euclidean distance between the center of the target defect and the topological reference point on the two-dimensional pixel plane, and calculates the topological risk weight by nonlinear mapping in combination with the engineering stress coefficient. S130 concatenates high-dimensional visual features with topological risk weights into a joint feature vector, and applies a dynamic distance metric algorithm to compare and retrieve the feature vector database to extract the prior level and the evolutionary anchor image with the highest similarity. S140, under the set triggering conditions, the original image of the transmission line and the evolution anchor point image are mapped to the latent space. The binary space mask matrix is ​​used to decouple the feature data in the latent space to generate the interpolation trajectory. The transition state synthetic image sequence is generated by inverse mapping along the interpolation trajectory and input into the network for detection. The interpolation step size ratio for determining the step transition is recorded and output as the gradient transition point. S150 inputs the prior level of retrieval, topological risk weight, and gradient jump point into the risk scoring equation to calculate the comprehensive risk score, and outputs the final defect severity level by combining the preset interval threshold of the project.

[0028] See attached document Figure 3 To achieve accurate extraction and spatial feature separation of the basic characteristics of transmission line defects, step S110 further includes the following steps: Step S111: Feature extraction and topology parsing module 201 acquires the original image of the transmission line collected by the image sensing device.

[0029] As a preferred approach, considering the differences in hardware sensing devices, data alignment is required before input model processing. Specifically, the feature extraction and topology parsing module 201 performs format alignment on the original transmission line image, converting the image to a preset RGB color channel format and a uniform pixel resolution, such as scaling it proportionally to a preset 640×640 pixel size, to generate a standardized input image. .

[0030] Step S112: For the feature extraction network involved in this scheme, in order to take into account the recognition accuracy of multi-scale targets, a convolutional neural network structure including a backbone feature extraction architecture, a feature pyramid network, and a multi-task detection head can be adopted.

[0031] The feature extraction and topology parsing module 201 standardizes the input image. The input is used for forward propagation inference in this feature extraction network. The detection branch within the feature extraction network is responsible for outputting the two-dimensional bounding box of the corresponding target defect. 2D bounding box The coordinate data structure is defined as follows .parameter and The parameter represents the coordinates of the center point of the target defect region in a two-dimensional pixel plane. The pixel width of the 2D bounding box, parameter Represents the pixel height of the two-dimensional bounding box.

[0032] The feature extraction and topology parsing module 201 extracts the feature map output from the deep convolutional layer of the feature extraction network, and performs a global average pooling operation on the feature map to output a one-dimensional high-dimensional visual feature vector. ,constant The dimension of the feature vector is usually determined by the number of output channels of the selected backbone feature extraction architecture, and is typically set to empirical values ​​such as 512 or 1024.

[0033] The classification branch embedded in the feature extraction network synchronously outputs a preliminary classification of the target defect based on high-dimensional visual features. and the corresponding preliminary classification confidence level Preliminary classification confidence level The probability value of the representation model in determining the current category is limited to the range of [0,1].

[0034] In step S113, the feature extraction and topology parsing module 201 obtains the topological reference points associated with the target defect through the key point detection branches set in parallel in the feature extraction network.

[0035] From a physical mechanics perspective, the severity of a defect in a transmission line component often depends on whether that component bears the primary mechanical tension or structural support. Based on this, the topology reference point represents the core mechanical connection node in the transmission component structure that bears tensile or gravitational forces, such as the hanging hole of a suspension clamp or the center of the steel cap of an insulator string. The coordinates of the topology reference point output by the key point detection branch on the two-dimensional pixel plane are defined as follows: .

[0036] Considering that actual inspection images typically contain multiple power transmission components or multiple defect targets, the feature extraction and topology parsing module 201 further calculates the spatial affinity field between the center point of each two-dimensional bounding box and each topological reference point, mapping the target defect to the topological reference point of its associated component one-to-one, and outputting the paired coordinate set to avoid mismatches in subsequent topological distance calculations. To enable this feature extraction network to output the aforementioned coordinates and classification data, the training process of the network model is described here.

[0037] To enable the feature extraction network to output the aforementioned coordinates and classification data, the training process of the network model is described here. Those skilled in the art can collect a large number of historical transmission line inspection images to construct a training set, and business experts can manually annotate the defect bounding boxes, defect category labels, and key point physical coordinates of core components in the images.

[0038] Based on this, a joint loss function is constructed, which includes a bounding box regression loss function, a classification cross-entropy loss function, and a mean squared error loss function for key point coordinates. The gradient descent algorithm is used to continuously backpropagate to update the network weights until the model converges.

[0039] Step S114, the feature extraction and topology parsing module 201, based on the two-dimensional bounding box... The coordinate parameters are used to generate a binary space mask matrix.

[0040] The underlying principle is that simply changing image features in the latent space often leads to disordered variation of pixels across the entire image. By establishing a binary index matrix based on the original position, it is possible to isolate the defect itself from the surrounding background environment at the feature level.

[0041] The feature extraction and topology parsing module 201 initializes and constructs a model in memory that is identical to the standardized input image. A zero matrix with the same pixel resolution, with its width set to... The height is .

[0042] Combining the size coordinate parameters obtained from the previous detection, the feature extraction and topology parsing module 201 traverses the pixel coordinate space within the matrix, based on the two-dimensional bounding box. The coordinate boundaries are binarized and calculated to generate the final binary space mask matrix. .

[0043] Binary space mask matrix The formula for assigning values ​​to inner pixels is defined as follows:

[0044] ; In the formula, The horizontal coordinate of the binary space mask matrix is... And the ordinate is The quantization assignment result of the pixels. To avoid boundary dead zone determination during algorithm operation, when the spatial position coordinates A pixel is considered to fall within a 2D bounding box if the absolute difference between its x-coordinate and the center point's x-coordinate is less than or equal to half the box width, and the absolute difference between its y-coordinate and the center point's y-coordinate is less than or equal to half the box height. Within the defined rectangular physical area, the corresponding pixel value is assigned a constant 1; When spatial position coordinates Within a two-dimensional bounding box When outside the defined rectangular area, the corresponding pixel value is kept constant at 0.

[0045] The feature extraction and topology parsing module 201 physically maps regions with values ​​of 1 in the matrix to active defect regions. These active defect regions represent the spatial range within which feature data is allowed to evolve and change during subsequent processing. The same module also physically maps regions with values ​​of 0 in the matrix to frozen background regions. These frozen background regions represent the spatial range within which the initial feature state is forcibly maintained. Together, the active defect regions and the frozen background regions constitute the complete structural topology of the binary space mask matrix, directly serving as spatial position constraints in the subsequent feature decoupling stage.

[0046] See attached document Figure 4 In this embodiment, in order to solve the technical problem that it is difficult to reflect the degree of physical failure of power transmission components by relying solely on visual features, step S120 further includes the following steps to establish a quantitative correlation between geometric spatial location and mechanical risk.

[0047] Step S121, the risk weight calculation module 202 extracts the coordinate data obtained from the previous step, specifically covering the center point coordinates of the target defect area. , ) and the coordinates of the topological reference point closely related to the mechanical structure of the component ( , ).

[0048] To quantify the spatial span between the defect origin point and the critical stress node, the risk weight calculation module 202 calculates the spatial geometric Euclidean distance between these two points on a two-dimensional pixel plane. The relevant calculation formulas are as follows:

[0049] ; In the formula, The linear pixel spacing between the target defect location and the load-bearing core of the transmission component is represented from the image viewpoint. From the perspective of physical damage assessment principles, the impact of a defect on the main load-bearing structure of the component is often limited by its location. The geometric Euclidean distance in this space is obtained. The purpose is to provide a basic geometric metric for subsequent analysis of the potential damage of defects to the overall mechanical strength of components. The smaller the distance value, the closer the defect is to the core stress area, and the higher the potential risk of it causing structural fracture or failure.

[0050] In step S122, the risk weight calculation module 202 introduces a nonlinear attenuation mechanism that incorporates engineering mechanical properties to calculate the spatial geometric Euclidean distance. Mapped to topological risk weights .

[0051] As a preferred approach, different types of transmission components experience significant differences in the mechanical tension or supporting load they bear during actual power grid operation. In light of this objective engineering reality, the risk weight calculation module 202 retrieves the engineering stress coefficient of the component to which the current target defect belongs from a preset configuration database. Engineering stress coefficient Used to characterize the fundamental importance of a component, its value range is typically set to a constant between (0,1).

[0052] In specific settings, for major load-bearing hardware such as tension clamps and suspension clamps, their primary function is to maintain conductor tension. Damage to these components can easily lead to conductor derailment. Therefore, their engineering stress coefficient... The value is set relatively high, ranging from 0.8 to 1.0; for non-load-bearing or auxiliary fittings such as vibration dampers and equalizing rings, the value is relatively low, for example, ranging from 0.1 to 0.3.

[0053] To accurately reflect the weakening effect of increased spatial distance on risk transmission, the risk weight calculation module 202 also introduces a distance attenuation scale factor. This factor is a positive-zero empirical engineering constant used to adjust the steepness of the spatial risk decay curve.

[0054] Those skilled in the art can pre-configure the system based on the statistical laws governing the physical dimensions of conventional components in the power grid transmission lines. For example, when using high-definition inspection images, to prevent excessive attenuation of pixel values ​​at distant locations, The empirical values ​​are usually distributed in the range of 0.01 to 0.1 to adapt to images taken under different focal lengths or resolutions.

[0055] After obtaining the above parameters, the risk weight calculation module 202 calculates the topological risk weight based on the exponential mapping equation. The formula is defined as follows: ; In the structural design of this formula, a negative sign is introduced into the exponent of the natural constant, constructing a nonlinear mapping relationship that monotonically decreases for the distance parameter. The calculation logic shows that when the location of the target defect is closer in physical space to the core force reference point (i.e.,...),... The calculated topological risk weights are close to zero. The value exhibits a non-linear increasing characteristic and eventually approximates the value derived from the engineering stress coefficient. The theoretical upper limit determined.

[0056] Through this quantitative mapping step, the risk assessment system incorporates considerations of the mechanical sensitivity of defect locations at the data level, generating topological risk weights that reflect the likelihood of physical structural degradation. This weighted data will be used as multimodal feature input and output to the downstream comparison and retrieval process.

[0057] See attached document Figure 5 In this embodiment, to achieve deep fusion of visual representation features and physical position constraints in the multimodal feature space, step S130 further includes the following steps: In step S131, the feature database retrieval module 203 receives the high-dimensional visual features output by the feature extraction and topology parsing module 201. and the topological risk weights output by the risk weight calculation module 202 To eliminate the interference of different physical dimensions and numerical scales on the subsequent distance calculation mechanism, the feature database retrieval module 203 performs normalization processing on these two types of cross-modal data.

[0058] Specifically, the feature database retrieval module 203 uses L2 norm operations to perform high-dimensional visual features. By constraining the vector so that the sum of squares of all its elements is 1, visual features can be obtained. Simultaneously, the extreme value normalization method is used to adjust the topological risk weights. Mapping to the [0,1] interval yields normalized weights. .

[0059] After unifying the numerical scale, the feature database retrieval module 203 performs visual feature retrieval. with normalized weights Perform a tensor concatenation operation. This operation directly appends the topological feature dimension data to the end of the original visual feature dimension, constructing a tensor with a dimension of [missing value]. One-dimensional joint eigenvector Its vector structure expression is defined as By using this tensor concatenation method, the joint feature vector can simultaneously carry visual attributes representing local pixel degradation and positional attributes representing spatial mechanical risks, thus providing underlying data support for subsequent multidimensional similarity calculations.

[0060] After obtaining the joint feature vector, the process proceeds to step S132, where the feature database retrieval module 203 retrieves the joint feature vector. As a query parameter, a similarity comparison retrieval is performed in the external feature vector database 301.

[0061] The feature vector database 301 pre-stores a large number of historical manually reviewed samples containing precise physical classification labels. Each historical sample corresponds to a historical joint feature vector. subscript This represents the storage index number of historical samples in the feature vector database 301.

[0062] Considering the failure patterns of actual transmission lines, when defects occur in high-risk load-bearing areas, their physical location attributes should play a more decisive role in determining the severity level than their surface visual features. To reflect this objective law at the algorithm level, the feature database retrieval module 203 constructs and applies a dynamic distance metric algorithm to calculate the current joint feature vector. Joint feature vectors of each history similarity distance between The relevant formulas are defined as follows: ; In the formula, This represents the Euclidean distance calculation result at the visual feature level. This represents the calculated absolute difference at the topological risk characteristic level. Parameters The dynamic adjustment factor, used to control the distribution of feature weights, directly depends on the topological risk level of the current defect. To achieve adaptive weight transitions, the dynamic adjustment factor... Based on a nonlinear smoothing function, the specific mathematical expression is as follows: ; In the formula, This represents a preset high-risk topology threshold. As a preferred approach, this threshold is typically set between 0.6 and 0.8 to define the physical hazard threshold of a component in the feature space. Parameters The scaling factor for adjusting the smoothness of the mapping curve is usually configured as a positive real number between 10 and 50, which is used to control the steepness of the logistic curve near the threshold.

[0063] Based on the calculation principle of the above formula, it can be seen that when the normalized weight of the current defect... Climbing and exceeding the high-risk topology threshold At that time, this mechanism led to The value of approaches zero. This makes the coefficient representing the weight of the topological position difference in the formula (1- The forced increase, and thus the calculation of the comprehensive similarity distance At that time, the dominant proportion of the topological location difference is forcibly increased, thereby guiding the retrieval results of the feature database retrieval module 203 to prioritize historical samples with similar high-risk physical locations.

[0064] Compared to the fixed modality weight ratio in conventional multimodal retrieval, the risk drift-based metric mechanism introduced in this embodiment can more flexibly adapt to complex engineering scenarios. For the underlying fast retrieval engine architecture of massive high-dimensional vector data, those skilled in the art can use existing approximate nearest neighbor search frameworks to perform calculations. The specific index tree construction and memory addressing allocation methods are well-known technologies in the field and will not be described in detail here.

[0065] After completing the full database distance comparison, the module calls the processing logic of sub-step S133, and the feature database retrieval module 203 performs the similarity distance comparison. The numerical values ​​are sorted in ascending order for all historical samples. The feature database retrieval module 203 extracts the samples from the top of the sorted list. Historical tag information of bits, constants The system's preset upper limit for the number of samples retrieved can be configured based on the total number of samples in the feature vector database 301. As a preferred approach, a constant... The value can be set to a positive integer between 5 and 15.

[0066] Feature database retrieval module 203 targets the extracted... The system performs a weighted mode operation on each historical tag information. Specifically, the operation logic is as follows: the system uses the similarity distance between each historical sample... The reciprocal of the numerical value is used as a statistical weight to weight and accumulate the frequency of occurrence of each physical classification category. The classification level with the highest weighted cumulative value is selected as the prior retrieval level. To avoid the situation where the distance is zero when extremely similar samples are detected, which could lead to computational overflow when calculating the reciprocal, a very small positive real constant (e.g., 1 × 10⁻⁶) is added to the denominator when performing the reciprocal operation. -6 The prior level of the retrieval. This reflects preliminary empirical judgments based on historical maintenance data.

[0067] While generating the retrieval prior level, the feature database retrieval module 203 locks the similarity distance in ascending order. Historical samples that reach the minimum value. The feature database retrieval module 203 retrieves the original image data corresponding to the specific sample from the feature vector database 301 and defines it as the evolution anchor image. Evolution of anchor point images The image possesses a high overall similarity to the original image of the current transmission line in the feature space, and can serve as the physical evolution endpoint for subsequent feature-oriented interpolation operations in the latent space.

[0068] See attached document Figure 6 In this embodiment, to address the technical challenge that some defective images exhibit visual features at multiple severity levels, step S140 introduces a directed evolution stress test mechanism based on a generative model. This step further includes the following steps: During the execution of step S141, the decoupled interpolation generation verification module 204 obtains the preliminary classification confidence score output from the previous step. It is then compared with a preset confidence threshold.

[0069] As a preferred approach, the confidence threshold can be set to 0.85. When the initial classification confidence level... If the values ​​fall below this threshold, or if the distribution of categories output by the detection network contains multiple candidate labels with similar probabilities, the system will trigger a latent spatial verification mechanism. This is mainly because in actual inspections, many defects, such as metal corrosion and insulator contamination, are in a gradual state, and it is often difficult to give a definite severity level based on a single image. Therefore, a dynamic evolution test is needed to determine their state boundaries.

[0070] The decoupled interpolation generation verification module 204 incorporates an image generation model based on an autoencoder architecture. This model includes an encoder network responsible for dimensionality reduction and compression, and a decoder network responsible for reconstruction. To ensure the transparency of feature representation, the encoder network in this embodiment is configured to receive image input with a size of 256×256×3. Internally, it stacks three feature downsampling modules, each consisting of a two-dimensional convolutional layer (stride set to 2), a batch normalization layer, and a LeakyReLU nonlinear activation layer. As the network deepens, the number of feature channels increases sequentially from 3 to 64, 128, and 256.

[0071] The decoupling interpolation generation verification module 204 generates the original image of the transmission line. and the evolutionary anchor point image output by the feature database retrieval module 203 The encoder network is input synchronously. Before inputting the encoder network, to avoid subsequent mask truncation errors caused by the different physical locations of defects in the two images, the decoupled interpolation generation verification module 204 retrieves the evolution anchor point image. The built-in historical bounding box coordinates are used to align their spatial positions to the 2D bounding box Bd region of the original image of the current transmission line using affine transformation, generating spatially aligned anchor point images. Subsequently, the encoder network performs feature extraction and spatial dimensionality reduction operations on the input pixel spatial data, mapping them into high-order and low-dimensional latent feature tensors respectively. With latent feature tensor This ensures that the defect activation regions in the two sets of potential feature tensors perfectly overlap in spatial dimensions.

[0072] The dimensional structure of the aforementioned latent feature tensor is defined as follows: ,in Represents the number of potential channels. and The spatial height and width representing the potential space are typically much smaller than the original image resolution, thus filtering out high-frequency interference details in the image while preserving structural semantics.

[0073] After obtaining the latent feature tensors of independent mappings, step S142 addresses the background interference problem in the feature mixing process. The decoupled interpolation generation verification module 204 introduces the binary space mask matrix generated by the feature extraction and topology analysis module 201. Since the spatial size of the latent feature tensor differs from the size of the original pixel space, the decoupled interpolation generation verification module 204 pairs of binary space mask matrices... Perform adaptive max-pooling downsampling to align its spatial resolution to Generate a latent mask matrix .

[0074] To ensure that the defect evolution process does not cause disordered distortion of the environmental background, the decoupled interpolation generation verification module 204 utilizes a latent mask matrix. Spatially decoupled interpolation equations were constructed. The decoupled interpolation generation and verification module 204 sets an interpolation step size variable that is discretely increasing within the interval [0,1]. (For example, incrementing the value at fixed intervals of 0.1), and generating the transition state tensor according to the following interpolation equation. : ; In the above equation, the symbol This represents the Hadamard product operation, which is the bitwise multiplication of matrix elements.

[0075] Through the structural design of this decoupled interpolation equation, it can be seen that in the latent mask matrix... Defect active regions with a value of 1, feature data along the path from Towards The direction is linearly interpolated and evolved; In the latent mask matrix The background freeze area has a value of 0. The term locks the result of the latter part of the calculation as the initial input feature. This mechanism mathematically achieves the forced separation of target defect feature evolution from background environmental features, effectively avoiding feature entanglement in the generated image.

[0076] For the generated transition state tensor sequence, step S143 is responsible for converting it back to a visually visible form and performing a level crossing determination.

[0077] The decoupled interpolation generation verification module 204 calls the decoder network in the image generation model. This decoder network structure is mirror-symmetric to the encoder, consisting of multiple transposed convolutional layers (responsible for upsampling to restore spatial dimensions) and activation layers. The decoupled interpolation generation verification module 204 determines the interpolation generation verification based on the discrete step size. Incrementing the set sequence nodes will change the corresponding transition state tensor. The inputs are sequentially fed into the decoder network, where an inverse mapping operation is performed. The decoder network outputs a transitional synthetic image sequence consisting of a series of synthesized images. .

[0078] The decoupling interpolation generation verification module 204 synthesizes the transition state image sequence. Based on step size The data are fed back into the detection network of the feature extraction and topology parsing module 201 in ascending order for classification.

[0079] The system continuously monitors the classification results output by the detection network for this sequence. When the classification judgment output by the detection network first jumps from a low level to a high level, such as from a general defect to a severe defect, the decoupled interpolation generation verification module 204 stops detecting subsequent frames and records the interpolation step size ratio corresponding to the current image. The system defines and outputs this recorded value as the gradient transition point. .

[0080] Considering that some defects in a relatively mild state may still fail to cross the judgment level threshold even after evolving to the interpolation endpoint, the decoupled interpolation generation verification module 204 will identify gradient jump points in cases where no level transition is detected. The default value is 1.0.

[0081] Meanwhile, regarding the initial classification confidence level in the preceding step S141 If the confidence threshold is greater than or equal to the threshold but the potential space verification mechanism is not triggered, it indicates that the detection network has a very high degree of confidence in the current defect state. In this case, the system directly assigns a value based on the preliminary classification and outputs the gradient transition point. For example, if the initial classification is a general defect, the default value is 1.0; if the initial classification is a serious or critical defect, the default value is 0.0, to ensure the integrity of the parameters required for the operation of the downstream comprehensive risk scoring equation.

[0082] The smaller the jump point value, the more likely the characterization system only needs to introduce a small amount of high-risk features to cause the judgment model to go out of bounds, indicating that the original image itself is already on the critical edge of being extremely close to a deterioration state.

[0083] The construction and training mechanism of the autoencoder generative model involved in this step needs to be reorganized based on specific scenario data. During the network construction and training preparation phase, images of transmission components covering both intact states and defect states at various levels need to be collected to construct an unsupervised training set. During training, the model aims to reconstruct the original input image via forward propagation, without relying on paired label guidance.

[0084] To ensure that the generated images conform to the real physical appearance, a pixel-level mean squared error loss is used to constrain the model's ability to recover large smooth areas. Simultaneously, a feature-aware loss based on a pre-trained network is introduced to preserve complex texture edges. The weight parameters of the encoder and decoder are iteratively updated through backpropagation until the total reconstruction loss function converges, thus enabling the model's internal latent space to reasonably encode and map the real physical evolution process.

[0085] See attached document Figure 7 In this embodiment, to overcome the limitations of traditional rating methods that rely solely on a single visual judgment or a single data dimension, step S150 establishes a quantitative evaluation mechanism that integrates prior knowledge, physical topology, and evolutionary potential. This step further includes the following steps: During step S151, the comprehensive decision rating module 205 gathers the heterogeneous parameters output by the preprocessing module, specifically including the retrieval prior level output by the feature database retrieval module. Topological risk weights output by the risk weight calculation module and the gradient jump points output by the decoupled interpolation generation verification module. .

[0086] From the engineering principles of multidimensional joint evaluation, a single visual classification confidence score is prone to misjudgment when dealing with environmental lighting interference or long-tailed samples in the early stages of degradation. By orthogonally fusing empirical judgments based on historical matching, physical judgments based on spatial force characteristics, and evolutionary trend judgments based on generative models, the target state can be cross-verified from three dimensions: experience, current status, and trend, thereby significantly improving the reliability of the final rating result. To perform weighted fusion of data with different dimensions within the same mathematical space, normalization and polarity alignment are typically performed first.

[0087] For retrieval prior levels The comprehensive decision rating module 205 divides it by the maximum rating value defined in the business specifications. For example, under a common three-level classification system A value of 3 yields the normalized prior level mapped to the interval [0,1]. For gradient jump points Considering that a smaller jump step size indicates that the defect is closer to the deterioration boundary, i.e., a higher potential risk, this is inversely proportional to the conventional positive correlation assessment logic of risk. Based on this, the comprehensive decision rating module 205 performs an inverse operation to construct a risk gain term. .

[0088] After completing the underlying data alignment, the comprehensive decision rating module 205 calculates the final comprehensive risk score based on the preset comprehensive risk scoring equation. The formula is defined as follows: ; In the formula, , and These represent the weight coefficients of the corresponding feature parameters, and satisfy the following conditions: + + =1 constraint condition.

[0089] As a preferred approach, these weighting coefficients can be obtained through the analytic hierarchy process (AHP) or by regression fitting using a large amount of historical maintenance work order data. In conventional transmission line applications, the physical load-bearing condition significantly impacts grid operational safety; therefore, the weight of topology risk is typically set relatively high, for example... , and Empirical values ​​can be taken as 0.3, 0.4 and 0.3 respectively.

[0090] After completing the above quantitative scoring calculation, the comprehensive decision rating module 205 proceeds to step S152 to determine the final physical rating of the target defect.

[0091] The comprehensive decision rating module 205 calls the built-in logic comparator unit to calculate the comprehensive risk score. Compared with the system's preset severity threshold and critical threshold Parallel comparisons are performed. The setting of these two thresholds is based on the engineering experience of power grid business experts. They are determined by statistically analyzing the probability distribution of actual component breakage or line drop accidents in different scoring intervals over the years. Their values ​​are usually set in the range of 0.4 to 0.5 and 0.75 to 0.85, respectively.

[0092] The comprehensive decision rating module 205 executes conditional logic equations based on the comparison results from the logic comparator unit, and outputs the final defect severity level. : ; Through the design of the segmented conditional judgment structure described above, the algorithm logic achieves dead-zone-free coverage of the entire data range, effectively avoiding judgment anomalies and program crashes when the score is at boundary values. The comprehensive decision rating module 205 will output the defect severity level. The original images and location coordinates are packaged together into a structured data frame and pushed to an external business operation and maintenance platform to provide direct physical decision support for subsequent on-site troubleshooting and line maintenance operations.

[0093] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent identification and classification of defects in transmission line images, characterized in that, Includes the following steps: Acquire raw images of transmission lines, extract high-dimensional visual features, preliminary classification confidence scores, and binary space mask matrices; Extract the coordinates of the center point and the topological reference point of the target defect region, and calculate the topological risk weight by combining the nonlinear attenuation mechanism and engineering mechanical properties; The high-dimensional visual features and topological risk weights are concatenated to obtain a joint feature vector, which is then searched in a preset feature vector database to obtain the prior retrieval level and evolution anchor point image. When the initial classification confidence is lower than the confidence threshold, based on the original image of the transmission line, the evolution anchor point image and the binary space mask matrix, the generative model is used to perform decoupled interpolation and cross-level determination in the latent space, and output the gradient jump point. The prior retrieval level, topological risk weight, and gradient transition point are aligned and weighted to calculate a comprehensive risk score. The comprehensive risk score is then compared with a preset level threshold to output the final defect severity level.

2. The intelligent identification and classification method for transmission line defect images according to claim 1, characterized in that, The calculation of topological risk weights by combining nonlinear attenuation mechanisms and engineering mechanical properties specifically includes: Calculate the spatial geometric Euclidean distance between the coordinates of the center point and the coordinates of the topological reference point on the two-dimensional pixel plane; Retrieve the engineering stress coefficient of the component to which the current target defect belongs, as well as the preset distance attenuation scale factor, from the preset configuration database; Using a nonlinear mapping equation that includes a negative natural constant exponent, the spatial geometric Euclidean distance, engineering stress coefficient, and distance attenuation scale factor are mapped into topological risk weights. The smaller the spatial geometric Euclidean distance, the closer the topological risk weight is to the theoretical upper limit determined by the engineering stress coefficient.

3. The intelligent identification and classification method for transmission line defect images according to claim 1, characterized in that, The process of performing feature concatenation on high-dimensional visual features and topological risk weights to obtain a joint feature vector specifically includes: The high-dimensional visual features are numerically constrained by L2 norm operations to obtain normalized visual features. The topological risk weights are mapped to a set interval using the extreme value normalization method to obtain normalized weights; The dimension data of the normalized weights are directly appended to the end of the dimension of the normalized visual features, and a tensor concatenation operation is performed to construct a one-dimensional joint feature vector.

4. The intelligent identification and classification method for transmission line defect images according to claim 3, characterized in that, The retrieval in the preset feature vector database specifically includes: Based on the current defect topology risk weight and the preset high-risk topology threshold, a dynamic adjustment factor is calculated through a nonlinear smoothing function; The dynamic adjustment factor is used to control the adaptive transition of the feature weight distribution; By applying a dynamic distance metric algorithm and combining it with the dynamic adjustment factor, the Euclidean distance between the joint feature vector of the current defect and each historical joint feature vector in the feature vector database at the visual feature level, as well as the absolute difference at the topological risk feature level, are calculated to obtain the comprehensive similarity distance.

5. The intelligent identification and classification method for transmission line defect images according to claim 4, characterized in that, The acquisition of the retrieval prior level and evolution anchor point image specifically includes: Based on the similarity distance values, the historical samples in the feature vector database are sorted in ascending order, and the historical label information of the top K positions is extracted. The reciprocal of the similarity distance values ​​of each extracted historical sample is used as the statistical weight. The frequency of occurrence of each physical classification category is weighted and accumulated. The classification level with the largest weighted accumulated value is selected as the retrieval prior level. Lock the historical samples whose similarity distance reaches the minimum value in ascending order, and retrieve their corresponding original image data as the evolution anchor image.

6. The intelligent identification and classification method for transmission line defect images according to claim 1, characterized in that, The process of performing decoupling interpolation in the latent space using a generative model specifically includes: The historical bounding box coordinates of the evolved anchor point image are retrieved, and the spatial position is aligned to the two-dimensional bounding box range of the original image of the transmission line using an affine transformation operation to generate a spatially aligned anchor point image. The original image of the transmission line and the spatially aligned anchor point image are input into the encoder network to perform feature extraction and spatial dimensionality reduction operations, which are respectively mapped into independent high-order low-dimensional latent feature tensors. An adaptive max-pooling downsampling operation is performed on the binary spatial mask matrix to align its spatial resolution with the latent feature tensor, thereby generating a latent mask matrix.

7. The intelligent identification and classification method for transmission line defect images according to claim 6, characterized in that, The specific process of performing decoupling interpolation is as follows: Set a discretely increasing interpolation step size variable; Based on the potential mask matrix, construct spatially decoupled interpolation equations; In the region where the latent mask matrix is ​​numerically represented as a defect active area, the feature data is linearly interpolated along the latent feature tensor of the original image of the transmission line to the latent feature tensor of the evolved anchor image. The region in the potential mask matrix that is numerically represented as the background frozen area is locked as the initial input feature of the original image of the transmission line; This generates a transition state tensor sequence with feature stripping.

8. The intelligent identification and classification method for transmission line defect images according to claim 7, characterized in that, The cross-level determination and output gradient jump point specifically include: The generated transition state tensor sequence is sequentially input into the decoder network in ascending order of step size to perform inverse mapping operation, and the output is a transition state synthetic image sequence composed of synthetic images; The transition state synthesized image sequence is fed into the detection network frame by frame for classification. When the classification category output by the detection network first jumps from a low level to a high level, the interpolation step size ratio of the corresponding image is recorded as the gradient jump point. If no level jump is detected, the gradient jump point is assigned a default upper limit value; For cases where the initial classification confidence level is greater than or equal to the confidence threshold, the gradient transition point is directly assigned a value based on the initial classification category.

9. The intelligent identification and classification method for transmission line defect images according to claim 1, characterized in that, The process of aligning and weighting the prior retrieval levels, topological risk weights, and gradient transition points to calculate a comprehensive risk score specifically includes: Divide the retrieval prior level by the preset maximum level value to obtain the normalized prior level; Invert the gradient jump points to construct a risk gain term; The normalized prior level, topological risk weight, and risk gain term are multiplied by their respective preset weight coefficients and then summed to calculate the comprehensive risk score.

10. The intelligent identification and classification method for transmission line defect images according to claim 9, characterized in that, The step of comparing the comprehensive risk score with a preset level threshold and outputting the final defect severity level specifically includes: The built-in logic comparator unit is invoked to perform a parallel comparison between the comprehensive risk score and the preset severity threshold and critical threshold. When the overall risk score is less than the severity threshold, a general defect is determined and output. When the comprehensive risk score is greater than or equal to the severity threshold and less than the critical threshold, a severe defect is determined and output. When the comprehensive risk score is greater than or equal to the critical threshold, a critical defect is determined and output.

Images