Intelligent inspection methods, systems, equipment and storage media for power grid equipment defects
By using a convolutional neural network model that fuses multispectral images and 3D point cloud data, and dynamically adjusting the topology and feature weights, the efficiency and accuracy issues of existing technologies in power grid equipment defect inspection are solved, achieving more efficient defect detection.
Patent Information
- Application Number
- CN202510635800.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-16
- Publication Date
- 2026-11-17
AI Technical Summary
Existing technologies for power grid equipment defect inspection suffer from low efficiency due to manual inspection, difficulty in identifying early micro-defects, reliance on single spectral features and insufficient three-dimensional geometric information for image detection, poor generalization ability of traditional image processing algorithms, and inability of convolutional neural network models to dynamically adjust topology to adapt to multi-scale features, resulting in low detection accuracy and efficiency.
We employ a convolutional neural network model based on feature importance feedback. By fusing multispectral images and 3D point cloud data, we dynamically adjust the topology and feature weights of the convolutional neural network. Combined with implicit second-derivative estimation and manifold learning methods, we optimize the network training process and enhance its adaptability to complex image data.
It improves the accuracy and efficiency of power grid equipment defect detection, can better capture complex texture and structural information, avoids overfitting, enhances the adaptability to complex image data, and strengthens the stability and detection accuracy of the model.
Smart Images

Figure CN123119084A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of power grid equipment defect inspection technology, and in particular to an intelligent inspection method, system, equipment and storage medium for power grid equipment defects. Background Technology
[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.
[0003] Power grid equipment includes substation equipment, transmission lines, and distribution equipment. Defect inspection of power grid equipment is an important part of ensuring the safe and stable operation of the power system.
[0004] In existing technologies, the inspection of power grid equipment defects often involves handheld portable devices scanning the equipment point by point, combined with experience to judge the defect status; or, images of the equipment are acquired through fixed monitoring cameras or drones, and image processing algorithms (such as edge detection and template matching) are used for defect detection. However, manual inspection is limited by human subjectivity and inspection efficiency, making it difficult to cover all equipment areas, and its recognition rate for early micro-defects (such as cracks <0.1mm) is low. Image detection methods are limited by the imaging capabilities of image acquisition equipment, and the acquired data lacks multispectral features and three-dimensional geometric information, failing to meet the needs of refined diagnosis. Furthermore, traditional image processing algorithms heavily rely on preset thresholds for detecting surface defects, exhibiting poor generalization ability when faced with complex defect morphological changes, resulting in a high false alarm rate.
[0005] While existing technologies have disclosed solutions for intelligent defect detection of power grid equipment using neural networks (such as ResNet and YOLO), these models typically employ fixed topology structures and static feature extraction methods. They cannot dynamically adjust the network structure and feature weights based on the complexity of the equipment image and changes in the state of local image regions. Consequently, they struggle to adapt to the unique multi-scale characteristics of power grid equipment (such as ranging from millimeter-level bolts to meter-level structures). When faced with equipment images with complex textures or structures, both the accuracy and efficiency of model recognition decrease significantly. Summary of the Invention
[0006] To address the aforementioned issues, this invention proposes an intelligent inspection method, system, device, and storage medium for power grid equipment defects. Based on a convolutional neural network model with feature importance feedback, the topology of the convolutional neural network model is dynamically adjusted according to the complexity of the equipment image data. This enables the convolutional neural network model to better capture complex texture and structural information in the image, enhancing the network's adaptability to complex image data. Simultaneously, it can dynamically optimize the weight of each feature in the image data during training, resulting in more accurate final classification.
[0007] In some implementations, the following technical solutions are adopted:
[0008] A method for intelligent inspection of defects in power grid equipment includes:
[0009] Multispectral image data and three-dimensional point cloud image data of the power grid equipment under test are acquired and preprocessed. The preprocessed three-dimensional point cloud image data and spectral image data are aligned in time and space.
[0010] The aligned multispectral image data and 3D point cloud image data are input into the trained convolutional neural network model, which outputs the type of power grid equipment and the defect detection results.
[0011] During the training process, the convolutional neural network model uses the implicit estimation method of second derivative based on the input image data, while also considering nonlinear correction terms to dynamically update the connection weights between neurons in the convolutional neural network model; and automatically adjusts the weights of each feature in the model according to the importance of different features in the input image data.
[0012] As a further approach, during the training process of the convolutional neural network model, the connection weights between each neuron are first randomly initialized. Specifically, the connection weight matrix is randomly generated, the randomly generated matrix is biased and corrected using a bias correction factor, and the bias-corrected matrix is adjusted using an activation function based on the input image data.
[0013] Specifically, the initialized connection weights are as follows:
[0014] W p =λ1·B p ·rand(a,b)+λ2·sig(X p );
[0015] Among them, W p The concatenation weight matrix is defined by λ1, the first learning rate parameter, and λ2, the second learning rate parameter; rand(a,b) generates an a×b random matrix; sig() is the sigmoid activation function; X p Device image data is input to the convolutional neural network model.
[0016] As a further solution, the bias correction factor is specifically:
[0017]
[0018] Among them, B p X is the bias correction factor, where n is the number of samples input into the convolutional neural network model in the current batch; pi For the i-th device image data feature, μ pλ3 represents the cluster center of the device image data; λ3 is the weighting parameter of the correction factor; |||| is the L2 norm.
[0019] As a further solution, during the training process of the convolutional neural network model, based on the input image data, an implicit second-derivative estimation method is used, while also considering nonlinear correction terms, to dynamically update the connection weights between neurons in the convolutional neural network model. Specifically:
[0020] The updated connection weight matrix is equal to the sum of the original connection weight matrix, the weight update amount implicitly estimated by the second derivative, and the nonlinear correction term.
[0021] Specifically, the weight update amount of the implicit estimation of the second derivative is the negative of the product of the learning rate of the convolutional neural network model, the inverse of the Hessian matrix of the loss function of the convolutional neural network model, and the gradient of the loss function of the convolutional neural network model; wherein, the Hessian matrix of the loss function of the convolutional neural network model is calculated based on the feature weight correction factor.
[0022] The nonlinear correction term is as follows:
[0023]
[0024] Where λ5 is the weighting parameter of the nonlinear correction term; σ p L is a nonlinear exponent that controls the strength of the correction term's influence. p Let X be the loss function of the convolutional neural network model. pi is the feature of the i-th device image data; n is the number of samples input into the convolutional neural network model in the current batch.
[0025] As a further embodiment, the loss function of the convolutional neural network model is the sum of the topology evolution loss function and the manifold learning loss function; the topology evolution loss function introduces a topology factor, which is calculated based on the changes in the features of the input image data, so that the topology evolution loss function can be dynamically adjusted according to the changes in the features of the input image data; the manifold learning loss function introduces the importance of the device image data features, and automatically adjusts the weight of each feature in the model according to the importance of different features in the input image data.
[0026] As a further approach, the method for determining the importance of different features in the input image data is as follows:
[0027]
[0028] Among them, I p (X pi ) represents the importance of the i-th device image data feature, λ6 is the weighting parameter of the second derivative term, and L pLet X be the loss function of the convolutional neural network model. pi Let i be the feature of the i-th device image data.
[0029] In other embodiments, the following technical solutions are adopted:
[0030] An intelligent inspection system for defects in power grid equipment, comprising:
[0031] The data acquisition module is configured to acquire multispectral image data and three-dimensional point cloud image data of the power grid equipment under test, respectively, and preprocess them, and align the preprocessed three-dimensional point cloud image data with the spectral image data in time and space.
[0032] The defect detection module is configured to input aligned multispectral image data and 3D point cloud image data into a trained convolutional neural network model and output the defect detection results of the power grid equipment.
[0033] During the training process, the convolutional neural network model dynamically updates the connection weights between neurons in the model based on the input image data using an implicit second-derivative estimation method and considering nonlinear correction terms. Furthermore, it automatically adjusts the weights of each feature in the model according to the importance of different features in the input image data.
[0034] In other embodiments, the following technical solutions are adopted:
[0035] A terminal device includes a processor and a memory, wherein the processor is used to implement instructions; and the memory is used to store multiple instructions, which are adapted to be loaded and executed by the processor to perform the above-described intelligent inspection method for defects in power grid equipment.
[0036] In other embodiments, the following technical solutions are adopted:
[0037] A computer-readable storage medium storing a plurality of instructions adapted for loading and execution by a processor of a terminal device of the above-described intelligent inspection method for defects in power grid equipment.
[0038] In other embodiments, the following technical solutions are adopted:
[0039] A computer program product includes a computer program / instructions that, when executed by a processor, implement the aforementioned intelligent inspection method for defects in power grid equipment.
[0040] Compared with the prior art, the beneficial effects of the present invention are:
[0041] (1) The present invention constructs a multimodal data acquisition system, which acquires multispectral image data and three-dimensional point cloud image data through different image acquisition devices, and ensures the accurate spatial position of point cloud data through data alignment, so that the three-dimensional point cloud and the two-dimensional spectral image are consistent in time and space, providing accurate geometric information and complete data information for subsequent feature extraction and fusion, thereby improving the performance and reliability of the entire intelligent inspection system.
[0042] (2) The convolutional neural network model proposed in this invention can dynamically adjust the topology of the neural network according to the feature distribution of the input image data to adapt to image data from different types of devices and working environments. This dynamic adjustment mechanism enables the network to extract features more efficiently when faced with highly uneven image features, thereby improving the overall detection accuracy. In addition, the dynamically adjusted topology also enhances the network's adaptability to complex image data, making the model's performance more stable in different scenarios.
[0043] In the early stages of model training, a heuristic method is used to set the initial connection weights between neurons. At the same time, a bias correction factor for the local data distribution of the device image is introduced. Through weighted bias correction, the connection response of neurons is effectively adjusted, enabling the network to better adapt to local changes in the device image data and improve the accuracy of defect detection.
[0044] During model training, the connections and positions of neurons are adaptively adjusted by minimizing the topological evolution loss function to ensure that the network structure dynamically adapts to the data distribution and avoids overfitting. Second-order derivative implicit estimation is used to update and optimize the neuron connection weights, accelerating parameter optimization and improving the efficiency of capturing high-dimensional subtle features. The synergistic effect of these two methods allows for more precise adjustment of connection weights, which not only speeds up convergence but also solves the gradient vanishing problem. In high-dimensional feature space, it can effectively capture subtle local feature changes in images. At the same time, a nonlinear correction term is introduced to enhance the network's ability to model nonlinear device image data.
[0045] (3) The convolutional neural network model of the present invention adopts the manifold learning method. By constructing the manifold learning loss function, it can maintain the low-dimensional manifold structure of the device image data and avoid the curse of dimensionality in high-dimensional data processing. By mapping high-dimensional data to low-dimensional space and maintaining the local structural characteristics of the data, the model can effectively capture the nonlinear characteristics of the device image data in high-dimensional space, avoid the loss of important information, and promote the improvement of defect detection accuracy.
[0046] Other features and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description
[0047] Figure 1 This is a flowchart of the intelligent inspection method for power grid equipment defects in an embodiment of the present invention;
[0048] Figure 2 This is a schematic diagram of the convolution operation process of multiple convolutional layers in an embodiment of the present invention. Detailed Implementation
[0049] It should be noted that the following detailed descriptions are illustrative and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.
[0050] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.
[0051] Example 1
[0052] In one or more embodiments, an intelligent inspection method for defects in power grid equipment is disclosed, combining... Figure 1 Specifically, it includes the following process:
[0053] S101: Acquire the multispectral image data and three-dimensional point cloud image data of the power grid equipment under test, respectively.
[0054] First, it should be noted that the power grid equipment in this embodiment can be substation equipment, transmission lines, or distribution equipment, etc. This embodiment takes substation equipment defect detection as an example for explanation, but the method is also applicable to defect detection of other power grid equipment.
[0055] Existing power grid inspection technologies often rely on data from a single perspective, which cannot comprehensively acquire information on different defects or structures of power grid equipment. They lack multispectral features and three-dimensional geometric information, which can easily lead to incomplete data, fail to meet the needs of refined diagnosis, and cannot accurately diagnose various faults in equipment.
[0056] This embodiment addresses the scenario of visual inspection of substation equipment. It utilizes high-definition cameras and related sensors to acquire image data of the substation equipment under inspection, ensuring that the acquired images are clear and possess sufficient detail. The specific data acquisition equipment is as follows:
[0057] (1) An inspection drone equipped with a multispectral imaging module (wavelength range 400-1700nm) is used to acquire multispectral image data of substation equipment and capture micro-defects such as rust and cracks on the surface of substation equipment.
[0058] (2) A fixed wide-angle HDR camera array is deployed at key nodes in the substation to record the overall appearance of the substation equipment with a dynamic range of 120dB.
[0059] (3) A lidar point cloud scanning device is used to acquire the three-dimensional geometric features of substation equipment to assist in the alignment of two-dimensional images.
[0060] Data acquisition is achieved through a spatiotemporal coupled sampling protocol, establishing a collaborative relationship between UAV trajectory planning and the fixed camera triggering mechanism. When the UAV enters the electromagnetic shielding area of the substation, the fixed camera triggers the UAV with a Δt signal. a =1 / F a (F a For adaptive frame rate, F a =10+5·sin(θ) a ), θ a The frequency of data acquisition is synchronously triggered to ensure the spatiotemporal consistency of multi-view data (based on the illumination tilt angle).
[0061] The collected raw data adopts a layered encapsulation structure, specifically including:
[0062] (1) Store the multispectral image sequence in 12-bit RAW format and embed the high dynamic range metadata of the ISO / IEC 23008-12 standard.
[0063] (2) The point cloud data is encoded using Octree compression, and each voxel contains a reflection intensity R. a and polarization phase D a Dual-channel characteristics.
[0064] All data streams are synchronized using spatiotemporal stamps via the IEEE 1857.4 protocol, forming a multidimensional tensor. (H a W a These are the length and width of the spatial dimension, C. a =3 represents the number of spectral channels, S a (Time series depth).
[0065] S102: Preprocess the acquired data and align the preprocessed 3D point cloud image data with the spectral image data in time and space.
[0066] In this embodiment, data preprocessing mainly includes operations such as image denoising, cropping, and standardization to ensure the quality and consistency of the images input to the convolutional neural network model.
[0067] As a further implementation, target alignment of point cloud data is performed. The target alignment is performed using feature points in the point cloud data (such as the edges and corners of the equipment). The point cloud data is aligned with a pre-built 3D model of the power grid equipment or a reference point cloud through an alignment algorithm (such as the Iterative Closest Point Algorithm, ICP) to ensure the accurate spatial position of the point cloud data.
[0068] After alignment, the point cloud data will be spatially consistent with the two-dimensional image data, providing accurate geometric information for subsequent feature extraction and fusion, thereby improving the performance and reliability of the entire intelligent inspection system; it solves the problem in existing technologies that the lack of multispectral features and three-dimensional geometric information in the acquired data makes it impossible to meet the needs of refined diagnosis.
[0069] S103: Input the aligned multispectral image data and 3D point cloud image data into the trained convolutional neural network model, and output the defect detection results of the power grid equipment.
[0070] In this embodiment, a convolutional neural network model is used to classify the appearance defects of substation equipment.
[0071] The structure of a convolutional neural network model specifically includes an input layer, a convolutional layer, a pooling layer, and a fully connected layer connected in sequence; specifically:
[0072] The input layer data format is a multidimensional array of device image data, for example: the device image data input to the convolutional neural network model is H cnn ×W cnn ×C cnn The tensor of H, where H cnn It is the height of the multidimensional array of device image data, W cnn It is the width of the multidimensional array of device image data, C cnn It is the number of channels in the multidimensional array of device image data.
[0073] Convolutional layers perform convolution operations on the device image data from the input layer to extract local features; combined with Figure 2 In this embodiment, multiple convolutional layers are used to perform continuous convolution operations on the input data, which enables more comprehensive feature extraction.
[0074] Specifically, convolution operations are implemented through convolution kernels. The stride of the convolution operation is preset by the user. For example, if the stride is 1, zeros are padded at the edges of the device image data in the input layer to control the size of the output.
[0075] Pooling layers reduce the size of feature maps by downsampling, thereby reducing computational complexity while preserving key features. Pooling layers can be either max pooling (taking the maximum value in the pooling window) or average pooling (taking the average value in the pooling window). The pooling window size is preset by the user, such as 2×2.
[0076] The fully connected layer flattens the high-dimensional feature map into a one-dimensional vector, performs calculations with the fully connected neurons, and obtains the defect type identification result based on the output of the fully connected layer.
[0077] In this embodiment, the training process for the convolutional neural network model is as follows:
[0078] S1031: First, acquire historical datasets, collecting historical multispectral image data and 3D point cloud data of substation equipment.
[0079] S1032: Manually label the data in the dataset, specifically labeling the equipment type (e.g., transformer, switchgear, instrument transformer, etc.) and the equipment defect type.
[0080] As a specific example, five types of equipment defects can be defined, including: normal condition, minor corrosion, structural cracks, insulation contamination, and missing parts; the criteria for each type are shown in Table 1.
[0081] Table 1 Defect Types and Judgment Criteria
[0082] Defect types Judgment criteria Normal state <![CDATA[surface reflectance R a ∈[0.65,0.85]]]> Slight rust <![CDATA[Rust spot area ratio η a ∈(0, 15%]]]> Structural cracks <![CDATA[Crack aspect ratio ρ a >8:1]]> Insulation contamination <![CDATA[Gray level co-occurrence matrix entropy E a >4.2]]> Missing parts <![CDATA[Contour Hu moment difference ΔH a >0.3]]>
[0083] Of course, those skilled in the art can choose the type of defect to be marked and the judgment criteria according to actual needs.
[0084] It should be noted that the surface reflectivity R a , rust spot area ratio η a Crack length-to-width ratio ρ a Gray-level co-occurrence matrix entropy E a The difference between the contour Hu moment and ΔH a The specific calculation methods are readily known to those skilled in the art and will not be detailed here.
[0085] S1033: After the data is labeled, target alignment of the point cloud data is performed so that the point cloud data will be spatially consistent with the two-dimensional image data.
[0086] S1034: Input the aligned training dataset from S1033 into the convolutional neural network model to train the convolutional neural network model.
[0087] In existing technologies, fault diagnosis systems based on convolutional neural network models often suffer from overfitting of training data, especially when faced with high-dimensional and complex device images. The models have poor generalization ability and are prone to a decrease in recognition accuracy. In addition, existing models cannot dynamically adjust the network topology according to the data distribution during training, resulting in poor adaptability of the network to complex data distribution, which affects the accuracy and training efficiency of the model.
[0088] In this embodiment, the topology of the convolutional neural network model is dynamically adjusted according to the complexity of the device image data, enabling the network to better adapt to the characteristics of complex device image data. It can accurately extract features from power grid equipment images with different lighting, textures, and structures, thereby effectively avoiding overfitting. At the same time, it enables the network to automatically optimize its structure when faced with highly uneven image features, improving its adaptability to complex device image data and feature extraction accuracy.
[0089] As a specific implementation method, the training process of the convolutional neural network model in this embodiment is as follows:
[0090] (1) In the early stages of training, the topology of the convolutional neural network model is initialized as a randomly connected network. Each neuron represents a certain feature dimension of the input device image data. The neurons form an initial connection matrix according to certain rules. In order to make the network more adaptable, a heuristic method is used to set the initial connection weights and biases between neurons. During initialization, the connection weights between neurons are randomly initialized and adjusted in combination with the Sigmoid activation function. The connection weight matrix during the initialization process is expressed as follows:
[0091] W p =λ1·B p ·rand(a,b)+λ2·sig(X p (1)
[0092] Among them, W p λ1 is the connection weight matrix of the convolutional neural network model; λ2 is the first learning rate parameter and λ1 is the second learning rate parameter; rand(a,b) generates an a×b random matrix; sig() is the sigmoid activation function used to adjust the feature bias; X p The input device image data is used as input to the convolutional neural network model; 'a' is the first dimension of the connection weight matrix of the convolutional neural network model, and 'b' is the second dimension of the connection weight matrix of the convolutional neural network model. As an optional example, λ1 is set to 0.3 and λ2 is set to 0.1.
[0093] Among them, B p The bias correction factor is used to adjust the local data distribution of the device image. It is related to the distance between the current device image data features and the cluster centers of all device image data. The larger the distance, the larger the bias correction factor, allowing the network to better adapt to local variations in the device image data. Device image data often contains certain local regions with different textures, lighting, or shapes. These variations can affect the accuracy of feature extraction. Through weighted bias correction, the network can effectively adjust the responses of neurons, avoiding over-reliance on any one type of feature and improving classification accuracy. This is expressed as:
[0094]
[0095] Where n is the number of samples input into the convolutional neural network model in the current batch; X pi Let μ be the feature of the i-th device image data; p λ3 is the cluster center of the device image data, which is the center of all input device image data; |||| is the weighting parameter of the correction factor; |||| is the L2 norm; as an optional example, λ3 is set to 0.5.
[0096] (2) During training, the connection relationship of neurons will be dynamically adjusted according to the local structure of the device image data. Specifically, the neurons selectively adjust their connection strength and the relationship with neighboring neurons according to the current distribution of the device image data, so as to better adapt to the changes in the device image data, avoid overfitting and improve the generalization ability. By minimizing the topological evolution loss function, the neuron connection and position will be adaptively adjusted to optimize the extraction of device image data features.
[0097] Specifically, the topology evolution loss function is:
[0098]
[0099] In the formula, L evo X is the topology evolution loss function; pi Y represents the features of the i-th device image data. i The label corresponding to the i-th device image data; N e N represents the location of the e-th neuron; e-1 γ represents the position of the (e-1)th neuron; γ is the weight parameter for topology adjustment; m is the number of neurons; Nd is the dynamically adjusted topology factor. As an optional example, γ is set to 2.
[0100] As a further solution, device image data often contains highly uneven feature distributions, requiring the network to adjust its topology in real time according to different data distributions, thereby improving its adaptability to the complexity of image data. In this embodiment, the dynamically adjusted topology factor Nd changes continuously according to data variations, thus continuously optimizing the network topology during training, enabling the model to accurately extract important features from the image data. The calculation method of the dynamically adjusted topology factor is expressed as follows:
[0101]
[0102] Where, θ p To control the distance item The variable parameter, for example, can be set to 0.1; X pkX is the feature of the kth device image data; pk-1 This is the feature of the (k-1)th device image data.
[0103] (3) During training, the topology of the convolutional neural network model is further optimized. The number of neurons and the connection method are adaptively adjusted according to the complexity of the device image data, thereby improving the feature extraction capability. For example, an image may contain multiple objects or complex background textures. By optimizing the topology, the network can extract the features of these objects or textures at a higher level, improving the accuracy of image classification. The weight update method of the convolutional neural network model is expressed as follows:
[0104] W′ p ←W p +ΔW p +ΔL non (5)
[0105] Among them, W′ p For the updated connection weight matrix, W p The connection weight matrix before the update is ΔW. p The weight update amount, ΔL, is the implicit estimate of the second derivative. non This is a nonlinear correction term.
[0106] Specifically, this embodiment uses implicit second-derivative estimation for network optimization. By employing implicit second-derivative estimation, the second-order information of the loss function of the convolutional neural network model is considered, thereby allowing for more precise adjustment of weights and improving training efficiency. Device image data typically has high dimensionality; using implicit second-derivative estimation not only accelerates convergence but also solves the gradient vanishing problem. In high-dimensional feature spaces, it can effectively capture subtle local feature changes in images. The weight update method of implicit second-derivative estimation is expressed as follows:
[0107]
[0108] Wherein, ΔW p η represents the weight update amount of the convolutional neural network model. p L represents the learning rate of the convolutional neural network model. p L is the loss function for the convolutional neural network model. p =L evo +L man L evo Let L be the topology evolution loss function. man The loss function is the manifold learning function. The gradient of the loss function of the convolutional neural network model; η is the Hessian matrix (i.e., the second derivative matrix) of the loss function of the convolutional neural network model. As an optional example, η pSet to 0.01.
[0109] As a further solution, to improve the accuracy of feature extraction, the Hessian matrix is calculated based on a feature weight correction factor to address the gradient vanishing problem in high-dimensional device image data, and is expressed as:
[0110]
[0111] In the formula, This is a feature weight correction factor; For X pi transpose; λ is the identity matrix; λ4 is the feature weight hyperparameter. As an optional example, λ4 is set to 0.2.
[0112] In this embodiment, the nonlinear correction term enhances the network's ability to model nonlinear device image data, and its calculation method is expressed as follows:
[0113]
[0114] Where λ5 is the weighting parameter of the nonlinear correction term; σ p λ5 is a nonlinear exponent that controls the strength of the correction term's influence. As an optional example, λ5 is set to 0.3, and σ... p Set it to 3.
[0115] In this embodiment, a manifold learning method is employed to maintain the low-dimensional manifold structure of the device image data, ensuring that the locality of the device image data is not lost during feature extraction. Device image data typically exhibits nonlinear characteristics in high-dimensional space, and manifold learning can effectively capture these characteristics, avoiding information loss and improving classification accuracy. The manifold learning loss function is calculated as follows:
[0116]
[0117] In the formula, f p () represents the mapping of device image data samples in a low-dimensional space, specifically achieved through principal component analysis; X pi X is the feature of the i-th device image data; pj The j-th device image data feature;
[0118] Let be the second derivative of the mapping function at the i-th device image data point; Let be the second derivative of the mapping function at the j-th device image data point.
[0119] I p (X piThe importance of the i-th device image data feature is represented by ; key features that may exist in each device image data (such as texture features, gradient features, or local features) often have different weights, with some features having a greater impact on the classification results, while others are less important.
[0120] In this embodiment, through a feature importance feedback mechanism, the algorithm automatically adjusts the weight of each input feature in the model based on its importance, and expands this by using a second-order derivative term. Through this feedback mechanism, the network can dynamically optimize the weight of each feature during training, resulting in more accurate final classification. p (X pi The calculation method for ) is expressed as follows:
[0121]
[0122] Here, λ6 is a weighting parameter for the second derivative term, controlling the influence of the second derivative on the importance of features. As an optional example, λ6 is set to 0.2.
[0123] This embodiment utilizes manifold learning to maintain the low-dimensional manifold structure of device image data, avoiding the curse of dimensionality in high-dimensional data processing. Manifold learning maps high-dimensional data to a low-dimensional space while preserving the local structural characteristics of the data, ensuring that important information is not lost during feature extraction. With the same training data, the number of iterations required for the model to reach 95% accuracy in this embodiment is reduced by 62%, from 1200 iterations in traditional methods to 450 iterations. This demonstrates that the implicit estimation method of the second derivative significantly accelerates the model's convergence speed and improves training efficiency. Simultaneously, the manifold learning method effectively improves the accuracy of feature extraction, enabling the model to better capture local features in high-dimensional image data, thereby improving classification accuracy.
[0124] (4) Repeat the above process until the preset stopping iteration condition is met, which indicates that the model training is complete. In one embodiment, the preset stopping iteration condition is reaching the preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 1000.
[0125] The convolutional neural network model topology in this embodiment is dynamically adjusted according to the complexity of the device image data. For simultaneous detection of millimeter-level bolt defects and meter-level structural deformation, the recall rate is improved by 31%. Compared with the traditional single-scale convolutional neural network model, the dynamic adjustment mechanism in this embodiment enables the network to extract features more efficiently when facing highly uneven image features, improve the overall detection accuracy, and enhance the network's adaptability to complex image data, making the model's performance more stable in different scenarios.
[0126] Example 2
[0127] In one or more embodiments, an intelligent inspection system for defects in power grid equipment is disclosed, specifically including:
[0128] The data acquisition module is configured to acquire multispectral image data and three-dimensional point cloud image data of the power grid equipment under test, respectively, and preprocess them, and align the preprocessed three-dimensional point cloud image data with the spectral image data in time and space.
[0129] The defect detection module is configured to input aligned multispectral image data and 3D point cloud image data into a trained convolutional neural network model and output the type of power grid equipment and the defect detection results.
[0130] During the training process, the convolutional neural network model uses the implicit estimation method of second derivative based on the input image data, while also considering nonlinear correction terms to dynamically update the connection weights between neurons in the convolutional neural network model; and automatically adjusts the weights of each feature in the model according to the importance of different features in the input image data.
[0131] It should be noted that the specific implementation methods of the above modules are exactly the same as those in Example 1, and will not be described in detail again.
[0132] Example 3
[0133] In one or more embodiments, a terminal device is disclosed, which includes a processor and a memory. The processor is used to implement instructions; the memory is used to store multiple instructions, which are adapted to be loaded by the processor and executed by the intelligent inspection method for power grid equipment defects described in Embodiment 1.
[0134] It should be understood that in this embodiment, the processor can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc.
[0135] Memory may include read-only memory and random access memory, and provides instructions and data to the processor. A portion of memory may also include non-volatile random access memory. For example, memory may also store information about the device type.
[0136] In the implementation process, each step of the above method can be completed by the integrated logic circuits in the processor hardware or by software instructions.
[0137] Example 4
[0138] In one or more embodiments, a computer-readable storage medium is disclosed, wherein a plurality of instructions are stored, the instructions being adapted to be loaded by a processor of a terminal device and executed as described in Embodiment 1, the intelligent inspection method for power grid equipment defects.
[0139] Example 5
[0140] In one or more embodiments, a computer program product is disclosed, including a computer program / instruction that, when executed by a processor, implements the intelligent inspection method for power grid equipment defects described in Embodiment 1.
[0141] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.
Claims
1. A method for intelligent inspection of defects in power grid equipment, characterized in that, include: Multispectral image data and 3D point cloud image data of the power grid equipment under test are acquired and preprocessed. The preprocessed 3D point cloud image data is then aligned with the multispectral image data in time and space. The aligned multispectral image data and 3D point cloud image data are input into the trained convolutional neural network model, and the defect detection results of the power grid equipment are output. During the training process, the convolutional neural network model dynamically updates the connection weights between neurons in the model based on the input image data using an implicit second-derivative estimation method and considering nonlinear correction terms. Furthermore, it automatically adjusts the weights of each feature in the model according to the importance of different features in the input image data.
2. The intelligent inspection method for defects in power grid equipment as described in claim 1, characterized in that, During the training process, the convolutional neural network model first randomly initializes the connection weights between neurons. Specifically, it randomly generates a connection weight matrix, performs bias correction on the randomly generated matrix using a bias correction factor, and adjusts the bias-corrected matrix based on the input image data using an activation function.
3. The intelligent inspection method for defects in power grid equipment as described in claim 2, characterized in that, The initialized connection weights are specifically as follows: W p =λ1·B p ·rand(a,b)+λ2·sig(X p ); Among them, W p The concatenation weight matrix is defined by λ1, the first learning rate parameter, and λ2, the second learning rate parameter; rand(a,b) generates an a×b random matrix; sig() is the sigmoid activation function; X p Image data that serves as input to a convolutional neural network model.
4. A method for intelligent inspection of defects in power grid equipment as described in claim 2 or 3, characterized in that, The bias correction factor is specifically: Among them, B p X is the bias correction factor, where n is the number of samples input into the convolutional neural network model in the current batch; pi For the i-th device image data feature, μ p λ is the cluster center of the device image data; λ3 is the weighting parameter of the correction factor; || is the L2 norm.
5. The intelligent inspection method for defects in power grid equipment as described in claim 1, characterized in that, During the training process, the convolutional neural network model dynamically updates the connection weights between neurons based on the input image data using an implicit second-derivative estimation method, while also considering nonlinear correction terms. Specifically: The updated connection weight matrix is equal to the sum of the original connection weight matrix, the weight update amount implicitly estimated by the second derivative, and the nonlinear correction term.
6. The intelligent inspection method for defects in power grid equipment as described in claim 5, characterized in that, The weight update ΔW implicitly estimated by the second derivative p Specifically, it is the negative of the product of the learning rate of the convolutional neural network model, the inverse of the Hessian matrix of the loss function of the convolutional neural network model, and the gradient of the loss function of the convolutional neural network model; wherein, the Hessian matrix of the loss function of the convolutional neural network model is calculated based on the feature weight correction factor.
7. The intelligent inspection method for defects in power grid equipment as described in claim 5, characterized in that, The nonlinear correction term is specifically: Where λ5 is the weighting parameter of the nonlinear correction term; σ p L is a nonlinear exponent that controls the strength of the correction term. p Let X be the loss function of the convolutional neural network model. pi Let X be the feature of the i-th device image data. p The device image data is input to the convolutional neural network model; n is the number of samples input to the convolutional neural network model in the current batch.
8. A method for intelligent inspection of defects in power grid equipment as described in claim 6 or 7, characterized in that, The loss function of the convolutional neural network model is the sum of the topology evolution loss function and the manifold learning loss function. The topology evolution loss function introduces a topology factor, which is calculated based on the changes in the features of the input image data, so that the topology evolution loss function can be dynamically adjusted according to the changes in the features of the input image data. The manifold learning loss function introduces the importance of the device image data features, and automatically adjusts the weight of each feature in the model according to the importance of different features in the input image data.
9. The intelligent inspection method for defects in power grid equipment as described in claim 1, characterized in that, The method for determining the importance of different features in input image data is as follows: Among them, I p (X pi ) represents the importance of the i-th device image data feature, λ6 is the weighting parameter of the second derivative term, and L p Let X be the loss function of the convolutional neural network model. pi Let i be the feature of the i-th device image data.
10. An intelligent inspection system for defects in power grid equipment, characterized in that, include: The data acquisition module is configured to acquire multispectral image data and three-dimensional point cloud image data of the power grid equipment under test, respectively, and preprocess them, and align the preprocessed three-dimensional point cloud image data with the spectral image data in time and space. The defect detection module is configured to input aligned multispectral image data and 3D point cloud image data into a trained convolutional neural network model and output the type of power grid equipment and the defect detection results. During the training process, the convolutional neural network model dynamically updates the connection weights between neurons in the model based on the input image data using an implicit second-derivative estimation method and considering nonlinear correction terms. Furthermore, it automatically adjusts the weights of each feature in the model according to the importance of different features in the input image data.
11. A terminal device comprising a processor and a memory, the processor for implementing instructions; the memory for storing multiple instructions, characterized in that, The instructions are adapted to be loaded by a processor and executed as described in any one of claims 1-9 for intelligent inspection of defects in power grid equipment.
12. A computer-readable storage medium storing a plurality of instructions, characterized in that, The instructions are adapted to be loaded by the processor of the terminal device and executed as described in any one of claims 1-9, the intelligent inspection method for defects in power grid equipment.
13. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instruction is executed by the processor, it implements the intelligent inspection method for defects in power grid equipment as described in any one of claims 1-9.
