Secondary-circuit terminal image recognition method based on contrastive learning and improved crnn

Abstract

The present invention belongs to the technical field of secondary circuit health state evaluation of power systems, and specifically relates to a secondary-circuit terminal image recognition method based on contrastive learning and an improved CRNN. The method comprises: step 1: by means of contrastive learning, pre-training terminal block sample data of a secondary circuit of a power system; step 2: by means of using a residual neural network, improving a feature extraction layer of a CRNN; and step 3: on the basis of step 2, introducing an ECA-Net high-efficiency channel attention network, and constructing a secondary-circuit terminal recognition model based on contrastive learning and the improved CRNN. The method of the present invention can accurately recognize secondary-circuit terminal block images, and greatly improves the image recognition accuracy, detection accuracy and detection efficiency.

Description

A Secondary Circuit Terminal Image Recognition Method Based on Contrast Learning and Improved CRNN Technical Field

[0001] This invention belongs to the field of power system secondary circuit health status assessment technology, specifically involving a secondary circuit terminal image recognition method based on contrastive learning and improved CRNN. Background Technology

[0002] In power system secondary equipment, terminal blocks are crucial transition components for signal transmission between monitoring and control equipment and electrical information acquisition points. The secondary signal lines from the sampling points must correspond one-to-one with the terminal block labels. The correctness of these secondary signal line connections is vital to system stability and safety. During power system equipment maintenance, checking the correctness of the connections between the secondary circuit signal lines and terminal blocks in the junction box is a massive undertaking. Traditional identification methods rely on manual visual inspection, which is prone to errors such as misidentification and omissions. If image recognition technology can automatically identify features and extract information from the secondary circuit terminal blocks, intelligent operation and maintenance of the secondary circuits can be achieved, providing fundamental support for the safe operation of substations.

[0003] Therefore, there is an urgent need to develop an efficient and accurate method for identifying the terminal blocks and signal lines of secondary circuits in order to overcome the shortcomings of existing technologies. Summary of the Invention

[0004] The purpose of this invention is to provide a method for recognizing secondary circuit terminal images based on contrastive learning and improved convolutional recurrent neural networks (CRNN). This method can accurately recognize secondary circuit terminal block images, greatly improving image recognition accuracy, detection accuracy and detection efficiency.

[0005] Technical solution to achieve the purpose of this invention:

[0006] A method for recognizing secondary circuit terminals based on contrastive learning and improved CRNN, the method comprising:

[0007] Step 1: Pre-train the sample data of the secondary circuit terminal blocks of the power system through comparative learning;

[0008] Step 2: Improve the feature extraction layer of CRNN using a residual neural network;

[0009] Step 3: Based on Step 2, introduce the ECA-Net high-efficiency channel attention network to construct a secondary loop terminal recognition model based on contrastive learning and improved CRNN.

[0010] Step 1 includes:

[0011] Data augmentation: Data augmentation is performed by applying techniques such as cropping, flipping, rotating, adding random noise, random Gaussian blurring, and color transformation to unlabeled data. Data augmented from the same source will be treated as similar sample pairs, while the remaining data will be treated as different sample pairs.

[0012] Encoder network: Maps similar and different sample pairs to the feature space, extracts features and similarities that can distinguish samples, and outputs high-dimensional feature vectors through average pooling layers;

[0013] Projection network: Projects the high-dimensional feature vector output by the encoder network onto a low-dimensional space to obtain the feature vector output by the projection network.

[0014] The formula for calculating the feature vector output by the projection network in step 1 is: z i '=g(h i ) = W (2) σ(W (1) h i )

[0015] In the formula: z i ' is the feature vector output by the projection network; h i W is the feature vector output by the average pooling layer. (1) W (2) Let σ be the weight matrix of the two fully connected layers in the projection network; σ(·) is the activation function ReLU.

[0016] Step 1 further includes: constructing a loss function, pre-training the encoder using the loss function to enable it to extract key features that distinguish terminal block images, outputting a high-quality feature embedding vector X1, and obtaining a trained encoder.

[0017] The formula for calculating the loss function is: L i,j =λL NT +(1-λ)L sup

[0018] In the formula: L i,j L is the loss function; NT For standardized temperature-scaled cross-entropy loss; L sup To extend the contrastive loss function; λ is the weighting parameter;

[0019] The formula for calculating the normalized temperature-scale cross-entropy loss is:

[0020] In the formula: L NT The standard temperature-scaled cross-entropy loss is used; N is the total number of samples; Γ(·) is a binary function, which is 1 if the condition is met, and 0 otherwise; z j(i) To be with z iEigenvectors of the same origin; z g To be with z i Feature vectors of different types; τ is the temperature parameter; i and g are both counting units; sim(·) is the cosine similarity;

[0021] The formula for calculating the extended contrastive loss function is:

[0022] In the formula: L sup To expand the contrastive loss function; For the label y i The number of samples; z j To be with z i Feature vectors of the same type but from different sources; i, g, j are all counting units; τ is a temperature parameter; Γ(·) is a binary function, which is 1 when the condition is met and 0 otherwise; sim(·) is the cosine similarity;

[0023] The formula for calculating cosine similarity is:

[0024] In the formula: z i and z j These are the feature vectors output by the two projection networks, and the feature vectors after normalization.

[0025] Step 2 specifically involves: transferring the encoder trained in Step 1 to the improved CRNN model as the initialization parameters for the feature extraction layer; optimizing the feature extraction layer using a residual neural network structure, and continuing to use the contrastive learning loss function L from Step 1. i,j Supervised training is performed to further enhance the model's ability to extract features from terminal block images, and the feature vector X2 is output.

[0026] The formula for calculating the residual unit in the residual neural network in step 2 is: r l =p(s l )+F(s l J l ) s l+1 =R(r) l )

[0027] In the formula, s l and s l+1 Let F and J represent the input and output of the l-th residual unit, respectively; F represents the learned residual; J is the convolution operation; r l Represents a residual unit; p is the input signal s l An identity mapping is performed; R is the activation function of a linear rectified neural network; p(s) l ) = s l Represents an identity mapping;

[0028] The learning features from shallow layer l to deep layer L are:

[0029] In the formula, s L This represents the output of the residual unit in the deep L layer.

[0030] The gradient calculation formula for the reverse process in the residual neural network in step 2 is as follows:

[0031] In the formula, The first factor represents the gradient of the loss function to l. This represents the gradient of the loss function to reach L. This represents the gradient from deep layer L to shallow layer l. This represents the gradient of the sum of residuals learned from layer l to layer L-1 with respect to l.

[0032] Step 3 specifically involves introducing an ECA-Net efficient channel attention network based on the feature vector X2 output in step 2, adaptively assigning weights to each feature channel to obtain an improved feature vector X3. The improved feature vector X3 is then processed through the recurrent layer of a CRNN for sequence modeling and contextual information processing, ultimately completing the feature classification and text recognition task.

[0033] The formula for calculating the channel weights in the efficient channel attention network in step 3 is: ω=σ(Mq)

[0034] In the formula, ω is the channel weight; σ is the Sigmoid function; M is the C×C parameter matrix; ω=σ(C1D k (q))

[0035] In the formula, C1D is a 1D convolution operation; k is the kernel size.

[0036] The formula for calculating the kernel size in the efficient channel attention network in step 3 is as follows:

[0037] In the formula, k is the kernel size; |t| odd Let t represent the nearest odd number; C is the channel dimension; f and γ are both constants.

[0038] The beneficial technical effects of this invention are as follows:

[0039] 1. The present invention provides a method for recognizing secondary circuit terminal images based on contrastive learning and improved CRNN. ​​By introducing contrastive learning, it is possible to perform effective pre-training using unlabeled or a small amount of labeled data without relying on a large amount of labeled data, thereby improving the sample feature representation ability.

[0040] 2. This invention provides a method for recognizing secondary circuit terminal images based on contrastive learning and an improved CRNN. ​​By establishing "short-circuit connections" between the preceding and following layers using a ResNet residual neural network, the speed of gradient backpropagation is accelerated, effectively solving the gradient vanishing problem and achieving efficient feature extraction with low complexity. The extracted feature maps retain key information from the terminal block images, while identity mapping and skip connections ensure the stability of deep network training.

[0041] 3. This invention provides a method for recognizing secondary circuit terminal images based on contrastive learning and an improved CRNN. ​​By introducing the ECA-Net efficient channel attention network, the model achieves better results in fine-grained feature capture and overall feature fusion, effectively improving the accuracy of terminal block image recognition. ECA-Net aggregates convolutional features through global average pooling, and then uses one-dimensional convolution and the Sigmoid activation function to establish the dependencies between channels without performing dimensionality reduction operations, thereby reducing the computational complexity and number of parameters of the model. Attached Figure Description

[0042] Figure 1 is a flowchart of the pre-training process for sample data of secondary circuit terminal blocks of power system through comparative learning in the method of the present invention.

[0043] Figure 2 is a structural diagram of a secondary circuit terminal image recognition model based on contrastive learning and improved CRNN provided by the present invention. Detailed Implementation

[0044] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments.

[0045] A method for recognizing secondary circuit terminals based on contrastive learning and improved CRNN specifically includes the following steps:

[0046] Step 1: Pre-train the sample data of the secondary circuit terminal blocks of the power system through comparative learning;

[0047] In this step, after pre-training the sample data of the secondary circuit terminal blocks of the power system through contrastive learning, the optimized sample feature embedding vectors are output. These embedding vectors are high-dimensional feature representations of the samples, and after contrastive learning, they can better reflect the similarity and differences between samples.

[0048] When automatically identifying and detecting terminal block labels and wire numbers in secondary circuits, image pre-training is required for different types of terminal blocks. However, since some terminal blocks may contain dozens or even hundreds of terminals, manually labeling these samples is time-consuming. Therefore, a contrastive learning method is introduced. This method can effectively pre-train using unlabeled or minimally labeled data without relying on a large amount of labeled data. Its core is to minimize the distance between similar samples, bringing similar samples closer together in the feature space, and maximize the distance between different samples, making different types of samples far apart in the feature space, thereby improving the sample feature representation ability. This method consists of three parts: data augmentation, encoder network, and projection network.

[0049] The contrastive learning process begins with data augmentation. Unlabeled data is augmented using techniques such as cropping, flipping, rotating, adding random noise, random Gaussian blurring, and color transformation. Data augmented from the same source is treated as similar pairs, while the remaining data are treated as different pairs. The encoder network maps similar and different pairs to a feature space, extracting features and similarities that distinguish the samples—features that play a crucial role in the model's performance, are closely related to the target, and have strong discriminative power. The selection of the encoder is critical to feature extraction performance. The encoder is typically a deep neural network that outputs high-dimensional feature vectors through average pooling layers. The projection network, composed of fully connected layers, projects the high-dimensional feature vectors output by the encoder into a low-dimensional space, reducing data complexity and redundancy, helping to quantify sample similarity and supporting subsequent tasks. Its flowchart is shown in Figure 1.

[0050] The feature vector output by the projection network is as follows: z i '=g(h i ) = W (2) σ(W (1) h i Formula 1

[0051] In the formula: z i ' is the feature vector output by the projection network; h i W is the feature vector output by the average pooling layer. (1) W (2) Let σ be the weight matrix of the two fully connected layers in the projection network; σ(·) is the activation function ReLU.

[0052] After the above three parts, the network will use cosine similarity to quantify the similarity between two feature vectors output by the projection network, thereby helping to distinguish between similar and different sample pairs, enabling the model to effectively learn features that distinguish different sample categories. The cosine similarity formula is:

[0053] In the formula: z i and z j These are the feature vectors output by the two projection networks, and the feature vectors after normalization.

[0054] Contrastive learning guides the model to distinguish the similarity of samples by constructing a loss function, and maximizes the similarity of feature vectors of similar samples by minimizing the contrastive loss function. Therefore, choosing an appropriate loss function is crucial. This paper uses the Normalized Temperature Scale Cross-Entropy Loss (NT-Xent) as the contrastive loss function, and its formula is as follows:

[0055] In the formula: L NT The standard temperature-scaled cross-entropy loss is used; N is the total number of samples; Γ(·) is a binary function, which is 1 if the condition is met, and 0 otherwise; z j(i) To be with z i Eigenvectors of the same origin; z g To be with z i Different types of eigenvectors; τ is the temperature parameter. i and g are both components of the ∑ function, and are both counting units with no physical meaning (e.g., ...). );Γ g≠i A binary function when g≠i

[0056] Standardized temperature-scale cross-entropy loss can effectively distinguish between similar and dissimilar sample pairs, but its ability to identify samples of the same type but from different sources is weak. Therefore, this paper uses the extended contrastive loss function to account for the differences between these samples, and its formula is as follows:

[0057] In the formula: L sup To expand the contrastive loss function; For the label y i The number of samples; z j To be with z i Eigenvectors of the same type but from different sources. j is a component of the ∑ function, a counting unit, and has no physical meaning; Γ i≠j It is a binary function when i ≠ j; For y i ≠y j The binary function at time; the weighted sum of the above loss function and the cross-entropy loss: L i,j =λL NT +(1-λ)L sup Formula 5

[0058] In the formula: λ is the weighting parameter. i,j This is the loss function.

[0059] Using loss function L i,j The encoder is pre-trained to extract key features that distinguish terminal block images and output a high-quality feature embedding vector X1, thus obtaining a trained encoder and providing a foundation for the next training step.

[0060] Step 2: Improve the feature extraction layer of CRNN using a residual neural network;

[0061] The encoder trained in step 1 is transferred to the improved CRNN model as the initialization parameters for the feature extraction layer, thus inheriting the pre-trained feature knowledge. The feature extraction layer is optimized using a Residual Neural Network (ResNet) structure, and the contrastive learning loss function L from step 1 is continued. i,j Supervised training is performed to further enhance the model's ability to extract features from terminal block images, and the feature vector X2 is output.

[0062] Residual Neural Networks (ResNets) improve the vanishing gradient problem by establishing "short-circuit connections" between consecutive layers, thus accelerating gradient backpropagation. Compared to other neural networks, ResNets have lower complexity and higher computational efficiency at the same depth. This makes ResNets more flexible and practical in real-world applications; therefore, this paper uses ResNets to extract feature maps from input images.

[0063] The residual unit in a residual neural network can be represented as: r l =p(s l )+F(s l J l Formula 6 s l+1 =R(r) l Formula 7

[0064] Among them, s l and s l+1 Let represent the input and output of the l-th residual unit, respectively; F represents the learned residual; J is the convolution operation; and p(s) l ) = s l Let r denote the identity mapping, and R be the activation function of a linear rectified neural network. l Represents a residual unit; p is the input signal s l An identity mapping is performed.

[0065] Based on the above equations, the learned features from the shallow layer l to the deep layer L can be obtained:

[0066] In the formula, s L This represents the output of the residual unit in the deep L layer.

[0067] Using the chain rule, the gradient of the reverse process can be obtained:

[0068] Among them, the first factor This represents the gradient of the loss function to reach L. This represents the gradient of the loss function as it approaches l. This represents the gradient from deep layer L to shallow layer l. This represents the gradient of the sum of residuals learned from layer l to layer L-1 with respect to l. The 1 in parentheses indicates that the gradient can propagate without loss through skip connections, while the other residual gradient (i.e., ...) represents the gradient of the sum of residuals learned from layer l to layer L-1 with respect to l. Then it needs to go through a layer with weights, and the gradient is not directly passed over.

[0069] By improving step 2, the vanishing gradient problem is effectively solved using the residual structure of ResNet, and efficient feature extraction is achieved with lower complexity. The extracted feature maps retain the key information of the terminal block image, while identity mapping and skip connections ensure the stability of deep network training. The output of step 2 is the image feature vector X2 processed by the improved feature extraction layer, which will serve as the input for step 3.

[0070] Step 3: Based on Step 2, introduce the ECA-Net high-efficiency channel attention network to construct a secondary loop terminal recognition model based on contrastive learning and improved CRNN.

[0071] During data collection, the captured images may lack significant features, causing the ResNet network to extract invalid features during training, thus affecting the overall performance of the model. To further enhance the correlation between features, step 3 introduces the ECA-Net efficient channel attention network based on the feature vector X2 output in step 2. Through adaptive weight allocation, it highlights key feature channels, enabling the model to achieve better results in fine-grained feature capture and overall feature fusion, ultimately achieving the goal of high-precision recognition of terminal block images.

[0072] Efficient Channel Attention Networks (ECA-Net) is an efficient channel attention module that involves fewer parameters, thus improving model performance. ECA-Net aggregates convolutional features through global average pooling, and then uses one-dimensional convolution and the sigmoid activation function to establish dependencies between channels without performing dimensionality reduction, thereby reducing the model's computational complexity and number of parameters.

[0073] First, for a given aggregate feature q∈R C(Where C is the number of channels), the attention channel weights are learned as follows: ω=σ(Mq) Equation 10

[0074] Where ω is the channel weight; σ is the Sigmoid function; and M is a C×C parameter matrix:

[0075] To further improve model performance, the ECA module allows all channels to share the same learning parameters, thereby reducing computational complexity. This sharing is achieved through one-dimensional convolution: ω = σ(C1D) k (q)) Equation 12

[0076] C1D is a 1D convolution operation, and its core parameter is the kernel size k. Through experiments and experience, it has been found that when k=3, the model has lower complexity and achieves the best results.

[0077] Secondly, the core of the ECA-Net module lies in cross-channel interaction, therefore the coverage of the interaction information needs to be determined. Since this interaction range is proportional to the channel dimension C, the following mapping relationship exists: C = φ(k) Equation 13

[0078] The simplest mapping method is a linear function φ(k) = γ*kf. However, considering the limitations of features and the fact that the channel dimension is usually an exponential multiple of 2, ECA-Net uses an exponential function with base 2 to represent the nonlinear mapping relationship: C = φ(k) = 2. (γ*k-f) Formula 14

[0079] Where f and γ are constants, typically f = 1 and γ = 2. k is the kernel size.

[0080] Given the channel dimension C, the kernel size k can be calculated:

[0081] Among them, |t| odd This represents the nearest odd number to t.

[0082] By introducing the ECA-Net efficient channel attention network, weights are adaptively assigned to each feature channel, enhancing the contribution of key channels to the model's final decision. Then, the improved feature vector X3 is processed through the recurrent layers of a CRNN for sequence modeling and contextual information processing, ultimately completing the feature classification and text recognition tasks. Thus, a complete image recognition model is constructed, as shown in Figure 2.

[0083] The present invention has been described in detail above with reference to the accompanying drawings and embodiments. However, the present invention is not limited to the above embodiments, and various changes can be made within the scope of knowledge possessed by those skilled in the art without departing from the spirit of the present invention. All contents not described in detail in the present invention can be derived from existing technologies.

Claims

1. A method for recognizing secondary circuit terminal images based on contrastive learning and improved CRNN, characterized in that, The method includes: Step 1: Pre-train the sample data of the secondary circuit terminal blocks of the power system through comparative learning; Step 2: Improve the feature extraction layer of CRNN using a residual neural network; Step 3: Based on Step 2, introduce the ECA-Net high-efficiency channel attention network to construct a secondary loop terminal recognition model based on contrastive learning and improved CRNN.

2. The method for secondary circuit terminal image recognition based on contrastive learning and improved CRNN according to claim 1, characterized in that, Step 1 includes: Data augmentation: Data augmentation is performed by applying techniques such as cropping, flipping, rotating, adding random noise, random Gaussian blurring, and color transformation to unlabeled data. Data augmented from the same source will be treated as similar sample pairs, while the remaining data will be treated as different sample pairs. Encoder network: Maps similar and different sample pairs to the feature space, extracts features and similarities that can distinguish samples, and outputs high-dimensional feature vectors through average pooling layers; Projection network: Projects the high-dimensional feature vector output by the encoder network onto a low-dimensional space to obtain the feature vector output by the projection network.

3. The method for secondary circuit terminal image recognition based on contrastive learning and improved CRNN according to claim 2, characterized in that, The formula for calculating the feature vector output by the projection network in step 1 is: z i '=g(h i ) = W( 2 )σ(W( 1 )h i ) In the formula: z i ' is the feature vector output by the projection network; h i W is the feature vector output by the average pooling layer. (1) W (2) Let σ be the weight matrix of the two fully connected layers in the projection network; σ(·) is the activation function ReLU.

4. The method for secondary circuit terminal image recognition based on contrastive learning and improved CRNN according to claim 2, characterized in that, Step 1 further includes: constructing a loss function, pre-training the encoder using the loss function to enable it to extract key features that distinguish terminal block images, outputting a high-quality feature embedding vector X1, and obtaining a trained encoder.

5. The method for secondary circuit terminal image recognition based on contrastive learning and improved CRNN according to claim 4, characterized in that, The formula for calculating the loss function is: L i,j =λL NT +(1-λ)L sup In the formula: L i,j L is the loss function; NT For standardized temperature-scaled cross-entropy loss; L sup To extend the contrastive loss function; λ is the weighting parameter; The formula for calculating the normalized temperature-scale cross-entropy loss is: In the formula: L NT The standard temperature-scaled cross-entropy loss is used; N is the total number of samples; Γ(·) is a binary function, which is 1 if the condition is met, and 0 otherwise; z j(i) To be with z i Eigenvectors of the same origin; z g To be with z i Feature vectors of different types; τ is the temperature parameter; i and g are both counting units; sim(·) is the cosine similarity; The formula for calculating the extended contrastive loss function is: In the formula: L sup To expand the contrastive loss function; For the label y i The number of samples; z j To be with z i Feature vectors of the same type but from different sources; i, g, j are all counting units; τ is a temperature parameter; Γ(·) is a binary function, which is 1 when the condition is met and 0 otherwise; sim(·) is the cosine similarity; The formula for calculating cosine similarity is: In the formula: z i and z j These are the feature vectors output by the two projection networks, and the feature vectors after normalization.

6. The method for secondary circuit terminal image recognition based on contrastive learning and improved CRNN according to claim 4, characterized in that, Step 2 specifically involves: transferring the encoder trained in Step 1 to the improved CRNN model as the initialization parameters for the feature extraction layer; optimizing the feature extraction layer using a residual neural network structure, and continuing to use the contrastive learning loss function L from Step 1. i,j Supervised training is performed to further enhance the model's ability to extract features from terminal block images, and the feature vector X2 is output.

7. The method for secondary circuit terminal image recognition based on contrastive learning and improved CRNN according to claim 6, characterized in that, The formula for calculating the residual unit in the residual neural network in step 2 is: r l =p(s l )+F(s l J l ) s l+1 =R(r) l ) In the formula, s l and s l+1 Let F and J represent the input and output of the l-th residual unit, respectively; F represents the learned residual; J is the convolution operation; r l Represents a residual unit; p is the input signal s l An identity mapping is performed; R is the activation function of a linear rectified neural network; p(s) l ) = s l Represents an identity mapping; The learning features from shallow layer l to deep layer L are: In the formula, s L This represents the output of the residual unit in the deep L layer.

8. The method for secondary circuit terminal image recognition based on contrastive learning and improved CRNN according to claim 6, characterized in that, The gradient calculation formula for the reverse process in the residual neural network in step 2 is as follows: In the formula, The first factor represents the gradient of the loss function to l. This represents the gradient of the loss function to reach L. This represents the gradient from deep layer L to shallow layer l. This represents the gradient of the sum of residuals learned from layer l to layer L-1 with respect to l.

9. The method for secondary circuit terminal image recognition based on contrastive learning and improved CRNN according to claim 6, characterized in that, Step 3 specifically involves introducing an ECA-Net efficient channel attention network based on the feature vector X2 output in step 2, adaptively assigning weights to each feature channel to obtain an improved feature vector X3. The improved feature vector X3 is then processed through the recurrent layer of a CRNN for sequence modeling and contextual information processing, ultimately completing the feature classification and text recognition task.

10. The method for secondary circuit terminal image recognition based on contrastive learning and improved CRNN according to claim 9, characterized in that, The formula for calculating the channel weights in the efficient channel attention network in step 3 is: ω=σ(Mq) In the formula, ω is the channel weight; σ is the Sigmoid function; and M is the C×C parameter matrix. ω=σ(C1D k (q)) In the formula, C1D is a 1D convolution operation; k is the kernel size.

11. The method for secondary circuit terminal image recognition based on contrastive learning and improved CRNN according to claim 9, characterized in that, The formula for calculating the kernel size in the efficient channel attention network in step 3 is as follows: In the formula, k is the kernel size; |t| odd Let t represent the nearest odd number; C is the channel dimension; f and γ are both constants.

Images