Power distribution network high-resistance grounding fault detection method
By constructing a neural network framework with multi-scale parallel analysis and adaptive weight fusion, the problem of insufficient accuracy and robustness in high-resistance grounding fault detection in distribution networks is solved, and efficient and accurate fault identification is achieved.
Patent Information
- Application Number
- CN202511376470.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-25
- Publication Date
- 2025-12-05
AI Technical Summary
Existing technologies are insufficient to accurately detect high-resistance grounding faults in distribution networks, leading to missed or false detections. The detection difficulty increases, especially after the widespread integration of distributed energy resources. Neural network models lack accuracy and robustness when processing weak feature signals.
A neural network framework with multi-scale parallel analysis, adaptive weight fusion, and channel attention mechanism is constructed. The signal is converted into an image by Walsh function basis, and the weights are dynamically adjusted by combining Hit-or-Miss neurons and channel attention modules to optimize feature extraction and classification.
It significantly improves the accuracy and robustness of high-resistance grounding fault detection, adapts to various complex working conditions, reduces the number of training iterations, improves computational efficiency and model interpretability, and enhances robustness in noisy environments.
Smart Images
Figure CN121069098A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of power system fault detection, and particularly relates to a high-resistance grounding fault detection method for a distribution network. BACKGROUND
[0002] The high-resistance grounding fault of a distribution network is a common fault type in a power system. Because the contact resistance at the fault point is large, the fault current is weak and the characteristics are not obvious. Traditional fault detection methods often cannot accurately detect the fault, which easily causes fault omission or misjudgment, and seriously affects the safe and stable operation of the distribution network. With the wide access of distributed energy, the topology structure and operating characteristics of the distribution network become more complex, which further increases the difficulty of high-resistance grounding fault detection. The neural network technology has been applied to the field of fault detection to some extent, but the existing neural network model has problems such as insufficient detection accuracy and robustness when processing weak characteristic signals of high-resistance grounding faults. Therefore, there is an urgent need for a method that can effectively extract weak characteristics of high-resistance grounding faults and improve detection accuracy. SUMMARY
[0003] The present application aims to overcome the shortcomings and deficiencies of the prior art, and provides a high-resistance grounding fault detection method for a distribution network. By innovatively integrating multi-scale parallel analysis, adaptive weight fusion and channel attention mechanism, a high-resistance grounding fault detection framework for a distribution network that can adapt to various complex working conditions is constructed.
[0004] To achieve the above-mentioned purpose, the technical scheme provided by the present application is as follows: a high-resistance grounding fault detection method for a distribution network, comprising the following steps:
[0005] Step 1: Collect one-dimensional fault signal data after a high-resistance grounding fault and a disturbance event occurs, and convert the one-dimensional fault signal data into a two-dimensional image through a sequence of Walsh function bases, and ensure that the high-energy area of the image is clear and the local features are prominent;
[0006] Step 2: Label the two-dimensional image obtained in step 1 and divide it into a training set and a test set; use the training set to train the constructed multi-scale morphological adaptive weight neural network until the maximum number of iterations is reached, save the network parameters, and make the network performance optimal; the multi-scale morphological adaptive weight neural network comprises a feature extraction module, a channel attention module, an adaptive weight fusion module and a classification network; wherein the feature processing module adopts a multi-scale parallel processing strategy to comprehensively capture different scale features of the signal; the channel attention module integrates a channel attention mechanism to further enhance the selection ability of key features; the adaptive weight fusion module dynamically adjusts the contribution of each scale according to the feature importance and outputs the fusion weight; the classification network calculates the classification probability of each sample according to the fusion weight and outputs the classification result;
[0007] Step 3: Call the multi-scale morphological adaptive weight neural network trained in step 2 to perform high resistance ground fault identification on the test set in step 2, and output the classification results.
[0008] Further, in step 1, the one-dimensional fault signal data is the zero sequence current signal data after the high resistance ground fault and disturbance time in the distribution network line, collected from the fault monitoring point of the distribution network.
[0009] Further, in step 2, the feature extraction module includes multiple parallel morphological channels, each channel running at different scales, and each scale path uses Hit-or-Miss neurons for local feature extraction. The Hit-or-Miss neuron is a neuron that performs Hit-or-Miss transformation. Each Hit-or-Miss neuron performs Hit-or-Miss transformation on the local image output by the previous layer. Hit-or-Miss transformation is defined as:
[0010]
[0011] In the formula, is the operation result of Hit-or-Miss transformation, f is the input image, is the structure element pair, is the dilation operation, is the erosion operation, and the intermediate variable is obtained by operation is the structure element The value at -x, x is the position index variable;
[0012] The value of the structure element pair is directly used as the connection weight between Hit-or-Miss neurons; to improve efficiency, Hit-or-Miss neurons at the same scale share the same structure element pair The structure element pair refers to the use of two different structure elements in morphological processing for erosion and dilation operations, respectively, to enhance the flexibility and accuracy of feature extraction. The value of the structure element pair is directly used as the connection weight between Hit-or-Miss neurons; in addition, no bias term and activation function are introduced in the Hit-or-Miss neuron to retain the nonlinear feature extraction characteristics of morphological operations;
[0013] For the input image block a, the output of neuron y is:
[0014]
[0015] In the formula, a y is the neuron output, nety is the Hit-or-Miss transform operation result, is the erosion operation result, is the dilation operation result, and are the definition domains of the structuring elements and respectively, x is the position index variable, a(x) is the value of the input image block at x, and are the values of the structuring elements and at x, is the value of the structuring element at -x.
[0016] Further, in step 2, the process of the channel attention module generating a channel attention vector is:
[0017] a = σ(MLP(AvgPool(F)) + MLP(MaxPool(F)))
[0018] wherein a is the channel attention vector, σ is the Sigmoid activation function, MLP is the multi-layer perceptron, F is the input fault feature map, AvgPool and MaxPool are respectively the global average pooling and global maximum pooling operations performed on the fault feature map F;
[0019] Subsequently, the refined feature map F' is obtained by applying the channel attention vector a to each channel of the original feature:
[0020] F' = a c ⊙F c c e {1, 2, …, C}
[0021] wherein a c is the attention weight of the c-th channel in the channel attention vector a, Fc is the feature map of the c-th channel in the fault feature map F, and C is the total number of channels.
[0022] Further, in step 2, the multi-scale morphological adaptive weight neural network performs the following operations:
[0023] Step 2.1: The feature extraction module calculates the feature maps by morphological operations:
[0024]
[0025] wherein M ki represents the i-th feature map corresponding to the feature map at scale k, X represents the input image, denotes the structural element pair corresponding to the i-th feature map at scale k, k = 1, 2, …, m, i = 1, 2, …, n, m is the number of parallel scales of the feature extraction module, and n is the number of generated feature maps at each scale;
[0026] Step 2.2: Using the feature maps obtained in step 2.1, enhanced feature maps are generated through a channel attention mechanism:
[0027] F k = α k ⊙M k
[0028] In the formula, F k denotes the enhanced feature map at scale k, α k is the learning attention weight of each feature channel, which is used to adaptively adjust the feature importance; M k is the tensor stacked by all feature maps M ki at scale k;
[0029] Step 2.3: Using the enhanced feature maps obtained in step 2.2, fusion weights are calculated through a two-layer transformation network, and the final fusion feature is constructed:
[0030] w k = softmax(φ2(δ(φ1(F k ))))
[0031] Y = ∑w k ·F k
[0032] In the formula, w k is the fusion weight of the k-th scale, φ1 and φ2 are learnable linear transformations, δ is the ReLU activation function, and Y is the fusion feature; softmax is a normalized exponential function, and its core function is to convert a set of numerical values into a probability distribution;
[0033] Step 2.4: Using the fusion feature obtained in step 2.3, the classification label is calculated through a fully connected classification layer:
[0034] y = argmax(Classifier(Y))
[0035] In the formula, y is the classification label, Classifier(Y) is the output probability value of the classifier for the input fusion feature Y, and the function argmax is used to find the independent variable that maximizes the objective function value from a set of data.
[0036] Further, in step 2, the multi-scale morphological adaptive weight neural network is trained using the training set, and the goal is to minimize the loss function, which is defined as:
[0037]
[0038] In the formula, E is a loss function, O i' represents the network output, d i' represents the expected output, i' is the i'th sample in the training set, n' is the number of samples in the training set; the set of trainable parameters includes structural elements and the attention weight α k and the linear transformations φ1 and φ2 in the adaptive weight fusion module.
[0039] Further, the update of the structural elements is obtained by the gradient descent method, and the gradient is:
[0040]
[0041] In the formula, and are the gradients of and respectively, ρ is the learning rate for controlling the step size of parameter update; net y is the Hit-or-Miss transformation operation result, is the erosion operation result, is the dilation operation result.
[0042] Further, the update of the attention weight α k is obtained by the gradient descent method, and the gradient is:
[0043]
[0044] In the formula, F k represents the enhanced feature map at scale k, M k is the tensor stacked by all feature maps M ki at scale k.
[0045] Further, the update of the linear transformations φ1 and φ2 in the adaptive weight fusion module is obtained by the gradient descent method, and the gradients are:
[0046]
[0047] In the formula, w k is the fusion weight at the kth scale, δ is the ReLU activation function, F k represents the enhanced feature map at scale k.
[0048] Compared with the prior art, the present application has the following advantages and beneficial effects:
[0049] 1、The application can dynamically optimize the structural elements, greatly reducing the dependence on artificial prior knowledge, and significantly improving the adaptability of the network under diversified working conditions.
[0050] 2、The application shows high detection accuracy and robustness under strong noise, low sampling rate and different power distribution network topologies, and has excellent computing efficiency, fewer training iteration times and shorter time consumption.
[0051] 3、The application realizes end-to-end optimization of structural elements and morphological operators by introducing a neural network architecture, further enhancing the accuracy and adaptability of feature extraction.
[0052] 4、The application adopts parallel multi-scale analysis and adaptive fusion strategy, which significantly improves the computing efficiency and model interpretability while maintaining high-precision feature extraction capability.
[0053] 5、The multi-scale morphological adaptive weight neural network in the application embeds a channel attention mechanism and a multi-scale parallel analysis structure compared with a conventional neural network model, further strengthening the extraction ability of key information and improving the robustness and discriminability of the network in a noisy environment. BRIEF DESCRIPTION OF DRAWINGS
[0054] Figure 1 The figure is a schematic diagram of the logical flow of the method of the application.
[0055] Figure 2 The figure is a structural schematic diagram of the multi-scale morphological adaptive weight neural network. DETAILED DESCRIPTION
[0056] The application will be further described in detail below in conjunction with the embodiments and the drawings, but the embodiments of the application are not limited thereto.
[0057] As shown in Figure 1 and Figure 2 , the embodiment discloses a high-resistance ground fault detection method for a power distribution network, which has the following conditions:
[0058] Step 1: Convert the one-dimensional power signal (i.e., the zero-sequence current signal data after a high-impedance ground fault and disturbance time in the distribution network line, collected from the fault monitoring point of the distribution network) into a two-dimensional image more suitable for neural network analysis, and mine potential high-dimensional information. This embodiment uses two-dimensional Walsh transform to convert the signal. To achieve this goal, this study uses two-dimensional Walsh transform as a preprocessing method. Two-dimensional Walsh transform rearranges one-dimensional time series signals into two-dimensional matrix form through the orthogonal characteristics of Walsh functions, not only retaining the energy characteristics of the signal, but also highlighting the periodicity and local pattern characteristics. The specific implementation steps are as follows: first, normalize the original one-dimensional fault signal to eliminate the dimension effect, then convert the signal to two-dimensional form through the sequence of Walsh function bases, and ensure that the high-energy area of the image is clear and the local features are prominent. This processing method not only improves the recognition of signal features, but also provides high-quality input data for subsequent neural networks.
[0059] Step 2: Add labels to the two-dimensional image obtained in step 1 and divide it into training set and test set; use the training set to train the constructed multi-scale morphological adaptive weight neural network until the maximum number of iterations is reached, then save the network parameters to make the network performance optimal; the multi-scale morphological adaptive weight neural network includes a feature extraction module, a channel attention module, an adaptive weight fusion module, and a classification network; wherein the feature processing module uses a multi-scale parallel processing strategy to comprehensively capture different scale features of the signal; the channel attention module integrates the channel attention mechanism to further enhance the selection ability of key features; the adaptive weight fusion module dynamically adjusts the contribution of each scale according to the feature importance and outputs the fusion weight; the classification network calculates the classification probability of each sample according to the fusion weight and outputs the classification result;
[0060] Specifically, the feature extraction module includes multiple parallel morphological channels, each channel runs at different scales, and a Hit-or-Miss neuron is used in each scale path for local feature extraction. Each Hit-or-Miss neuron performs Hit-or-Miss transformation on the local image output by the previous layer. Hit-or-Miss transformation is defined as:
[0061]
[0062] wherein, is the Hit-or-Miss transformation operation result, f is the input image, is the structure element pair, is the dilation operation, is the erosion operation, and the intermediate variable is obtained by operation obtained, is a structuring element the value at -x, x is a position index variable;
[0063] where the values of the structuring element pair are directly taken as the connection weights between Hit-or-Miss neurons; to improve efficiency, Hit-or-Miss neurons under the same scale share the same structuring element pair The structuring element pair refers to the use of two different structuring elements in morphological processing respectively used for erosion operation and dilation operation to enhance the flexibility and accuracy of feature extraction, the values of the structuring element pair are directly taken as the connection weights between Hit-or-Miss neurons; in addition, no bias term and activation function are introduced in the Hit-or-Miss neuron to retain the nonlinear feature extraction characteristics of morphological operation;
[0064] For the input image block a, the output of neuron y is:
[0065]
[0066] In the formula, a y is the neuron output, net y is the Hit-or-Miss transform operation result, is the erosion operation result, is the dilation operation result, and are the definition domains of structuring elements and respectively, x is a position index variable, and a(x) is the value of the input image block at x, and are the values of structuring elements and at x, is a structuring element at -x.
[0067] Specifically, the process of generating a channel attention vector by the channel attention module is:
[0068] α=σ(MLP(AvgPool(F))+MLP(MaxPool(F)))
[0069] In the formula, α is a channel attention vector, σ is a Sigmoid activation function, MLP is a multi-layer perceptron, F is an input fault feature map, AvgPool and MaxPool perform global average pooling and global maximum pooling operations on the fault feature map F respectively;
[0070] Subsequently, the refined feature map F' is obtained by applying the channel attention vector a to each channel of the original feature:
[0071] F' = a c ⊙F c , c e {1, 2, …, C}
[0072] wherein a c is the attention weight corresponding to the c-th channel in the channel attention vector a, F ki is the feature map of the c-th channel in the failure feature map F, and C is the total number of channels.
[0073] Specifically, the multi-scale morphological adaptive weight neural network performs the following operations:
[0074] Step 2.1: The feature extraction module calculates the feature map by morphological operation:
[0075]
[0076] wherein M ki represents the feature map corresponding to the i-th feature map at scale k, X represents the input image, represents the structural element pair corresponding to the i-th feature map at scale k, k = 1, 2, …, m, i = 1, 2, …, n, m is the number of parallel scales of the feature extraction module, and n is the number of feature maps generated at each scale;
[0077] The feature extraction process proposed in this embodiment adopts a sliding window mechanism, and the specific implementation steps are as follows: taking the structural element and as the basis, sliding in the horizontal and vertical directions with a step size of l and n", respectively. Among them, the horizontal sliding step size l is set to 1 / 2 of the width of the structural element, which ensures the sufficiency of local region feature extraction; the vertical sliding step size n" is usually selected as a factor of the image height N, when n" = N, then only sliding in the horizontal direction to reduce the computational complexity, to meet the requirements of specific signal analysis scenarios;
[0078] Step 2.2: Using the feature map obtained in step 2.1, enhanced feature map is generated by channel attention mechanism:
[0079] F k = a k ⊙M k
[0080] wherein F k represents the enhanced feature map at scale k, a k is the learning attention weight of each feature channel, which is used to adaptively adjust the feature importance; M k is a tensor stacked by all feature maps M ki at scale k.
[0081] Step 2.3: Using the enhanced feature maps obtained in step 2.2, calculate the fusion weights through a two-layer transformation network, and construct the final fusion feature:
[0082] w k = softmax(φ2(δ(φ1(F k )))
[0083] Y = ∑w k · F k
[0084] where w k is the fusion weight of the k-th scale, φ1 and φ2 are learnable linear transformations, δ is the ReLU activation function, Y is the fusion feature; softmax is a normalization exponential function, and its core function is to convert a set of numerical values into a probability distribution;
[0085] Step 2.4: Using the fusion feature obtained in step 2.3, calculate the classification label through the fully connected classification layer:
[0086] y = argmax(Classifier(Y))
[0087] where y is the classification label, Classifier(Y) is the output probability value of the classifier for the input fusion feature Y, and the function argmax is used to find the independent variable that maximizes the objective function value from a set of data.
[0088] Specifically, the multi-scale morphological adaptive weight neural network is trained using the training set, and the goal is to minimize the loss function, which is defined as:
[0089]
[0090] where E is the loss function, O i' represents the network output, d i' represents the expected output, i' is the i'th sample in the training set, and n' is the number of samples in the training set; the trainable parameter set includes the structural elements and attention weights α k and linear transformations φ1 and φ2 in the adaptive weight fusion module.
[0091] Specifically, for the training of the structural element, the network output net y is a piecewise function for the partial derivative of the structural element and
[0092]
[0093] wherein, and Only piecewise differentiable.
[0094] When the derivative exists, and is:
[0095]
[0096] Based on the above partial derivatives, the update of the structure element is obtained by gradient descent method, and the gradient is:
[0097]
[0098] wherein, and are the gradients of and respectively, and p is the learning rate for controlling the step size of parameter update. y is the Hit-or-Miss transform operation result, is the erosion operation result, is the dilation operation result. and In the training process, step-by-step optimization is performed according to the above formula.
[0099] Specifically, the update of the attention weight a k is obtained by gradient descent method, and the gradient is:
[0100]
[0101] This gradient reflects the contribution of the attention mechanism to the loss, which is used to guide the adjustment of a k .
[0102] Specifically, the update of the linear transformations f1 and f2 in the adaptive weight fusion module is obtained by gradient descent method, and the gradients are:
[0103]
[0104] These gradients ensure that the adaptive weight fusion module can effectively update the transformations f1 and f2 according to the changes of the loss function, thereby optimizing the feature fusion process;
[0105] The above neural network training process iteratively optimizes the set of trainable parameters θ through steps of feature extraction, attention weighting, feature fusion, and parameter updating, and finally outputs the optimized set of parameters. This process ensures that the multiscale morphologically adaptive weight neural network can efficiently converge and improve performance in training. The learning process has several advantageous properties. First, the piecewise nature of morphological operations leads to sparsity in gradient updates, causing optimization efforts to focus on the most relevant parts of the structuring element. Second, the integration of attention mechanisms and fusion weights creates a multi-level optimization that gradually improves the network's feature extraction and decision-making capabilities. Finally, the end-to-end training method ensures that all components of the network work together to minimize the overall loss function.
[0106] Step 3: Call the multiscale morphologically adaptive weight neural network trained in step 2 to perform high-resistance ground fault identification on the test set in step 2, and output the classification results.
[0107] The above embodiments are the preferred embodiments of the present application, but the embodiments of the present application are not limited to the above embodiments, and any changes, modifications, substitutions, combinations, simplifications made without departing from the spirit and principles of the present application should be equivalent replacement methods and are included in the protection scope of the present application.
Claims
1. A method for detecting high impedance ground fault in a power distribution network, characterized in that, The method comprises the following steps: Step 1: collecting one-dimensional fault signal data after a high-resistance grounding fault and a disturbance event occurs, and converting the one-dimensional fault signal data into a two-dimensional image through a sequence of Walsh function bases, and ensuring that the high-energy area of the image is clear and the local features are prominent; Step 2: adding labels to the two-dimensional image obtained in step 1 and dividing it into a training set and a test set; using the training set to train the constructed multi-scale morphological adaptive weight neural network until the maximum number of iterations is reached, then saving the network parameters to make the network performance optimal; the multi-scale morphological adaptive weight neural network comprises a feature extraction module, a channel attention module, an adaptive weight fusion module and a classification network; wherein the feature processing module adopts a multi-scale parallel processing strategy to comprehensively capture different scale features of the signal; the channel attention module integrates the channel attention mechanism to further enhance the selection ability of key features; the adaptive weight fusion module dynamically adjusts the contribution of each scale according to the feature importance to output the fusion weight; the classification network calculates the classification probability of each sample according to the fusion weight and outputs the classification result; Step 3: calling the multi-scale morphological adaptive weight neural network trained in step 2 to identify the high-resistance grounding fault of the test set in step 2, and outputting the classification result.
2. The method for detecting high impedance ground fault of power distribution network according to claim 1, characterized in that, In step 1, the one-dimensional fault signal data is zero sequence current signal data after a high-resistance grounding fault and a disturbance event occurs in a distribution network line, collected from a fault monitoring point of the distribution network.
3. The method of claim 1, wherein the method further comprises: In step 2, the feature extraction module includes multiple parallel morphological channels, each channel runs at different scales, and a Hit-or-Miss neuron is used in each scale path for local feature extraction; each Hit-or-Miss neuron performs Hit-or-Miss transformation on the local image output by the previous layer, and the Hit-or-Miss transformation is defined as: wherein is the Hit-or-Miss transform operation result, f is the input image, is the structuring element pair, is the dilation operation, is the erosion operation, the intermediate variable is obtained from the operation is obtained from the operation is the structuring element the value at -x, x being the position index variable; where the value of the pair of structuring elements is directly taken as the connection weight between Hit-or-Miss neurons; to improve efficiency, Hit-or-Miss neurons under the same scale share the same pair of structuring elements The pair of structuring elements refers to the use of two different structuring elements in morphological processing The pair of structuring elements is directly taken as the connection weight between Hit-or-Miss neurons; in addition, no bias term and activation function are introduced in the Hit-or-Miss neuron to retain the nonlinear feature extraction characteristics of morphological operations For an input image block a, the output of neuron y is: where a y is the neuron output, net y is the result of the Hit-or-Miss transform operation, is the result of the erosion operation, is the result of the dilation operation, and are the definition domains of the structuring elements and respectively, x is the position index variable, and a(x) is the value of the input image block at x, and are the values of the structuring elements and at x, is the value of the structuring element at -x.
4. The method of claim 1, wherein, In step 2, the process of generating a channel attention vector by the channel attention module is as follows: α=σ(MLP(AvgPool(F))+MLP(MaxPool(F))) In the formula, α is the channel attention vector, σ is the Sigmoid activation function, MLP is the multilayer perceptron, F is the input fault feature map, AvgPool and MaxPool perform global average pooling and global maximum pooling operations on the fault feature map F respectively; Then, the refined feature map F' is obtained by applying the channel attention vector α to each channel of the original feature: F' = a c F c c e {1,2,...,C} In the formula, α c is the attention weight of the corresponding cth channel in the channel attention vector α, is the feature map of the cth channel in the fault feature map F, and C is the total number of channels.
5. The method of claim 1, wherein, In step 2, the multi-scale morphological adaptive weight neural network performs the following operations: Step 2.1: The feature extraction module calculates the feature map through morphological operation: In the formula, M ki denotes the feature map corresponding to the i-th feature map at scale k, X denotes the input image, denotes the structural element pair corresponding to the i-th feature map at scale k, k = 1, 2, …, m, i = 1, 2, …, n, m is the number of parallel scales of the feature extraction module, and n is the number of generated feature maps at each scale; Step 2.2: using the feature map obtained in step 2.1, an enhanced feature map is generated through the channel attention mechanism: F k = a k ⊙M k In the formula, F k denotes the enhanced feature map at scale k, α k is the learning attention weight of each feature channel, which is used to adaptively adjust the feature importance; M k is the tensor of all feature maps M ki stacked tensor; Step 2.3: using the enhanced feature map obtained in step 2.2, a fusion weight is calculated through a two-layer transformation network, and a final fusion feature is constructed: w k = softmax(φ2(δ(φ1(F k )))) Y =∑w k • F k where w k is the fusion weight of the kth scale, φ1and φ2are learnable linear transformations, δ is the ReLU activation function, Y is the fused feature; softmax is a normalization exponential function, and its core role is to convert a set of numerical values into a probability distribution; Step 2.4: Using the fusion features obtained in step 2.3, the classification label is calculated by a fully connected classification layer: y = arg max(Classifier(Y)) where y is the classification label, Classifier(Y) is the output probability value of the classifier on the input fusion feature Y, and the function argmax is used to find the argument that maximizes the value of the objective function from a set of data.
6. The method of claim 1, wherein, In step 2, the multi-scale morphological adaptive weight neural network is trained using the training set, with the goal of minimizing the loss function, which is defined as: where E is a loss function, O i' denotes the network output, d i' denotes the expected output, i' is the i'th sample in the training set, n' is the number of samples in the training set; a set of trainable parameters includes the structural elements and the attention weight a k and linear transformations φ1 and φ2 in the adaptive weight fusion module.
7. The method of claim 6, wherein, The update of the structuring element is obtained by gradient descent, with the gradient being: wherein and are gradients of and respectively, p is a learning rate for controlling the step size of parameter update; net y is a Hit-or-Miss transform operation result, is an erosion operation result, is a dilation operation result.
8. The method of claim 6, wherein the method further comprises: The attention weight a k The update of a is derived by gradient descent with the gradient In the formula, F k denotes the enhanced feature map at scale k, M k is the stack of all feature maps M ki at scale k.
9. The method of claim 6, wherein the method further comprises: The update of the linear transformations φ1 and φ2 in the adaptive weight fusion module is obtained by gradient descent, with the gradients being: In the formula, w k is the fusion weight of the kth scale, δ is the ReLU activation function, F k represents the enhanced feature map at scale k.