Power transmission line equipment three-dimensional model segmentation method based on deep learning

Abstract

A power transmission line equipment three-dimensional model segmentation method based on deep learning comprises the following steps: performing over-segmentation on a three-dimensional model of power transmission line equipment to obtain a plurality of sub-patches, performing feature extraction on each sub-patch, clustering the extracted features, and inputting a clustering result into a neural network prediction model to obtain a neural network prediction model; the neural network prediction model is a bidirectional long-short-term memory network improved based on a principal component analysis network and an ScSR sparse dictionary, and hyper-parameters of the bidirectional long-short-term memory network are optimized by adopting a whale optimization algorithm improved by a hybrid strategy; and carrying out segmentation prediction by using the obtained optimal hyper-parameter, and finally adjusting the segmentation labels of the sub-patches by using a graph cut algorithm. According to the invention, on the basis of the prediction result of the neural network prediction model, the segmentation labels of the sub-patches are adjusted by using the graph cut algorithm, and the segmentation is further optimized, so that the segmentation effect of the segmentation edge is ensured.

Description

Technical Field

[0001] The present invention belongs to the technical field of three-dimensional model segmentation, and specifically relates to a three-dimensional model segmentation method for power transmission line equipment based on deep learning. Background Art

[0002] Traditional power transmission line equipment detection and segmentation often rely on rule-based or feature engineering methods, which have limited effectiveness in dealing with complex scenes, lighting changes, and equipment morphology diversity. In recent years, with the development and popularization of deep learning technology, its application in the field of computer vision has made significant progress. Deep learning models have become the mainstream method for image processing, object detection, and segmentation with their ability to learn large-scale data and excellent feature extraction capabilities. Introducing deep learning algorithms into the segmentation of three-dimensional models of power transmission line equipment provides new possibilities for improving the accuracy and efficiency of equipment identification, positioning, and segmentation. However, traditional deep learning algorithms use split segmentation labels for training, which reduces the learning ability of neural networks, resulting in poor prediction performance of deep learning neural networks at segmentation edges and difficulty in reducing loss functions. Summary of the invention

[0003] The purpose of the present invention is to provide a three-dimensional model segmentation method for power transmission line equipment based on deep learning, which can improve the segmentation effect of segmentation edges in response to the above-mentioned problems existing in the prior art.

[0004] To achieve the above objectives, the technical solution of the present invention is as follows:

[0005] A method for segmenting a three-dimensional model of a power transmission line device based on deep learning, the method comprising:

[0006] S1, over-segmenting the three-dimensional model of the power transmission line equipment to obtain a plurality of sub-surfaces;

[0007] S2, extracting features from each sub-facet and clustering the extracted features;

[0008] S3, inputting the clustering results into the neural network prediction model to obtain the segmentation results output by the neural network prediction model; the neural network prediction model is a bidirectional long short-term memory network improved based on the principal component analysis network and the ScSR sparse dictionary, and the whale optimization algorithm improved by the hybrid strategy is used to optimize the hyperparameters of the bidirectional long short-term memory network;

[0009] S4. Use the graph cut algorithm to trim the segmentation result output by the neural network prediction model to obtain the final segmentation result.

[0010] The graph cut algorithm is specifically:

[0011] The optimal segmentation label of the sub-face is obtained by minimizing the edge segmentation objective function; the objective function is:

[0012]

[0013]

[0014] In the above formula, E(l) represents the edge segmentation objective function; e1 and e2 represent the prediction label and segmentation label optimization items respectively; λ is the weight, which is a constant; l f represents the segmentation label of the sub-face f; (P(f|l f ) indicates that the segmentation label of the sub-face f is l f The probability of g represents the segmentation label of sub-facet g; F represents the set of all sub-facets on the 3D model; G represents the set consisting of any two adjacent sub-facets; ε is the probability value, which is a very small number; θ(f,g) represents the dihedral angle between sub-facets f and sub-facets g.

[0015] Specifically, S1 includes: firstly cutting the three-dimensional model by normalization to obtain a plurality of sub-faces, and then aligning the boundaries of the sub-faces by fuzzy cutting.

[0016] The S2 is specifically:

[0017] S21, select five feature descriptors to describe the sub-face from different angles;

[0018] S22. Use a Gaussian mixture model to cluster the feature vector composed of five feature descriptors, and use the clustering results as input data for the neural network prediction model.

[0019] The five feature descriptors are respectively scale-invariant heat kernel feature, shape diameter function, average geodesic distance, spectral wavelet feature, and conformal factor.

[0020] The neural network prediction model is specifically as follows: first, the input data is screened using the principal component analysis network, and then the low-resolution feature data in the screened feature data is reconstructed using the ScSR sparse dictionary to obtain high-quality feature data, and finally the obtained high-quality feature data is input into the bidirectional long short-term memory network for segmentation prediction.

[0021] The method of reconstructing the low-resolution feature data in the screened feature data by using the ScSR sparse dictionary is specifically as follows:

[0022] A1. Input the training data into the principal component analysis network to obtain the feature data set P = {F h ,F l},in represents the high-resolution feature blocks in the set P, represents the i-th high-resolution feature in the high-resolution feature block, represents the low-resolution feature blocks in the set P, Represents the i-th low-resolution feature in the low-resolution feature block;

[0023] A2. Construct sparse expressions of high-resolution feature blocks and low-resolution feature blocks over-complete dictionaries:

[0024]

[0025]

[0026] In the above formula, B h , B l They represent the sparse expressions of the over-complete dictionary of the high-resolution feature block and the low-resolution feature block respectively; λ is the equalization coefficient; g is the sparse matrix; g i is the sparse expression coefficient; E l Represents a sparse representation matrix, which is represented by the sparse expression coefficient g i It is obtained by multiplying with the sparse matrix g; N and M are the vector dimensions of the high-resolution block and the low-resolution image block respectively;

[0027] A3. Assume B c For a given dictionary, S c For a high-resolution given dictionary, transform the above formula into:

[0028]

[0029] It is concluded that each training data is in a given dictionary B c Sparse expression coefficient g i , based on the sparse expression coefficient g i Calculate the sparse representation matrix E l , and then update the given dictionary B c ;

[0030] A4. Iterate the calculation until the stopping condition is reached;

[0031] A5. Based on the high-resolution feature dictionary B obtained in the end h and low-resolution feature dictionary B l , obtain high-quality feature data, and use the high-quality feature data as the input of the bidirectional long short-term memory network.

[0032] The whale optimization algorithm improved by hybrid strategy is used to optimize the hyperparameters of the bidirectional long short-term memory network as follows:

[0033] B1. Initialize the whale population;

[0034] B2. Calculate the fitness value of individual whales in the population and obtain the optimal individual whale position;

[0035] B3. Calculate the search threshold; randomly select a whale individual, and compare the random number randomly generated by the whale individual with the search threshold. If the random number randomly generated by the whale individual is less than the search threshold, the position is updated according to the adaptive adjustment formula, otherwise the position is updated according to the basic formula; the positions of the remaining whale individuals in the whale population are updated according to the basic formula, which includes the position update formulas for the encirclement stage, the bubble net predation stage and the prey search stage;

[0036] The calculation formula of the search threshold is:

[0037]

[0038] In the above formula, is the average fitness value of the current whale population; F min is the optimal fitness value of the current whale population; F max is the worst fitness value of the current whale population;

[0039] The adaptive adjustment formula is:

[0040] X rand =X min +α(X max -X min );

[0041] In the above formula, X rand represents the position of a randomly selected whale; α is a random number between [0,1]; X min , X max X rand The minimum and maximum values ​​of

[0042] B4. Return to step B2 to iterate the calculation until the termination condition is met, and output the optimal whale individual, which is the optimal hyperparameter of the bidirectional long short-term memory network.

[0043] The position update formula in the round-up phase is:

[0044] X(t+1)=X * (t)-A×D;

[0045] A = 2 × a × rand - a;

[0046]

[0047] D=|C×X * (t)-X(t)|;

[0048] C = 2 × rand;

[0049] In the above formula, a is the nonlinear convergence factor; X(t+1) represents the updated whale position; X * (t) is the position of the optimal whale individual; A, C, and D are all coefficients; X(t) is the current whale position; rand is a generated random number; t is the current iteration number; T represents half of the maximum iteration number;

[0050] The position update formula of the bubble net predation stage is:

[0051]

[0052]

[0053] D * =|X * (t)-X(t)|;

[0054] In the above formula, D * represents the distance between the current whale individual and the optimal whale individual; ω(t) represents the adaptive weight of the whale; Iter max is the maximum number of iterations; k is the adjustment coefficient used to adjust the whale's weight; l is a random number between [-1,1];

[0055] The position update formula in the prey search phase is:

[0056] X(t+1)=X rand -A×D';

[0057] D'=|C×X rand -X(t)|;

[0058] In the above formula, X rand Represents the positions of random whale individuals other than the best whale individual.

[0059] Specifically, B1 generates a feasible solution, and then generates the inverse solution of the feasible solution according to the following formula, and uses the inverse solution as the initial population:

[0060]

[0061] In the above formula, Represents X i The inverse solution of i is the i-th feasible solution; lb i ,ub i Respectively represent X i The upper and lower limits of .

[0062] Compared with the prior art, the present invention has the following beneficial effects:

[0063] 1. The present invention discloses a method for segmenting a three-dimensional model of a power transmission line equipment based on deep learning. The method first performs over-segmentation on the three-dimensional model of the power transmission line equipment to obtain a plurality of sub-facets, then extracts features from each sub-facet, clusters the extracted features, and inputs the clustering results into a neural network prediction model to obtain a segmentation result output by the neural network prediction model. The neural network prediction model is a bidirectional long short-term memory network improved based on a principal component analysis network and an ScSR sparse dictionary, and a whale optimization algorithm improved by a hybrid strategy is used to optimize the hyperparameters of the bidirectional long short-term memory network. The obtained optimal hyperparameters are used for segmentation prediction, and finally the segmentation labels of the sub-facets are adjusted using a graph cut algorithm, thereby further optimizing the segmentation on the basis of the prediction results of the neural network prediction model, and ensuring the segmentation effect of the segmentation edge. Therefore, the present invention further improves the segmentation effect on the basis of the prediction results of the neural network prediction model.

[0064] 2. In the method for segmenting a three-dimensional model of power transmission line equipment based on deep learning described in the present invention, the input data is first screened using a principal component analysis network to make the input data more representative, and then the low-resolution feature data in the screened feature data is reconstructed using a ScSR sparse dictionary to obtain high-quality feature data, and finally the obtained high-quality feature data is input into a bidirectional long short-term memory network for segmentation prediction, thereby improving the prediction accuracy and ultimately improving the segmentation accuracy. Therefore, the present invention can improve the prediction accuracy of the neural network prediction model and ultimately improve the segmentation accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 The present invention is a flowchart of the method.

[0066] Figure 2 This is a comparison chart of RI values ​​of the method described in the present invention, the MeshCNN deep learning model, and the ShapePFCN deep learning model. DETAILED DESCRIPTION

[0067] The present invention is further described in detail below in conjunction with specific implementations and drawings.

[0068] Example:

[0069] See also Figure 1 , a 3D model segmentation method for power transmission line equipment based on deep learning, is performed in the following steps:

[0070] S1. First, construct a three-dimensional model of the transmission line equipment; specifically, taking the 220kv double-circuit drum-type SJ2 tension tower as an example, for the design parameters of the transmission line, the minimum gap vertical distance in the suspension insulator string and the minimum vertical distance between the transmission line and the ground are used; for the environment, select areas with relatively small populations, that is, non-residential areas; for climate condition parameters, select according to the meteorological conditions of typical meteorological zones; 3DS Inax is used to complete the solid modeling of the tower. In 3DS max, since the tower is a steel frame structure, the corresponding model of the AutoCAD linear tower is selected with the help of lofting. The lofting figure is a two-dimensional section, but the two-dimensional section needs to have a certain thickness; during the lofting, first select the path and then select the figure, otherwise the generated model will be lofted with the help of a two-dimensional section, which is contrary to the specific requirements; the insulator is lofted with the help of 3DS The hose shapes of the extended basic bodies in max are used to obtain the consistency between the suspension conductor and the tower model. A straight line is the lofting path, and the circular cross-section is the lofting figure. For the base of the tower, two cylinders can be used for simulation. It should be noted that the diameters of the two cylinders should be different. In the three-dimensional scene of large-scale transmission lines, only the string length of the suspension insulator string needs to be reflected in the model without reflecting its detailed structure. The design parameters of the transmission line mainly involve the minimum vertical distance between the suspension insulator string, the minimum vertical distance between the conductor and the ground, and the span of the transmission line. In addition, the design of all related equipment must also consider the maximum ice thickness and maximum wind speed that can be tolerated, so as to provide a basis for the alarm threshold of the subsequent design status warning. The design of the transmission line is affected by external conditions such as terrain and weather. The meteorological conditions and related parameters are selected according to the meteorological conditions of the typical meteorological zone specified in "GB50545-2010 Technical Regulations for the Design of 110kV-750kV Overhead Transmission Lines".

[0071] Then the established three-dimensional model of the power transmission line equipment is first over-segmented, and the over-segmentation means that the three-dimensional model is first cut by normalization to obtain multiple sub-surfaces, and then the boundaries of the sub-surfaces are aligned by fuzzy cutting; after the over-segmentation, feature extraction is performed on each of the obtained sub-surfaces, and the feature vector composed of five feature descriptors is clustered using a Gaussian mixture model, and the clustering result is used as the input data of the neural network prediction model;

[0072] The steps of feature extraction are:

[0073] S21, select five feature descriptors to describe the sub-face from different angles; the five feature descriptors are scale-invariant heat kernel feature (SIHKS), shape diameter function (SDF), average geodesic distance (AGD), spectral graph wavelet feature (SGWS), and conformal factor (CF), and these features can express the three-dimensional model from different angles; for SDF, AGD, and CF, a feature distribution map can be directly obtained, and for SIHK and SGWS, this embodiment uses a bag of features model (BOF) algorithm to extract features;

[0074] S22, clustering the feature vectors composed of the five feature descriptors, and using the clustering results as input data of the neural network prediction model;

[0075] S3, inputting the clustering result into the neural network prediction model to obtain the segmentation result output by the neural network prediction model;

[0076] The neural network prediction model is a bidirectional long short-term memory network improved based on the principal component analysis network and the ScSR sparse dictionary. Specifically, the principal component analysis network is used to filter the input data, and then the ScSR sparse dictionary is used to reconstruct the low-resolution feature data in the filtered feature data to obtain high-quality feature data. Finally, the obtained high-quality feature data is input into the bidirectional long short-term memory network for segmentation prediction.

[0077] The method of reconstructing the low-resolution feature data in the screened feature data by using the ScSR sparse dictionary is specifically as follows:

[0078] A1. Input the training data into the principal component analysis network to obtain the feature data set P = {F h ,F l},in represents the high-resolution feature blocks in the set P, represents the i-th high-resolution feature in the high-resolution feature block, represents the low-resolution feature blocks in the set P, Represents the i-th low-resolution feature in the low-resolution feature block;

[0079] A2. Construct sparse expressions of high-resolution feature blocks and low-resolution feature blocks over-complete dictionaries:

[0080]

[0081] In the above formula, B h , B l They represent the sparse expressions of the over-complete dictionary of the high-resolution feature block and the low-resolution feature block respectively; λ is the equalization coefficient; g is the sparse matrix; g i is the sparse expression coefficient; El Represents a sparse representation matrix, which is represented by the sparse expression coefficient g i It is obtained by multiplying with the sparse matrix g; N and M are the vector dimensions of the high-resolution block and the low-resolution image block respectively;

[0082] A3. Assume B c For a given dictionary, S c For a high-resolution given dictionary, transform the above formula into:

[0083]

[0084] It is concluded that each training data is in a given dictionary B c Sparse expression coefficient g i , based on the sparse expression coefficient g i Calculate the sparse representation matrix E l , and then update the given dictionary B c ;

[0085] A4. Iterate the calculation until the stopping condition is reached; the stopping condition refers to checking the sparse representation matrix E l If the change in the most recent iterations is small, stop the iteration;

[0086] A5. The high-resolution feature dictionary B h and low-resolution feature dictionary B l The high-quality feature data is obtained by fusion, and the high-quality feature data is used as the input of the bidirectional long short-term memory network;

[0087] In order to prevent the whale optimization algorithm from falling into the local optimum in the later stage and leading to premature convergence, the whale optimization algorithm improved by the hybrid strategy is used to optimize the hyperparameters (learning rate, training time, regularization, hidden layer units) of the bidirectional long short-term memory network. Specifically:

[0088] B1. Initialize the whale population;

[0089] Specifically, a feasible solution is generated, and then the inverse solution of the feasible solution is generated according to the following formula. The inverse solution is used as the initial population. The population distribution obtained by reverse learning is more uniform, and the solution space coverage is more complete:

[0090]

[0091] In the above formula, Represents X i The inverse solution of i is the i-th feasible solution; lb i ,ub i Respectively represent X i The upper and lower limits of

[0092] B2. Calculate the fitness value of individual whales in the population and obtain the optimal individual whale position;

[0093] B3. Calculate the search threshold according to the following formula:

[0094]

[0095] In the above formula, is the average fitness value of the current whale population; F min is the optimal fitness value of the current whale population; F max is the worst fitness value of the current whale population; by setting the search threshold, the algorithm can randomly generate a set of solutions with a considerable probability in the global scope at the beginning of the iteration, avoiding the loss of population diversity caused by whale aggregation and improving the global search ability of the algorithm;

[0096] A whale individual is randomly selected, and the random number randomly generated by the whale individual is compared with the search threshold. If the random number randomly generated by the whale individual is less than the search threshold, the position is updated according to the adaptive adjustment formula, otherwise the position is updated according to the basic formula; the remaining whale individuals in the whale population update their positions according to the basic formula, and the adaptive adjustment formula is:

[0097] X rand =X min +α(X max -X min );

[0098] In the above formula, X rand represents the position of a randomly selected whale; α is a random number between [0,1]; X min , X max X rand The minimum and maximum values ​​of

[0099] The basic formula includes the position update formulas of the encirclement stage, bubble net predation stage and prey search stage; in the encirclement stage, after determining the position of the target prey (assuming that the position of the optimal whale individual is the position of the target prey, and the position of the optimal whale individual in each iteration), other whales will try to surround the position of the target prey, and the position update formula is:

[0100] X(t+1)=X*(t)-A×D;

[0101] A = 2 × a × rand - a;

[0102]

[0103] D = |C × X*(t) - X(t)|;

[0104] C = 2 × rand;

[0105] In the above formula, a is the nonlinear convergence factor; X(t+1) represents the updated whale position; X*(t) is the position of the optimal whale individual; A, C, and D are all coefficients; X(t) is the current whale position; rand is a generated random number; t is the current iteration number; T represents half of the maximum iteration number;

[0106] In the bubble net predation stage, in order to simulate the whale's shrinking encirclement and spiral update mechanism, it is assumed that the execution probability of the two mechanisms is equal, and the position update formula is:

[0107]

[0108] D*=|X*(t)-X(t)|;

[0109] In the above formula, D* represents the distance between the current whale individual and the optimal whale individual; ω(t) represents the adaptive weight of the whale; Iter max is the maximum number of iterations; k is the adjustment coefficient used to adjust the whale's weight; l is a random number between [-1,1]; in the initial stage, the weight coefficient is large, which makes the algorithm have a robust global search capability. As the number of iterations increases, the weight coefficient is relatively small, so that the algorithm can perform a refined search in a specific area, avoiding falling into the local optimum, thereby improving the accuracy of the solution;

[0110] In the prey search phase, the position of a whale individual (except the best whale individual) is randomly selected as the reference position for updating the next position. The update formula is:

[0111] X(t+1)=X rand -A×D';

[0112] D'=|C×X rand -X(t)|;

[0113] In the above formula, X rand represents the position of random whale individuals other than the best whale individual;

[0114] B4, return to step B2 to iterate the calculation until the termination condition is met, and output the optimal whale individual, which is the optimal hyperparameter of the bidirectional long short-term memory network;

[0115] S4. Solve the graph cut algorithm using the α-expansion algorithm. Specifically, the optimal segmentation label of the sub-face is obtained by minimizing the edge segmentation objective function. The objective function is:

[0116]

[0117] In the above formula, E(l) represents the edge segmentation objective function; e1 and e2 represent the prediction label and segmentation label optimization items respectively; λ is the weight, which is a constant; l f represents the segmentation label of the sub-face f; (P(f|l f ) indicates that the segmentation label of the sub-face f is l f The probability of g represents the segmentation label of sub-facet g; F represents the set of all sub-facets on the 3D model; G represents the set of any two adjacent sub-facets; ε is the probability value, which is a very small number; θ(f,g) represents the dihedral angle between sub-facet f and sub-facet g; the angle of the facets should be along the concave edge as much as possible;

[0118] S5. Re-trim the segmentation result output by the neural network prediction model based on the optimal segmentation label to obtain the final segmentation result.

[0119] Performance Verification:

[0120] To verify the effectiveness of the method described in the present invention, the following tests were performed:

[0121] 1. Randomly select the huamn model, cup model, ant model, chair model, table model, fish model, and hand model from the Princeton segmentation benchmark (PSB) data set, and apply the method of the present invention, the MeshCNN deep learning model, and the ShapePFCN deep learning model to segment them respectively, and calculate the average accuracy of each model; the calculation results are shown in Table 1:

[0122] Table 1 Average accuracy of each model

[0123] Model The method of the present invention MeshCNN ShapePFCN Human Model 99.78％ 97.36％ 98.64％ Cup Model 97.68％ 96.67％ 97.54％ Ant Model 98.36％ 94.25％ 96.48％ Chair Model 96.95％ 92.64％ 90.45％ Table Model 97.89％ 95.19％ 96.81％ Fish Model 94.69％ 91.32％ 89.36％ Hand Model 99.84％ 97.54％ 97.19％ average value 97.88％ 94.99％ 95.21％

[0124] It can be seen from Table 1 that the average accuracy of the method proposed in the present invention is better than that of the other two equally advanced models.

[0125] 2. Construct a total of six three-dimensional models of the tower, lightning arrester, transmission line, conduit, insulator and tower foundation in the transmission line equipment, and apply the method of the present invention, the MeshCNN deep learning model and the ShapePFCN deep learning model for segmentation, calculate the RI value of each model, and evaluate the model segmentation effect based on the RI value. The smaller the RI value, the more similar the model segmentation effect is to the benchmark result, that is, the better the segmentation result. The calculation formula of the RI value is:

[0126]

[0127] In the above formula, P represents the benchmark segmentation; Q represents the segmentation to be evaluated; n is the number of sub-facets on the model; when the facet f i And dough j When the segmentation labels in the baseline results are consistent, P ij =1; otherwise P ij = 0; Similarly, when the face fi and the face f j When the segmentation labels in the model segmentation results are consistent, Q ij =1; otherwise Q ij =0;

[0128] The calculation results are as follows Figure 2 As shown, groups 1-6 in the figure represent the tower, lightning arrester, transmission line, conduit, insulator and foundation pile in the transmission line equipment respectively; it can be seen that the RI value of the method proposed in the present invention is smaller, indicating that the segmentation result is better.

Claims

1. A method for segmenting a three-dimensional model of a power transmission line equipment based on deep learning, characterized in that: The three-dimensional model segmentation method comprises: S1, over-segmenting the three-dimensional model of the power transmission line equipment to obtain a plurality of sub-surfaces; S2, extracting features from each sub-facet and clustering the extracted features; S3, inputting the clustering results into the neural network prediction model to obtain the segmentation results output by the neural network prediction model; the neural network prediction model is a bidirectional long short-term memory network improved based on the principal component analysis network and the ScSR sparse dictionary, and the whale optimization algorithm improved by the hybrid strategy is used to optimize the hyperparameters of the bidirectional long short-term memory network; S4. Use the graph cut algorithm to trim the segmentation result output by the neural network prediction model to obtain the final segmentation result.

2. According to claim 1, a method for segmenting a three-dimensional model of a power transmission line equipment based on deep learning is characterized in that: The graph cut algorithm is specifically: The optimal segmentation label of the sub-face is obtained by minimizing the edge segmentation objective function; the objective function is: In the above formula, E(l) represents the edge segmentation objective function; e1 and e2 represent the prediction label and segmentation label optimization terms respectively; λ is the weight, which is a constant; lf represents the segmentation label of sub-facet f; (P(f|lf) represents the probability that the segmentation label of sub-facet f is lf; lg represents the segmentation label of sub-facet g; F represents the set of all sub-facets on the three-dimensional model; G represents the set composed of any two adjacent sub-facets; ε is the probability value, which is a very small number; θ(f,g) represents the dihedral angle between sub-facets f and sub-facets g.

3. A method for segmenting a three-dimensional model of a power transmission line equipment based on deep learning according to claim 1 or 2, characterized in that: Specifically, S1 includes: firstly cutting the three-dimensional model by normalization to obtain a plurality of sub-faces, and then aligning the boundaries of the sub-faces by fuzzy cutting.

4. A method for segmenting a three-dimensional model of a power transmission line equipment based on deep learning according to claim 1 or 2, characterized in that: The S2 is specifically: S21, select five feature descriptors to describe the sub-face from different angles; S22. Use a Gaussian mixture model to cluster the feature vector composed of five feature descriptors, and use the clustering results as input data for the neural network prediction model.

5. The method for segmenting a three-dimensional model of a power transmission line equipment based on deep learning according to claim 4, characterized in that: The five feature descriptors are respectively scale-invariant heat kernel feature, shape diameter function, average geodesic distance, spectral wavelet feature, and conformal factor.

6. A method for segmenting a three-dimensional model of a power transmission line equipment based on deep learning according to claim 1 or 2, characterized in that: The neural network prediction model is specifically as follows: first, the input data is screened using the principal component analysis network, and then the low-resolution feature data in the screened feature data is reconstructed using the ScSR sparse dictionary to obtain high-quality feature data, and finally the obtained high-quality feature data is input into the bidirectional long short-term memory network for segmentation prediction.

7. The method for segmenting a three-dimensional model of a power transmission line equipment based on deep learning according to claim 6, characterized in that: The method of reconstructing the low-resolution feature data in the screened feature data by using the ScSR sparse dictionary is specifically as follows: A1. Input the training data into the principal component analysis network to obtain the feature data set P = {Fh, Fl}, where represents the high-resolution feature block in the set P, fih represents the i-th high-resolution feature in the high-resolution feature block, represents the low-resolution feature block in the set P, and fil represents the i-th low-resolution feature in the low-resolution feature block; A2. Construct sparse expressions of high-resolution feature blocks and low-resolution feature blocks over-complete dictionaries: In the above formula, Bh and Bl represent the sparse expressions of the over-complete dictionary of the high-resolution feature block and the low-resolution feature block respectively; λ is the equalization coefficient; g is the sparse matrix; gi is the sparse expression coefficient; El represents the sparse expression matrix, which is obtained by multiplying the sparse expression coefficient gi and the sparse matrix g; N and M are the vector form dimensions of the high-resolution block and the low-resolution image block respectively; A3. Let Bc be the given dictionary and Sc be the high-resolution given dictionary, and transform the above formula into: Obtain the sparse expression coefficient gi of each training data in the given dictionary Bc, calculate the sparse representation matrix El based on the sparse expression coefficient gi, and then update the given dictionary Bc; A4. Iterate the calculation until the stopping condition is reached; A5. According to the high-resolution feature dictionary Bh and the low-resolution feature dictionary Bl finally obtained, high-quality feature data is obtained, and the high-quality feature data is used as the input of the bidirectional long short-term memory network.

8. A method for segmenting a three-dimensional model of a power transmission line equipment based on deep learning according to claim 1 or 2, characterized in that: The whale optimization algorithm improved by hybrid strategy is used to optimize the hyperparameters of the bidirectional long short-term memory network as follows: B1. Initialize the whale population; B2. Calculate the fitness value of individual whales in the population and obtain the optimal individual whale position; B3. Calculate the search threshold; randomly select a whale individual, and compare the random number randomly generated by the whale individual with the search threshold. If the random number randomly generated by the whale individual is less than the search threshold, the position is updated according to the adaptive adjustment formula, otherwise the position is updated according to the basic formula; the positions of the remaining whale individuals in the whale population are updated according to the basic formula, which includes the position update formulas for the encirclement stage, the bubble net predation stage and the prey search stage; The calculation formula of the search threshold is: In the above formula, is the average fitness value of the current whale population; Fmin is the best fitness value of the current whale population; Fmax is the worst fitness value of the current whale population; The adaptive adjustment formula is: Xrand = Xmin + α (Xmax - Xmin); In the above formula, Xrand represents the position of a randomly selected individual whale; α is a random number between [0,1]; Xmin and Xmax are the minimum and maximum values ​​of Xrand respectively; B4. Return to step B2 to iterate the calculation until the termination condition is met, and output the optimal whale individual, which is the optimal hyperparameter of the bidirectional long short-term memory network.

9. The method for segmenting a three-dimensional model of a power transmission line equipment based on deep learning according to claim 8, characterized in that: The position update formula in the round-up phase is: X(t+1)=X*(t)-A×D; A = 2 × a × rand - a; D = |C × X*(t) - X(t)|; C = 2 × rand; In the above formula, a is the nonlinear convergence factor; X(t+1) represents the updated whale position; X*(t) is the position of the optimal whale individual; A, C, and D are all coefficients; X(t) is the current whale position; rand is a generated random number; t is the current iteration number; T represents half of the maximum iteration number; The position update formula of the bubble net predation stage is: D*=|X*(t)-X(t)|; In the above formula, D* represents the distance between the current whale individual and the optimal whale individual; ω(t) represents the adaptive weight of the whale; Iter max is the maximum number of iterations; k is the adjustment coefficient used to adjust the whale's weight; l is a random number between [-1,1]; The position update formula in the prey search phase is: X(t+1)=Xrand-A×D'; D' = |C × Xrand - X(t)|; In the above formula, Xrand represents the position of random whale individuals except the optimal whale individual.

10. The method for segmenting a three-dimensional model of a power transmission line equipment based on deep learning according to claim 8, characterized in that: Specifically, B1 generates a feasible solution, and then generates the inverse solution of the feasible solution according to the following formula, and uses the inverse solution as the initial population: In the above formula, represents the inverse solution of Xi; Xi is the i-th feasible solution; lbi and ubi represent the upper and lower limits of Xi respectively.

Images