Three-dimensional modeling and virtual reality data processing method for underground cable pipe gallery

Abstract

The invention discloses a three-dimensional modeling and virtual reality data processing method for an underground cable pipe gallery, and relates to the technical field of data processing. Comprising the steps of collecting data, labeling the data, carrying out standardization processing, capturing local and global similar information between samples, carrying out feature learning on high-dimensional pipe gallery data, improving the robustness of pipe gallery data noise, calculating the update quantity of encoder weight and decoder weight, and reducing redundant information. Meanwhile, a dynamic similarity matrix is adopted in the dimension reduction process, the similarity relation can be flexibly adjusted according to local and global features of data points, feature learning based on an auto-encoder is adopted, constraint of dynamic similarity maximization is added in training, and it is guaranteed that structural information of data obtained after dimension reduction is kept, and meanwhile the dynamic similarity maximization effect is achieved. The method can effectively process low-quality or high-noise data, optimizes the aggregation effect of similar samples through a similarity error loss function, and enhances the robustness of low-precision data.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and particularly to a three-dimensional modeling and virtual reality data processing method for underground cable galleries. Background Art

[0002] With the rapid development of cities, the laying of urban underground pipe networks has been continuously strengthened. Urban underground pipe networks are responsible for tasks such as information transmission, energy transportation, and water supply and drainage, and are known as the lifelines of cities. Underground cable galleries, also known as cable tunnels and power tunnels, are important municipal infrastructure in cities. Cable tunnels are deep underground and extend to every corner of the city. Once problems occur, they will have a serious impact on the normal operation of the city. Establishing a three-dimensional model and virtual reality data processing system for underground cable galleries helps to achieve visual and intelligent management of underground cables, and improve the operation efficiency and safety of urban infrastructure.

[0003] The existing technical solutions have the following disadvantages: 1. Existing technologies usually rely on a single data source, such as laser scanning or image data, and it is difficult to comprehensively capture complex environments and gallery structures, lacking adaptability to dynamic environmental changes; 2. Existing dimensionality reduction methods use static similarity matrices and cannot effectively capture local and global differences between data, which will lead to the loss of key information in the dimensionality reduction results and affect the accuracy of subsequent analysis and modeling; 3. Existing autoencoder methods mainly focus on reconstruction errors and fail to effectively process low-precision or noisy data, resulting in inaccurate feature extraction during the dimensionality reduction process; 4. Traditional methods may not be able to effectively handle the processing requirements of large-scale gallery data, with low classification efficiency and limited accuracy for detailed classification, and it is difficult to handle complex gallery structures. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to overcome the above-mentioned disadvantages of the prior art and provide a three-dimensional modeling and virtual reality data processing method for underground cable galleries.

[0005] The technical solution adopted to solve the above technical problem is: A three-dimensional modeling and virtual reality data processing method for underground cable galleries, including the following steps:

[0006] S1. Collect on-site laser scanning data, sensor network data, internal and external image data of the underground cable gallery, satellite images, and remote sensing data of the underground cable gallery, and label the data;

[0007] S2. Standardize all the collected data;

[0008] S3. Calculate a dynamic similarity matrix based on the feature vectors of the standardized samples to fully capture local and global similarity information between samples;

[0009] S4. Calculate the total loss function of the autoencoder based on the reconstruction error loss function of the autoencoder and the similarity error loss function of the autoencoder, and achieve feature learning for high-dimensional utility tunnel data;

[0010] S5. Calculate the perturbation term of the autoencoder based on the projection matrix obtained from the outer product of the standardized utility tunnel data and the low-dimensional representation, break the risk of falling into local optimal solutions, and improve the robustness to utility tunnel data noise;

[0011] S6. Calculate the update amounts of the encoder weights and decoder weights based on the gradients of the encoder and decoder weights with respect to the loss function;

[0012] S7. Calculate the influence factor considering low-rank matrix factorization based on the rank penalty term, sparsity regularization term, and reconstruction error term, and achieve reduction of redundant information;

[0013] S8. Implement autoencoder training according to steps S3 - S7 until the preset stop iteration condition is satisfied, which indicates that the model training is completed;

[0014] S9. After the autoencoder training is completed, use a classifier to classify the dimensionality-reduced data, and the classifier uses a support vector machine;

[0015] S10. Adopt a cascaded classification architecture. In the first layer, use a lightweight classifier to perform coarse-grained classification on the dimensionality-reduced features, and in the second layer, perform fine-grained recognition on the structure of the underground cable utility tunnel based on a convolutional neural network;

[0016] S11. Map the classification results to the virtual reality scene in real time and receive manual correction feedback through the VR interaction interface.

[0017] Further, the method of standardization processing in S2 is as follows:

[0018] The standardized utility tunnel data calculated based on the input utility tunnel data is expressed as:

[0019]

[0020] In the formula, is the standardized utility tunnel data, X r is the input high-dimensional utility tunnel data, is the weighted mean, σ r is the standard deviation vector of each column feature of the utility tunnel data;

[0021] The weighted mean is calculated based on the maximum and minimum values of the sample features to achieve mean adjustment, which is expressed as:

[0022]

[0023] In the formula, is the weighted mean, μr is the mean vector of each column feature of the pipe gallery data, γ r is the mean adjustment coefficient, and are the maximum and minimum values of the features of the i a -th sample respectively, where i a is a positive integer.

[0024] Furthermore, the dynamic similarity matrix in S3 is expressed as:

[0025]

[0026] In the formula, S r (i, j) is the element in the i-th row and j-th column of the dynamic similarity matrix, and both i and j are positive integers. is the feature vector of the i-th sample after standardization, is the feature vector of the j-th sample after standardization, |||| is the L2 norm, and α r is the adjustment parameter of the first non-linear similarity extension term, and β r is the adjustment parameter of the second non-linear similarity extension term.

[0027] Furthermore, the total loss function of the autoencoder in S4 is expressed as:

[0028] L r = L recon + λ r L similarity + η r ||Δ r || 2

[0029] In the formula, L r is the total loss function of the autoencoder, L recon is the reconstruction error loss function of the autoencoder, λ r is the balance coefficient, L similarity is the similarity error loss function of the autoencoder, η r is the perturbation balance coefficient of the autoencoder, Δ r is the perturbation term of the autoencoder, and |||| is the L2 norm;

[0030] Based on the dynamic similarity matrix, calculate the similarity error loss function of the autoencoder to measure the similarity more accurately and highlight the sample pairs with larger errors, which is expressed as:

[0031]

[0032] In the formula, is the element in the i-th row and j-th column of the similarity matrix calculated according to the pipe gallery data after dimensionality reduction, γ ris the similarity adjustment coefficient, n r is the number of samples, δ r is the similarity error threshold is the indicator function, and sign is the sign function

[0033] Furthermore, the perturbation term of the autoencoder in S5 is expressed as:

[0034]

[0035] In the formula, γ rer is the perturbation intensity is the tensor operation of the high-dimensional pipe gallery data and the low-dimensional representation, Z r is the feature representation after dimensionality reduction, θ r is the perturbation enhancement coefficient is the projection matrix obtained from the outer product of the standardized pipe gallery data and the low-dimensional representation. The calculation method of the outer product of the standardized pipe gallery data and the low-dimensional representation is is the transpose of the standardized pipe gallery data

[0036] Based on the encoder mapping function, calculate the feature representation after dimensionality reduction to achieve feature dimensionality reduction, which is expressed as:

[0037]

[0038] In the formula, Z r is the feature representation after dimensionality reduction, f r is the encoder mapping function

[0039] Based on the features of the standardized pipe gallery data, perform the tensor operation of the high-dimensional pipe gallery data and the low-dimensional representation to achieve the effective fusion of the low-dimensional feature representation and the standardized high-dimensional pipe gallery data, which is expressed as:

[0040]

[0041] In the formula is the k-th column feature of the standardized pipe gallery data, m r is the feature dimension of the standardized pipe gallery data, Z r,k is the k-th column feature of the low-dimensional representation, and k is a positive integer is the tensor product operation

[0042] Furthermore, the update amounts of the encoder weights and decoder weights in S6 are expressed as:

[0043]

[0044] In the formula, ΔW r and ΔV rare the update amounts of the encoder weights and decoder weights respectively, Cea is the influence factor considering low-rank matrix decomposition, η rea is the learning rate of the autoencoder, and are the gradients of the loss function with respect to the encoder and decoder weights, respectively.

[0045] Furthermore, the impact factor of low-rank matrix decomposition considered in S7 is expressed as:

[0046] Cea=C rank +C sparsity +C recon

[0047] In the formula, C rank is the rank penalty term, C sparsity is the sparsity regularization term, C recon is the reconstruction error term;

[0048] Since the goal of low-rank matrix decomposition is to express the weight matrix of the autoencoder as the product of two low-rank matrices to explore the potential structure in the pipeline corridor data and avoid overfitting by controlling the matrix rank, the weight matrix of the encoder is decomposed into two matrices through low-rank decomposition. The decomposition method is expressed as:

[0049]

[0050] Where U r is the first decomposition matrix, W r is the weight of the encoder, V r is the second decomposition matrix, is the transpose of the second decomposition matrix, and U r and V r The rank is much smaller than W r ;

[0051] The rank penalty term is calculated based on the rank of the encoder weight matrix to ensure that the rank of the matrix decomposition does not exceed the predetermined target value r target , expressed as:

[0052] C rank =λ rank ||W r || rank

[0053] In the formula, ||W r || rank is the rank of the encoder weight matrix, λ rank is the adjustment coefficient;

[0054] The sparsity regularization term is calculated based on the rank of the first decomposition matrix and the L1 norm of the second decomposition matrix to ensure that the decomposed matrix is as sparse as possible, expressed as:

[0055] C sparsity = λ spc ||U r ||₁ + λ spe ||V r ||₁

[0056] In the formula, ||·||₁ is the L1 norm, and λ spc is the first norm weight, and λ spe is the second norm weight;

[0057] Based on the Frobenius norm, the reconstruction error term is calculated to ensure that the decomposed matrix can reconstruct the original matrix well, which is expressed as:

[0058]

[0059] In the formula, ||·|| F represents the Frobenius norm of the matrix.

[0060] Furthermore, the following steps are also included:

[0061] S12. Convert the classification result into a VR-parsable semantic three-dimensional model, and different types of components are rendered with different materials and colors;

[0062] S13. The classification result is transmitted to the VR terminal through a lightweight protocol for visual display.

[0063] The beneficial effects of the present invention are as follows: (1) The present invention performs standardized processing on the collected data, and at the same time adopts a dynamic similarity matrix in the dimensionality reduction process, which can flexibly adjust the similarity relationship according to the local and global features of the data points. Compared with the traditional static similarity matrix, it can more accurately reflect the differences between data in different regions and avoid key information that may be lost during the dimensionality reduction process.

[0064] (2) The present invention adopts feature learning based on an autoencoder and adds a constraint of maximizing dynamic similarity during training, which ensures that the data after dimensionality reduction can effectively process low-quality or noisy data while retaining the structural information. The clustering effect of similar samples is optimized through the similarity error loss function, and the robustness to low-precision data is enhanced.

[0065] (3) In the dimensionality reduction process of the present invention, low-rank matrix decomposition is combined with sparsity regularization, which effectively reduces the interference of redundant information and noise, can improve the robustness of the model to the noise of the pipe gallery data, and avoid the computational complexity and dimensionality reduction error caused by redundant information.

[0066] (4) In the 3D modeling of the underground cable corridor, the data sources collected are not limited to laser scanning and image data, but also include multi-dimensional sensor data and satellite remote sensing data. Through the integration of multiple data sources, a comprehensive modeling of the complex structure and dynamic environment of the corridor is achieved. Different from the prior art that usually relies on a single data source, it can capture the dynamic changes of the corridor structure and its environment more comprehensively. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 It is a comparative curve graph showing the influence of the traditional single-source method and multi-source data fusion of the present invention on the modeling accuracy.

[0068] Figure 2 It is a comparative curve graph of the anti-noise performance of various methods. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0069] In order to make the objectives, technical solutions and advantages of the present invention clearer and more understandable, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0070] A 3D modeling and virtual reality data processing method for an underground cable corridor includes the following steps:

[0071] S1. Collect the on-site laser scanning data, sensor network data, internal and external image data of the underground cable corridor, satellite images and remote sensing data of the underground cable corridor, and label the data.

[0072] On-site laser scanning of the underground cable corridor: High-precision measurement of the structure inside the cable corridor is carried out through a laser scanner to obtain the 3D spatial data of the corridor.

[0073] Sensor network: Multi-dimensional sensors (such as temperature, humidity, vibration, etc.) installed in the corridor provide real-time environmental monitoring data, providing a basis for subsequent analysis.

[0074] Collection of internal and external images of the underground cable corridor: Internal and external images of the corridor are captured through a high-resolution camera and a drone to capture detailed information.

[0075] Satellite images and remote sensing data: The collection of the macro environment is combined with satellite data to provide large-scale spatial data support.

[0076] The collected data is stored in various formats, including but not limited to the point cloud data format and the image data format.

[0077] The collected data includes the following attributes:

[0078] The number or identifier of the data point;

[0079] The type of data points, which distinguishes different categories of spatial points (such as pipes, walls, cables, etc.);

[0080] The coordinate values of the data points in three-dimensional space;

[0081] The measured values of the data points (such as temperature, humidity, etc.);

[0082] The timestamp, indicating the time of data acquisition;

[0083] The sensor accuracy of the data points, indicating the accuracy of the acquisition device;

[0084] The density of the data points, indicating the sparsity of sampling;

[0085] The category of environmental conditions, indicating the environment of the acquisition point (such as light, humidity, etc.).

[0086] Annotate the collected data, and the annotation category of the data is the category of each component in the pipe gallery, such as pipes, joints, brackets, etc.

[0087] S2, perform standardization processing on all the collected data.

[0088] The method of standardization processing is:

[0089] The standardized pipe gallery data calculated based on the input pipe gallery data is expressed as:

[0090]

[0091] In the formula, is the standardized pipe gallery data, X r is the input high-dimensional pipe gallery data, is the weighted mean, σ r is the standard deviation vector of each column feature of the pipe gallery data, σ r is set to 0.1.

[0092] During the data acquisition process, the obtained data includes multi-dimensional sensor data (such as temperature, humidity, etc.) and image information. These data often have a high dimension, and different data points have different sampling densities and measurement accuracies. This requires standardization processing during data dimensionality reduction to solve the problem of importance imbalance caused by inconsistent different feature scales and make the subsequent dimensionality reduction process more stable. For example, the scale difference between temperature (measured value) and three-dimensional coordinate values (such as meter level) may cause the model to bias towards large-range numerical features. Through data standardization, the dynamic adjustment of the difference is realized, and the sensitivity to the category of environmental conditions (such as humidity gradient) is enhanced. If the humidity sensor data of a certain section of the pipe gallery fluctuates violently in a short period of time (dense timestamp area), the weighted mean will adaptively expand, enabling the subsequent dimensionality reduction process to better retain such dynamic change patterns.

[0093] To better handle the differences between features and enhance the feature expression ability during the standardization process, the mean value is adjusted according to the difference between the maximum and minimum values.

[0094] Calculate the weighted mean value based on the maximum and minimum values of the sample features to achieve mean value adjustment, which is expressed as:

[0095]

[0096] In the formula, is the weighted mean value, μ r is the mean vector of each column feature of the pipe gallery data, γ r is the mean adjustment coefficient, γ r is set to 0.5, and are respectively the maximum and minimum values of the i a th sample feature, where i a is a positive integer.

[0097] S3. Calculate the dynamic similarity matrix based on the feature vectors of the standardized samples to fully capture the local and global similarity information between samples.

[0098] The dynamic similarity matrix is expressed as:

[0099]

[0100] In the formula, S r (i, j) is the element in the i-th row and j-th column of the dynamic similarity matrix, both i and j are positive integers, is the feature vector of the i-th standardized sample, is the feature vector of the j-th standardized sample, |||| is the L2 norm, α r is the adjustment parameter of the first non-linear similarity extension term, α r is set to 0.1, β r is the adjustment parameter of the second non-linear similarity extension term, β r is set to 0.5.

[0101] The coordinate values of data points in three-dimensional space are key features in utility tunnel data. This characteristic directly determines the positional relationship of spatial points. A dynamic similarity matrix is constructed based on similarity and combined with non-linear expansion terms to more fully capture the local and global similarity information between samples during the dimensionality reduction process. By measuring the distance between samples, the similarity relationship is dynamically adjusted. In the case where the data points of different components (such as pipes, joints, etc.) in the utility tunnel have large differences, the dynamic similarity can better reflect the correlation between them, capture local and global spatial features. Existing dimensionality reduction methods usually rely on static similarity matrices and cannot consider the local feature differences of utility tunnel data in different regions, which will lead to the loss of local information during the dimensionality reduction process, unable to fully reflect the changes and differences between different data points, and may result in a decrease in the accuracy of the dimensionality reduction results.

[0102] S4. Calculate the total loss function of the autoencoder based on the reconstruction error loss function of the autoencoder and the similarity error loss function of the autoencoder to achieve feature learning for high-dimensional utility tunnel data.

[0103] The total loss function of the autoencoder is expressed as:

[0104] L r =L recon +λ r L similarity +η r ||Δ r || 2

[0105] In the formula, L r is the total loss function of the autoencoder, L recon is the reconstruction error loss function of the autoencoder, λ r is the balance coefficient, λ r is set to 0.3, L similarity is the similarity error loss function of the autoencoder, η r is the perturbation balance coefficient of the autoencoder, η r is set to 0.1, Δ r is the perturbation term of the autoencoder, |||| is the L2 norm.

[0106] Combined with the autoencoder structure, feature learning is performed on high-dimensional utility tunnel data, and a constraint method of maximizing dynamic similarity is adopted in the loss function to solve the problem of the loss of similarity relationship between samples after dimensionality reduction, and enhance the aggregation effect of similar samples while maintaining the reconstruction accuracy.

[0107] The similarity error term, through the sign function and threshold, focuses on optimizing sample pairs that exceed the error tolerance range. For example, for data points with low sensor accuracy (such as large noise in vibration sensors), there may be biases in the similarity calculation. At this time, the similarity error term will impose stronger error correction constraints on such samples to prevent low-quality data from affecting the dimensionality reduction result. Traditional autoencoder training does not particularly focus on the similarity between samples and the impact of noise. It usually only focuses on the reconstruction error and lacks effective management of data quality. Ignoring similarity information and noisy data may cause the autoencoder to fail to correctly capture the key features in the utility tunnel data during training. In particular, low-quality or noisy data may bring negative results to dimensionality reduction, leading to a decline in the performance of the autoencoder model.

[0108] Based on the dynamic similarity matrix, calculate the similarity error loss function of the autoencoder to measure the similarity more accurately and highlight sample pairs with larger errors, which is expressed as:

[0109]

[0110] In the formula, is the element in the \(i\)-th row and \(j\)-th column of the similarity matrix calculated based on the dimensionality-reduced utility tunnel data, and \(\gamma\) r is the similarity adjustment coefficient, and \(\gamma\) r is set to 0.3, \(n\) r is the number of samples, \(\delta\) r is the similarity error threshold, and \(\delta\) r is set to 0.5, is the indicator function, and sign is the sign function.

[0111] S5. Calculate the perturbation term of the autoencoder based on the projection matrix obtained from the outer product of the standardized utility tunnel data and the low-dimensional representation, break the risk of falling into local optimal solutions, and improve the robustness to the noise of the utility tunnel data.

[0112] The perturbation term of the autoencoder is expressed as:

[0113]

[0114] In the formula, \(\gamma\) rer is the perturbation intensity, and \(\gamma\) rer is set to 0.3, is the tensor operation of the high-dimensional utility tunnel data and the low-dimensional representation, \(Z\) r is the feature representation after dimensionality reduction, \(\theta\) r is the perturbation enhancement coefficient, and \(\theta\) r is set to 0.5, is the projection matrix obtained from the outer product of the standardized utility tunnel data and the low-dimensional representation. The calculation method of the outer product of the standardized utility tunnel data and the low-dimensional representation is It is the transpose of the standardized utility tunnel data.

[0115] During the training process, a tensor random perturbation mechanism is used to add a perturbation term to the low-dimensional representation, breaking the risk of falling into a local optimal solution and enhancing the robustness to the noise of the utility tunnel data. First, a tensor product operation is performed on the high-dimensional utility tunnel data and the low-dimensional representation, and then perturbations are applied in multiple dimensions through an expansion method, effectively enhancing the robustness to noise and avoiding falling into a local optimal solution. Traditional dimensionality reduction methods usually use a fixed learning rate and perturbation strategy, fail to consider the multi-dimensional characteristics of the data, and the processing of data noise is relatively simple. Simple perturbation methods are likely to cause the model to fall into a local optimal solution during training, and cannot effectively improve the robustness of the model when facing noise or sparse data.

[0116] Map the high-dimensional utility tunnel data to a space with a smaller dimension but concentrated information, achieve feature dimensionality reduction, and make subsequent analysis such as classification or clustering more efficient.

[0117] Calculate the feature representation after dimensionality reduction based on the encoder mapping function to achieve feature dimensionality reduction, expressed as:

[0118]

[0119] In the formula, Z r is the feature representation after dimensionality reduction, and f r is the encoder mapping function.

[0120] In areas where the data point density is sparse, the sensor network may have insufficient sampling, resulting in blurred local features. The perturbation term enhances the correlation between the low-dimensional representation and the original high-dimensional features through a tensor product operation, so that the features after dimensionality reduction can still reflect the implicit structure of the sparse area (such as the geometric continuity of the wall joints).

[0121] Based on the features of the standardized utility tunnel data, perform a tensor operation on the high-dimensional utility tunnel data and the low-dimensional representation to achieve an effective fusion of the low-dimensional feature representation and the standardized high-dimensional utility tunnel data, expressed as:

[0122]

[0123] In the formula, is the k-th column feature of the standardized utility tunnel data, m r is the feature dimension of the standardized utility tunnel data, Z r,k is the k-th column feature of the low-dimensional representation, k is a positive integer, is the tensor product operation.

[0124] S6. Calculate the update amounts of the encoder weights and the decoder weights based on the gradients of the encoder and decoder weights with respect to the loss function.

[0125] The update amounts of the encoder weights and the decoder weights are expressed as:

[0126]

[0127] In the formula, ΔW r and ΔV r are the update amounts of the encoder weights and the decoder weights respectively, Cea is the influence factor considering low-rank matrix decomposition, and η rea is the learning rate of the autoencoder. η rea is set to 0.01. and are the gradients of the loss function with respect to the encoder weights and the decoder weights respectively.

[0128] The encoder and decoder of the autoencoder are iteratively optimized by the backpropagation algorithm with a perturbation term to reduce the reconstruction error and the similarity error simultaneously and maintain sufficient perturbation intensity during the training phase. In each iteration, the weight parameters of the encoder and decoder are updated using the update amounts of the encoder weights and the decoder weights.

[0129] S7. Calculate the influence factor considering low-rank matrix decomposition based on the rank penalty term, the sparsity regularization term, and the reconstruction error term to achieve the reduction of redundant information.

[0130] The objective of the influence factor considering low-rank matrix decomposition is to reduce the risk of the model falling into local optimal solutions and improve the stability of training. The influence factor of low-rank matrix decomposition helps to ensure the sparsity of feature representations and reduce redundant information, making the dimensionality reduction process more effective. Taking the data point identifier as an example, if the numbering system contains redundant coding (such as partial overlap between the area code and the coordinate value), low-rank decomposition can strip such repeated features, so that the identifier after dimensionality reduction only retains unique information. At the same time, the sparsity constraint forces the non-zero elements of the decomposition matrix to concentrate on key attributes (such as the cable type code), improving the subsequent classification efficiency. However, the existing dimensionality reduction methods do not consider combining low-rank matrix decomposition with sparsity regularization, resulting in insufficient removal of redundant information in the data. The features after dimensionality reduction may contain unnecessary noise. Especially when dealing with large-scale utility tunnel data, the failure to effectively remove redundant information will increase the computational complexity of the model and may affect the accuracy of subsequent classification.

[0131] The influence factor considering low-rank matrix decomposition is expressed as:

[0132] Cea = C rank + C sparsity + C recon

[0133] In the formula, C rank is the rank penalty term, C sparsity is the sparsity regularization term, C reconis the reconstruction error term.

[0134] Since the goal of low-rank matrix decomposition is to express the weight matrix of the autoencoder as the product of two low-rank matrices to explore the potential structure in the pipeline corridor data and avoid overfitting by controlling the matrix rank, the weight matrix of the encoder is decomposed into two matrices through low-rank decomposition. The decomposition method is expressed as:

[0135]

[0136] Where U r is the first decomposition matrix, W r is the weight of the encoder, V r is the second decomposition matrix, is the transpose of the second decomposition matrix, and U r and V r The rank is much smaller than W r .

[0137] The matrix rank after low-rank decomposition needs to be controlled to avoid excessive compression of information. The target rank r is set target , and then calculate the rank of the first decomposition matrix and the rank of the second decomposition matrix.

[0138] The rank penalty term is calculated based on the rank of the encoder weight matrix to ensure that the rank of the matrix decomposition does not exceed the predetermined target value r target , expressed as:

[0139] C rank =λ rank ||W r || rank

[0140] In the formula, ||W r || rank is the rank of the encoder weight matrix, λ rank is the adjustment coefficient, λ rank Set to 0.1.

[0141] The sparsity regularization term requires that the elements in the low-rank matrix maintain a certain sparsity. By calculating the sparsity measure of the matrix elements, the proportion of non-zero elements in the matrix is controlled to avoid redundant information.

[0142] The sparsity regularization term is calculated based on the rank of the first decomposition matrix and the L1 norm of the second decomposition matrix to ensure that the decomposed matrix is as sparse as possible, expressed as:

[0143] C sparsity =λ spc ||U r ||1+λ spe ||V r ||1

[0144] In the formula, ||·||1 is the L1 norm, and λ spc is the first norm weight, and λ spc is set to 0.3, and λ spe is the second norm weight, and λ spe is set to 0.5.

[0145] Calculate the reconstruction error term after low-rank decomposition to measure the reconstruction accuracy of the decomposed matrix for the original matrix, ensuring that the decomposed matrix can reconstruct the original matrix well. The sensor accuracy and the density of data points have a direct impact on the dimensionality reduction process. Lower sensor accuracy may lead to inaccurate measurement values, while higher data sparsity may lead to inaccurate calculation of the similarity between samples. The reconstruction error term helps to perform robust processing on noisy data, enhance the adaptability to low-precision or sparse data, and avoid the failure of dimensionality reduction caused by these factors.

[0146] Calculate the reconstruction error term based on the Frobenius norm to ensure that the decomposed matrix can reconstruct the original matrix well, which is expressed as:

[0147]

[0148] In the formula, ||·|| F represents the Frobenius norm of the matrix.

[0149] S8. Implement the training of the autoencoder according to steps S3 - S7 until the preset stop iteration condition is met, which means the model training is completed.

[0150] S9. After the training of the autoencoder is completed, use the classifier to classify the dimensionality-reduced data.

[0151] First, perform multimodal fusion on the dimensionality-reduced low-dimensional feature representation and the key attributes in the original data (such as environmental condition categories, sensor accuracy, etc.) to form enhanced classification input data. For sparse regions or low-precision data points, complete the missing features through interpolation algorithms and align the dynamic change patterns based on timestamps.

[0152] S10. Adopt a cascaded classification architecture. In the first layer, use a lightweight classifier to perform coarse-grained classification on the dimensionality-reduced features, and in the second layer, perform fine-grained recognition on the complex structure of the underground cable gallery based on a convolutional neural network.

[0153] The lightweight classifier uses a support vector machine. Coarse-grained classification includes distinguishing large categories such as pipes, cables, and brackets, and fine-grained classification includes details such as joints and cracks.

[0154] S11. Map the classification results to the virtual reality scene in real time, receive manual correction feedback through the VR interaction interface. If the classification confidence is lower than the threshold (e.g., <85%), automatically trigger the data transmission mechanism and update the classification model using incremental learning.

[0155] S12. Convert the classification results into a semantic three-dimensional model that can be parsed by VR, and render components of different categories with different materials and colors.

[0156] For example, components classified as "high-risk cables" are displayed as red flashing models in VR, and a heat map overlay is generated by associating the measured temperature data.

[0157] S13. Transmit the classification results to the VR terminal through a lightweight protocol for visual display, ensuring low-latency visualization. For large-scale pipe gallery networks, design a distributed classification task scheduling algorithm to dynamically allocate computing resources based on the data point density and improve the classification efficiency.

[0158] To verify the advantages of the technology of the present invention, the following experimental analysis is carried out:

[0159] (1) To verify the effectiveness of the multi-source data fusion strategy, by analyzing the modeling accuracy of different data source combinations and comparing with traditional single-source methods. The experimental results show that when fusing 4 data sources, the modeling accuracy of the present invention reaches 95% (75% for traditional methods), indicating that there is a significant improvement in the multi-source heterogeneous data fusion ability in this embodiment, as Figure 1 shown.

[0160] (2) To verify the anti-noise performance of the tensor perturbation mechanism, by analyzing the correlation between the noise level and the classification accuracy and comparing with traditional dimensionality reduction methods such as PCA / t-SNE. The experimental results show that when the noise level is 0.8, the accuracy of the present invention remains at 82%, leading other comparison algorithms, indicating that there is a significant improvement in the robustness ability in a noisy environment in this embodiment.

[0161] The above is only a preferred embodiment of the present invention and is not intended to limit the protection scope of the present invention.

Claims

1. A three-dimensional modeling and virtual reality data processing method for underground cable galleries, characterized in that The following steps are involved: S1, collects on-site laser scanning data, sensor network data, internal and external image data of underground cable corridors, satellite images and remote sensing data, and annotates the data; S2, standardize all collected data; S3, calculates the dynamic similarity matrix based on the feature vectors of the standardized samples to fully capture the local and global similarity information between samples; S4, calculating the total loss function of the autoencoder based on the reconstruction error loss function of the autoencoder and the similarity error loss function of the autoencoder, so as to realize feature learning of the high-dimensional pipeline corridor data; S5, based on the projection matrix obtained by the outer product of the standardized corridor data and the low-dimensional representation, calculates the perturbation term of the autoencoder, eliminating the risk of falling into the local optimal solution and improving the robustness to corridor data noise; S6, calculating the update amount of the encoder weight and the decoder weight based on the gradient of the loss function with respect to the encoder weight and the decoder weight; S7, based on the rank penalty term, the sparsity regularization term and the reconstruction error term, the influence factor of the low-rank matrix decomposition is calculated to reduce the redundant information; S8, implementing the autoencoder training according to steps S3 to S7 until the preset stop iteration condition is met, indicating that the model training is completed; S9, after the autoencoder training is completed, the classifier is used to classify the reduced-dimensional data, and the classifier adopts a support vector machine; S10 adopts a cascade classification architecture. The first layer uses a lightweight classifier to perform coarse-grained classification on the features after dimensionality reduction, and the second layer uses a convolutional neural network to perform fine-grained recognition of the structure of the underground cable gallery; S11, mapping the classification results with the virtual reality scene in real time, and receiving manual correction feedback through the VR interactive interface.

2. The three-dimensional modeling and virtual reality data processing method for the underground cable corridor according to claim 1, characterized in that, The method of standardization in S2 is: The standardized pipeline corridor data calculated based on the input pipeline corridor data is expressed as: In the formula, is the standardized data of the utility tunnel, and X r is the input high-dimensional utility tunnel data, is the weighted mean, and σ r is the standard deviation vector of each column feature of the utility tunnel data; The weighted mean is calculated based on the maximum and minimum values of the sample features to achieve mean adjustment, which is expressed as: In the formula, is the weighted mean, μ r is the mean vector of each column feature of the utility tunnel data, γ r is the mean adjustment coefficient, and are respectively the maximum and minimum values of the features of the i a -th sample, where i a is a positive integer.

3. The three-dimensional modeling and virtual reality data processing method for the underground cable corridor according to claim 1, characterized in that The dynamic similarity matrix in S3 is expressed as: Where S r (i, j) is the element in the i-th row and j-th column of the dynamic similarity matrix, where both i and j are positive integers, is the feature vector of the i-th sample after standardization, is the feature vector of the j-th sample after standardization, ∥∥ is the L2 norm, and α r is the adjustment parameter of the first non-linear similarity extension term, and β r is the adjustment parameter of the second non-linear similarity extension term.

4. The three-dimensional modeling and virtual reality data processing method for the underground cable corridor according to claim 1, characterized in that The total loss function of the autoencoder in S4 is expressed as: L r = L recon + λ r L similarity + η r ∥ Δ r ∥ 2 where, L r is the total loss function of the autoencoder, L recon is the reconstruction error loss function of the autoencoder, λ r is the balance coefficient, L similarity is the similarity error loss function of the autoencoder, η r is the perturbation balance coefficient of the autoencoder, Δ r is the perturbation term of the autoencoder, and ∥∥ is the L2 norm; The similarity error loss function of the autoencoder is calculated based on the dynamic similarity matrix, which measures the similarity more accurately and highlights the sample pairs with large errors, expressed as: In the formula, is the element in the i-th row and j-th column of the similarity matrix calculated according to the dimension-reduced utility tunnel data, γ r is the similarity adjustment coefficient, n r is the number of samples, δ r is the similarity error threshold, is the indicator function, and sign is the sign function.

5. The three-dimensional modeling and virtual reality data processing method for the underground cable corridor according to claim 1, characterized in that, The perturbation term of the autoencoder in S5 is expressed as: where γ rer is the disturbance intensity, is the tensor operation of high-dimensional utility tunnel data and low-dimensional representation, Z r is the feature representation after dimensionality reduction, θ r is the disturbance enhancement coefficient, is the projection matrix obtained from the outer product of the standardized utility tunnel data and the low-dimensional representation. The calculation method of the outer product of the standardized utility tunnel data and the low-dimensional representation is is the transpose of the standardized utility tunnel data; The feature representation after dimensionality reduction is calculated based on the encoder mapping function to achieve feature dimensionality reduction, which is expressed as: Where, Z r is the feature representation after dimensionality reduction, and f r is the encoder mapping function; Based on the standardized pipeline corridor data features, tensor operations are performed on high-dimensional pipeline corridor data and low-dimensional representations to achieve effective fusion of low-dimensional feature representation and standardized high-dimensional pipeline corridor data, which can be expressed as: In the formula, is the k-th column feature of the standardized utility tunnel data, m r is the feature dimension of the standardized utility tunnel data, Z r,k is the k-th column feature of the low-dimensional representation, k is a positive integer, is the tensor product operation.

6. The three-dimensional modeling and virtual reality data processing method for the underground cable corridor according to claim 1, wherein: The update amount of the encoder weight and the decoder weight in S6 is expressed as: where, ΔW r and ΔV r are the update amounts of the encoder weights and the decoder weights respectively, Cea is the influence factor considering low-rank matrix factorization, and η rea is the learning rate of the autoencoder, and are the gradients of the loss function with respect to the encoder and decoder weights respectively.

7. The three-dimensional modeling and virtual reality data processing method for the underground cable corridor according to claim 1, characterized in that, The impact factor of low-rank matrix decomposition considered in S7 is expressed as: Cea = C rank + C sparsity + C recon where C rank is the rank penalty term, C sparsity is the sparsity regularization term, and C recon is the reconstruction error term; Since the goal of low-rank matrix decomposition is to express the weight matrix of the autoencoder as the product of two low-rank matrices to explore the potential structure in the pipeline corridor data and avoid overfitting by controlling the matrix rank, the weight matrix of the encoder is decomposed into two matrices through low-rank decomposition. The decomposition method is expressed as: where, U r is the first decomposition matrix, W r is the weight of the encoder, V r is the second decomposition matrix, is the transpose of the second decomposition matrix, and the ranks of U r and V r are much smaller than that of W r ; Calculate the rank penalty term based on the rank of the encoder weight matrix to ensure that the rank of the matrix factorization does not exceed the predetermined target value r target , which is expressed as: C rank = λ rank ∥W r ∥ rank where ∥W r ∥ rank is the rank of the encoder weight matrix, and λ rank is the adjustment coefficient; Calculate the sparsity regularization term based on the rank of the first decomposition matrix and the L1 norm of the second decomposition matrix to ensure that the decomposed matrix is as sparse as possible, expressed as: C sparsity = λ spc ∥U r ∥1 + λ spe ∥V r ∥1 where ∥∥1 is the L1 norm, and λ spc is the first norm weight, and λ spe is the second norm weight; Calculate the reconstruction error term based on the Frobenius norm to ensure that the decomposed matrix can reconstruct the original matrix well, expressed as: where, ∥·∥ F denotes the Frobenius norm of the matrix.

8. The three-dimensional modeling and virtual reality data processing method for the underground cable corridor according to claim 1, characterized in that, It also includes the following steps: S12, convert the classification result into a VR-parsable semantic three-dimensional model, and render components of different categories with different materials and colors; S13, transmit the classification result to the VR terminal through a lightweight protocol for visual display.

Images