Deep learning-based structure body analysis method

Abstract

The invention discloses a deep learning-based structure body analysis method, and relates to the technical field of structure body analysis, and the method comprises the following steps: S1, data collection; s2, data enhancement and preprocessing; s3, constructing a deep learning network for processing diversity data; s4, designing a multi-task learning framework for the deep learning network; S5, adding an unsupervised learning module in the multi-task learning framework; s6, introducing the spatio-temporal variation factor into the deep learning network; and S7, model generation. Global and local features of point cloud data are dynamically captured through a PointNet + + architecture, multi-modal features are fused at the same time, the analysis ability of the model for complex shape data is remarkably improved, a time sequence convolutional network is adopted to construct a long-term trend and a sudden change point in modeling shape change, the recognition ability of the model for key events is enhanced, and the recognition efficiency of the model is improved. Meanwhile, high calculation efficiency is kept, multi-task collaborative optimization is achieved, and the overall performance and adaptability of the model are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of structure shape analysis, in particular to a structure shape analysis method based on deep learning. BACKGROUND

[0002] With the development of sensing technology and computing technology in structural engineering, dynamic analysis of complex structure shapes has gradually become a research hotspot. In modern infrastructure (such as bridges, buildings, wind power towers, etc.), by monitoring the shape changes and damage conditions, the accurate assessment of the structural health state can be realized, and then the basis for safety maintenance is provided. However, most of the traditional analysis methods rely on finite element modeling and physical formula derivation, and these methods face difficulties in processing high complexity, multi-modal, and time-varying characteristics of structure shape data.

[0003] 1. In the prior art, the feature extraction method mainly relies on traditional manual feature design (such as morphological parameters or frequency domain features), and these methods have limited adaptability and limited modal fusion effect;

[0004] 2. In the prior art, the existing time series modeling method has long time dependence, local information loss, and poor model generalization when processing complex structure shape data. SUMMARY

[0005] In view of the deficiencies of the prior art, the present application provides a structure shape analysis method based on deep learning to solve the problems raised in the background art.

[0006] To achieve the above purpose, the present application provides the following technical scheme:

[0007] In a first aspect, the present application provides a structure shape analysis method based on deep learning, comprising the following steps:

[0008] S1, data acquisition: collecting and labeling the shape information of the structure, and obtaining the original data of the structure;

[0009] S2, data enhancement and preprocessing: enhancing and preprocessing the original data under different environments or collection angles to obtain the diversity data of the structure;

[0010] S3, constructing a deep learning network for processing diversity data: combining multi-scale convolutional layers with graph convolutional networks, and constructing a deep learning network based on PointNett++ architecture;

[0011] S4, designing a multi-task learning framework for the deep learning network: the multi-task learning framework includes a shape change prediction branch and a damage location detection branch;

[0012] S5, adding an unsupervised learning module in the multi-task learning framework: introducing an unsupervised learning module using an autoencoder to process structural object shape changes or damage conditions that are difficult to label;

[0013] S6, introducing a spatiotemporal change factor into the deep learning network: introducing a time series convolution network to predict the shape changes of the structural object at different time points and under different conditions;

[0014] S7, model generation: bringing diverse data into the deep learning network to obtain the shape model of the structural object through the above steps.

[0015] Further optimize the technical solution, the original data enhancement method in step S2 includes:

[0016] geometric transformation;

[0017] random noise injection;

[0018] data difference and resampling;

[0019] simulation data generation.

[0020] Further optimize the technical solution, the preprocessing method in step S2 includes:

[0021] outlier rejection;

[0022] modal data alignment;

[0023] feature extraction and unified representation;

[0024] multimodal data fusion;

[0025] data consistency verification;

[0026] output unified data set.

[0027] Further optimize the technical solution, the deep learning network in step S3 includes:

[0028] input layer:

[0029] the input diversity data is the enhanced point cloud, each point contains three-dimensional coordinates , color and normal vector ;

[0030] The point cloud is processed by normalization and downsampling to adapt the data scale to the hardware limit;

[0031] multi-scale feature extraction layer:

[0032] extract the geometric features of the local area of the point cloud through multi-scale grouping method;

[0033] Using layer-by-layer downsampling to capture structure details of different sizes;

[0034] Multi-modal fusion module:

[0035] For multi-modal data, in the global feature layer, the embedding features from other modalities are fused;

[0036] Using attention mechanism to weight multi-modal features;

[0037] Damage focusing module:

[0038] An adaptive focusing mechanism is introduced to focus on potential damage areas in the point cloud;

[0039] Using point-level attention weights;

[0040] Output layer:

[0041] Damage identification classification;

[0042] Damage localization to shape regression.

[0043] Further optimize the technical solution, the layer-by-layer downsampling in the multi-scale feature extraction layer includes:

[0044] Hierarchical sampling;

[0045] Local feature learning;

[0046] Global feature aggregation.

[0047] Further optimize the technical solution, the multi-task learning framework in step S4 includes:

[0048] Input module:

[0049] The pre-processed point cloud or multi-modal data as input, including three-dimensional coordinates, normal vectors, colors;

[0050] Shared feature extraction module:

[0051] Based on the PointNet++ architecture, global and local features of the point cloud are extracted;

[0052] Branch 1, shape change prediction branch: focusing on global geometric shape change analysis and regression;

[0053] Branch 2, damage site detection branch: focusing on local damage area classification and positioning;

[0054] Among them,

[0055] The input form of the input module is:

[0056] ;

[0057] each point contains the following features: three-dimensional coordinates , normal vector and color value ;

[0058] The shared feature extraction module is a hierarchical structure, including:

[0059] Layer 1 local feature , using spherical radius query and shared MLP to extract local geometric features of each sampling point, for damage site detection;

[0060] Layer 2 global feature , aggregate local features, generate global feature vector through fully connected layer and global pooling, for shape change prediction;

[0061] Branch 1: shape change prediction branch:

[0062] The prediction error of shape change is calculated using the mean square error method:

[0063] ;

[0064] Where, is the predicted value, is the true value;

[0065] The mean square error calculated is used to guide the model to learn the accurate deformation mode and stress distribution;

[0066] Branch 2: damage site detection branch:

[0067] The classification error is calculated using the cross-entropy loss:

[0068] ;

[0069] Where, is the true class label, is the predicted probability;

[0070] The cross-entropy loss calculated is used for damage site classification, and strengthens the sensitivity of the model to damage points.

[0071] Further optimize the technical solution, the unsupervised learning module in the step S5, its architecture includes:

[0072] Encoder: responsible for extracting multi-scale features from input data, compressed into low-dimensional latent representation;

[0073] Its input is: point cloud data , each point contains three-dimensional coordinates , normal vector , and color value ;

[0074] The output is: low-dimensional latent feature ;

[0075] Latent space representation: used to learn a compact global feature representation, representing the core information of the point cloud, and avoiding overfitting through regularization;

[0076] Decoder: maps latent features back to the original space for reconstructing input point cloud data;

[0077] The input is: low-dimensional latent feature ;

[0078] The output is: reconstructed point cloud .

[0079] Further optimize the technical solution, the time sequence convolution network in step S6 includes:

[0080] Data preprocessing:

[0081] Convert point cloud or other spatio-temporal sequence data into time step sequence:

[0082] ;

[0083] Where is the feature representation of time ;

[0084] Causal convolution layer:

[0085] Ensure that the output only depends on the data of the current and past time points:

[0086] ;

[0087] Extended convolution layer:

[0088] Use extended convolution to expand the receptive field and capture long-time dependence:

[0089] ;

[0090] Where is the step length, is the convolution kernel weight, and the extended convolution expands the distance between points to , achieving efficient long-time feature capture;

[0091] Residual connection:

[0092] Residual connection makes the network deeper and more stable, avoiding gradient disappearance:

[0093] ;

[0094] Align the feature dimension using 1x1 convolution;

[0095] Network overall structure.

[0096] Further optimize the technical solution, the loss function in the time series convolution network comprises:

[0097] Prediction error loss:

[0098] For measuring the gap between the predicted output and the true value , the calculation formula is:

[0099] ;

[0100] Among them, the prediction error loss represents the accuracy of the model in predicting the future output of the time step sequence input, that is, the difference between the prediction result and the true label;

[0101] Regularization loss:

[0102] Limit network weight to avoid overfitting, the calculation formula is:

[0103] ;

[0104] Among them, the regularization loss represents the size of the model weight used to limit the phenomenon of overfitting, that is, the model over-reliance on training data, leading to the decline of generalization performance;

[0105] Total loss:

[0106] ;

[0107] The calculation of the loss function plays a trade-off role for the time series convolution network, which reduces the deviation between the damage positioning and the actual situation through trade-off, and at the same time, the regularization enhances the generalization ability of the model to unknown scenes.

[0108] Further optimize the technical solution, the step S7 generates the structure object analysis model, at the same time, set up a test set under multiple scenes, including simulation data and real acquisition data, comprehensive evaluation model performance, to provide reference for network structure optimization in analysis method.

[0109] In a second aspect, an embodiment of the present application provides a computer device, comprising a memory and a processor, and the memory stores a computer program, wherein the computer program instructions are executed by the processor to implement the steps of the structural object shape analysis method based on deep learning according to the first aspect of the present application.

[0110] In a third aspect, an embodiment of the present application provides a computer readable storage medium, which stores a computer program, wherein the computer program instructions are executed by the processor to implement the steps of the structural object shape analysis method based on deep learning according to the first aspect of the present application.

[0111] Compared with the prior art, the present application provides a structural object shape analysis method based on deep learning, which has the following beneficial effects:

[0112] The structural object shape analysis method based on deep learning solves the deficiencies of the prior art in feature extraction and time series modeling by combining PointNet++ and time series convolution network and other deep learning technologies. The PointNet++ architecture dynamically captures the global and local features of point cloud data, and simultaneously fuses multi-modal features, significantly improving the model's ability to analyze complex shape data. The time series convolution network is used to construct long-term trends and mutation points in the modeling shape changes, enhancing the model's ability to identify key events while maintaining high computational efficiency. Moreover, multi-task collaborative optimization is adopted, and a multi-task learning framework is used in shape change prediction and damage detection. Through information sharing and weight dynamic adjustment between tasks, the overall performance and adaptability of the model are improved. BRIEF DESCRIPTION OF DRAWINGS

[0113] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0114] Figure 1 A flowchart of a structural object shape analysis method based on deep learning according to the present application is shown in the figure.

[0115] Figure 2 A flowchart of a raw data enhancement method of a structural object shape analysis method based on deep learning according to the present application is shown in the figure.

[0116] Figure 3 A flowchart of a raw data preprocessing method of a structural object shape analysis method based on deep learning according to the present application is shown in the figure.

[0117] Figure 4A deep learning network structure schematic diagram of a structure shape analysis method based on deep learning is provided in the present application.

[0118] Figure 5 A multi-task learning framework structure schematic diagram of a structure shape analysis method based on deep learning is provided in the present application.

[0119] Figure 6 An unsupervised learning module structure schematic diagram of a structure shape analysis method based on deep learning is provided in the present application.

[0120] Figure 7 A time sequence convolution network structure schematic diagram of a structure shape analysis method based on deep learning is provided in the present application. DETAILED DESCRIPTION

[0121] In order to make the above objectives, features and advantages of the present application more apparent and comprehensible, the specific embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0122] In the following description, a lot of specific details are set forth in order to facilitate a thorough understanding of the present application, but the present application can also be implemented in other ways different from the description, and those skilled in the art can make similar generalizations without departing from the concept of the present application, so the present application is not limited to the specific embodiments disclosed below.

[0123] Secondly, the "one embodiment" or "embodiment" referred to herein means that the specific features, structures or characteristics can be included in at least one implementation of the present application. "In one embodiment" appearing in different places in the specification does not mean the same embodiment, nor does it mean an embodiment that is separate or selectively excluded from other embodiments.

[0124] Embodiment one:

[0125] Reference Figures 1-7 For the first embodiment of the present application, the embodiment provides a structure shape analysis method based on deep learning, including the following steps:

[0126] S1, data acquisition: collecting and labeling the shape information of the structure, obtaining the original data of the structure;

[0127] S2, data enhancement and preprocessing: bringing in different environments or collection angles, enhancing and preprocessing the original data, obtaining the diversity data of the structure;

[0128] S3, constructing a deep learning network for processing diversity data: combining multi-scale convolution layers with graph convolution networks, constructing a deep learning network based on PointNett++ architecture;

[0129] S4, design a multi-task learning framework for deep learning network: the multi-task learning framework contains a shape change prediction branch and a damage location detection branch;

[0130] S5, add an unsupervised learning module in the multi-task learning framework: introduce an unsupervised learning module using autoencoder to handle the structure shape change or damage condition that is difficult to label;

[0131] S6, introduce a spatio-temporal change factor into the deep learning network: introduce a time series convolution network to predict the shape change of the structure at different time points and under different conditions;

[0132] S7, model generation: bring diverse data into the deep learning network, and obtain the shape model of the structure through the above steps.

[0133] Specifically, the original data enhancement method in step S2 includes:

[0134] Geometric transformation: introduce task-related constraints in geometric transformation, such as maintaining the key morphological features of the structure (such as the integrity of the beam-column connection), to avoid misleading model learning

[0135] Random noise injection: randomly inject noise into point cloud data to simulate errors or environmental disturbances that may occur during acquisition

[0136] Data difference and resampling: sparse or dense operation on point cloud to generate data with different resolutions

[0137] Simulation data generation: generate structure simulation data using noise generation method.

[0138] Specifically, the preprocessing method in step S2 includes:

[0139] Outlier rejection: use statistical methods (such as based on mean and variance) or point cloud filtering algorithms (such as statistical filtering, radius filtering) to remove outliers in the collected data;

[0140] Modal data alignment: align all modal data to a unified spatial coordinate system through geometric alignment and registration methods;

[0141] Feature extraction and unified representation: extract key features in each modality and map them to a unified representation space;

[0142] Multi-modal data fusion: fuse features of different modalities in a unified feature space to achieve information complementation and enhancement;

[0143] Data consistency verification: ensure the consistency of the fused data in geometry, topology and semantics;

[0144] Output unified dataset: Generate a high-quality, multi-modal unified dataset for subsequent deep learning model use.

[0145] Specifically, the deep learning network in step S3 includes:

[0146] Input layer:

[0147] The input diversity data is the enhanced point cloud, each point containing three-dimensional coordinates , color and normal vector ;

[0148] The point cloud is processed by normalization and downsampling to adapt the data scale to hardware limitations.

[0149] Multi-scale feature extraction layer:

[0150] The multi-scale feature extraction layer is the core part of PointNet++.

[0151] Local geometric features of the point cloud are extracted through a multi-scale grouping method.

[0152] Layer-by-layer downsampling is used to capture structural details of different sizes.

[0153] Multi-modal fusion module:

[0154] For multi-modal data, embed features from other modalities are fused in the global feature layer.

[0155] Attention mechanisms are used to weight multi-modal features.

[0156] Damage focusing module:

[0157] An adaptive focusing mechanism is introduced to focus on potential damage areas in the point cloud.

[0158] Point-level attention weights are used.

[0159] Weight masks are generated according to local feature saliency, highlighting damage areas.

[0160] Weights are optimized through loss functions to improve the model's sensitivity to damage areas.

[0161] Output layer:

[0162] Damage identification classification; determine whether there is damage in the point cloud (classification task).

[0163] Damage localization shape regression; accurately predict the location and shape of the damage area (regression task).

[0164] In the implementation process of the deep learning network:

[0165] Implement the basic structure of PointNet++ using TensorFlow or PyTorch framework;

[0166] Existing PointNet++ open source implementation (such as PyTorchGeometric or TensorFlow implementation) can be used as a basis to extend multi-modal fusion and damage focusing modules;

[0167] Training data can be generated by finite element simulation combined with deep learning generated model, ensuring diversity and authenticity;

[0168] Specifically, the layer-by-layer down-sampling in the multi-scale feature extraction layer includes:

[0169] Layered sampling:

[0170] Use the farthest point sampling method to select representative points;

[0171] Perform KNN or ball radius query on the neighborhood point cloud of the sampled points to determine the local region;

[0172] Local feature learning:

[0173] Learn local structural information through a local feature extraction module (such as a shared MLP and pooling operation);

[0174] Features include geometric curvature, point cloud density, and local normal vector variation;

[0175] Global feature aggregation:

[0176] Step-by-step aggregation of multi-layer features to form a global feature representation of the point cloud.

[0177] Specifically, the multi-task learning framework in step S4 includes:

[0178] Input module:

[0179] Pre-processed point cloud or multi-modal data as input, including three-dimensional coordinates, normal vectors, and colors;

[0180] Shared feature extraction module:

[0181] Based on the PointNet++ architecture, extract global and local features of the point cloud;

[0182] Branch 1, shape change prediction branch: focuses on global geometric shape change analysis and regression;

[0183] Branch 2, damage site detection branch: focuses on local damage area classification and positioning;

[0184] Wherein,

[0185] The input form of the input module is:

[0186] ;

[0187] Each point contains the following features: three-dimensional coordinates , normal vector and color value ;

[0188] The shared feature extraction module is a hierarchical structure, including:

[0189] Layer 1 local feature , using spherical radius query and shared MLP to extract local geometric features of each sampling point, used for damage site detection;

[0190] Layer 2 global feature , aggregate local features, generate global feature vector through fully connected layer and global pooling, used for shape change prediction;

[0191] Branch 1: shape change prediction branch

[0192] Its architecture is:

[0193] Global feature input: input into a sequence of fully connected layers;

[0194] Deformation mode regression:

[0195] Fully connected layer + activation function (ReLU);

[0196] Output shape change parameters (such as deformation matrix, strain tensor, etc.);

[0197] The prediction error of shape change is calculated using the mean square error method:

[0198] ;

[0199] Where, is the predicted value, is the true value;

[0200] The mean square error calculated is used to guide the model to learn accurate deformation mode and stress distribution;

[0201] Branch 2: damage site detection branch

[0202] Its architecture is:

[0203] Local feature input: input into multiple PointNet layers, gradually extracting high-dimensional features at the point level;

[0204] Damage classification and localization:

[0205] Transform point features using shared MLP layers to generate damage probabilities;

[0206] Output probability distribution of each point belonging to damage categories , where represents non-damage, represents damage;

[0207] Calculate classification error using cross-entropy loss:

[0208] ;

[0209] where, is the true class label, is the predicted probability

[0210] Calculate the cross-entropy loss for damage location classification, to strengthen the model's sensitivity to damage points;

[0211] Mean squared error and cross-entropy loss Two losses complement each other in the multi-task framework, through joint optimization of the total loss, to ensure the model achieves the best balance between prediction accuracy and damage localization ability.

[0212] The multi-task learning framework is summarized in Table 1 below:

[0213] Table 1

[0214] Module Input / Output Function Input Module Pre-process point cloud and unify format Shared feature extraction module Extract local and global features Shape variation prediction branch Regression shape variation parameters (e.g. deformation amount) Damage site detection branch Classify and locate damage area Multi-task loss function Optimize performance of two tasks

[0215] Specifically, the unsupervised learning module in step S5 has the following architecture:

[0216] Encoder: responsible for extracting multi-scale features from input data and compressing them into low-dimensional latent representations;

[0217] Its input is: point cloud data , each point contains three-dimensional coordinates , normal vectors and color values ;

[0218] Its output is: low-dimensional latent features ;

[0219] Latent space representation: used to learn compact global feature representation, representing the core information of the point cloud, and avoiding overfitting through regularization;

[0220] Decoder: maps latent features back to original space for reconstructing input point cloud data;

[0221] Input: low-dimensional latent features ;

[0222] Output: reconstructed point cloud ;

[0223] Where:

[0224] PointNet++: extracts local and global features as encoder module;

[0225] Chamfer Distance or Earth Mover's Distance (EMD): can be used as a supplement to point cloud reconstruction loss to improve reconstruction accuracy;

[0226] PyTorch / TensorFlow: provides efficient neural network implementation and GPU acceleration.

[0227] Specifically, the time sequence convolution network in step S6 includes:

[0228] Data preprocessing:

[0229] Convert point cloud or other spatio-temporal sequence data into time step sequence:

[0230] ;

[0231] Where is the feature representation at time (such as point cloud global feature, damage indicator, etc.);

[0232] Causal convolution layer:

[0233] Ensure that the output only depends on the data at the current and past time points:

[0234] ;

[0235] Extended convolution layer:

[0236] Use extended convolution to expand the receptive field and capture long-time dependence:

[0237] ;

[0238] Where is the step length, is the convolution kernel weight, and the extended convolution expands the distance between points to , achieving efficient long-time feature capture;

[0239] Residual connection:

[0240] Residual connections make the network deeper and more stable, avoiding gradient vanishing.

[0241] ;

[0242] Align the feature dimensions using 1x1 convolutions;

[0243] Overall network structure:

[0244] This is achieved through the TCN module.

[0245] Specifically, the loss function in the temporal convolutional network includes:

[0246] Prediction error loss:

[0247] Used to measure predicted output and the true value The difference is calculated using the following formula:

[0248] ;

[0249] Among them, prediction error loss This represents a measure of the model's response to the input time-step sequence. Predict future output The accuracy of the prediction, i.e., the difference between the predicted result and the true label;

[0250] This calculation formula represents the average prediction error (usually the mean squared error) for each time step, which directly reflects the model's fitting ability, i.e., the model's prediction performance.

[0251] In this calculation, Mainly for:

[0252] Shape change prediction: The model predicts the shape change trend in future time steps, and the loss reflects the deviation between the predicted value and the actual value.

[0253] Damage detection: The model identifies damage-related features and generates localization results, measuring the difference between the loss results and the actual annotations.

[0254] Regularization loss:

[0255] To limit network weights and avoid overfitting, the calculation formula is as follows:

[0256] ;

[0257] Among them, regularization loss This represents the size of the model weights to prevent overfitting, which is the model's excessive reliance on the training data, leading to a decrease in generalization performance.

[0258] The calculation formula is the L2 norm of the network parameters, which balances the parameter complexity, i.e., the weight constraint is a weight coefficient of the regularization strength.

[0259] In this task, the regularization part can:

[0260] improve the stability of the model;

[0261] prevent the model from learning too much noise under limited training data.

[0262] Total loss:

[0263]

[0264] Total loss combines the model prediction performance and the weight constraint:

[0265] by minimizing to maximize the prediction accuracy and reasonably constrain the model complexity, if is small, the regularization constraint is weak, and the model may be prone to overfitting, if is large, the regularization constraint is strong, and the model may be underfitting, making it difficult to capture complex patterns, by optimizing to improve the future pattern prediction accuracy;

[0266] The calculation of this loss function plays a balancing role for the time series convolutional network, reduces the deviation between the damage positioning and the actual situation through balancing, and simultaneously regularizes the generalization ability of the model to unknown scenarios.

[0267] Specifically, the step S7 generates the structure body analysis model while setting up a test set under multiple scenarios, including simulation data and real acquisition data, comprehensively evaluates the model performance, and provides a reference for optimizing the network structure in the analysis method.

[0268] Embodiment Two:

[0269] The embodiment also provides a computer device suitable for the case of the structure body analysis method based on deep learning, including a memory and a processor; the memory is used to store computer executable instructions, and the processor is used to execute the computer executable instructions to realize the structure body analysis method based on deep learning as proposed in the above embodiment.

[0270] The embodiment also provides a storage medium having a computer program stored thereon, and the program is executed by a processor to realize the structure body analysis method based on deep learning as proposed in the above embodiment.

[0271] ​The computer device can be a terminal, and the computer device includes a processor, a memory, a communication interface, a display screen and an input device connected by a system bus. The processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for running the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is configured to perform wired or wireless communication with an external terminal. The wireless communication can be achieved by WIFI, an operator network, NFC (Near Field Communication) or other technologies. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen. The input device of the computer device can be a touch layer overlaid on the display screen, or can be a key, a trackball or a touchpad arranged on the shell of the computer device, or can be an external keyboard, a touchpad or a mouse, etc.

[0272] If the functions are implemented in the form of software function units and sold or used as independent products, the functions can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application or the parts of the technical solutions that essentially contribute to the prior art can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the embodiments of the present application. The aforementioned storage medium includes a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various media that can store program codes.

[0273] The logic and / or steps represented in the flowcharts or otherwise described herein, for example, can be considered as a list of executable instructions for implementing logic functions, which can be specifically embodied in any computer readable medium for use by or in conjunction with an instruction execution system, device or apparatus, such as a computer-based system, a system including a processor or other system that can fetch and execute instructions from the instruction execution system, device or apparatus. For the purpose of this specification, the "computer readable medium" can be any device that can contain, store, communicate, propagate or transport programs for use by or in conjunction with the instruction execution system, device or apparatus, or in conjunction with these instruction execution systems, devices or apparatus.

[0274] More specific examples (a non-exhaustive list) of the computer-readable medium include the following: an electrical connection (electronic) having one or more wires, a portable computer diskette (magnetic), a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or Flash memory), an optical fiber device, and a portable compact disc read-only memory (CDROM). Additionally, the computer-readable medium can also be paper or another suitable medium upon which the program is printed, as the program can be electronically captured, for example, via an optical scanner, then compiled, interpreted, or otherwise processed, as necessary, and stored in a computer memory.

[0275] It should be understood that aspects of the application can be implemented in hardware, software, firmware or combinations thereof. In the embodiments described above, various steps or methods can be implemented, for example, in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, and in another embodiment, any of the following technologies, known in the art, or combinations thereof, can be used: discrete logic circuitry having logic gates for implementing logic functions upon data signals, application specific integrated circuits having appropriate combinational logic gates, programmable gate arrays (PGA), field programmable gate arrays (FPGA), and the like.

[0276] It should be noted that the above-mentioned embodiments are merely used to illustrate the technical solutions of the present application but not to limit the present application, and although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or equivalent replaced without departing from the spirit and scope of the present application, and all of them should be covered in the scope of the claims of the present application.

Claims

1. A deep learning-based structural object shape analysis method, characterized by, The method comprises the following steps: S1, data collection: collecting and labeling the shape information of the structure to obtain the original data of the structure; S2, data enhancement and preprocessing: introducing different environments or collection angles, enhancing and preprocessing the original data to obtain diversity data of the structure; S3, constructing a deep learning network for processing diversity data: combining multi-scale convolutional layers with graph convolutional networks to construct a deep learning network based on PointNett++ architecture; S4, designing a multi-task learning framework for the deep learning network: the multi-task learning framework includes a shape change prediction branch and a damage location detection branch; S5, adding an unsupervised learning module in the multi-task learning framework: introducing an unsupervised learning module using an autoencoder to process the shape change or damage of the structure which is difficult to label; S6, introducing a time-space change factor into the deep learning network: introducing a time-series convolutional network to predict the shape change of the structure at different time points and under different conditions; S7, model generation: introducing diversity data into the deep learning network to obtain the shape model of the structure through the above steps.

2. The method of claim 1, wherein the method is based on deep learning. The original data enhancement method in step S2 includes: Geometric transformation; Random noise injection; Data difference and resampling; Simulation data generation.

3. The method of claim 1, wherein the method is based on deep learning. The preprocessing method in step S2 includes: Outlier rejection; Modal data alignment; Feature extraction and unified representation; Multi-modal data fusion; Data consistency verification; Output unified data set.

4. The structure shape analysis method based on deep learning according to claim 1, characterized by, The deep learning network in step S3 includes: Input layer: The input diversity data is an enhanced point cloud, each point containing a three-dimensional coordinate , a color , and a normal vector ; Point cloud is processed by normalization and downsampling to adapt the data size to hardware limitations; Multi-scale feature extraction layer: Extract the geometric features of the local area of the point cloud through multi-scale grouping method; Use layer-by-layer downsampling to capture structure details of different sizes; Multi-modal fusion module: For multi-modal data, embed features from other modalities in the global feature layer; Use attention mechanism to weight multi-modal features; Damage focusing module: Introduce an adaptive focusing mechanism to focus on potential damage areas in the point cloud; Use point-level attention weights; Output layer: Damage identification classification; Damage localization regression.

5. The method of claim 4, wherein the method is based on deep learning. The layer-by-layer downsampling in the multi-scale feature extraction layer includes: Hierarchical sampling; Local feature learning; Global feature aggregation.

6. The structure shape analysis method based on deep learning according to claim 1, characterized by, The multi-task learning framework in step S4 includes: Input module: Preprocessed point cloud or multi-modal data as input, including three-dimensional coordinates, normal vectors, and colors; Shared feature extraction module: Based on PointNet++ architecture, extract global and local features of point cloud; Branch 1, shape change prediction branch: focus on global geometric shape change analysis and regression; Branch 2, damage location detection branch: focus on local damage area classification and positioning; Wherein, The input form of the input module is: ； Each point contains the following features: a three-dimensional coordinate , a normal vector , and a color value ; The shared feature extraction module is a hierarchical structure, including: Layer 1 local features Local geometric features of each sampling point are extracted using a ball radius query and a shared MLP for lesion site detection; Layer 2 global features , aggregate local features, generate global feature vectors through a fully connected layer and global pooling for shape variation prediction; Branch 1: shape change prediction branch: Use mean square error method to calculate the prediction error of shape change: ； wherein is the predicted value, is the true value; computed mean square error to guide the model to learn accurate deformation patterns and stress distributions; Branch 2: damage location detection branch: Use cross-entropy loss to calculate classification error: ； wherein, is a true class label, is a predicted probability; The computed cross-entropy loss For lesion site classification, the model is reinforced for sensitivity to lesion sites.

7. The method of claim 1, wherein the method is based on deep learning. The architecture of the unsupervised learning module in step S5 includes: Encoder: responsible for extracting multi-scale features from input data and compressing them into low-dimensional latent representations; Its input is: point cloud data: Each point contains three-dimensional coordinates , normal vector and color value ; The output is: low-dimensional latent features ; Latent space representation: used to learn a compact global feature representation that represents the core information of the point cloud and avoids overfitting through regularization; Decoder: maps latent features back to the original space for reconstructing input point cloud data; Its input is: low-dimensional latent features ; Its output is: reconstructed point cloud . 8.The method of claim 1, wherein, The time series convolution network in step S6 includes: Data preprocessing: Convert point cloud or other spatio-temporal sequence data into time step sequences: ； wherein is a characteristic representation of time ; Causal convolution layer: Ensure that the output only depends on data at the current and past time points: ； Extended convolution layer: Use extended convolution to expand the receptive field and capture long-term dependencies: ； wherein is the stride length, is the convolution kernel weight, the extended convolution expands the distance between points to , achieving efficient long-time feature capture; Residual connection: Residual connection makes the network deeper and more stable, avoiding gradient vanishing: ； Use 1x1 convolution to align feature dimensions; Overall structure of the network. 9.The method of claim 8, wherein, The loss function in the time series convolution network includes: Prediction error loss: For measuring the gap between the predicted output and the true value is calculated as ； where the prediction error loss represents a measure of how well the model is capturing the temporal step sequence of inputs; and the accuracy of the prediction of future outputs, i.e., the difference between the predicted result and the true label. Regularization loss: Limit network weights to avoid overfitting, with the formula: ； wherein the regularization loss represents for limiting the scale of model weights, to prevent overfitting phenomenon, that is, the model over-reliance on training data, leading to the decline of generalization performance; Total loss: ； The calculation of this loss function plays a trade-off role in the time series convolution network, reducing the deviation between the damage location and the actual situation while regularizing the model's generalization ability to unknown scenarios. 10.The method of claim 1, wherein, While generating the structure object analysis model in step S7, a test set under multiple scenarios is established, including simulation data and real acquisition data, to comprehensively evaluate the model performance and provide a reference for optimizing the network structure in the analysis method.

Images