Adaptive training method for complex environment of underground disease recognition model using 3D ground penetrating radar

Abstract

The present invention discloses a complex environment adaptive training method for a three-dimensional ground penetrating radar underground disease recognition model, which belongs to the field of computer model technology, including obtaining a source domain data set D s and the target domain dataset D t ; Construct a three-dimensional encoding and decoding network; use D S Train the 3D codec network to get the basic model; use D t The student model is trained to optimize pseudo-labels; a mixed-domain dataset is constructed based on hybrid weights to train the student model. During training, the teacher model is used to generate pseudo-labels for samples within the mixed-domain dataset. The student model is updated to minimize the joint loss, and the teacher model is updated using the EMA method. This invention is used to train and update underground disease data with domain differences, meeting the complex requirements of cross-domain underground disease detection. It features high 3D semantic segmentation accuracy, strong boundary recognition capabilities, breakthrough enhancements in cross-domain environmental adaptability and generalization performance, and optimized computational efficiency and resource consumption.

Description

Technical Field

[0001] The present invention relates to the technical field of specific computer models, and in particular to a complex environment adaptive training method for a three-dimensional ground penetrating radar underground hazard recognition model. Background Art

[0002] Hidden structural defects and media anomalies in urban underground spaces (such as roads, tunnels, pipelines, foundations, and surrounding soils) caused by natural geological processes or human engineering activities are characterized by their hidden location and dangerous evolution. As the intensity of urban underground space development continues to increase, the previously stable strata beneath roads gradually suffer structural damage or deterioration of their media properties due to the coupled influence of multiple factors, including repeated vehicle loads, vibration disturbances from underground engineering projects, and dynamic changes in groundwater flow (such as pipeline leakage and rainwater erosion). This leads to typical defects such as voids (material-free areas caused by soil loss), voids (detachment between structural layers), loose bodies (loose soil with insufficient compaction, categorized as severely loose or generally loose), water-rich bodies (water-saturated seepage zones), and crack networks (fracture zones within rock, soil, or structures). These diseases often have no obvious symptoms in the early stages, but will continue to expand through stress redistribution (such as longitudinal extension of cracks and formation of honeycomb-like erosion channels in cavities), leading to a gradual loss of the bearing capacity of the rock and soil; when the disease develops upward and breaks through the critical state, it may trigger sudden disasters such as road collapse, tunnel deformation, pipeline rupture, and building foundation instability, posing a serious threat to urban public safety.

[0003] Three-dimensional ground-penetrating radar (GPR) technology is a non-destructive inspection method that uses electromagnetic wave reflection signals to obtain three-dimensional information about underground structures. It is widely used for underground defect detection. Traditional detection methods rely on manual experience to interpret radar images, which is inefficient and has high false positive rates. In recent years, the integration of deep learning and 3D GPR has provided new approaches for intelligent defect identification. Deep learning automatically extracts multi-scale features and, through data training, learns complex feature expressions and underlying patterns. This allows for precise localization of defect morphology and distribution from complex radar data, significantly improving detection efficiency and accuracy. Furthermore, the spatial continuity and high resolution of 3D GPR data enable accurate capture of detailed information such as the morphology, size, and depth of underground defects. When combined with deep learning models (such as encoder-decoder architectures), it enables three-dimensional defect analysis, adapting to the requirements of cross-domain generalization across diverse geological conditions and equipment parameters, providing reliable support for defect detection in complex environments.

[0004] However, in real-world applications, varying geological conditions, radar equipment parameter differences, and environmental interference can cause data distribution shifts, severely impacting the model's ability to generalize across scenarios. Domain adaptation is a transfer learning technique designed to address the issue of inconsistent data distribution between the source domain (training data) and the target domain (test data). Domain adaptation adjusts the model or data representation to effectively transfer knowledge learned in the source domain to the target domain, thereby improving the model's generalization capabilities.

[0005] As mentioned in the paper "Apple Leaf Disease Recognition Based on a Self-Supervised Domain Adaptation Network," a model for apple leaf disease recognition introduces a domain adaptation method and a self-supervised module. This model utilizes joint training of source and target domain datasets to reduce domain bias and enhance the generalization of the pre-trained model. A contrastive loss is used to narrow the feature distance between positive and negative samples, enhancing the model's ability to distinguish similar diseases. This allows the model to learn more detailed representations of diseased areas, improving classification performance and generalization in real agricultural scenarios. However, this method primarily targets two-dimensional image data and lacks the ability to model continuous three-dimensional features. Subsurface diseases (such as the longitudinal extension of cracks and the honeycomb structure of cavities) require stereoscopic analysis using 3D radar data. However, the self-supervised module in this paper focuses solely on two-dimensional transformations such as image rotation and brightness enhancement, failing to effectively capture long-range dependencies in three-dimensional space. Furthermore, during the source domain training phase, optimization is performed solely using a cross-entropy loss, without dynamic adjustments to address class imbalance. This limits the model's ability to recognize sparse diseases (such as cavities and cracks). In addition, its data mixing strategy does not combine the uncertainty of the target domain data with the local noise assessment. For example, it does not introduce a three-dimensional neighborhood confidence screening mechanism, resulting in insufficient reliability of the pseudo-labels generated under complex geological interference (such as metal pipeline reflection artifacts), which is prone to false detection due to noise propagation.

[0006] For example, the paper "Domain-Adaptive Remote Sensing Image Segmentation Based on Hierarchical Attention" addresses the task of semantic segmentation in remote sensing images. It employs a hierarchical multi-head self-attention mechanism combined with multi-scale feature fusion, combined with a self-trained domain adaptation framework. This framework optimizes cross-domain knowledge transfer through local pseudo-label quality (LPQ) and dynamic class weighting (DCW), significantly improving the accuracy and robustness of remote sensing image segmentation. Despite the introduction of hierarchical attention and multi-scale fusion, its domain adaptation framework is designed for 2D images, making it difficult to directly transfer to 3D radar data. For example, the local pseudo-label quality (LPQ) algorithm uses 2D neighborhood convolution to assess confidence, ignoring the spatial continuity between 3D voxels. This results in large pseudo-label filtering errors in regions with blurred boundaries, such as the transition zone between fractures and normal media. Furthermore, its self-training strategy relies on 2D sliding window slicing, making it inefficient for extracting vertically layered features from 3D radar data. This makes it ineffective in handling high-dimensional computational complexity, limiting real-time detection capabilities. Furthermore, its data blending strategy fails to fully consider the target domain data, making pseudo-labels susceptible to false signals in the presence of complex geological interference, further reducing the stability of cross-domain knowledge transfer.

[0007] Therefore, existing technologies are limited to two-dimensional data processing or lack three-dimensional spatial modeling capabilities, making it difficult to adapt to the challenges of complex environments in underground disease detection, such as parameter differences among multiple devices and geological noise interference.

[0008] Glossary:

[0009] InfoNCE (Information Noise Contrastive Estimation) is a contrastive learning loss function that maximizes the mutual information between positive samples (matching samples) while distancing negative samples (non-matching samples).

[0010] The three-dimensional neighborhood confidence calculation formula of the LPQ (Local Phase Quantization) function is as follows:

[0011] ,

[0012] This formula indicates that in three-dimensional space, the phase quantization value Q of each position is averaged within a neighborhood centered at position (d, h, w) to obtain the LPQ value of that position, where K is the neighborhood radius.

[0013] Monte Carlo Dropout (MC Dropout) is a deep learning inference method that combines the Monte Carlo method and the Dropout regularization technique. It is mainly used to quantify the prediction uncertainty of neural network models. The key innovation of Monte Carlo Dropout is that it still keeps the Dropout layer activated during the test phase and performs multiple forward propagations on the same input. Each forward propagation randomly drops different neurons, resulting in different outputs. The results of multiple forward propagations can be used to estimate the prediction distribution of the model and thus quantify the uncertainty of the prediction. Its workflow is as follows: (1) Enable Dropout for inference: Keep the Dropout layer activated during the test phase and randomly drop some neurons during each forward propagation; (2) Perform multiple forward propagations for each input sample to obtain multiple prediction results; (3) Statistical analysis: By analyzing the outputs of multiple forward propagations, the mean and variance of the prediction results are calculated. The mean represents the model's prediction, and the variance represents the model's uncertainty about the prediction. Summary of the Invention

[0014] The purpose of the present invention is to provide a complex environment adaptive training method for a three-dimensional ground penetrating radar underground disease recognition model to solve the problems of the existing technology being limited to two-dimensional data processing or lacking three-dimensional spatial modeling capabilities, making it difficult to adapt to complex environments in underground disease detection, such as differences in multi-device parameters and geological noise interference.

[0015] In order to achieve the above object, the technical solution adopted by the present invention is as follows: a complex environment adaptive training method for a three-dimensional ground penetrating radar underground disease recognition model, comprising the following steps;

[0016] S1, obtain several samples with category annotations to form the source domain dataset D s , several unlabeled samples constitute the target domain dataset D t , the sample is three-dimensional ground penetrating radar data, and the category is disease category;

[0017] D s Sample X s The category labeling diagram is P s , D t One of the samples is X t , X s ∈R D×H×W×Cin 、X t ∈R D×H×W×Cin 、P s ∈R D×H×W×C , D, H, W are the depth, length and width of the sample resolution respectively, Cin is the number of channels of radar reflection intensity, Cin=1, C is the total number of disease categories, P sThe value at position (d,h,w) is X s The true disease category at the middle position (d, h, w), d, h, w are the depth, length and width of the position (d, h, w) respectively;

[0018] S2, constructing a 3D encoding and decoding network, including a 3D encoder, a 3D decoder, and a self-supervision module;

[0019] The three-dimensional encoder is used to encode the input sample to obtain the encoding feature, and the three-dimensional decoder is used to decode the encoding feature and then perform boundary enhancement, and output the predicted category probability of each position to form a probability prediction map , the self-supervisory module is used to construct positive samples and negative samples according to the encoding features of a batch of samples and calculate the InfoNCE loss;

[0020] S3, use D S To minimize the total loss L of source domain training s Train the 3D encoding and decoding network to obtain the basic model, L s =λ cn1 L cn1 +λ cls L cls , L cn1 is the InfoNCE loss, L cls is the classification loss, λ cn1 ,λ cls L cn1 and L cls The weight of

[0021] S4, copy two base models and use them as teacher model and student model respectively;

[0022] S5, use D t Training the student model, each iteration optimizing the student model based on self-supervised contrastive learning and double-screened pseudo-label supervised learning, wherein one optimization includes steps S51 to S54;

[0023] S51, use the teacher model to output D t The probability prediction map of each sample in is used as a pseudo label, X t The pseudo label is , preset quality threshold λ and uncertainty threshold ϵ;

[0024] S52, from D t Choose a batch of samples from the set, where the bth sample X b , use the student model to output the probability prediction map of each sample, and use the self-supervised module to calculate the InfoNCE loss;

[0025] S53, calculation of X based on double screening b The pseudo-label supervision loss Lp,b ;

[0026] To X b Pseudo labels , based on the local phase quantization calculation position (d, h, w) LPQ value LPQ (d, h, w), for X b , calculate the uncertainty U(d,h,w) of the position (d,h,w) based on MC Dropout, and form X with the position where LPQ(d,h,w)>λ and U(d,h,w)<ϵ b Areas of interest , calculate L p,b ;

[0027] ,

[0028] Where, for Total number of positions within is the normalized value of U(d,h,w), For X b Probability prediction chart of The value at position (d,h,w);

[0029] S54, calculate the total loss L of the target domain t , and to minimize L t Adjust student model parameters;

[0030] ,

[0031] Where B is the batch size, L cn2 is the InfoNCE loss of the self-supervised module in the student model, λ cn2 For L cn2 The weight of

[0032] S6, use D s and D t Construct multiple batches of mixed-domain datasets;

[0033] S7, train the student model with the mixed domain dataset to minimize the joint loss L total Update the student model and update the teacher model using EMA. total =+L s +λ t L t +λ c L consist , L consist is the consistency loss of a batch of samples.

[0034] Preferably, the disease categories include background, voids, voids, loose bodies, water-rich bodies and crack networks.

[0035] As a preferred method, the self-supervisory module calculates the InfoNCE loss as follows:

[0036] Get the encoding features of a batch of samples, marked as F1~F B , calculate the InfoNCE loss of the b-th encoded feature and the InfoNCE loss L of a batch of samples cn1 ;

[0037] ,

[0038] ,

[0039] Where, For F b Randomly add data enhancement to get positive samples, F b The vth negative sample, 1≤v≤2B-2, sim(⋅) is the cosine similarity, τ is the temperature coefficient, and exp(⋅) is the exp function.

[0040] Preferably, in S3, the classification loss is calculated according to the following formula:

[0041] ,

[0042] Where, α c is the weight of the c-th disease category, 1≤c≤C, γ is the modulation factor to reduce the weight of easy-to-classify samples, for the position (d, h, w) of the b-th sample, 、 are the normalized value of the predicted category probability at the location and the actual disease category label, respectively.

[0043] Preferably, LPQ(d,h,w) in S53 is calculated according to the following formula:

[0044] ,

[0045] Where K is the neighborhood size, i, j, and k are the offsets of the position (d, h, w) in depth, length, and width, respectively, and Q(d+i,h+j,w+k) is the pseudo label The value at position (d+i,h+j,w+k).

[0046] Preferably, the method for calculating U(d,h,w) in S53 is:

[0047] Use MC Dropout method to t Perform M forward propagations, each generating a probability prediction map. Calculate the uncertainty U(d, h, w) at the position (d, h, w) according to the following formula:

[0048] ,

[0049] Where, The probability prediction graph obtained by the mth forward propagation The value at position (d,h,w), is the mean of the M probability prediction maps at position (d, h, w).

[0050] As a preferred method, in S6, a method for constructing a batch mixed domain dataset is: s and D t After mixing, select any batch of samples, or select from D according to the mixing weight. s and D t A batch of samples is selected, and the mixing weight is preset or obtained according to steps S61 to S63;

[0051] S61, from D s 、D t Each batch of samples in constitutes a set 、 ;

[0052] S62, calculation The LPQ value and uncertainty of each sample at each position in the batch are averaged to obtain the batch average local quality score. , all uncertainties are averaged to get the batch average uncertainty ;

[0053] S63, calculation The mixed weight ;

[0054] ,

[0055] α and β are and The adjustment coefficient, Sigmoid(⋅) is the Sigmoid function;

[0056] according to , generate a batch of mixed domain datasets .

[0057] As a preference, in S7, L consist Calculated according to the following formula:

[0058] ,

[0059] ,

[0060] Where, is the square of the L2 norm, for the b-th sample in the mixed domain, is the probability prediction graph output by the student model, Pseudo labels generated for the teacher model, is the consistency loss of the b-th sample.

[0061] Preferably, in the 3D encoding and decoding network, the 3D encoder includes a fixed block embedding module, a first multi-scale sampling module, and a second multi-scale sampling module connected in sequence;

[0062] The fixed-block embedding module includes a three-dimensional convolutional layer and a three-dimensional position encoding layer;

[0063] The three-dimensional convolution layer is used to perform three-dimensional convolution on the input sample to reduce the dimension and obtain the reduced dimension feature F cn , where the convolution kernel of the three-dimensional convolution is 3×3×3 and the stride is 2;

[0064] The three-dimensional position encoding layer is used to cn Generate a three-dimensional sinusoidal position code for each position and obtain the position code feature F p , and then F cn and F p The embedded feature F0 is obtained by adding each element, and the scale of the sample is D×H×W×C in , F cn 、F p , the scale of F0 is ;

[0065] The first multi-scale sampling module includes a global attention layer, a local convolution layer, a layer normalization layer and a three-dimensional downsampling layer;

[0066] The multi-scale sampling module is used to perform global attention on the input features to obtain the global feature F global ;

[0067] The local convolution layer is used to global Perform depth-separable 3D convolution to obtain local features F local , the convolution kernel is 3×3×3, and the stride is 1;

[0068] The layer normalization layer is used according to the formula F mix =LayerNorm(F global +F local ) Get the mixed feature F mix ;

[0069] The three-dimensional downsampling layer is used to mix Perform three-dimensional convolution downsampling and obtain The output feature F d1 ;

[0070] The structure of the second multi-scale sampling module is the same as that of the first multi-scale sampling module, and the input feature is F d1 , the output scale is The output feature F d2 , and F d2 is the encoding feature of the three-dimensional encoder.

[0071] Preferably, in the 3D encoding and decoding network, the 3D decoder comprises a first upsampling module, a second upsampling module, a third upsampling module and a boundary enhancement layer connected in sequence;

[0072] The first upsampling module includes a transposed convolution upsampling layer, a cross-layer splicing layer, a hole residual module, and a residual connection activation layer;

[0073] The transposed convolution upsampling layer is used to perform a transposed three-dimensional convolution on the encoded features to obtain the first feature F u1 , the convolution kernel of the transposed three-dimensional convolution is 2×2×2, and the step size is 2; the cross-layer splicing layer is used to convert F u1 The second feature F is obtained by splicing it with the cross-layer feature along the channel dimension u2 , where the cross-layer feature is F d1 ;

[0074] The hole residual module is used to calculate F u2 Perform dilation convolution with dilation rates of 2 and 4, and then add them together to obtain the third feature F u3 ;

[0075] The residual connection activation layer is used to convert F u2 and F u3 Perform residual connection and activate it with GELU function to get the output F of the first upsampling module uout1 ;

[0076] Duplicate the first upsampling module and add a channel compression layer at the input end to compress the input feature channel by half before feeding it into the transposed convolution upsampling layer. Modify the cross-layer feature to F0 and the expansion rate of the void residual module to 1 or 2 to obtain the second upsampling module.

[0077] Copy the second upsampling module and modify the cross-layer features into feature F 01 , the hole residual module only performs hole convolution with a dilation rate of 1 on the input features, and obtains the third upsampling module, feature F 01 Obtained by upsampling the F0 transposed convolution, the convolution kernel is 2×2×2, and the stride is 2;

[0078] The classifier is used to uout3 Classify by position and get the initial predicted category probability P for each position base ;

[0079] The boundary enhancement layer F uout3 Use the 3D Sobel operator to detect edges and output the boundary response map P Sobel , then according to the formula , generate a probability prediction map , where λ Sobel P Sobel The weight of .

[0080] In this invention, underground disease body types primarily include voids (areas without material formed by soil loss), voids (detachment between structural layers), loose bodies (loose soil with insufficient compaction, categorized as severely loose or moderately loose), water-rich bodies (water-saturated seepage zones), and fracture networks (fracture zones within rock, soil, or structures). These are, of course, not limited to these types and can be defined and labeled as needed.

[0081] This paper addresses the problem of degraded model generalization performance caused by differences in GPR data distribution across different domains (e.g., varying geological conditions and equipment parameters). By doing so, we propose a complex environment adaptive training method for a 3D GPR underground defect recognition model. This method implements cross-domain knowledge transfer through three training phases: source domain training, target domain training, and joint optimization.

[0082] The first stage: source domain training, which combines supervised learning and self-supervised contrastive learning to enhance the model's generalization ability to sparse diseases and use Focal Loss to alleviate the class imbalance problem in the data.

[0083] The first stage of supervised learning is to train the 3D encoding and decoding network with the source domain dataset, input the sample into the model, output the probability prediction map and calculate the classification loss L cls The process of L cls In the formula, FocalLoss is used to replace cross entropy loss to solve the problem of class imbalance. By dynamically adjusting the weights of difficult and easy samples, the classification accuracy of sparse categories is improved, which significantly alleviates the model bias problem caused by class imbalance. In general, set ,make The frequency of category c Inversely proportional, It is used to reduce the weight of easy-to-classify samples and focus on difficult samples (such as sparse diseases).

[0084] The first stage of self-supervised contrastive learning involves using the self-supervised module to construct positive and negative samples based on a batch of source domain samples using the InfoNCE method. The InfoNCE loss is then calculated through contrastive learning. This is because ground penetrating radar data can suffer from domain shift due to differences in the acquisition environment, such as changes in soil moisture. Contrastive learning using the InfoNCE loss enhances feature discriminability, forcing the model to bring enhanced positive samples closer together in feature space while pushing negative samples further away. This increases feature robustness to geometric deformation and noise, mitigating distribution differences between the source and target domains.

[0085] The second stage: target domain training, optimizing the student model based on self-supervised contrastive learning and double-screened pseudo-label supervised learning.

[0086] The second stage of self-supervised contrastive learning is similar to the first stage, except that the input batch of samples comes from the target domain dataset.

[0087] The second stage of double-screening pseudo-label supervised learning refers to the pseudo-label supervision loss L of the sample calculated by double-screening the focus area through LPQ and MC Dropout. p,b , and based on L p,b Calculate the batch pseudo-label supervision loss L ped This phase first calculates an LPQ value for each position of the sample's pseudo-label. The MC Dropout method then calculates the uncertainty of each position. Finally, the positions are double-filtered using the conditions LPQ(d,h,w)>λ and U(d,h,w)<ϵ. The benefit of LPQ filtering is that it filters out low-quality pseudo-label areas through three-dimensional neighborhood confidence assessment, suppressing the error propagation of noisy labels. The benefit of MCDropout filtering is that it quantifies model prediction uncertainty, eliminates high-variance areas, and ensures the reliability of knowledge transfer. This double filtering mechanism improves the reliability of pseudo-labels and the robustness of cross-domain knowledge transfer. It also enhances feature consistency through comparative learning, autonomously strengthening the ability to discriminate between similar diseases.

[0088] The third stage: mixed domain training, refers to the process of constructing a mixed domain training student model and updating the teacher model parameters through exponential moving average (EMA). This stage introduces mixed weights, which are based on the samples in a batch of target domains. and The calculation shows that the model dynamically adapts to data quality: high-quality target data prioritizes the target domain (γ→1) to accelerate adaptation, while low-quality data prioritizes the source domain (γ→0) to avoid negative transfer. This improves the model's adaptability in complex environments. Consistency constraints are introduced at this stage to reduce noise interference.

[0089] Compared with the prior art, the advantages of the present invention are:

[0090] (1) The constructed 3D encoding and decoding network significantly improves the 3D semantic segmentation accuracy and boundary recognition capabilities:

[0091] When dealing with underground diseases with complex three-dimensional distributions such as the network expansion of cracks and the longitudinal extension of cavities, traditional methods often lead to blurred boundaries and misjudgments due to insufficient local feature extraction or inadequate global dependency modeling. The three-dimensional encoding and decoding network proposed in this invention models global long-range dependencies through a three-dimensional multi-head self-attention mechanism, combines deep separable convolution to extract local detail features, and effectively integrates multi-scale contextual information; at the same time, it introduces progressive void residual upsampling and three-dimensional Sobel operator boundary enhancement modules to enhance the distinguishability of disease boundaries through gradient response. Theoretical derivation shows that this method can accurately capture the three-dimensional distribution characteristics of underground diseases, significantly improve the positioning capability of small targets and the robustness of boundary segmentation, and is particularly suitable for the refined identification of cracks and cavities in complex geological structures.

[0092] (2) Breakthrough enhancement of cross-domain environment adaptability and generalization performance:

[0093] Existing models often suffer from performance degradation in cross-domain scenarios, such as those involving varying geological conditions and radar equipment parameters, due to data distribution shifts. The proposed three-stage training method can autonomously adapt to noise interference and domain shifts in complex environments, reducing reliance on labeled data while maintaining high-precision detection capabilities in unlabeled target domain data. This significantly outperforms traditional models trained in a single domain.

[0094] (3) Optimization of computing efficiency and resource consumption:

[0095] To address the high computational complexity of 3D data processing, the model employs depthwise separable 3D convolutions and progressive dilated residual upsampling techniques, reducing computational complexity while maintaining efficient fusion of multi-scale features. Furthermore, the teacher-student framework, combined with an exponential moving average (EMA) parameter update mechanism, reduces training fluctuations through smooth iteration of pseudo-labels, further optimizing training resource consumption. Theoretical derivation demonstrates that compared to the frequent parameter adjustments and high computing power requirements of traditional domain adaptation methods, this method achieves more efficient computational resource utilization while maintaining accuracy, making it suitable for the real-time processing of large-scale 3D radar data in practical engineering applications.

[0096] In summary, the present invention can maintain high-precision detection capabilities in complex scenarios with different geological conditions and equipment parameters, significantly reducing dependence on target domain annotation data, and providing reliable support for intelligent and universal detection of underground diseases. BRIEF DESCRIPTION OF THE DRAWINGS

[0097] Figure 1 Flowchart of the present invention;

[0098] Figure 2 This is a diagram of the architecture of the present invention during training;

[0099] Figure 3 Schematic diagram of the three-dimensional encoder structure;

[0100] Figure 4 Schematic diagram of the structure of a three-dimensional decoder;

[0101] Figure 5 Schematic diagram of the structure of the first upsampling layer;

[0102] Figure 6 Schematic diagram of the structure of the second upsampling layer;

[0103] Figure 7 Schematic diagram of cross-layer feature access. DETAILED DESCRIPTION

[0104] The present invention will be further described below with reference to the embodiments and accompanying drawings.

[0105] Example 1: See Figure 1 and Figure 2 ,A complex environment adaptive training method for a 3D ground penetrating radar underground hazard recognition model, ,including the following steps;

[0106] S1, obtain several samples with category annotations to form the source domain dataset D s , several unlabeled samples constitute the target domain dataset D t , the sample is three-dimensional ground penetrating radar data, and the category is disease category;

[0107] D s Sample X s The category labeling diagram is P s , D t One of the samples is X t , X s ∈R D×H×W×Cin 、X t ∈R D×H×W×Cin 、P s ∈R D×H×W×C , D, H, W are the depth, length and width of the sample resolution respectively, Cin is the number of channels of radar reflection intensity, Cin=1, C is the total number of disease categories, P s The value at position (d,h,w) is X s The true disease category at the middle position (d, h, w), d, h, w are the depth, length and width of the position (d, h, w) respectively;

[0108] S2, constructing a 3D encoding and decoding network, including a 3D encoder, a 3D decoder, and a self-supervision module;

[0109] The three-dimensional encoder is used to encode the input sample to obtain the encoding feature, and the three-dimensional decoder is used to decode the encoding feature and then perform boundary enhancement, and output the predicted category probability of each position to form a probability prediction map , the self-supervisory module is used to construct positive samples and negative samples according to the encoding features of a batch of samples and calculate the InfoNCE loss;

[0110] S3, use D S To minimize the total loss L of source domain training s Train the 3D encoding and decoding network to obtain the basic model, L s =λ cn1 L cn1 +λ cls L cls , L cn1 is the InfoNCE loss, L cls is the classification loss, λ cn1 ,λ cls L cn1 and L cls The weight of

[0111] S4, copy two base models and use them as teacher model and student model respectively;

[0112] S5, use D t Training the student model, each iteration optimizing the student model based on self-supervised contrastive learning and double-screened pseudo-label supervised learning, wherein one optimization includes steps S51 to S54;

[0113] S51, use the teacher model to output D t The probability prediction map of each sample in is used as a pseudo label, X t The pseudo label is , preset quality threshold λ and uncertainty threshold ϵ;

[0114] S52, from D t Choose a batch of samples from the set, where the bth sample X b , use the student model to output the probability prediction map of each sample, and use the self-supervised module to calculate the InfoNCE loss;

[0115] S53, calculation of X based on double screening b The pseudo-label supervision loss L p,b ;

[0116] To X b Pseudo labels , based on the local phase quantization calculation position (d, h, w) LPQ value LPQ (d, h, w), for X b, calculate the uncertainty U(d,h,w) of the position (d,h,w) based on MC Dropout, and form X with the position where LPQ(d,h,w)>λ and U(d,h,w)<ϵ b Areas of interest , calculate L p,b ;

[0117] ,

[0118] Where, for Total number of positions within is the normalized value of U(d,h,w), For X b Probability prediction chart of The value at position (d,h,w);

[0119] S54, calculate the total loss L of the target domain t , and to minimize L t Adjust student model parameters;

[0120] ,

[0121] Where B is the batch size, L cn2 is the InfoNCE loss of the self-supervised module in the student model, λ cn2 For L cn2 The weight of

[0122] S6, use D s and D t Construct multiple batches of mixed-domain datasets;

[0123] S7, train the student model with the mixed domain dataset to minimize the joint loss L total Update the student model and update the teacher model using EMA. total =+L s +λ t L t +λ c L consist , L consist is the consistency loss of a batch of samples.

[0124] The damage categories include background, voids, voids, loose bodies, water-rich bodies and crack networks.

[0125] The method for calculating InfoNCE loss in the self-supervised module is:

[0126] Get the encoding features of a batch of samples, marked as F1~F B , calculate the InfoNCE loss of the b-th encoded feature and the InfoNCE loss L of a batch of samples cn1 ;

[0127] ,

[0128] ,

[0129] Where, For F b Randomly add data enhancement to get positive samples, F b The vth negative sample, 1≤v≤2B-2, sim(⋅) is the cosine similarity, τ is the temperature coefficient, and exp(⋅) is the exp function.

[0130] In S3, the classification loss is calculated according to the following formula;

[0131] ,

[0132] Where, α c is the weight of the c-th disease category, 1≤c≤C, γ is the modulation factor to reduce the weight of easy-to-classify samples, for the position (d, h, w) of the b-th sample, 、 are the normalized value of the predicted category probability at the location and the actual disease category label, respectively.

[0133] In S53, LPQ (d, h, w) is calculated according to the following formula:

[0134] ,

[0135] Where K is the neighborhood size, i, j, and k are the offsets of the position (d, h, w) in depth, length, and width, respectively, and Q(d+i,h+j,w+k) is the pseudo label The value at position (d+i,h+j,w+k).

[0136] The method for calculating U(d,h,w) in S53 is:

[0137] Use MC Dropout method to t Perform M forward propagations, each generating a probability prediction map. Calculate the uncertainty U(d, h, w) at the position (d, h, w) according to the following formula:

[0138] ,

[0139] Where, The probability prediction graph obtained by the mth forward propagation The value at position (d,h,w), is the mean of the M probability prediction maps at position (d, h, w).

[0140] In S6, the method of constructing a batch mixed domain dataset is: s and D t After mixing, select any batch of samples, or select from D according to the mixing weight. s and D t A batch of samples is selected, and the mixing weight is preset or obtained according to steps S61 to S63;

[0141] S61, from D s 、D t Each batch of samples in constitutes a set 、 ;

[0142] S62, calculation The LPQ value and uncertainty of each sample at each position in the batch are averaged to obtain the batch average local quality score. , all uncertainties are averaged to get the batch average uncertainty ;

[0143] S63, calculation The mixed weight ;

[0144] ,

[0145] α and β are and The adjustment coefficient, Sigmoid(⋅) is the Sigmoid function;

[0146] according to , generate a batch of mixed domain datasets .

[0147] In S7, L consist Calculated according to the following formula:

[0148] ,

[0149] ,

[0150] Where, is the square of the L2 norm, for the b-th sample in the mixed domain, is the probability prediction graph output by the student model, Pseudo labels generated for the teacher model, is the consistency loss of the b-th sample.

[0151] Example 2: See Figures 1 to 7Based on Example 1, a more specific three-dimensional encoder structure is given as follows: Figure 3 As shown, a more specific three-dimensional decoder structure is as follows Figure 4 Of course, the structure of the 3D encoder and the 3D decoder is not limited to the structure in this embodiment, and can encode and decode 3D samples.

[0152] In this embodiment, the three-dimensional encoder includes a fixed block embedding module, a first multi-scale sampling module, and a second multi-scale sampling module connected in sequence;

[0153] The fixed-block embedding module includes a three-dimensional convolutional layer and a three-dimensional position encoding layer;

[0154] The three-dimensional convolution layer is used to perform three-dimensional convolution on the input sample to reduce the dimension and obtain the reduced dimension feature F cn , where the convolution kernel of the three-dimensional convolution is 3×3×3 and the stride is 2;

[0155] The three-dimensional position encoding layer is used to cn Generate a three-dimensional sinusoidal position code for each position and obtain the position code feature F p , and then F cn and F p The embedded feature F0 is obtained by adding each element, and the scale of the sample is D×H×W×C in , F cn 、F p , the scale of F0 is ;

[0156] The first multi-scale sampling module includes a global attention layer, a local convolution layer, a layer normalization layer and a three-dimensional downsampling layer;

[0157] The multi-scale sampling module is used to perform global attention on the input features to obtain the global feature F global ;

[0158] The local convolution layer is used to global Perform depth-separable 3D convolution to obtain local features F local , the convolution kernel is 3×3×3, and the stride is 1;

[0159] The layer normalization layer is used according to the formula F mix =LayerNorm(F global +F local ) Get the mixed feature F mix ;

[0160] The three-dimensional downsampling layer is used to mix Perform three-dimensional convolution downsampling and obtain The output feature F d1 ;

[0161] The structure of the second multi-scale sampling module is the same as that of the first multi-scale sampling module, and the input feature is F d1 , the output scale is The output feature F d2 , and F d2 is the encoding feature of the three-dimensional encoder.

[0162] The working principle of the 3D encoder is as follows: in the downsampling process of the 3D encoder, global attention and local convolution are used alternately to extract features and then downsample, where global attention obtains the global feature F global , then F global Perform depth-separable 3D convolution to obtain local features F local , and then fused with F global and F local The fused features reduce the computational complexity while taking into account long-range dependencies and local details. Downsampling reduces the resolution and increases the number of channels. In this invention, the input sample size is D×H×W×C in , after the first multi-scale sampling module, the output feature F d1 , the scale is , after the second multi-scale sampling module, the output feature F d2 , the scale is ,The output of the second multi-scale sampling module is also used as the encoding feature,for the fundamental deep representation.

[0163] In this embodiment, the 3D decoder includes a first upsampling module, a second upsampling module, a third upsampling module and a boundary enhancement layer connected in sequence;

[0164] The first upsampling module includes a transposed convolution upsampling layer, a cross-layer splicing layer, a hole residual module, and a residual connection activation layer;

[0165] The transposed convolution upsampling layer is used to perform a transposed three-dimensional convolution on the encoded features to obtain the first feature F u1 , the convolution kernel of the transposed three-dimensional convolution is 2×2×2, and the stride is 2;

[0166] The cross-layer splicing layer is used to convert F u1 The second feature F is obtained by splicing it with the cross-layer feature along the channel dimension u2 , where the cross-layer feature is F d1 ;

[0167] The hole residual module is used to calculate F u2 Perform dilation convolution with dilation rates of 2 and 4, and then add them together to obtain the third feature F u3 ;

[0168] The residual connection activation layer is used to convert Fu2 and F u3 Perform residual connection and activate it with GELU function to get the output F of the first upsampling module uout1 ;

[0169] Duplicate the first upsampling module and add a channel compression layer at the input end to compress the input feature channel by half before feeding it into the transposed convolution upsampling layer. Modify the cross-layer feature to F0 and the expansion rate of the void residual module to 1 or 2 to obtain the second upsampling module.

[0170] Copy the second upsampling module and modify the cross-layer features into feature F 01 , the hole residual module only performs hole convolution with a dilation rate of 1 on the input features, and obtains the third upsampling module, feature F 01 Obtained by upsampling the F0 transposed convolution, the convolution kernel is 2×2×2, and the stride is 2;

[0171] The classifier is used to uout3 Classify by position and get the initial predicted category probability P for each position base ;

[0172] The boundary enhancement layer F uout3 Use the 3D Sobel operator to detect edges and output the boundary response map P Sobel , then according to the formula , generate a probability prediction map , where λ Sobel P Sobel The weight of .

[0173] The 3D decoder works by performing progressive upsampling, followed by edge enhancement, using the first to third upsampling modules. Each upsampling module involves cross-layer splicing and dilated residuals. Cross-layer splicing doubles the number of channels while preserving fine-grained structural features; dilated residuals effectively expand the receptive field of features while preserving local details, thereby enhancing the features.

[0174] In the first upsampling module, the input encoding feature is , first transpose the 3D convolution upsampling and adjust the channel to 128, the obtained F u1 The scale is , and then with the scale F d1 Splicing along the channel dimension, the obtained F u2 The scale is , and then pass through the void residual module to get F u3 , use the residual connection activation layer to process F u2 and F u3 , and get F uout1, the scale is .

[0175] In the second upsampling module, first F uout1 Channel compression is performed to obtain a scale of The features are obtained by transposing the 3D convolution upsampling. , and then with the scale The cross-layer feature F0 is spliced along the channel dimension to obtain a scale of The features of the second upsampling module are obtained after activation by the hole residual and residual connection. uout2 , the scale is .

[0176] In the third upsampling module, first uout2 Channel compression is performed to obtain a scale of The features are obtained by transposing the 3D convolution upsampling. , and then with the scale Cross-layer features F 01 Splicing along the channel dimension, the scale is The features of the third upsampling module are obtained after activation by the hole residual and residual connection. uout3 , the scale is .

[0177] Through the three-up sampling module, it has the characteristics of progressive resolution recovery, symmetric jump connection and multi-scale feature fusion. Then the classifier is used to obtain P base , for F uout3 Boundary enhancement to obtain P Sobel , and then according to the formula Calculate probability prediction graph ,λ Sobel It is usually set to 0.5. The 3D decoder restores the resolution through multi-level upsampling, fuses the context through dilated convolution, and optimizes the details through boundary enhancement, ultimately obtaining a high-precision probability prediction map.

[0178] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A complex environment adaptive training method for a 3D ground penetrating radar underground disease recognition model, characterized in that: The following steps are included: S1, obtain several samples with category annotations to form the source domain dataset D s , several unlabeled samples constitute the target domain dataset D t , the sample is three-dimensional ground penetrating radar data, and the category is disease category; D s Sample X s The category labeling diagram is P s , D t One of the samples is X t , X s ∈R D×H×W×Cin 、X t ∈R D×H×W×Cin 、P s ∈R D ×H×W×C , D, H, W are the depth, length and width of the sample resolution respectively, Cin is the number of channels of radar reflection intensity, Cin=1, C is the total number of disease categories, P s The value at position (d,h,w) is X s The true disease category at the middle position (d, h, w), d, h, w are the depth, length and width of the position (d, h, w) respectively; S2, constructing a 3D encoding and decoding network, including a 3D encoder, a 3D decoder, and a self-supervision module; The three-dimensional encoder is used to encode the input sample to obtain the encoding feature, and the three-dimensional decoder is used to decode the encoding feature and then perform boundary enhancement, and output the predicted category probability of each position to form a probability prediction map , the self-supervisory module is used to construct positive samples and negative samples according to the encoding features of a batch of samples and calculate the InfoNCE loss; S3, use D S To minimize the total loss L of source domain training s Train the 3D encoding and decoding network to obtain the basic model, L s =λ cn1 L cn1 +λ cls L cls , L cn1 is the InfoNCE loss, L cls is the classification loss, λ cn1 ,λ cls L cn1 and L cls The weight of S4, copy two base models and use them as teacher model and student model respectively; S5, use D t Training the student model, each iteration optimizing the student model based on self-supervised contrastive learning and double-screened pseudo-label supervised learning, wherein one optimization includes steps S51 to S54; S51, use the teacher model to output D t The probability prediction map of each sample in is used as a pseudo label, X t The pseudo label is , preset quality threshold λ and uncertainty threshold ϵ; S52, from D t Choose a batch of samples from the set, where the bth sample X b , use the student model to output the probability prediction map of each sample, and use the self-supervised module to calculate the InfoNCE loss; S53, calculation of X based on double screening b The pseudo-label supervision loss L p,b ; To X b Pseudo labels , based on the local phase quantization calculation position (d, h, w) LPQ value LPQ (d, h, w), for X b , calculate the uncertainty U(d,h,w) of the position (d,h,w) based on MC Dropout, and form X with the position where LPQ(d,h,w)>λ and U(d,h,w)<ϵ b Areas of interest , calculate L p,b ; ， Where, for Total number of positions within is the normalized value of U(d,h,w), For X b Probability prediction chart of The value at position (d,h,w); S54, calculate the total loss L of the target domain t , and to minimize L t Adjust student model parameters; ， Where B is the batch size, L cn2 is the InfoNCE loss of the self-supervised module in the student model, λ cn2 For L cn2 The weight of S6, use D s and D t Construct multiple batches of mixed-domain datasets; S7, train the student model with the mixed domain dataset to minimize the joint loss L total Update the student model and update the teacher model using EMA. total =+L s +λ t L t +λ c L consist , L consist is the consistency loss of a batch of samples.

2. The complex environment adaptive training method of the 3D ground penetrating radar underground disease recognition model according to claim 1 is characterized in that: The damage categories include background, voids, voids, loose bodies, water-rich bodies and crack networks.

3. The complex environment adaptive training method of the 3D ground penetrating radar underground disease recognition model according to claim 1 is characterized in that: The method for calculating InfoNCE loss in the self-supervised module is: Get the encoding features of a batch of samples, marked as F1~F B , calculate the InfoNCE loss of the b-th encoded feature and the InfoNCE loss L of a batch of samples cn1 ; ， ， Where, For F b Randomly add data enhancement to get positive samples, F b The vth negative sample, 1≤v≤2B-2, sim(⋅) is the cosine similarity, τ is the temperature coefficient, and exp(⋅) is the exp function.

4. The complex environment adaptive training method of the 3D ground penetrating radar underground disease recognition model according to claim 1 is characterized in that: In S3, the classification loss is calculated according to the following formula; ， Where, α c is the weight of the c-th disease category, 1≤c≤C, γ is the modulation factor to reduce the weight of easy-to-classify samples, for the position (d, h, w) of the b-th sample, 、 are the normalized value of the predicted category probability at the location and the actual disease category label, respectively.

5. The complex environment adaptive training method of the 3D ground penetrating radar underground disease recognition model according to claim 1 is characterized in that: In S53, LPQ (d, h, w) is calculated according to the following formula: ， Where K is the neighborhood size, i, j, and k are the offsets of the position (d, h, w) in depth, length, and width, respectively, and Q(d+i,h+j,w+k) is the pseudo label The value at position (d+i,h+j,w+k).

6. The complex environment adaptive training method of the 3D ground penetrating radar underground disease recognition model according to claim 1 is characterized in that: The method for calculating U(d,h,w) in S53 is: Use MC Dropout method to t Perform M forward propagations, each generating a probability prediction map. Calculate the uncertainty U(d, h, w) at the position (d, h, w) according to the following formula: ， Where, The probability prediction graph obtained by the mth forward propagation The value at position (d,h,w), is the mean of the M probability prediction maps at position (d, h, w).

7. The complex environment adaptive training method of the 3D ground penetrating radar underground disease recognition model according to claim 1 is characterized in that: In S6, the method of constructing a batch mixed domain dataset is: s and D t After mixing, select any batch of samples, or select from D according to the mixing weight. s and D t A batch of samples is selected, and the mixing weight is preset or obtained according to steps S61 to S63; S61, from D s 、D t Each batch of samples in constitutes a set 、 ; S62, calculation The LPQ value and uncertainty of each sample at each position in the batch are averaged to obtain the batch average local quality score. , all uncertainties are averaged to get the batch average uncertainty ; S63, calculation The mixed weight ; ， α and β are and The adjustment coefficient, Sigmoid(⋅) is the Sigmoid function; according to , generate a batch of mixed domain datasets .

8. The complex environment adaptive training method of the 3D ground penetrating radar underground disease recognition model according to claim 1 is characterized in that: In S7, L consist Calculated according to the following formula: ， ， Where, is the square of the L2 norm, for the b-th sample in the mixed domain, is the probability prediction graph output by the student model, Pseudo labels generated for the teacher model, is the consistency loss of the b-th sample.

9. The complex environment adaptive training method of the 3D ground penetrating radar underground disease recognition model according to claim 1 is characterized in that: In the 3D encoding and decoding network, the 3D encoder includes a fixed block embedding module, a first multi-scale sampling module, and a second multi-scale sampling module connected in sequence; The fixed-block embedding module includes a three-dimensional convolutional layer and a three-dimensional position encoding layer; The three-dimensional convolution layer is used to perform three-dimensional convolution on the input sample to reduce the dimension and obtain the reduced dimension feature F cn , where the convolution kernel of the three-dimensional convolution is 3×3×3 and the stride is 2; The three-dimensional position encoding layer is used to cn Generate a three-dimensional sinusoidal position code for each position and obtain the position code feature F p , and then F cn and F p The embedded feature F0 is obtained by adding each element, and the scale of the sample is D×H×W×C in , F cn 、F p , the scale of F0 is ; The first multi-scale sampling module includes a global attention layer, a local convolution layer, a layer normalization layer and a three-dimensional downsampling layer; The multi-scale sampling module is used to perform global attention on the input features to obtain the global feature F global ; The local convolution layer is used to global Perform depth-separable 3D convolution to obtain local features F local , the convolution kernel is 3×3×3, and the stride is 1; The layer normalization layer is used according to the formula F mix =LayerNorm(F global +F local ) Get the mixed feature F mix ; The three-dimensional downsampling layer is used to mix Perform three-dimensional convolution downsampling and obtain The output feature F d1 ; The structure of the second multi-scale sampling module is the same as that of the first multi-scale sampling module, and the input feature is F d1 , the output scale is The output feature F d2 , and F d2 is the encoding feature of the three-dimensional encoder.

10. The complex environment adaptive training method of the 3D ground penetrating radar underground disease recognition model according to claim 9 is characterized in that: In the 3D encoding and decoding network, the 3D decoder includes a first upsampling module, a second upsampling module, a third upsampling module and a boundary enhancement layer connected in sequence; The first upsampling module includes a transposed convolution upsampling layer, a cross-layer splicing layer, a hole residual module, and a residual connection activation layer; The transposed convolution upsampling layer is used to perform a transposed three-dimensional convolution on the encoded features to obtain the first feature F u1 , the convolution kernel of the transposed three-dimensional convolution is 2×2×2, and the stride is 2; The cross-layer splicing layer is used to convert F u1 The second feature F is obtained by splicing it with the cross-layer feature along the channel dimension u2 , where the cross-layer feature is F d1 ; The hole residual module is used to calculate F u2 Perform dilation convolution with dilation rates of 2 and 4, and then add them together to obtain the third feature F u3 ; The residual connection activation layer is used to convert F u2 and F u3 Perform residual connection and activate it with GELU function to get the output F of the first upsampling module uout1 ; Duplicate the first upsampling module and add a channel compression layer at the input end to compress the input feature channel by half before feeding it into the transposed convolution upsampling layer. Modify the cross-layer feature to F0 and the expansion rate of the void residual module to 1 or 2 to obtain the second upsampling module. Copy the second upsampling module and modify the cross-layer features into feature F 01 , the hole residual module only performs hole convolution with a dilation rate of 1 on the input features, and obtains the third upsampling module, feature F 01 Obtained by upsampling the F0 transposed convolution, the convolution kernel is 2×2×2, and the stride is 2; The classifier is used to uout3 Classify by position and get the initial predicted category probability P for each position base ; The boundary enhancement layer F uout3 Use the 3D Sobel operator to detect edges and output the boundary response map P Sobel , then according to the formula , generate a probability prediction map , where λ Sobel P Sobel The weight of .

Images