Nuclear magnetic resonance-based method and system for evaluating freeze-thaw damage of concrete

Abstract

The present application relates to the technical field of concrete durability detection, and particularly relates to a concrete freeze-thaw damage evaluation method and system based on nuclear magnetic resonance, which comprises the following steps: obtaining a nuclear magnetic resonance transverse relaxation time spectrum of a concrete test piece, and recording the mix proportion parameters and the number of freeze-thaw cycles; performing feature extraction on the spectrum to obtain a spectrum feature vector; constructing a concrete freeze-thaw knowledge graph; performing comparative learning alignment on the spectrum feature vector and the entity embedding in the knowledge graph to establish a joint embedding space; constructing a graph structure with each test piece as a node and the spectrum feature similarity and the mix proportion parameter encoding vector similarity as edge weights, and performing message passing by using a graph neural network; performing semi-supervised training by using a small amount of labeled samples and a large amount of unlabeled samples to obtain a damage grade prediction model; collecting the spectrum and the mix proportion parameters of the concrete to be measured, and inputting the model to output the freeze-thaw damage grade.

Description

Technical Field

[0001] This invention relates to the field of concrete durability testing technology, specifically to a method and system for assessing concrete freeze-thaw damage based on nuclear magnetic resonance. Background Technology

[0002] Freeze-thaw damage to concrete is one of the major structural defects in hydraulic structures, roads, and bridges in cold regions. Under freeze-thaw cycles, the pore water inside the concrete repeatedly freezes and thaws, generating expansion and osmotic pressures, leading to deterioration of the pore structure, expansion of microcracks, and ultimately, a decrease in compressive strength and structural failure. Accurately assessing the degree of freeze-thaw damage to concrete structures is of great significance for safe operation, maintenance, and repair decisions.

[0003] Nuclear magnetic resonance (NMR) technology, as a non-destructive testing method, measures the transverse relaxation time of hydrogen protons in the pore water inside concrete. NMR spectrum can characterize the distribution of pore structure and moisture state, and has been gradually applied to freeze-thaw damage detection. However, existing freeze-thaw damage assessment methods based on NMR have the following shortcomings: (1) The interpretation of NMR signals depends on a large number of labeled samples (i.e., samples with both NMR data and measured compressive strength values), and obtaining such labels requires destructive testing, which is costly and time-consuming, making it difficult to obtain sufficient training data; (2) The concrete mix proportion (water-cement ratio, aggregate type, admixtures, etc.) has a significant impact on the NMR response characteristics, and the freeze-thaw damage patterns under different mix proportions are different, making it difficult to establish a universal model using traditional methods; (3) Existing methods mostly use manual feature extraction or simple statistical regression, failing to make full use of the rich correlation information contained in historical detection data, resulting in low data utilization.

[0004] Therefore, there is an urgent need for a technical solution that can achieve semi-supervised freeze-thaw damage assessment with a small number of labels. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention aims to provide a method and system for assessing freeze-thaw damage in concrete based on nuclear magnetic resonance, solving the problems of existing assessment methods being highly dependent on label data, difficult to adapt to different mix proportions, and having low utilization of historical data.

[0006] The basic solution provided by this invention is a method for assessing freeze-thaw damage in concrete based on nuclear magnetic resonance, comprising the following steps: Step S1: Obtain the transverse relaxation time of the concrete specimen via nuclear magnetic resonance. The spectrum was analyzed, and the mix proportion parameters and number of freeze-thaw cycles for each specimen were recorded. Step S2: For the Feature extraction is performed on the spectrum to obtain Spectral eigenvectors; Step S3: Construct a concrete freeze-thaw knowledge graph, wherein the nodes of the knowledge graph include mix proportion entities, freeze-thaw cycle entities, etc. Spectral feature entities and damage level entities, with nodes connected by associative edges; Step S4: Place the... The spectral feature vectors are compared and aligned with the entity embeddings in the knowledge graph to establish a joint embedding space between nuclear magnetic resonance signal features and knowledge graph entities. Step S5: Using each concrete specimen as a node, with The similarity of spectral features and the similarity of the coding vector of the commensurate ratio parameter are used to construct a graph structure with edge weights, and a graph neural network is used for message passing. The coding vector of the commensurate ratio parameter is obtained by one-hot encoding of the commensurate ratio parameter. Step S6: Divide all specimens collected in step S1 into labeled sample sets and unlabeled sample sets. The damage level labels of the labeled sample sets are determined by compressive strength tests. Use the labeled sample sets and unlabeled sample sets to perform semi-supervised training on the graph neural network to obtain a damage level prediction model. Step S7: Acquire nuclear magnetic resonance (NMR) data of the concrete to be tested. The spectrum and mixing ratio parameters are obtained by encoding the mixing ratio parameters using one-heat coding to obtain the mixing ratio parameter encoding vector, and then compared with the... The spectral feature vector is input into the damage level prediction model, and the freeze-thaw damage level is output.

[0007] Furthermore, the steps described in step S2 Spectral feature extraction includes: After preprocessing the spectrum by smoothing filtering, baseline correction, and amplitude normalization, multi-peak Gaussian fitting is used to obtain the peak position and peak area ratio in the relaxation time distribution, and the results are calculated. The geometric mean of the spectrum, where the number of peaks is determined based on the typical components of the concrete pore structure, is formed by combining the above parameters. Spectral eigenvectors.

[0008] Furthermore, the method for constructing the knowledge graph in step S3 includes: discretizing continuous parameters into symbolic nodes according to a preset threshold interval; the mix proportion entity includes at least water-cement ratio, aggregate type, air content, and mineral admixture content; the freeze-thaw cycle entity is divided according to the cumulative interval of freeze-thaw cycle counts; the... Spectral feature entities The damage level entities are divided by a combination of the geometric mean interval and the peak area ratio interval; the damage level entities are divided according to the compressive strength loss rate threshold interval; the correlation between the entities is obtained through statistical analysis of historical test data.

[0009] Furthermore, the contrastive learning alignment method described in step S4 includes: constructing a nuclear magnetic resonance encoder to... The spectral feature vector is mapped to the nuclear magnetic resonance embedding vector. A knowledge graph encoder is constructed to aggregate the knowledge graph entities corresponding to the specimens into knowledge graph embedding vectors. Positive sample pairs are formed by combining the nuclear magnetic resonance embedding vectors of the same specimens with the knowledge graph embedding vectors, and negative sample pairs are formed by combining different specimens or random combinations. By minimizing the contrastive loss function, positive sample pairs are made closer to each other in the embedding space, while negative sample pairs are made farther apart.

[0010] Furthermore, the contrastive loss function for:

[0011] in, For the set of positive sample pairs, For the set of negative sample pairs, The preset positive margin parameter, For nuclear magnetic resonance embedding vectors, Embed vectors into knowledge graphs. This represents the L2 norm.

[0012] Furthermore, the method for constructing the graph structure in step S5 includes: initial node features are derived from... The spectral feature vector is concatenated with the matching ratio parameter encoding vector; the edge weights between nodes are based on... The cosine distance between spectral feature vectors and the Euclidean distance between the coordinate ratio parameter encoding vectors are weighted and calculated. The K neighbors with the largest edge weights in the neighborhood of each node are retained to form a sparse graph, where K is a preset positive integer.

[0013] Furthermore, the formula for calculating the edge weight is as follows:

[0014] in, Cosine distance For Euclidean distance, and These are the weighting coefficients.

[0015] Furthermore, the graph neural network described in step S5 employs a graph convolutional propagation mechanism, and the node feature update formula for each layer is as follows:

[0016] in, For nodes The neighborhood group, For the first The learnable weight matrix of the layer, It is a non-linear activation function.

[0017] Furthermore, the loss function for semi-supervised training in step S6 includes: cross-entropy classification loss for labeled samples, graph Laplacian regularization loss for all nodes, and L2 regularization loss for network parameters. The graph Laplacian regularization loss is defined as the weighted sum of squares of the differences between the hidden layer features of adjacent nodes.

[0018] The principle of this invention lies in achieving accurate assessment of concrete freeze-thaw damage under minimal label conditions through the collaborative work of a nuclear magnetic resonance data acquisition module, a data processing module, a knowledge graph storage module, a joint embedding module, a graph neural network inference module, and an output module. Specifically, the following is a detailed explanation: The data processing module first performs smoothing filtering, baseline correction, and amplitude normalization preprocessing on the acquired T2 spectrum. Then, it uses multi-peak Gaussian fitting to extract the peak position and peak area ratio in the relaxation time distribution and calculates... The geometric mean of the spectrum, which combines the above parameters to form Spectral Feature Vectors. This module transforms the original relaxed signal into structured numerical features, providing quantitative input for subsequent knowledge graph construction and model inference.

[0019] The knowledge graph storage module constructs a concrete freeze-thaw knowledge graph, discretizing continuous parameters into symbolic nodes according to preset threshold intervals. Node types include mix proportion entities (water-cement ratio, aggregate type, air content, and mineral admixture dosage) and freeze-thaw cycle entities (divided according to the cumulative interval of freeze-thaw cycle counts). Spectral feature entities (by) The module categorizes data into a structured knowledge network, using a combination of geometric mean intervals and peak area ratio intervals, and damage level entities (categorized by compressive strength loss rate threshold intervals). The relationships between these entities are obtained through statistical analysis of historical test data.

[0020] The joint embedding module will The module performs comparative learning and alignment between spectral feature vectors and entity embeddings in the knowledge graph, establishing a joint embedding space between NMR signal features and knowledge graph entities. This module achieves this through collaborative training of the NMR encoder and the knowledge graph encoder. The spectral signal can be directly correlated with the corresponding mix ratio, freeze-thaw conditions and damage level in the spectrum, achieving semantic alignment between signal features and structured knowledge.

[0021] The graph neural network inference module uses each concrete specimen as a node and... The similarity of spectral features and the similarity of the matching ratio parameter encoding vectors are used to construct a graph structure for edge weights, and a graph convolutional network is used for message passing. This module is based on a semi-supervised learning strategy, using supervision information from a small number of labeled samples combined with regularization constraints on the graph structure to propagate label information from labeled nodes to unlabeled nodes along the graph edges. The loss function includes the cross-entropy classification loss of labeled samples, the graph Laplacian regularization loss of all nodes, and the L2 regularization loss of the network parameters, enabling the model to learn an effective classification boundary even with a small number of labels.

[0022] The beneficial effects of this invention are as follows: 1. At the label dependency level, this invention utilizes a semi-supervised learning mechanism of graph neural networks, requiring only a small number of labeled samples to train a high-precision evaluation model. Compared to existing technologies that rely on numerous destructive experiments to obtain labels, this invention fully leverages the structural information of a large amount of unlabeled historical NMR data. Through graph Laplacian regularization, label information is propagated from labeled nodes along graph edges to unlabeled nodes, reducing the dependence on the number of labeled samples and lowering the cost of destructive experiments.

[0023] 2. At the level of cross-mixing ratio generalization, this invention uses a knowledge graph storage module to structurally store mixing ratio parameters in entity form, and uses a joint embedding module to... The spectral feature vectors and the entity embeddings of the mixing ratios are aligned in the joint embedding space, enabling the model to distinguish between cases of "different mixing ratios but similar signals" and "same mixing ratios but different damage." Compared to existing methods that establish independent models for specific mixing ratios, this invention achieves cross-sample transfer of mixing ratio knowledge and supports cold-start evaluation of new mixing ratios.

[0024] 3. Regarding data utilization, this invention explicitly organizes and stores the relationships between material parameters, freeze-thaw conditions, and signal features contained in historical experimental data through a knowledge graph. Furthermore, it incorporates unlabeled samples into the training process using a graph neural network in a graph structure, enabling the model to learn from both labeled and unlabeled samples simultaneously. Compared to existing technologies that rely solely on labeled samples for independent modeling, this invention significantly improves the utilization rate of historical testing data.

[0025] Accordingly, the present invention also discloses a concrete freeze-thaw damage assessment system based on nuclear magnetic resonance, used to implement the above method, comprising: The nuclear magnetic resonance data acquisition module is used to obtain the transverse relaxation time of concrete specimens or test components. Spectrum; The data processing module is used to process the data. The spectrum is preprocessed and features are extracted to obtain Spectral eigenvectors; The knowledge graph storage module is used to store a concrete freeze-thaw knowledge graph. The nodes of the knowledge graph include entities for mix proportions and freeze-thaw cycles. Spectral characteristic entities and damage level entities; The joint embedding module is used to embed the The spectral feature vectors are compared and aligned with the entity embeddings in the knowledge graph to establish a joint embedding space between nuclear magnetic resonance signal features and knowledge graph entities. The graph neural network inference module is used to infer information from each concrete specimen as a node. The similarity of spectral features and the similarity of the encoding vectors of the combination ratio parameter are used to construct a graph structure with edge weights, and semi-supervised training is performed based on labeled and unlabeled sample sets to obtain a damage level prediction model. Output module, used to analyze the nuclear magnetic resonance of the concrete to be tested. The spectrum and mix proportion parameter encoding vector are used to output the freeze-thaw damage level through the damage level prediction model. Attached Figure Description

[0026] Figure 1 This is a flowchart of an embodiment of the concrete freeze-thaw damage assessment method based on nuclear magnetic resonance according to the present invention.

[0027] Figure 2 This is a system module diagram of an embodiment of the concrete freeze-thaw damage assessment system based on nuclear magnetic resonance of the present invention. Detailed Implementation

[0028] The following detailed description illustrates the specific implementation method: The basic implementation examples are as follows: Figure 1 As shown: This embodiment uses the freeze-thaw cycle test of standard concrete specimens as an application scenario to illustrate a method for assessing freeze-thaw damage in concrete based on nuclear magnetic resonance, including the following steps: Step S1: Obtain the transverse relaxation time of the concrete specimen via nuclear magnetic resonance. The spectra were analyzed, and the mixing ratio parameters and freeze-thaw cycle number of each specimen were recorded.

[0029] In this embodiment, three different water-cement ratios (e.g., 0.40, 0.50, and 0.60) were selected for concrete mix design, and multiple 100mm cube specimens were prepared for each mix design. The specimens were grouped and subjected to 0, 25, 50, 100, 150, and 200 freeze-thaw cycles, respectively. The temperature range and duration of the freeze-thaw cycles were performed according to GB / T50082 standard. At each freeze-thaw cycle node, the specimens were removed for nuclear magnetic resonance (NMR) testing, and the mix design parameters and the number of freeze-thaw cycles for each specimen were recorded.

[0030] NMR detection was performed using a low-field NMR spectrometer configured with a CPMG pulse sequence. The echo time TE was set to 0.15 ms, the number of echoes NECH to 2048, the number of accumulations NS to 64, and the repetition time TR to 3000 ms to ensure accurate acquisition. The spectral signal-to-noise ratio meets the requirements.

[0031] After nuclear magnetic resonance (NMR) testing, each specimen underwent a compressive strength test. The damage level was determined based on the compressive strength loss rate: minor damage corresponds to a compressive strength loss rate of <15%, moderate damage corresponds to 15%~30%, severe damage corresponds to 30%~50%, and dangerous state corresponds to ≥50%. This serves as a source of labels for a small number of labeled samples in subsequent semi-supervised training.

[0032] Step S2: For the Feature extraction is performed on the spectrum to obtain Spectral eigenvectors.

[0033] In this embodiment, Spectral feature extraction includes the following sub-steps: (1) Preprocessing: Preprocessing the collected data The spectrum undergoes smoothing filtering, baseline correction, and amplitude normalization preprocessing. Smoothing filtering uses a Savitzky-Golay filter (window size 11, polynomial order 3) to remove high-frequency noise; baseline correction... The mean of the signal at the tail end of the spectrum (>1000ms) was used as the baseline, and the baseline was subtracted; amplitude normalization was performed on all specimens. The maximum amplitude of the spectrum is normalized to 1.

[0034] (2) Multi-peak Gaussian fitting: The multi-peak Gaussian fitting method is used to fit the data. The spectrum is decomposed. The fitting function is:

[0035] in, The signal amplitude, The horizontal relaxation time is expressed in milliseconds (ms). The preset number of peaks, For the first The amplitude of each peak For the first Logarithmic peak position of each peak For the first The width parameter of each peak.

[0036] In this embodiment, the number of peaks is set. (Corresponding to gel pores, capillary pores, and macropores, respectively). Nonlinear least squares method was used for fitting, with initial peak positions set at 0.5ms, 30ms, and 200ms, allowing peak position drift within a certain range.

[0037] (3) Geometric mean calculation and feature composition: After fitting, extract the peak position of each peak. Peak area ratio and calculate Geometric mean of the spectrum:

[0038] in, This is the index of the discrete relaxation time point. relaxation time The corresponding signal amplitude.

[0039] The above parameters (peak position, peak area ratio, geometric mean) are combined to form Spectral eigenvector. For assessing the degree of freeze-thaw damage to concrete, the change in gel pore peak position reflects the degree of deterioration of the microporous structure, the proportion of capillary pore peak area reflects the impact of freeze-thaw on the distribution of freezeable water, and the geometric mean reflects the evolution of the overall pore structure. Therefore, this eigenvector can comprehensively characterize the changes in pore structure caused by freeze-thaw damage.

[0040] Step S3: Construct a concrete freeze-thaw knowledge graph, wherein the nodes of the knowledge graph include mix proportion entities, freeze-thaw cycle entities, etc. Spectral feature entities and damage level entities are connected by associative edges.

[0041] In this embodiment, the knowledge graph is constructed using a graph database for storage. Continuous parameters are discretized into symbolic nodes according to a preset threshold interval: (1) Mix proportion entity: The water-cement ratio is discretized into nodes at intervals of 0.1 (0.40, 0.50, 0.60), the aggregate type is coded by category (e.g., granite, limestone, basalt), the fly ash content is discretized into nodes by grade (0%, 15%, 30%), and the air content is discretized by grade; (2) Freeze-thaw cycle count entities: The cumulative interval of the freeze-thaw cycle count is divided into five intervals: [0,25), [25,50), [50,100), [100,200), and [200,∞). Each interval corresponds to an entity node. (3) Spectral feature entities: The geometric mean is divided into five intervals (<10ms, 10~30ms, 30~100ms, 100~300ms, >300ms), and the maximum peak area ratio is divided into five intervals ([0,0.2), [0.2,0.4), [0.4,0.6), [0.6,0.8), [0.8,1.0]). A maximum of 25 combinations of these two values ​​can be formed. Spectral feature entities; (4) Damage level entities: Based on the threshold range of compressive strength loss rate, entities are divided into four categories: minor damage (<15%), moderate damage (15%~30%), severe damage (30%~50%), and dangerous state (≥50%).

[0042] The relationships between the various entities were obtained through statistical analysis of historical test data: for each standard specimen, its mix proportion entity, freeze-thaw cycles entity, and Establish association edges between entities with spectral characteristics and entities with damage levels; at the same time, based on prior knowledge of materials science, establish similarity association edges between entities with similar mix proportions, such as establishing a similarity relationship between water-cement ratios of 0.40 and 0.45.

[0043] The knowledge graph embedding uses the TransE model, which maps each entity to a low-dimensional vector (dimension 50), so that the semantic relationships between entities are approximately preserved through vector operations (e.g., "water-cement ratio 0.50" corresponds to "moderate damage" in relation to "100 freeze-thaw cycles"). When updating the graph, the embedding of a new entity is initialized by constraints on its association with existing entities.

[0044] Step S4: Place the... The spectral feature vectors are compared and aligned with the entity embeddings in the knowledge graph to establish a joint embedding space between nuclear magnetic resonance signal features and knowledge graph entities.

[0045] In this embodiment, two encoders are constructed: Nuclear magnetic resonance encoder A 2-layer MLP is used, with the input dimensions corresponding to... Spectral feature vector, hidden layer dimension 64, output dimension 32, Spectral eigenvectors are mapped to nuclear magnetic resonance embedding vectors .

[0046] Knowledge Graph Encoder The average of the knowledge graph entity embedding vector sets corresponding to the specimens (including the embeddings of mix ratio entities, freeze-thaw cycles entities, and damage level entities, all with a dimension of 50) is then applied through a 2-layer MLP mapping to a 32-dimensional array to obtain the knowledge graph embedding vectors. .

[0047] For each labeled specimen, its NMR embedding vector and the knowledge graph embedding vector corresponding to that specimen are used to form a positive sample pair; the NMR embedding vector and the knowledge graph embedding vector of other randomly sampled specimens (or a randomly shuffled combination of entities) are used to form a negative sample pair.

[0048] The goal of contrastive learning is to narrow the distance between positive sample pairs in the embedding space while widening the distance between negative sample pairs. The contrastive loss function is defined as:

[0049] in, For the set of positive sample pairs, For the set of negative sample pairs, The preset positive margin parameter, For nuclear magnetic resonance embedding vectors, Embed vectors into knowledge graphs. This represents the L2 norm.

[0050] By minimizing this loss function, the NMR encoder is trained so that the NMR features and their corresponding knowledge graph entities are close to each other in the embedding space. The Adam optimizer is used with a learning rate of 0.001 and 200 iterations.

[0051] Taking a training sample in this embodiment as an example, a certain specimen's The spectral feature vector is obtained by the nuclear magnetic resonance encoder. The mix proportion entity "water-cement ratio 0.50", freeze-thaw cycle entity "100 times", and damage level entity "moderate" corresponding to this specimen were aggregated by a knowledge graph encoder. The two constitute a positive sample pair that is brought closer together; at the same time, this Compared with another randomly sampled specimen with a water-cement ratio of 0.40, 50 freeze-thaw cycles, and minor damage... Negative sample pairs are pushed away. Through comparative training with a large number of sample pairs, the NMR encoder learns to... The spectral signal is mapped to an embedding space that is aligned with the semantics of the knowledge graph.

[0052] Step S5: Using each concrete specimen as a node, with The similarity of spectral features and the similarity of the coding vector of the proportion parameter are used to construct a graph structure with edge weights, and a graph neural network is used for message passing. The coding vector of the proportion parameter is obtained by one-hot encoding of the proportion parameter.

[0053] In this embodiment, all specimens (e.g., 72) are considered as nodes. Initial feature vector of each node. Depend on Spectral eigenvectors Coordination ratio parameter encoding vector It is pieced together:

[0054] Among them, the mixing ratio parameter encoding vector It is assembled by combining the mix proportion parameters (water-cement ratio, aggregate type, fly ash content, and air content) through unique thermal coding.

[0055] The edge weights between nodes are based on The weighted calculation of the cosine distance between spectral eigenvectors and the Euclidean distance between the coordination ratio parameter encoding vectors is as follows:

[0056] in, Cosine distance For Euclidean distance, and For the weighting coefficient, in this embodiment, we take... It is 0.5. The value is 1.0 (determined via grid search [0.1, 2.0], with the goal of achieving the classification accuracy on the validation set).

[0057] Retain the node with the largest edge weight in its neighborhood. The neighbors form a sparse graph.

[0058] The graph neural network uses a graph attention network (GAT). The first layer has 8 attention heads and outputs 64 dimensions, while the second layer has a single head that outputs 4 dimensions (corresponding to four damage levels). The node feature update formula for each layer is as follows:

[0059] in, For nodes The neighborhood group, For the first The learnable weight matrix of the layer, It is the ReLU nonlinear activation function.

[0060] Step S6: Divide all specimens collected in step S1 into labeled sample sets and unlabeled sample sets. The damage level labels of the labeled sample sets are determined by compressive strength tests. Use the labeled sample sets and unlabeled sample sets to perform semi-supervised training on the graph neural network to obtain a damage level prediction model.

[0061] In this embodiment, all specimens are divided into labeled sample sets. and unlabeled sample set Labeled samples accounted for 10% of the total sample. The damage level labels for the labeled samples were determined through compressive strength testing (e.g., "moderate damage," "severe damage," etc.), and the loss function was defined as:

[0062] The first item is supervised classification loss. Let cross-entropy be the loss function. The nodes output by the graph neural network The probability distribution of predicted damage level, For nodes The first term is the true damage level label; the second term is the graph Laplacian regularization loss, which applies to all nodes. The regularization coefficient is . For nodes In the last hidden layer, the hidden features of adjacent nodes are encouraged to remain smooth, allowing label information to propagate from labeled nodes along the graph edges to unlabeled nodes; the third term is L2 weight decay regularization. This is the weight decay coefficient. These are all trainable parameters for the network.

[0063] The Adam optimizer was used with a learning rate of 0.001 for 500 training epochs. Early stopping was based on the accuracy of the validation set (which was a subset of the labeled samples). After training, the evaluation accuracy on the remaining test samples was over 90%.

[0064] Step S7: Acquire nuclear magnetic resonance (NMR) data of the concrete to be tested. The spectrum and mixing ratio parameters are obtained by encoding the mixing ratio parameters using one-heat coding to obtain the mixing ratio parameter encoding vector, and then compared with the... The spectral feature vector is input into the damage level prediction model, and the freeze-thaw damage level is output.

[0065] In this embodiment, for the concrete component to be tested, its nuclear magnetic resonance imaging (NMR) is first acquired. Spectrum, Extraction The spectral feature vector is obtained; simultaneously, its mix proportion parameters (water-cement ratio, aggregate type, fly ash content, and air content) are acquired, and the mix proportion parameter encoding vector is obtained through unique thermal encoding. The two are concatenated and input into a trained graph neural network model. The model outputs the probability distribution of four damage levels, and the level corresponding to the highest probability is taken as the evaluation result.

[0066] For example, a concrete bridge pier to be tested The spectrum showed that the capillary peak position shifted from 30ms to 45ms, and the geometric mean increased from 25ms to 52ms. The mix proportion was a water-cement ratio of 0.45, granite aggregate, and 15% fly ash. The model output damage level as "moderate damage" with a probability of 87%. Based on this, the engineers determined that the bridge pier had suffered significant freeze-thaw damage and required repair and reinforcement.

[0067] It should be noted that the continuous updating mechanism of the knowledge graph is as follows: With the accumulation of engineering practice, new concrete freeze-thaw test data are continuously obtained. This new data (including NMR features, mix proportions, freeze-thaw cycles, and measured strength) is added to the knowledge graph, creating new entities and relationships. After the graph is updated, the knowledge graph embedding algorithm (TransE) is re-run to obtain the updated entity vectors. Then, the NMR encoder and graph neural network are fine-tuned on the new data to allow the model to absorb the new knowledge.

[0068] Specifically, after new data is added to the knowledge graph, the system automatically identifies the type and relationships of the new entities, sets their initial embedding as a weighted average of the embeddings of their related entities, and then performs global optimization using a knowledge graph embedding algorithm. Subsequently, on a complete dataset containing both new and old data, the joint embedding module and the graph neural network inference module are trained again with a low initial learning rate (e.g., 0.0001), ensuring that new knowledge is smoothly absorbed without destroying existing knowledge. This incremental learning mechanism enables the evaluation model to continuously evolve, adapt to new materials and working conditions, and continuously improve evaluation accuracy.

[0069] Correspondingly, as shown in the appendix Figure 2 As shown, the present invention also provides a concrete freeze-thaw damage assessment system based on nuclear magnetic resonance (NMR) for performing the above-described method, comprising: The nuclear magnetic resonance data acquisition module is used to obtain the transverse relaxation time of concrete specimens or test components. Spectrum; The data processing module is used to process the data. The spectrum is preprocessed and features are extracted to obtain Spectral eigenvectors; The knowledge graph storage module is used to store a concrete freeze-thaw knowledge graph. The nodes of the knowledge graph include entities for mix proportions and freeze-thaw cycles. Spectral characteristic entities and damage level entities; The joint embedding module is used to embed the The spectral feature vectors are compared and aligned with the entity embeddings in the knowledge graph to establish a joint embedding space between nuclear magnetic resonance signal features and knowledge graph entities. The graph neural network inference module is used to infer information from each concrete specimen as a node. The similarity of spectral features and the similarity of the encoding vectors of the combination ratio parameter are used to construct a graph structure with edge weights, and semi-supervised training is performed based on labeled and unlabeled sample sets to obtain a damage level prediction model. Output module, used to analyze the nuclear magnetic resonance of the concrete to be tested. The spectrum and mix proportion parameter encoding vector are used to output the freeze-thaw damage level through the damage level prediction model.

[0070] The above are merely embodiments of the present invention. Commonly known structures and characteristics are not described in detail here. Those skilled in the art are aware of all common technical knowledge in the field prior to the application date or priority date, are aware of all existing technologies in that field, and have the ability to apply conventional experimental methods prior to that date. Those skilled in the art can, under the guidance of this application, improve and implement this solution in combination with their own capabilities. Some typical known structures or methods should not be obstacles for those skilled in the art to implement this application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the structure of the present invention. These should also be considered within the scope of protection of the present invention, and will not affect the effectiveness of the implementation of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. A method for assessing freeze-thaw damage in concrete based on nuclear magnetic resonance, characterized in that, Includes the following steps: Step S1: Obtain the transverse relaxation time of the concrete specimen via nuclear magnetic resonance. The spectrum was analyzed, and the mix proportion parameters and number of freeze-thaw cycles for each specimen were recorded. Step S2: For the Feature extraction is performed on the spectrum to obtain Spectral eigenvectors; Step S3: Construct a concrete freeze-thaw knowledge graph, wherein the nodes of the knowledge graph include mix proportion entities, freeze-thaw cycle entities, etc. Spectral feature entities and damage level entities, with nodes connected by associative edges; Step S4: Place the... The spectral feature vectors are compared and aligned with the entity embeddings in the knowledge graph to establish a joint embedding space between nuclear magnetic resonance signal features and knowledge graph entities. Step S5: Using each concrete specimen as a node, with The similarity of spectral features and the similarity of the coding vector of the commensurate ratio parameter are used to construct a graph structure with edge weights, and a graph neural network is used for message passing. The coding vector of the commensurate ratio parameter is obtained by one-hot encoding of the commensurate ratio parameter. Step S6: Divide all specimens collected in step S1 into labeled sample sets and unlabeled sample sets. The damage level labels of the labeled sample sets are determined by compressive strength tests. Use the labeled sample sets and unlabeled sample sets to perform semi-supervised training on the graph neural network to obtain a damage level prediction model. Step S7: Acquire nuclear magnetic resonance (NMR) data of the concrete to be tested. The spectrum and mixing ratio parameters are obtained by encoding the mixing ratio parameters using one-heat coding to obtain the mixing ratio parameter encoding vector, and then compared with the... The spectral feature vector is input into the damage level prediction model, and the freeze-thaw damage level is output.

2. The method for assessing freeze-thaw damage to concrete based on nuclear magnetic resonance according to claim 1, characterized in that, The steps described in step S2 Spectral feature extraction includes: After preprocessing the spectrum by smoothing filtering, baseline correction, and amplitude normalization, multi-peak Gaussian fitting is used to obtain the peak position and peak area ratio in the relaxation time distribution, and the results are calculated. The geometric mean of the spectrum, where the number of peaks is determined based on the typical components of the concrete pore structure, is formed by combining the above parameters. Spectral eigenvectors.

3. The method for assessing freeze-thaw damage to concrete based on nuclear magnetic resonance according to claim 1, characterized in that, The method for constructing the knowledge graph in step S3 includes: discretizing continuous parameters into symbolic nodes according to a preset threshold interval; the mix proportion entity includes at least water-cement ratio, aggregate type, air content, and mineral admixture content; the freeze-thaw cycle entity is divided according to the cumulative interval of freeze-thaw cycle counts; the... Spectral feature entities The damage level entities are divided by a combination of the geometric mean interval and the peak area ratio interval; the damage level entities are divided according to the compressive strength loss rate threshold interval; the correlation between the entities is obtained through statistical analysis of historical test data.

4. The method for assessing freeze-thaw damage to concrete based on nuclear magnetic resonance according to claim 1, characterized in that, The contrastive learning alignment method described in step S4 includes: constructing a nuclear magnetic resonance encoder to... The spectral feature vector is mapped to the nuclear magnetic resonance embedding vector. A knowledge graph encoder is constructed to aggregate the knowledge graph entities corresponding to the specimens into knowledge graph embedding vectors. Positive sample pairs are formed by combining the nuclear magnetic resonance embedding vectors of the same specimens with the knowledge graph embedding vectors, and negative sample pairs are formed by combining different specimens or random combinations. By minimizing the contrastive loss function, positive sample pairs are made closer to each other in the embedding space, while negative sample pairs are made farther apart.

5. The method for assessing freeze-thaw damage to concrete based on nuclear magnetic resonance according to claim 4, characterized in that, The contrast loss function for: in, For the set of positive sample pairs, For the set of negative sample pairs, The preset positive margin parameter, For nuclear magnetic resonance embedding vectors, Embed vectors into knowledge graphs. This represents the L2 norm.

6. The method for assessing freeze-thaw damage of concrete based on nuclear magnetic resonance according to claim 1, characterized in that, The method for constructing the graph structure in step S5 includes: initial node features are derived from... The spectral feature vector is concatenated with the matching ratio parameter encoding vector; the edge weights between nodes are based on... The cosine distance between spectral feature vectors and the Euclidean distance between the coordinate ratio parameter encoding vectors are weighted and calculated. The K neighbors with the largest edge weights in the neighborhood of each node are retained to form a sparse graph, where K is a preset positive integer.

7. The method for assessing freeze-thaw damage to concrete based on nuclear magnetic resonance according to claim 6, characterized in that, The formula for calculating the edge weight is: in, Cosine distance For Euclidean distance, and These are the weighting coefficients.

8. The method for assessing freeze-thaw damage of concrete based on nuclear magnetic resonance according to claim 1, characterized in that, The graph neural network described in step S5 employs a graph convolutional propagation mechanism, and the node feature update formula for each layer is as follows: in, For nodes The neighborhood group, For the first The learnable weight matrix of the layer, It is a non-linear activation function.

9. The method for assessing freeze-thaw damage of concrete based on nuclear magnetic resonance according to claim 1, characterized in that, The loss function for semi-supervised training in step S6 includes: cross-entropy classification loss for labeled samples, graph Laplacian regularization loss for all nodes, and L2 regularization loss for network parameters. The graph Laplacian regularization loss is defined as the weighted sum of squares of the differences between the hidden layer features of adjacent nodes.

10. A concrete freeze-thaw damage assessment system based on nuclear magnetic resonance, used to implement the method described in any one of claims 1 to 9, characterized in that, include: The nuclear magnetic resonance data acquisition module is used to obtain the transverse relaxation time of concrete specimens or test components. Spectrum; The data processing module is used to process the data. The spectrum is preprocessed and features are extracted to obtain Spectral eigenvectors; The knowledge graph storage module is used to store a concrete freeze-thaw knowledge graph. The nodes of the knowledge graph include entities for mix proportions and freeze-thaw cycles. Spectral characteristic entities and damage level entities; The joint embedding module is used to embed the The spectral feature vectors are compared and aligned with the entity embeddings in the knowledge graph to establish a joint embedding space between nuclear magnetic resonance signal features and knowledge graph entities. The graph neural network inference module is used to infer information from each concrete specimen as a node. The similarity of spectral features and the similarity of the encoding vectors of the combination ratio parameter are used to construct a graph structure with edge weights, and semi-supervised training is performed based on labeled and unlabeled sample sets to obtain a damage level prediction model. Output module, used to analyze the nuclear magnetic resonance of the concrete to be tested. The spectrum and mix proportion parameter encoding vector are used to output the freeze-thaw damage level through the damage level prediction model.

Images