A method, system, apparatus, and medium for relational completion of a cement-based material
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-05-13
- Publication Date
- 2026-06-09
AI Technical Summary
In the research and development of traditional cement-based materials, the forced formatting of knowledge graphs leads to irreversible semantic loss and sample imbalance, resulting in poor relation prediction ability and difficulty in effectively modeling the complex interactions between formulation and process.
By employing a self-adversarial loss function and a self-interference decoder, and through semi-supervised training and controllable perturbation, the generalization ability of the relation encoding model is enhanced, information loss is reduced, and overfitting is prevented, thus constructing a relation completion method for cement-based materials.
It improves the accuracy and generalization ability of predicting relationships in cement-based materials, enabling a more comprehensive characterization of the relationships between material composition, processing, and properties, and enhancing the robustness and discriminativeness of the model.
Smart Images
Figure CN122177321A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of cement-based composite materials technology, and particularly relates to a method, system, equipment and medium for completing relationships in cement-based materials. Background Technology
[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.
[0003] Cement-based materials are key materials for infrastructure construction. Their compressive strength, corrosion resistance, and other properties are not determined by a single factor, but are the result of the combined effects of formulation composition, preparation process, and microstructure. Traditional research and development of cement-based materials mainly relies on trial and error and empirical tables. This method has a long experimental cycle, high cost, and difficulty in effectively modeling the complex interactions between formulation and process.
[0004] Knowledge graphs can realize structured storage and relational reasoning of material knowledge. However, traditional knowledge graphs usually force the original data of cement-based materials to be formatted, resulting in irreversible loss of semantics. In addition, the sample of relation type in cement-based material knowledge graphs is unbalanced, which leads to poor predictive ability of material relations. Summary of the Invention
[0005] To overcome the shortcomings of the existing technologies, this invention proposes a relation completion method, system, equipment, and medium for cement-based materials. This addresses the problems of poor prediction performance caused by information loss and sample imbalance in relational reasoning during the research and development of cement-based materials. It constructs a data foundation for the entire chain of knowledge regarding the composition, process, structure, and performance of cement-based materials, which is of great value for the research and development of new cement-based materials.
[0006] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions: In a first aspect, the present invention discloses a method for completing relationships in cement-based materials, comprising: A set of ternary sets of material maps was constructed based on the original text samples of the formulation, process and properties of cement-based materials. The encoder is trained based on the set of triples as input and the relation category as output to obtain the first relation encoding model; Unlabeled samples are mixed into the original text samples, and a first relational encoding model is used as the model base. The first relational encoding model is semi-supervised by using a self-adversarial loss function to obtain a second relational encoding model. Based on the second relational encoding model, the source node embedding, target node embedding and relation embedding of all triples are extracted. Based on the source node embedding, target node embedding and relation embedding, a positive and negative sample pair input self-interference decoder is constructed and trained to predict the missing relations in the material spectrum. The self-interference decoder actively applies a controllable perturbation to the input, which is constrained by the overall loss. Traverse the material spectrum, input the embedding of any two nodes, and complete the missing relationship based on the self-interference decoder.
[0007] Secondly, this invention discloses a relationship completion system for cement-based materials, comprising: The atlas construction module is configured to: construct a set of triplets of material atlases based on the original text samples of the formulation, process and performance of cement-based materials; The pre-training module is configured to: train the encoder to obtain a first relation encoding model based on the set of triples as input and the relation category as output; The relation encoding module is configured to: mix unlabeled samples into the original text samples and use a first relation encoding model as the model base; use a self-adversarial loss function to semi-supervised train the first relation encoding model to obtain a second relation encoding model; and extract the source node embedding, target node embedding and relation embedding of all triples based on the second relation encoding model. The self-interference decoding module is configured to: construct a positive and negative sample pair input self-interference decoder based on the source node embedding, target node embedding and relation embedding and train to predict the missing relations in the material spectrum; the self-interference decoder actively applies a controllable perturbation to the input constrained by the overall loss. The relation completion module is configured to: traverse the material spectrum, input the embeddings of any two nodes, and complete the missing relations based on the self-interference decoder.
[0008] Thirdly, the present invention discloses an electronic device, including a memory and a processor, and computer instructions stored in the memory and running on the processor, wherein the computer instructions, when run by the processor, complete the steps of the above-mentioned method for completing the relationship of cement-based materials.
[0009] Fourthly, the present invention discloses a computer-readable storage medium for storing computer instructions, which, when executed by a processor, complete the steps of the above-mentioned method for completing the relationship of cement-based materials.
[0010] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention employs a semi-supervised self-adversarial loss function, enabling the encoder to automatically learn detailed features such as the conditional thresholds, collaborative factors, and quantification levels for relationship establishment from unlabeled raw text. This compensates for the irreversible semantic loss caused by the forced formatting in traditional knowledge graph construction of material maps, making link predictions closer to the real physical laws of materials. Furthermore, a self-interference decoder is used to actively apply controllable perturbations, constrained by the overall loss, to the input embedding during training. This disrupts the rigid correlation between similar relationship samples, allowing the decoder to learn the essential laws behind the relationships between material composition, processing, and performance, thus improving the generalization ability during relationship completion.
[0011] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description
[0012] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.
[0013] Figure 1 This is a flowchart of the relationship completion method for cement-based materials described in Embodiment 1 of the present invention. Detailed Implementation
[0014] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.
[0015] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of the present invention.
[0016] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.
[0017] Example 1 When knowledge graphs are applied to the research and development of cement-based materials, by constructing entities and relationships into triple networks and mapping the physical world to the digital space, structured storage and relational reasoning of material knowledge can be achieved. However, problems still exist in the actual construction and application process: First, the irreversible loss of semantics caused by forced formatting.
[0018] Raw data on cement-based materials typically contains rich conditional and degree semantics. In traditional atlas construction processes, this information is forcibly broken down into fixed-type triples, using uniform symbols to represent the complex physical world. This representation provides a clear logical structure for knowledge reasoning, but in the process of transforming material physical world information into structured knowledge, there is also a potential irreversible semantic loss. During entity relation extraction, raw data is forcibly categorized and projected to specific relation types, forcibly formatting the rich semantics inherent in the original data, such as the conditions for establishment and the degree of association, resulting in information loss. This irreversible formatting loss makes it impossible to express the conditional and gradual characteristics of material relationships, causing the model to be unable to accurately infer the existence of relationships when faced with similar but unseen situations.
[0019] Second, overfitting is caused by sample rigidity and data imbalance.
[0020] In the knowledge graph of cement-based materials, some relation types have abundant samples, such as "silica fume and improved compressive strength," while others have extremely sparse samples, such as "specific admixtures and improved sulfate resistance." Existing relation completion models are typically trained by randomly constructing positive and negative samples based on existing triples in the graph. Because instances under the same relation type are highly similar semantically (i.e., sample rigidity), the model is prone to over-memorizing specific sample patterns in the training set, resulting in a significant decrease in its ability to predict relations for unseen formulation combinations or process variations.
[0021] Therefore, to address the above problems, this invention proposes a relationship completion method for cement-based materials based on self-adversarial and self-disturbance, which reduces information loss during knowledge graph construction and prevents overfitting of the relationship completion model.
[0022] In one or more embodiments, a relation completion method for cement-based materials is disclosed, employing self-adversarial loss and self-interference loss to enhance the generalization embedding features of relations, thereby improving the accuracy of material graph relation prediction, such as... Figure 1 As shown, it includes the following steps: Step S1: Construct a set of triplets for the material spectrum based on the original text samples of the formulation, process and performance of cement-based materials to form the material spectrum.
[0023] Collect raw text samples of formulation composition data, process data, and performance data related to cement-based materials. Use the original experimental records and literature paragraphs as raw text samples S1, and extract subject-object-relationship triples from them to form a triple set T1, preserving the alignment relationship between S1 and T1.
[0024] The formula composition includes, but is not limited to, data such as cement grade, water-cement ratio, fly ash content, and silica fume content; the process includes, but is not limited to, data such as curing temperature, curing humidity, sealing conditions, mixing time, and molding pressure; the performance includes, but is not limited to, data such as 28-day compressive strength, flexural strength, corrosion resistance coefficient, and impermeability grade.
[0025] The formulation composition, process, and performance are associated with the subject-object relationship in a directed manner: formulation-process-performance. For example, <subject: formulation 1, relationship: whether this process is applicable, object: calcination temperature 1100 to 1200 degrees>; <subject: formulation 1, relationship: whether it is a necessary condition, object: 28-day compressive strength 50 MPa>; <subject: calcination temperature 1100 to 1200 degrees, relationship: whether it is a necessary condition, object: 28-day compressive strength 50 MPa>.
[0026] As one implementation method, triplet data can be extracted using tools such as existing industry-wide large models or dedicated extraction models.
[0027] Step S2: Train the encoder based on the set of triples as input and the relation category as output to obtain the first relation encoding model.
[0028] Using each triplet in the triplet set T1 and its corresponding original text sample S1 as an independent input, and the relation type label as the output, the Transformer model encoder is trained to obtain the first relation encoding model M1.
[0029] Step S3: Mix unlabeled sample S2 into the original text sample S1 and use the first relation encoding model as the model base. Use the self-adversarial loss function to semi-supervised train the first relation encoding model to obtain the second relation encoding model. Based on the second relation encoding model, extract the source node embedding S, target node embedding T and relation embedding E of all triples.
[0030] Specifically, a self-adversarial loss function based on batch processing (BATCH) was adopted. The definition is as follows:
[0031] In the formula, This refers to extracting samples based on the second relational coding model M2. Embedded features; D is the size of the batch group BATCH, D is the number of batches, and N is the number of samples per batch. As a sample, and , For the first The first in the group One sample, For the first The first in the group One sample, For the first The first in the group One sample; For tagged samples, ; In BATCH Quantity; , , , All are hyperparameters and satisfy the following conditions: In the loss function, BATCH is divided into D groups, with N samples in each group; Find the distance between the two closest samples within a group; similarly, To find the distance between the two least similar samples in the group.
[0032] Furthermore, to enhance relation representation capabilities, the self-adversarial loss function enables the encoding model to utilize both labeled and unlabeled data simultaneously. Unsupervised adversarial training is conducted by BATCHing the nearest and least nearest Euclidean samples; this is to prevent the training process from deviating from the feature center. The supervised training mode is still used to ensure stable convergence during the training process.
[0033] To enable the knowledge graph to represent the original data in more detail, this embodiment designs a self-adversarial loss function based on a semi-supervised approach, and this loss function only applies to the transformer encoder model. This loss function improves the discriminativeness and robustness of the embedded features by prompting adversarial learning between samples with similar and dissimilar features, thereby enhancing the accuracy of subsequent link prediction.
[0034] This embodiment utilizes three features—source node embedding, target node embedding, and relationship embedding—to more comprehensively represent the multidimensional semantics of relationships between nodes, effectively compensating for the semantic loss caused by information formatting in traditional material atlas construction. This allows the material atlas to retain richer information on the conditions for validity and the degree of association in the original data, significantly improving the detail and completeness of knowledge representation.
[0035] This embodiment utilizes a semi-supervised training mechanism and an innovative BATCH intra-group self-adversarial loss function, enabling the model to simultaneously utilize labeled and unlabeled data. This enhances the model's generalization ability while effectively avoiding overfitting.
[0036] Step S4: Construct a positive and negative sample pair input self-interference decoder based on the source node embedding, target node embedding and relation embedding, and train it to predict the missing relations in the material spectrum. The self-interference decoder actively applies a controllable perturbation to the input that is constrained by the overall loss.
[0037] Step S401: Based on source node embedding, target node embedding, and relation embedding, construct positive and negative sample pairs according to whether a certain type of relation exists in the graph.
[0038] Specifically, based on the existing material map, source nodes and target nodes that have a certain relationship are taken as positive samples, while source nodes and target nodes that do not have a certain relationship are taken as negative samples.
[0039] Step S402: Train the self-interference decoder with positive and negative samples. When taking the derivative of the self-interference loss function, interference is injected, which is equivalent to adding controllable perturbations to the decoder input terms (S, T, E) to prevent overfitting when training the decoder model.
[0040] The self-interference decoder is defined as follows:
[0041] In the formula, It is a self-interference decoder; Indicates the first Modular product of each pattern, For learnable parameters The matrix formed; S is the source node embedding, Let S be the mean of all S values within BATCH; T is the embedding of the target node. Let T be the mean of all T values within BATCH; E is the relation embedding. This is the mean of all E values within the BATCH.
[0042] When training the self-interference decoder, the derivative of the loss function is calculated during backpropagation as follows:
[0043] In the formula, To find the gradient, Indicates to Find the gradient. Let cross-entropy be the loss function. For learnable parameters of a neural network, This is a hyperparameter.
[0044] in, The basic principle is: with the parameters unchanged, the input is obtained by inverse differentiation. Superimposed deviations make Minimizing the overall loss is equivalent to minimizing the input. Random interference is superimposed, and the intensity of this interference is constrained by minimizing the overall loss. Since the sample types of relationships in the cement-based material knowledge graph are imbalanced, and knowledge graph link prediction aims to assign new relationships to certain negative sample pairs based on existing positive and negative sample pairs, it is prone to overfitting and poor generalization ability. Therefore, this invention incorporates noise interference to reduce overfitting and improve the model's generalization ability.
[0045] To prevent overfitting of the relation completion model due to sample rigidity, this embodiment designs a self-interference decoder model, which adds a penalty to the input to its loss function. This is equivalent to adding a reverse controllable random perturbation to the input, and the perturbation only applies to the decoder model.
[0046] As one implementation method, the encoder model representing knowledge graph nodes / relationships and the decoder model predicting missing relations can be trained separately in a segmented manner.
[0047] As another implementation method, two models can be cascaded for end-to-end synchronous training. During end-to-end synchronous training, the loss function of the encoder model is... The training function of the decoder model is And only through Adding noise to the derivative of the decoder model perturbs the input.
[0048] Step S5: Traverse the material spectrum, input the embedding of any two nodes, predict the probability that there is a certain relationship between the two nodes based on the self-interference decoder, and set a threshold to determine the missing relationship and fill in the missing relationship.
[0049] A decoder based on encoder graph embedding (S, T, E) is used to predict and complete the relationships between missing material composition, process, and properties.
[0050] Specifically, take the embeddings S and T of any two nodes in the knowledge graph, select the embedding E of a certain relationship, input (S, T, E) into the decoder to obtain the probability score. If the score is greater than a certain threshold, it is considered that there should be a relationship E between S and T, thereby completing the missing relationship.
[0051] Therefore, this invention provides a material graph relationship completion method based on self-adversarial and self-perturbation to complete the relationships between formulations, processes, and properties, such as the missing correlation between "a certain formulation + process" and a certain level of corrosion resistance. To enable the knowledge graph to represent the original data in more detail, a self-adversarial loss function is designed based on a semi-supervised approach. To prevent overfitting of the relationship completion model due to sample rigidity, a self-perturbation decoder model is designed, incorporating a penalty for the input into its loss function, which is equivalent to adding a reverse controllable random perturbation to the input.
[0052] Example 2 In one or more embodiments, a relationship completion system for cement-based materials is disclosed, specifically including: The atlas construction module is configured to: construct a set of triplets of material atlases based on the original text samples of the formulation, process and performance of cement-based materials; The pre-training module is configured to: train the encoder to obtain a first relation encoding model based on the set of triples as input and the relation category as output; The relation encoding module is configured to: mix unlabeled samples into the original text samples and use a first relation encoding model as the model base; use a self-adversarial loss function to semi-supervised train the first relation encoding model to obtain a second relation encoding model; and extract the source node embedding, target node embedding and relation embedding of all triples based on the second relation encoding model. The self-interference decoding module is configured to: construct a positive and negative sample pair input self-interference decoder based on the source node embedding, target node embedding and relation embedding and train to predict the missing relations in the material spectrum; the self-interference decoder actively applies a controllable perturbation to the input constrained by the overall loss. The relation completion module is configured to: traverse the material spectrum, input the embeddings of any two nodes, and complete the missing relations based on the self-interference decoder.
[0053] Example 3 This embodiment provides an electronic device, including a memory and a processor, as well as computer instructions stored in the memory and running on the processor. When the computer instructions are executed by the processor, they complete the steps of the above-mentioned method for completing the relationship of cement-based materials.
[0054] Example 4 This embodiment provides a computer-readable storage medium for storing computer instructions, which, when executed by a processor, complete the steps of the above-described method for completing the relationship of cement-based materials.
[0055] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0056] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0057] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment, whereby a series of operational steps are performed to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0058] The descriptions of each embodiment in the above embodiments have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.
[0059] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A method for completing relationships in cement-based materials, characterized in that, include: A set of ternary sets of material maps was constructed based on the original text samples of the formulation, process and properties of cement-based materials. The encoder is trained based on the set of triples as input and the relation category as output to obtain the first relation encoding model; Unlabeled samples are mixed into the original text samples, and a first relational encoding model is used as the model base. The first relational encoding model is semi-supervised by using a self-adversarial loss function to obtain a second relational encoding model. Based on the second relational encoding model, the source node embedding, target node embedding and relation embedding of all triples are extracted. Based on the source node embedding, target node embedding and relation embedding, a positive and negative sample pair input self-interference decoder is constructed and trained to predict the missing relations in the material spectrum. The self-interference decoder actively applies a controllable perturbation to the input, which is constrained by the overall loss. Traverse the material spectrum, input the embedding of any two nodes, and complete the missing relationship based on the self-interference decoder.
2. The method for completing relationships in cement-based materials as described in claim 1, characterized in that, Employing a batch-based self-adversarial loss function ,as follows: In the formula, This refers to extracting samples based on the second relational coding model M2. Embedded features; D is the size of the batch group BATCH, D is the number of batches, and N is the number of samples per batch. For tagged samples, ; In BATCH Quantity; , , , All are hyperparameters; As a sample, and , For the first The first in the group One sample, For the first The first in the group One sample, For the first The first in the group One sample; Find the distance between the two closest samples within a group; To find the distance between the two least similar samples in the group.
3. The method for completing relationships in cement-based materials as described in claim 1, characterized in that, The construction of positive and negative sample pairs specifically involves: Based on the constructed material map, source nodes and target nodes that already have a certain relationship are taken as positive samples, while source nodes and target nodes that do not have a certain relationship are taken as negative samples.
4. The method for completing relationships in cement-based materials as described in claim 1, characterized in that, The self-interference decoder is: In the formula, It is a self-interference decoder; Indicates the first Modular product of each pattern, For learnable parameters The matrix formed; S represents the source node embedding. The mean of all S values within the BATCH; T represents the embedding of the target node. The mean of all T values within the BATCH; E stands for relational embedding. This is the mean of all E values within the BATCH.
5. The method for completing relationships in cement-based materials as described in claim 1, characterized in that, When training the self-interference decoder, the derivative of the loss function during backpropagation is calculated as follows: In the formula, To find the gradient, Indicates to Find the gradient. Let cross-entropy be the loss function. For learnable parameters of a neural network, Here are the hyperparameters: S is the source node embedding, E is the relation embedding, and T is the target node embedding. It is a self-interference decoder.
6. The method for completing relationships in cement-based materials as described in claim 1, characterized in that, Training of the first relation encoding model and the self-interference decoder: Two models were trained separately using a segmented approach. Alternatively, two cascaded models can be used for end-to-end synchronous training. During end-to-end synchronous training, the loss function of the encoder model is the sum of the self-adversarial loss function based on batch processing and the cross-entropy loss function, while the training function of the decoder model is the cross-entropy loss function.
7. The method for completing relationships in cement-based materials as described in claim 1, characterized in that, The method of completing missing relationships based on a self-interference decoder involves predicting the probability of a relationship between two nodes using the self-interference decoder, setting a threshold to determine missing relationships, and then completing the missing relationships. Specifically: Take the source node embedding S and target node embedding T of any two nodes in the knowledge graph, select a certain relationship embedding E, input (S, T, E) into the decoder, and obtain the probability score of the existence of a certain relationship. If the score is greater than the threshold, it is considered that there should be a relationship E between the source node embedding S and the target node embedding T, thereby completing the missing relationship.
8. A relationship completion system for cement-based materials, characterized in that, include: The atlas construction module is configured to: construct a set of triplets of material atlases based on the original text samples of the formulation, process and performance of cement-based materials; The pre-training module is configured to: train the encoder to obtain a first relation encoding model based on the set of triples as input and the relation category as output; The relation encoding module is configured to: mix unlabeled samples into the original text samples and use a first relation encoding model as the model base; use a self-adversarial loss function to semi-supervised train the first relation encoding model to obtain a second relation encoding model; and extract the source node embedding, target node embedding and relation embedding of all triples based on the second relation encoding model. The self-interference decoding module is configured to: construct a positive and negative sample pair input self-interference decoder based on the source node embedding, target node embedding and relation embedding and train to predict the missing relations in the material spectrum; the self-interference decoder actively applies a controllable perturbation to the input constrained by the overall loss. The relation completion module is configured to: traverse the material spectrum, input the embeddings of any two nodes, and complete the missing relations based on the self-interference decoder.
9. An electronic device, characterized in that, It includes a memory and a processor, as well as computer instructions stored in the memory and running on the processor, which, when executed by the processor, complete the relationship completion method for cement-based materials as described in any one of claims 1-7.
10. A computer-readable storage medium, characterized in that, Used to store computer instructions, which, when executed by a processor, complete the relationship completion method for cement-based materials as described in any one of claims 1-7.