Cultural relic damage form evaluation method and device, electronic equipment and storage medium

Abstract

The application provides a cultural relic damage form evaluation method and device, electronic equipment and storage medium, the method comprises the steps of: acquiring multi-source multi-scale data of cultural relics at different times, the multi-source multi-scale data comprises: respectively processing the multi-source multi-scale data of each cultural relic cooperatively and learning by comparison to obtain the quadruple feature of each damage in each cultural relic; constructing a cultural relic knowledge graph based on the quadruple feature of each damage, the time and space of the cultural relic knowledge graph are aligned; training a target evaluation model according to the cultural relic knowledge graph, obtaining an evaluation result of a cultural relic to be evaluated according to the target evaluation model, sending the evaluation result to a repair decision model, and outputting repair suggestion information according to the evaluation result by the repair decision model, the evaluation result comprises: an anchor point sequence obtained by sorting the crack propagation rate and confidence of the damage in each anchor point. The accuracy of cultural relic damage form evaluation is improved.

Description

Technical Field

[0001] This application relates to the field of model training technology, and more specifically, to a method, device, electronic device, and storage medium for assessing the damage morphology of cultural relics. Background Technology

[0002] my country is an ancient civilization with a long history. Many precious cultural heritages have been preserved and passed down intact through historical screening. However, some cultural heritages have been damaged due to historical reasons or natural disasters, and it is necessary to conduct inspection and assessment of the damage to cultural relics.

[0003] Existing technologies face key challenges in assessing damage to cultural relics, including limitations of single-modal data, reliance on manual annotation, insufficient cross-modal information fusion, low accuracy in damage identification, and a lack of quantitative support for restoration decisions. Summary of the Invention

[0004] The purpose of this application is to address the shortcomings of the prior art by providing a method, device, electronic equipment, and storage medium for assessing the damage morphology of cultural relics, thereby improving the accuracy of assessing the damage morphology of cultural relics.

[0005] To achieve the above objectives, the technical solutions adopted in the embodiments of this application are as follows: In a first aspect, embodiments of this application provide a method for assessing the damage morphology of cultural relics, the method comprising: Acquire multi-source, multi-scale data of cultural relics at different times. The multi-source, multi-scale data includes point cloud data, image data, and text data of the cultural relics. The text data includes material spectrum, material texture, and environmental parameters. The multi-source, multi-scale data of each of the cultural relics were subjected to collaborative preprocessing and comparative learning to obtain the quadruple features of each damage in each of the cultural relics. A cultural relic knowledge graph is constructed based on the quadruple features of each of the aforementioned damages, and the cultural relic knowledge graph is aligned in time and space. A target evaluation model is trained based on the cultural relic knowledge graph. An evaluation result for the cultural relic to be evaluated is obtained based on the target evaluation model. The evaluation result is sent to the restoration decision model, which outputs restoration suggestions based on the evaluation result. The evaluation result includes an anchor point sequence sorted according to the crack propagation rate and confidence level of the damage at each anchor point.

[0006] Optionally, the collaborative preprocessing and comparative learning of the multi-source, multi-scale data of each of the cultural relics to obtain the quadruple features of each damage in each of the cultural relics includes: Heterogeneous pairing momentum contrast coding network is used to perform multimodal collaborative coding on multi-source multi-scale data to obtain encoded multi-source multi-scale data; Based on the encoded multi-source, multi-scale data, the damage characteristics of each damage are obtained by comparing and learning the momentum-updated teacher encoder with the cross-modal shared negative sample queue strategy. Confidence prediction is performed based on the damage characteristics of each damage to obtain the confidence level of each damage, and the quadruple features of each damage are obtained based on the damage characteristics and the confidence level of each damage.

[0007] Optionally, the construction of the cultural relic knowledge graph based on the quadruple features of each of the aforementioned damages includes: The ontology corresponding to the damage is determined based on the four-tuple characteristics of the damage. The ontology includes: cultural relics, components, and anchor points. The cultural relics are whole individuals, the components are stable structural units, and the anchor points are micro-area entities. The environmental parameters of each artifact are embedded into the damaged ontology to obtain a semantic-physical coupled ontology. The type of the coupled ontology is determined using a preset mechanical constraint strategy. A knowledge graph of artifacts is constructed based on each coupled ontology and its type.

[0008] Optionally, the types of the coupled ontology include: state class, process class, action class, and event class. The coupled ontology of the state class is used to characterize the state of damage in the anchor point corresponding to the ontology at a preset time. The coupled ontology of the process class is used to characterize the damage evolution process of damage in the anchor point corresponding to the ontology. The coupled ontology of the action class is used to characterize the external influencing factors of damage in the anchor point corresponding to the ontology. The coupled ontology of the event class is used to characterize the triggering event of damage in the anchor point corresponding to the ontology.

[0009] Optionally, the mechanically constrained strategy includes: By utilizing the uniqueness constraint of a single time and the closure constraint of the measure, the state of the damage at each time step in the anchor point corresponding to the coupled ontology is bound to the coupled ontology; By utilizing time monotonicity constraints, path additivity constraints, and residual upper bound constraints, the states of the coupled entity at adjacent or non-adjacent times are connected to obtain the state evolution information of the coupled entity. Threshold constraints are used to determine the external event triggering factors of the coupled ontology.

[0010] Optionally, the step of training the target evaluation model based on the cultural relic knowledge graph includes: The cultural relic knowledge graph is processed by a relation-aware graph convolutional network to obtain the message flow corresponding to each anchor point in the cultural relic knowledge graph, and the tensor information of each anchor point is obtained based on the message flow. The tensor information of each anchor point is calibrated at the equation level using a differentiable physics layer to obtain calibrated tensor information. The target prediction model is obtained by performing multi-step evolution prediction and uncertainty modeling on the calibrated tensor information using an event-gated time series algorithm.

[0011] Optionally, the step of the repair decision model outputting repair suggestion information based on the output results includes: Calculate the upper confidence boundary risk index of each anchor point, and sort the anchor points in descending order according to the upper confidence boundary risk index of each anchor point to obtain the disposal candidate set; For each anchor point in the candidate disposal set, environmental parameters from the cultural relic knowledge graph are spliced ​​together to obtain each constrained anchor point; Decision analysis is performed sequentially on each constrained anchor point in the candidate disposal set to obtain repair suggestion information corresponding to each anchor point.

[0012] Secondly, this application also provides a device for assessing the damage morphology of cultural relics, the device comprising: The acquisition module is used to acquire multi-source, multi-scale data of cultural relics at different times. The multi-source, multi-scale data includes point cloud data, image data, and text data of the cultural relics. The text data includes material spectrum, material texture, and environmental parameters. The preprocessing module is used to perform collaborative preprocessing and comparative learning on the multi-source, multi-scale data of each of the cultural relics to obtain the quadruple features of each damage in each of the cultural relics. A construction module is used to construct a cultural relic knowledge graph based on the quadruple features of each of the aforementioned damages, wherein the cultural relic knowledge graph is time- and space-aligned; The training module is used to train a target evaluation model based on the cultural relic knowledge graph, obtain the evaluation result of the cultural relic to be evaluated based on the target evaluation model, and send the evaluation result to the restoration decision model. The decision module is used to output repair suggestions based on the assessment results from the repair decision model. The assessment results include a sequence of anchor points ordered according to the crack propagation rate and confidence level of the damage at each anchor point.

[0013] Optionally, the preprocessing module is specifically used for: Heterogeneous pairing momentum contrast coding network is used to perform multimodal collaborative coding on multi-source multi-scale data to obtain encoded multi-source multi-scale data; Based on the encoded multi-source, multi-scale data, the damage characteristics of each damage are obtained by comparing and learning the momentum-updated teacher encoder with the cross-modal shared negative sample queue strategy. Confidence prediction is performed based on the damage characteristics of each damage to obtain the confidence level of each damage, and the quadruple features of each damage are obtained based on the damage characteristics and the confidence level of each damage.

[0014] Optionally, the building module is specifically used for: The ontology corresponding to the damage is determined based on the four-tuple characteristics of the damage. The ontology includes: cultural relics, components, and anchor points. The cultural relics are whole individuals, the components are stable structural units, and the anchor points are micro-area entities. The environmental parameters of each artifact are embedded into the damaged ontology to obtain a semantic-physical coupled ontology. The type of the coupled ontology is determined using a preset mechanical constraint strategy. A knowledge graph of artifacts is constructed based on each coupled ontology and its type.

[0015] Optionally, the types of the coupled ontology include: state class, process class, action class, and event class. The coupled ontology of the state class is used to characterize the state of damage in the anchor point corresponding to the ontology at a preset time. The coupled ontology of the process class is used to characterize the damage evolution process of damage in the anchor point corresponding to the ontology. The coupled ontology of the action class is used to characterize the external influencing factors of damage in the anchor point corresponding to the ontology. The coupled ontology of the event class is used to characterize the triggering event of damage in the anchor point corresponding to the ontology.

[0016] Optionally, the mechanically constrained strategy includes: By utilizing the uniqueness constraint of a single time and the closure constraint of the measure, the state of the damage at each time step in the anchor point corresponding to the coupled ontology is bound to the coupled ontology; By utilizing time monotonicity constraints, path additivity constraints, and residual upper bound constraints, the states of the coupled entity at adjacent or non-adjacent times are connected to obtain the state evolution information of the coupled entity. Threshold constraints are used to determine the external event triggering factors of the coupled ontology.

[0017] Optionally, the training module is specifically used for: The cultural relic knowledge graph is processed by a relation-aware graph convolutional network to obtain the message flow corresponding to each anchor point in the cultural relic knowledge graph, and the tensor information of each anchor point is obtained based on the message flow. The tensor information of each anchor point is calibrated at the equation level using a differentiable physics layer to obtain calibrated tensor information. The target prediction model is obtained by performing multi-step evolution prediction and uncertainty modeling on the calibrated tensor information using an event-gated time series algorithm.

[0018] Optionally, the decision module is specifically used for: Calculate the upper confidence boundary risk index of each anchor point, and sort the anchor points in descending order according to the upper confidence boundary risk index of each anchor point to obtain the disposal candidate set; For each anchor point in the candidate disposal set, the environmental parameters in the cultural relic knowledge graph are spliced ​​together to obtain each constrained anchor point; Decision analysis is performed sequentially on each constrained anchor point in the candidate disposal set to obtain repair suggestion information corresponding to each anchor point.

[0019] Thirdly, embodiments of this application also provide an electronic device, including: a processor, a storage medium, and a bus. The storage medium stores program instructions executable by the processor. When the application runs, the processor communicates with the storage medium via the bus, and the processor executes the program instructions to perform the steps of the cultural relic damage morphology assessment method described in the first aspect.

[0020] Fourthly, embodiments of this application also provide a computer-readable storage medium storing a computer program, which is read and executes the steps of the cultural relic damage morphology assessment method described in the first aspect.

[0021] The beneficial effects of this application are: This application provides a method, device, electronic equipment, and storage medium for assessing the damage morphology of cultural relics. By acquiring multi-source, multi-scale data of cultural relics at different times, it performs collaborative preprocessing and comparative learning on the multi-source, multi-scale data of each cultural relic to obtain the four-tuple features of each type of damage. A unified coordinate system and management standard are established, solving the problem of data misalignment across different devices, batches, and times. This achieves data reproducibility and comparability, enabling synchronous archiving and management of observational and interactive data. Furthermore, the obtained four-tuple features are a joint expression of image, geometry, material, and environment, overcoming the difficulty of detecting subtle changes with single information. This achieves stable identification and continuous tracking of small changes, and outputs the four-tuple features with undetermined confidence levels for subsequent use. Based on the four-tuple features of each type of damage, a cultural relic knowledge graph is constructed. This knowledge graph connects cultural relics, structures, anchor points, states, and events into a computable graph, solving the problems of scattered evidence and difficulty in tracing causal chains. A target assessment model is trained using a cultural relic knowledge graph. The assessment results of the cultural relic to be assessed are then obtained from this model and sent to a restoration decision-making model. This model then outputs restoration recommendations based on the assessment results. The training process employs a three-stage reasoning process: graph computation, physical calibration, and time-series prediction. This constructs a verifiable computational closed loop, overcoming biases caused by relying solely on data or patterns. The results achieve interpretable, verifiable, and re-examination-worthy predictions, and enable automatic writing and updating consistent with physical data. Furthermore, it realizes the implementation of a closed-loop restoration decision-making process and human-machine collaboration. Attached Figure Description

[0022] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 A flowchart illustrating a method for assessing the damage morphology of cultural relics provided in this application embodiment; Figure 2 A flowchart illustrating the second method for assessing the damage morphology of cultural relics provided in this application embodiment; Figure 3 A flowchart illustrating the third method for assessing the damage morphology of cultural relics provided in this application embodiment; Figure 4 A flowchart illustrating the fourth method for assessing the damage morphology of cultural relics provided in this application embodiment; Figure 5 A flowchart illustrating the fifth method for assessing the damage morphology of cultural relics provided in this application embodiment; Figure 6 A schematic diagram of a cultural relic provided in this application embodiment; Figure 7 A schematic diagram of the point cloud data acquisition results for cultural relics provided in this application embodiment; Figure 8 A schematic diagram of the geometric tensor characteristics of an artifact provided in this application; Figure 9 A schematic diagram of a damage marker provided in an embodiment of this application; Figure 10 This is a schematic diagram of a feature extraction result provided in an embodiment of this application; Figure 11 This is a schematic diagram of repair suggestion information obtained using the repair decision model of this application; Figure 12 A schematic diagram of an apparatus for assessing the damage morphology of cultural relics provided in an embodiment of this application; Figure 13 This is a structural block diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. It should be understood that the accompanying drawings in this application are for illustrative and descriptive purposes only and are not intended to limit the scope of protection of this application. Furthermore, it should be understood that the schematic drawings are not drawn to scale. The flowcharts used in this application illustrate operations implemented according to some embodiments of this application. It should be understood that the operations in the flowcharts may not be implemented in sequence, and steps without logical contextual relationships may be reversed or implemented simultaneously. In addition, those skilled in the art, guided by the content of this application, may add one or more other operations to the flowcharts, or remove one or more operations from the flowcharts.

[0025] Furthermore, the described embodiments are merely some, not all, of the embodiments of this application. The components of the embodiments of this application described and illustrated herein can typically be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of the application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0026] It should be noted that the term "comprising" will be used in the embodiments of this application to indicate the presence of the features declared thereafter, but does not exclude the addition of other features.

[0027] Current mainstream cultural heritage knowledge graphs are mostly text-based, lacking effective integration of images and other visual information, as well as digital display. This makes it difficult to provide the public or experts with intuitive visual perception, and presents a natural barrier to the identification and understanding of the damaged forms of cultural relics. Text-based knowledge graphs cannot directly present image clues about damaged areas, nor can they support downstream tasks such as visual question answering and cross-modal retrieval. This severely underestimates the potential value of multimodal data and makes it difficult to meet the high-precision visual information requirements of intelligent quantitative assessment.

[0028] Existing research often relies on manual annotation or predefined visualization categories when constructing multimodal knowledge graphs. The annotation of images or iconographic items is highly dependent on manual operation, which is time-consuming, labor-intensive, and difficult to scale up. In addition, the predefined visual entity types have limited coverage and cannot cope with the complex and varied damage forms in cultural heritage. This results in bottlenecks in human resources and insufficient scalability for large-scale graph construction.

[0029] In terms of image feature modeling, most methods only use global visual features, ignoring the local salience of the damaged area and the key semantics in the cultural relic description text. This single global representation is difficult to fully capture the fine-grained damage morphology, thus making it difficult to provide accurate morphological description and damage level discrimination support in subsequent quantitative assessments, which makes the assessment results unreliable.

[0030] Furthermore, technologies specifically designed for the detection and assessment of damage to cultural relics often focus on a single modality, such as image segmentation models based on deep learning or multi-source feature fusion methods based on sensor data. Image segmentation models lack generalization ability across different materials and damage types, and even after supervised training, they still struggle to stably locate damaged areas. While earthquake damage data processing can improve data quality, it lacks semantic association with knowledge graphs and intelligent reasoning capabilities, making it difficult to form an integrated closed loop encompassing morphological recognition, damage quantification, and repair decision-making.

[0031] Based on the technical problems existing in the prior art, this application proposes a method for assessing the damage morphology of cultural relics.

[0032] Optionally, the method for assessing the damage morphology of cultural relics provided in this application can be applied to electronic devices, such as mobile phones, tablets, laptops, handheld computers, desktop computers, and other terminal devices with computing and display capabilities, or servers. Specifically, it can be applied to applications in terminal devices, such as mobile phone applications (APPs) and computer application systems.

[0033] The following section will explain in detail the specific implementation process of the cultural relic damage assessment provided in the embodiments of this application.

[0034] Figure 1 This is a flowchart illustrating a method for assessing the damage morphology of cultural relics provided in an embodiment of this application. The subject executing this method is the aforementioned electronic device. Figure 1 As shown, the method includes: S101. Obtain multi-source, multi-scale data of cultural relics at different times.

[0035] This multi-source, multi-scale data can include point cloud data, image data, and text data of the cultural relic. The text data includes material spectra, texture data, environmental parameters, and user interaction data. Environmental parameters include, for example, the temperature, humidity, light intensity, and pollution index of the cultural relic at a certain time.

[0036] Optionally, high-resolution oblique photography can be used to acquire image data of cultural relics, airborne and ground-based LiDAR scanning can be used to acquire point cloud data of cultural relics, macro photography can be used to acquire material texture of cultural relics, and material spectral measurement methods can be used to acquire material spectral data of cultural relics. At the same time, immersive interactive devices, such as AR / VR headsets and haptic feedback devices, can be used to record the operation trajectory and focus point of tourists or restoration experts on cultural relics to supplement user interaction data.

[0037] S102. Perform collaborative preprocessing and comparative learning on the multi-source, multi-scale data of each cultural relic to obtain the quadruple features of each damage in each cultural relic.

[0038] Optionally, for multi-source, multi-scale data of cultural relics at a certain time, collaborative processing can refer to data format unification, spatial coordinate alignment, time synchronization, and encoding of the multi-source, multi-scale data.

[0039] Specifically, in the heterogeneous pairwise momentum contrastive coding network, collaborative representation learning is performed on image data, material spectra, material textures, and point cloud data under a unified coordinate system. When multi-source, multi-scale data are in a unified coordinate system, the point cloud data is divided using three-dimensional mesh triangular patches as anchor points to obtain damage micro-blocks / component macro-blocks. The scale of the damage micro-blocks is smaller than the scale of the component macro-blocks, and the damage exists in the anchor points.

[0040] Optionally, contrastive learning is used for feature learning of image data, point cloud data, and text data. By comparing positive and negative samples, quadruple features of various types of damage in the artifact are obtained. Specifically, the quadruple features include the geometric tensor F of deformation-crack. geo Material-chemical damage characterization F mat Context-interaction context embedding F ctx And confidence level σ.

[0041] S103. Construct a knowledge graph of cultural relics based on the quadruple features of each type of damage.

[0042] Among them, the cultural relic knowledge graph is aligned in terms of time and space.

[0043] Specifically, feature alignment and node generation can be performed based on the quadruple features of each damage generated in step S102 above. Specifically, anchor nodes are generated. Each anchor node has attribute information such as unique identifier, coordinates, scale, and dimension. The anchor node establishes a subordinate relationship with the component and cultural relic through the unique identifier, and expands the relationship according to three mechanical constraints, and constructs a damage causal hyperedge.

[0044] S104. The target evaluation model is trained based on the cultural relic knowledge graph. The evaluation results of the cultural relic to be evaluated are obtained based on the target evaluation model. The evaluation results are sent to the restoration decision model, which outputs restoration suggestions based on the evaluation results.

[0045] The assessment results include a sequence of anchor points ordered by the crack propagation rate and confidence level of the damage at each anchor point.

[0046] Specifically, after obtaining the cultural relic knowledge graph, the obtained cultural relic knowledge graph can be input into the initial evaluation model to train the initial evaluation model. During the training process, a three-stage inference process of graph computation, physical calibration and time series prediction can be used to construct a verifiable computational closed loop. The training objective can adopt the "evidence-physical-time series" joint loss and introduce a mechanical residual term as a core constraint to train the initial evaluation model. When the training objective is achieved, the evaluation model that achieves the training objective is used as the target evaluation model.

[0047] After obtaining the target assessment model, multi-source, multi-scale data of the cultural relic to be assessed is input into the target assessment model. The target assessment model preprocesses the multi-source, multi-scale data of the cultural relic to be assessed and constructs a knowledge graph of the cultural relic to be assessed. Then, the constructed knowledge graph of the cultural relic to be assessed is input into the target assessment model for reasoning to obtain the assessment result of the cultural relic to be assessed. Finally, the assessment result of the cultural relic to be assessed is sent to the restoration decision model, which outputs restoration suggestions based on the assessment result.

[0048] In this embodiment, multi-source, multi-scale data of cultural relics at different times are acquired. Collaborative preprocessing and comparative learning are performed on the multi-source, multi-scale data of each cultural relic to obtain the four-tuple features of each type of damage. A unified coordinate and management standard is constructed, solving the problem of data misalignment across different devices, batches, and times. This achieves data reproducibility and comparability, enabling synchronous archiving and management of observational and interactive data. Furthermore, the obtained four-tuple features are a joint expression of image, geometry, material, and environment, overcoming the difficulty of detecting subtle changes with single information. This achieves stable identification and continuous tracking of small changes, and outputs the four-tuple features with expected confidence levels for subsequent use. A cultural relic knowledge graph is constructed based on the four-tuple features of each type of damage. This knowledge graph connects cultural relics, structures, anchor points, states, and events into a computable graph, solving the problems of scattered evidence and difficulty in tracing causal chains. A target evaluation model is trained based on the cultural relic knowledge graph. The evaluation results of the cultural relics to be evaluated are obtained from the target evaluation model and sent to the restoration decision model. The restoration decision model then outputs restoration suggestions based on the evaluation results. The training process employs a three-stage inference process involving graph computation, physical calibration, and time-series prediction, constructing a verifiable computational closed loop. This overcomes the biases caused by relying solely on data or patterns, achieving interpretable, verifiable, and re-examination standards for prediction results. It also enables automatic writing and updating consistent with physical data. Furthermore, it realizes the implementation of a closed-loop repair decision-making process and human-machine collaboration.

[0049] Figure 2 This is a flowchart illustrating the second method for assessing the damage morphology of cultural relics provided in this application embodiment, as shown below. Figure 2 As shown, in step S102 above, collaborative preprocessing and comparative learning are performed on the multi-source, multi-scale data of each cultural relic to obtain the quadruple features of each damage in each cultural relic, which may include: S201. Use heterogeneous paired momentum contrast coding networks to perform multimodal collaborative coding on multi-source multi-scale data to obtain coded multi-source multi-scale data.

[0050] Optionally, this heterogeneous paired momentum contrast coding network consists of a pair of synchronously operating encoders, specifically an image-material encoder and a geometry encoder. The image-material encoder performs encoding and analysis on two-dimensional data from multi-source, multi-scale datasets, while the geometry encoder performs encoding and analysis on three-dimensional data from multi-source, multi-scale datasets. Furthermore, the image-material encoder employs a modified ResNet-50, maintaining a stride of 1 in the high-resolution stage and introducing spectral-adaptive convolutions to access multi-channel material spectra. The geometry encoder uses density-adaptive PointNet++, embedding curvature-roughness weights in the sampling layer to preferentially preserve high-stress concentration areas such as crack tips and erosion boundaries.

[0051] The multi-source, multi-scale data of cultural relics are processed by an image-material encoder to obtain encoded image data; the multi-source, multi-scale data of cultural relics are processed by a geometric encoder to obtain encoded point cloud data and encoded material texture data.

[0052] S202. Based on the encoded multi-source, multi-scale data, the momentum-updated teacher encoder and the cross-modal shared negative sample queue strategy are used for comparative learning to obtain the damage characteristics of each damage.

[0053] Optionally, the image-material encoder and the geometry encoder are respectively connected to a shared physical prior modulator. This physical prior modulator fuses environmental parameters such as temperature, humidity, illumination, pollution index, and material spectrum into the encoded image data, encoded point cloud data, and encoded material texture data, respectively, to obtain injected point cloud data, injected image data, and injected material texture data. Specifically, this physical prior modulator injects environmental parameters into each layer of features via FiLM-style channel scaling and gating, i.e., into the image data, point cloud data, and material texture data respectively. This makes each layer of features controllable in its sensitivity to environmental parameters, thereby forming a distinguishable representation between damaged morphology and normal weathering.

[0054] Optionally, after the physical prior modulator outputs the injected point cloud data, injected image data, and injected material texture data, a 128-dimensional local and global contrast vector is output via a hierarchical projection head to obtain the damage vector and the overall vector of the artifact. These vectors are then input into a momentum update teacher encoder mechanism, which performs comparative learning using a momentum update teacher encoder and a cross-modal shared negative sample queue strategy. Specifically, the momentum coefficient m of the momentum update teacher encoder is 0.999, and comparative learning is performed using a cross-modal shared negative sample queue with a length of 65536. Negative samples are grouped and cached according to timestamps and spatial adjacency, strengthening the rejection of "near-negative samples" from different sampling periods of the same component, thereby improving the sensitivity to subtle evolution (crack propagation 0.2–0.5 mm).

[0055] In this embodiment, a bidirectional cross-modal InfoNCE objective function with damage prior weighting is designed, given the image-material vector Z and the geometry vector Z at the same anchor point. P Its direct counterpart is ( )and( The negative pairs come from other anchor samples in the shared queue. The loss of the objective function is given by the following formula (I).

[0056] Formula (1) in =0.07, representing another mode, weight. The interactive heat map h and spectral anomaly s are jointly determined to amplify the contribution of areas of expert concern and high-risk areas of chemical damage to feature learning, thereby realizing "human-machine co-driven" damage-sensitive feature mining.

[0057] To ensure temporal consistency and physical accessibility, this application introduces a "temporally stable-evolutionary selective" regularization in addition to the comparison target, applying consistency constraints to cross-period samples at the same anchor point, so that the features of the two periods remain close when environmental conditions are similar and there is no real damage change; when the damage growth trend is predicted by the environment-material prior, a sparsity triggering penalty is applied to the feature difference, so that the network produces a higher response to the real evolution. At the same time, the geometric encoder adds alignment perturbation and point density jitter based on SE(3) perturbation in the sampling and aggregation stages, and the image-material encoder adopts physical augmentation consistent with spectrum-illuminance, such as humidity mapping color change, thin-layer corrosion synthesis, and micro-blurring-noise coupling, so as to obtain a discriminative characterization that is both pose-invariant and sensing-robust.

[0058] The heterogeneous pairing momentum contrastive coding network outputs three types of feature vectors, namely the damage characteristics of each damage, one of which is the deformation-crack sensitive geometric tensor F. geo The first is indexed by vertices / triangles at the mesh level; the second is material-chemical damage characterization F. mat The first is a one-to-one correspondence with texture-spectral pixel blocks; the second is the embedding of the environment-interaction context F. ctx This is used to characterize the spatial distribution of external drivers and human-machine interactions. The three elements are aligned losslessly using damaged IDs, generating node and edge features that can be directly injected into knowledge graph entities.

[0059] S203. Based on the damage characteristics of each damage, perform confidence prediction to obtain the confidence level of each damage, and obtain the quadruple characteristics of each damage based on the damage characteristics and the confidence level of each damage.

[0060] Optionally, after obtaining the damage characteristics of each damage, the confidence level of each damage is calculated using temperature adaptation and teacher-student consistency differences, and low-confidence samples are dynamically screened out during queue entry and loss weighting. The resulting quadruple features are (F geo F mat F ctx ,σ).

[0061] Figure 3 This is a flowchart illustrating the second method for assessing the damage morphology of cultural relics provided in this application embodiment, as shown below. Figure 3 As shown, the above-mentioned S103, constructing a cultural relic knowledge graph based on the quadruple features of each type of damage, may include: S301. Determine the ontology corresponding to each damage based on the quadruple characteristics of each damage.

[0062] The ontology can include: artifacts, components, and anchor points. Artifacts are whole entities, components are stable structural units, and anchor points are micro-area entities.

[0063] Specifically, anchor nodes can be generated based on the quadruple characteristics of each damage. Each anchor point has immutable identifiers, geometric / material scales, measurement units, and coordinate references, ensuring lossless alignment across devices, batches, and time periods. Anchor points establish a subordinate relationship with components and cultural relics through unique identifiers, resulting in the ontology of each damage. The resulting ontology is based on a three-layer object structure of cultural relic-component-anchor point, corresponding one-to-one with the quadruple characteristics of each damage.

[0064] S302. Embed the environmental parameters of each cultural relic into the damaged ontology to obtain a semantic-physical coupled ontology, and use a preset mechanical constraint strategy to determine the type of the coupled ontology. Construct a cultural relic knowledge graph based on each coupled ontology and the type of each coupled ontology.

[0065] The types of coupled entities can include: state classes, process classes, action classes, and event classes.

[0066] To ensure project usability and evidence compliance, a "Phys-ProvChain" is implemented in the knowledge graph, binding each state and relationship hop-by-hop to its source, model version, calibration file, environmental context, and human intervention records. Simultaneously, a "unit / dimensional axiom set" and a "coordinate consistency axiom set" are defined to automatically verify and correct dimensions, coordinates, and time references during graph loading and incremental update phases. The knowledge graph employs a dual-track versioning strategy of time-series snapshots and event logs. Snapshots ensure a reconstructable evaluation baseline at any given time, while event logs reflect minor evolutions and human-computer interactions; both maintain consistency through hasStateAt(t) and evolvesTo.

[0067] Optionally, to support the causal synthesis of cross-scale damage, this embodiment introduces a "damage causality superedge," which uses... Exposure Material, Geometry The high-level relation of the State points to the state change by considering multiple external factors, material spectra, and geometric stress concentration. A learnable mechanism kernel is embedded on the hyperedge to store the equivalent parameters (such as the crack tip stress intensity factor range and the propagation rate prior) obtained by reasoning in step (4). At the same time, the ontology defines the "risk measurement node" and its measurement edge to the anchor point state, and calculates and maintains the structural integrity index I and the damage probability p. damage Its value is derived from the fusion of evidence within the figure, avoiding the bias caused by single-source observations.

[0068] Through the aforementioned ontology design and relational extension, this embodiment achieves standardized integration of damage evidence, environmental drivers, mechanical mechanisms, and human-computer interaction within the same graph space. This allows subsequent three-level reasoning to directly read structured priors and temporal evolution chains from the graph. Simultaneously, the repair decision module can target components with high damage probability p. damage A well-defined and interpretable risk management plan for regions with high crack propagation rates and identified by triggers.

[0069] Optionally, the above-mentioned mechanically constrained strategies include: First, by utilizing the uniqueness constraint of a single time and the closure constraint of the measure, the state of the damage at each time step in the anchor point corresponding to the coupled ontology is bound to the coupled ontology.

[0070] Specifically, the hasStateAt(t) relation is used to bind state nodes to specific timestamps. Each state node contains a four-tuple feature and a confidence interval for the current time. Where hasStateAt(t) is used to link any object E to its state node S at timestamp t. t Binding. This relationship satisfies "single-time uniqueness" and "measure closure", that is... t, E has at most one normalized state S t And S t Must carry {F geo F mat F ctx , σ} and physical dimension annotation.

[0071] Second, by utilizing time monotonicity constraints, path additivity constraints, and residual upper bound constraints, the states of the coupled ontology at adjacent or non-adjacent times are connected to obtain the state evolution information of the coupled ontology.

[0072] Specifically, the "evolvesTo" relationship is used to connect adjacent states. The evolution chain must satisfy physical residual constraints to ensure that the evolution process conforms to material boundary conditions. Here, "evolvesTo" is used to connect adjacent or non-adjacent states S. t →S t +ΔS t It also forces the recording of the evolution-driving tensor ΔF and the prediction source (data-driven / physical solution / human-machine correction). This relationship has three constraints: "time monotonicity", "path additivity" and "residual upper bound". The residual upper bound is automatically generated by the PINN mechanical residual threshold to ensure that the evolution chain does not violate material and boundary conditions.

[0073] Third, threshold constraints are used to determine the external event triggering factors of the coupled ontology.

[0074] Specifically, the "triggers" relationship is used to model the triggering effect of external actions or interactive events on state transitions. Here, "triggers" is used to characterize the triggering effect of external actions or interactive events on state transitions. The triple (C, θ, η) represents the triggering factor (such as temperature and humidity pulse, pollution exposure, tool contact), the onset threshold, and the trigger confidence, respectively. When C crosses θ and η is higher than the dynamic graph threshold, controlled edges of the `evolvesTo` method are allowed to be generated in the graph with "mechanism labels" (stress corrosion, salt stress, fatigue, etc.).

[0075] Figure 4 A flowchart illustrating the third method for assessing the damage morphology of cultural relics provided in this application embodiment is shown below. Figure 4 As shown, the target evaluation model obtained by training the cultural relic knowledge graph in S104 above may include: S401. Use a relation-aware graph convolutional network to perform message perception processing on the cultural relic knowledge graph, obtain the message flow corresponding to each anchor point in the cultural relic knowledge graph, and obtain the tensor information of each anchor point based on the message flow.

[0076] Using a runtime graph consisting of three layers of objects—"artifacts—components—anchors" and their relationships (hasStateAt, evolvesTo, triggers, and hyperedges DCH)—as the computational domain, for objects with F... geo F mat F ctx The nodes and edges of the σ quadruple perform relationship-specific message passing. Each message is weighted by "mechanism kernel attention," and the attention weight is calculated by the matching degree between the current environment and material signature by the MechanismKernel (stress corrosion / salt stress / fatigue, etc.) embedded in the hyperedge, thereby realizing "mechanism-selective" information flow within the layer. After multi-layer aggregation, RT-GCN outputs two types of headers: (i) field surrogate quantity, crack tip stress intensity factor. J integral Corrosion driving force (ii) Calculation results of materials and boundary equivalent parameters such as elastic modulus. And the reliability of the elastic modulus, Poisson's ratio The reliability of Poisson's ratio and the moisture diffusivity. and the reliability of the moisture diffusivity and the thermal / moisture expansion coefficient. And the confidence intervals of thermal / moisture expansion coefficients, these quantities are kept consistent with the coordinates of specific anchor points in the graph.

[0077] S402. Use the differentiable physical layer to perform equation-level calibration on the tensor information of each anchor point to obtain the calibrated tensor information.

[0078] PINN Differential Constraint Layer (Differential Physical Calibration). It automatically extracts implicit geometry reconstructed from meshes / point clouds using components as subdomains Ω, adaptively distributes physical collocations in the anchor point neighborhood, applies residual constraints to the following coupled equations, and achieves linear elastic equilibrium. ·σ+b=0、Constituent σ=C( , , h, T): ε, humidity diffusion h / t ·(Dh h)=sh, and a differentiable approximation of a specific crack growth law selected by the mechanism kernel. Boundary conditions are interpreted from the triggers and Exposure nodes of the knowledge graph as differentiable constraints such as load, displacement, and convection / flux; parameters are estimated using RT-GCN and physical feasibility is guaranteed through soft positive projection. This layer uses neural fields u, σ, h, etc. as intermediate variables, outputs the equation residuals relas, rdiff, rcrackr at each collocation point and the boundary / initial value residuals rbc, ric, and acts inversely on the previous network segment to achieve closed-loop calibration of "prior-equation-evidence".

[0079] S403. The event-gated time-series algorithm performs multi-step evolution prediction and uncertainty modeling on each calibrated tensor information to obtain the target prediction model.

[0080] Specifically, the event-gated temporal Transformer (EG-T-Former) organizes the state embeddings of each anchor point at different timestamps (calibrated by RT-GCN and spliced ​​with PINN's physical field summary) into unequal-interval sequences. It employs "physical step size relative position encoding" (constructed as a function of Δt and environmental exogenous quantities) and "trigger factor gating" (from the (C, θ, η) triples of triggers) to jointly schedule attention. The decoder outputs multi-step prediction sequences damaget:t+H, at:t+H, crack_growth_ratet:t+H and corresponding confidence intervals in a "monotonically accumulated damage" structure. The soft threshold increment and integral ensure the physical monotonicity of the damage degree over time and provide explicit annotations for anomalous evolution (sudden transitions).

[0081] The training objective adopts the "evidence-physics-time series" joint loss and introduces a mechanical residual term as the core constraint, which is expressed by the following formula (II).

[0082]

[0083] Formula (II) in Align with multimodal observations (geometric deformation, spectral anomalies, thermal / humid field survey lines, etc.). Constraint-triggered evolutionary causal consistency ensures stable gradient propagation under unequal sampling intervals; weights are adaptively recalibrated based on sample uncertainty σ to reduce interference from low-confidence observations. During the inference phase, the predictions of EG-T-Former are cross-validated with the physical feasible region derived from PINN. Only when the consistency between the two exceeds a dynamic threshold are new `evolvesTo` edges and risk measurement nodes generated in the knowledge graph, forming a highly reliable prediction of crack propagation / material degradation. This prediction is then directly output to the repair decision module as a disposal list sorted by `crack_growth_rate` and confidence interval.

[0084] This embodiment proposes a three-stage Physically Provable Reasoning Stack (PGT-Stack) consisting of "GCN → PINN Differential Constraint Layer → Temporal Transformer". This stack enables closed-loop reasoning from evidence fusion and field variable reconstruction to temporal evolution prediction, built upon the tensor attachment and hyperedge causal structure of the knowledge graph. Its core innovation lies in generating "mechanism parameters-driving forces" priors using a relation-aware graph network, then performing equation-level calibration of these priors using a differentiable physical layer, and finally, using an event-gated temporal Transformer to perform multi-step uncertainty prediction and risk ranking.

[0085] Figure 5 A flowchart illustrating the fifth method for assessing the damage morphology of cultural relics provided in this application is shown below. Figure 5 As shown, the repair suggestion information output by the repair decision model based on the assessment results in S104 above may include: S501. Calculate the upper confidence boundary risk index of each anchor point, and sort the anchor points in descending order according to the upper confidence boundary risk index of each anchor point to obtain the disposal candidate set.

[0086] The "Upper Confidence Boundary Risk Index" (UCB-Risk) is calculated using anchor points as units. The Upper Confidence Boundary Risk Index is given by Rucb = μ(crack_growth_rate) + βσ(crack_growth_rate), where μ and σ are derived from the multi-step prediction and uncertainty assessment of the event-gated time series EG-T-Former, and β is a task adaptive factor (automatically determined by the value density, artifact grade, and accessibility within the map). RDEC sorts all ROIs (damaged micro-blocks) in descending order according to Rucb, obtaining the "disposal candidate set".

[0087] S502. To process the environmental parameters in the cultural relic knowledge graph of each anchor point in the candidate set, the constrained anchor points are obtained.

[0088] Specifically, the environmental parameters in the knowledge graph for splicing cultural relics at each anchor point in the candidate set refer to constructing a SPrompt prefix for each anchor point in the candidate set: [ROI_id|R_ucb|CI(95%)|PDE_residual|triggers|material_signature|env_window]. This means using growth rate, confidence interval, and equation residual as hard constraints as prerequisite information, and then splicing triggering factors, material signatures, and constructible environment windows from the knowledge graph. This achieves quantitative constraints and risk ranking for the generation process.

[0089] S503. Decision analysis is performed on each constrained anchor point in the candidate disposal set in turn to obtain the repair suggestion information corresponding to each anchor point.

[0090] The repair decision model sequentially performs decision analysis on the constrained anchor points to obtain repair suggestion information corresponding to each anchor point.

[0091] The repair decision model can standardize repair actions into composable atomic tokens, {CLEAN, DRY, DESALT, CONSOL, GROUT, FILL, PATCH, REPASS, SHIELD, ENVCTRL, ...}, where each token carries a parameter vector θ = {material compatibility C}. p Reversibility R v The IAO framework includes parameters such as penetration depth d, viscosity η, pH, curing window T / RH, cost c, ..., and constraint syntax (e.g., matching rules with material spectra and temperature / humidity windows). The IAO framework embeds the knowledge graph in the form of a DSL, serving as the "action syntax" for the large model and an executable interface for RDEC, ensuring that the generated results automatically align with protection principles (minimal intervention, reversibility, compatibility).

[0092] A constraint-aware dual-chain generation-simulation verification mechanism is adopted. Chain 1, consisting of a domain-fine-tuned large language model, outputs a repair strategy graph under the joint constraints of SPrompt's hard prefix and IAO syntax. This graph includes action sequences, parameter ranges, quality control points, monitoring sensor deployment, and environmental control plans. Chain 2 consists of "intervention influencing PINN (I 2 -PINN) performs differentiable comparative simulations of RSG to predict changes in crack_growth_rate, stress field, and moisture diffusion field after application, and calculates the expected risk reduction ΔR=R. ucb When ΔR falls below the dynamic threshold or when equation residuals rebound, the system automatically writes back negative feedback to Heritage-LLM and triggers restricted regeneration until the crack_growth_rate and material / environment compatibility rules are met.

[0093] To achieve optimal ranking of multiple objectives, this embodiment defines a priority index, the function of which is shown in formula (III) below.

[0094] Formula (3) Among them, the second item To represent the expected risk reduction, Φ represents the combined utility of reversibility and compatibility, Ψ represents the cost and downtime penalty, and the coefficients α, γ, ρ, and ξ are adaptively determined based on the cultural relic's grade and the exhibition task. RDEC selects the best among multiple feasible RSGs according to S and outputs a three-part result of "solution - justification - quantified benefits" for experts to compare with a single click.

[0095] To ensure on-site feasibility, this application compiles the selected RSGs into an "Execution Package," which includes a bill of materials and batch traceability, parameter tolerances and verification methods, environmental control curves (temperature, humidity / light), process schedule (avoiding time periods outside the env_window), sensor placement and thresholds, and retesting paths. The execution package also generates a "compliance proof chain" (interfacing with the Phys-Prov Chain of the graph) and sets "shutdown conditions" (pausing and re-evaluating when the online monitoring's crack_growth_rate or PINN residual exceeds the threshold).

[0096] In human-machine collaboration, this embodiment provides an expert injection port, allowing partial overriding of action parameters θ and process sequence; the overriding record is written back to the knowledge graph to trigger triggers, I 2 -PINN instantly recalculates ΔR to give the impact amount and marks it on RSG with confidence bands, realizing traceable "human-machine co-creation" decision-making.

[0097] Ultimately, RDEC outputs both the Heritage-LLM generated text and the structured RSG simultaneously. The text explains the underlying mechanisms and key operational points, while the structured results are used for integration with the monitoring system. After implementation, the system continuously updates μ(crack_growth_rate), σ, and the PDE residuals, and automatically rearranges the R values ​​for subsequent batches accordingly. ucb Together with S, a closed loop of "prediction-decision-verification-relearning" is formed. Through the restoration decision-making mechanism in this embodiment, this embodiment realizes risk ranking dominated by growth rate and confidence interval, effect review guaranteed by physical simulation, and scheme generation constrained by ontology syntax, which significantly improves the safety, interpretability, and feasibility of cultural relic restoration.

[0098] Figure 6 This is a schematic diagram of a cultural relic provided in an embodiment of this application. Figure 7 This is a schematic diagram of the point cloud data acquisition results for cultural relics provided in an embodiment of this application. Figure 8A schematic diagram illustrating the geometric features of an artifact provided in this application. Figure 9 This is a schematic diagram of a damage marker provided in an embodiment of this application. Figure 10 This is a schematic diagram of a feature extraction result provided in an embodiment of this application. Figure 11 This is a schematic diagram illustrating repair suggestion information obtained using the repair decision model of this application. It is worth noting that... Figure 11 The display of repair suggestions in the document is merely illustrative.

[0099] Figure 12 A schematic diagram of an apparatus for assessing the damage morphology of cultural relics provided in this application embodiment is shown below. Figure 12 As shown, the device includes: The acquisition module 601 is used to acquire multi-source, multi-scale data of cultural relics at different times. The multi-source, multi-scale data includes point cloud data, image data, and text data of the cultural relics. The text data includes material spectrum, material texture, and environmental parameters. Preprocessing module 602 is used to perform collaborative preprocessing and comparative learning on the multi-source and multi-scale data of each of the cultural relics to obtain the quadruple features of each damage in each of the cultural relics. Construction module 603 is used to construct a cultural relic knowledge graph based on the quadruple features of each of the damages, wherein the cultural relic knowledge graph is time- and space-aligned; The training module 604 is used to train a target evaluation model based on the cultural relic knowledge graph, obtain the evaluation result of the cultural relic to be evaluated based on the target evaluation model, and send the evaluation result to the restoration decision model. The decision module 605 is used to output repair suggestion information by the repair decision model based on the evaluation results, wherein the evaluation results include: a sequence of anchor points ordered according to the crack propagation rate and confidence level of the damage at each anchor point.

[0100] Optionally, the preprocessing module 602 is specifically used for: The heterogeneous pairing momentum contrast coding network (HPC-MoCo) is used to perform multimodal collaborative coding on multi-source multi-scale data to obtain the encoded multi-source multi-scale data. Based on the encoded multi-source, multi-scale data, the damage characteristics of each damage are obtained by comparing and learning the momentum-updated teacher encoder with the cross-modal shared negative sample queue strategy. Confidence prediction is performed based on the damage characteristics of each damage to obtain the confidence level of each damage, and the quadruple features of each damage are obtained based on the damage characteristics and the confidence level of each damage.

[0101] Optionally, the construction module 603 is specifically used for: The ontology corresponding to the damage is determined based on the four-tuple characteristics of the damage. The ontology includes: cultural relics, components, and anchor points. The cultural relics are whole individuals, the components are stable structural units, and the anchor points are micro-area entities. The environmental parameters of each artifact are embedded into the damaged ontology to obtain a semantic-physical coupled ontology. The type of the coupled ontology is determined using a preset mechanical constraint strategy. A knowledge graph of artifacts is constructed based on each coupled ontology and its type.

[0102] Optionally, the types of the coupled ontology include: state class, process class, action class, and event class. The coupled ontology of the state class is used to characterize the state of damage in the anchor point corresponding to the ontology at a preset time. The coupled ontology of the process class is used to characterize the damage evolution process of damage in the anchor point corresponding to the ontology. The coupled ontology of the action class is used to characterize the external influencing factors of damage in the anchor point corresponding to the ontology. The coupled ontology of the event class is used to characterize the triggering event of damage in the anchor point corresponding to the ontology.

[0103] Optionally, the mechanically constrained strategy includes: By utilizing the uniqueness constraint of a single time and the closure constraint of the measure, the state of the damage at each time step in the anchor point corresponding to the coupled ontology is bound to the coupled ontology; By utilizing time monotonicity constraints, path additivity constraints, and residual upper bound constraints, the states of the coupled entity at adjacent or non-adjacent times are connected to obtain the state evolution information of the coupled entity. Threshold constraints are used to determine the external event triggering factors of the coupled ontology.

[0104] Optionally, the training module 604 is specifically used for: The cultural relic knowledge graph is processed by a relation-aware graph convolutional network to obtain the message flow corresponding to each anchor point in the cultural relic knowledge graph, and the tensor information of each anchor point is obtained based on the message flow. The tensor information of each anchor point is calibrated at the equation level using a differentiable physics layer to obtain calibrated tensor information. The target prediction model is obtained by performing multi-step evolution prediction and uncertainty modeling on the calibrated tensor information using an event-gated time series algorithm.

[0105] Optionally, the decision module 605 is specifically used for: Calculate the upper confidence boundary risk index of each anchor point, and sort the anchor points in descending order according to the upper confidence boundary risk index of each anchor point to obtain the disposal candidate set; For each anchor point in the candidate disposal set, the environmental parameters in the cultural relic knowledge graph are spliced ​​together to obtain each constrained anchor point; Decision analysis is performed sequentially on each constrained anchor point in the candidate disposal set to obtain repair suggestion information corresponding to each anchor point.

[0106] Figure 13 This is a structural block diagram of an electronic device 700 provided in an embodiment of this application. (See diagram below.) Figure 13 As shown, the electronic device may include: a processor 701 and a memory 702.

[0107] Optionally, a bus 703 may also be included, wherein the memory 702 is used to store machine-readable instructions executable by the processor 701. When the electronic device 700 is running, the processor 701 and the memory 702 communicate via the bus 703. When the machine-readable instructions are executed by the processor 701, the method steps in the above method embodiments are performed.

[0108] This application also provides a computer-readable storage medium storing a computer program, which, when run by a processor, executes the method steps described in the above-described method for assessing the damage morphology of cultural relics.

[0109] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems and devices described above can be referred to the corresponding processes in the method embodiments, and will not be repeated here. In the several embodiments provided in this application, it should be understood that the disclosed systems, devices, and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the displayed or discussed mutual coupling or direct coupling or communication connection can be through some communication interfaces; the indirect coupling or communication connection of devices or modules can be electrical, mechanical, or other forms.

[0110] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. If the functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes: USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, optical disks, and other media capable of storing program code.

[0111] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for assessing the damage morphology of cultural relics, characterized in that, The method includes: Acquire multi-source, multi-scale data of cultural relics at different times. The multi-source, multi-scale data includes point cloud data, image data, and text data of the cultural relics. The text data includes material spectrum, material texture, environmental parameters, and user interaction data. The environmental parameters are the temperature, humidity, light, and pollution index of the cultural relics at a certain time. The multi-source, multi-scale data of each of the cultural relics are subjected to collaborative preprocessing and comparative learning to obtain the quadruple features of each damage in each of the cultural relics. The quadruple features include the geometric tensor of deformation-crack, material-chemical damage characterization, environment-interaction context embedding, and confidence level. A cultural relic knowledge graph is constructed based on the quadruple features of each of the aforementioned damages, and the cultural relic knowledge graph is aligned in time and space. A target evaluation model is trained based on the cultural relic knowledge graph. An evaluation result of the cultural relic to be evaluated is obtained based on the target evaluation model. The evaluation result is sent to the restoration decision model. The restoration decision model outputs restoration suggestions based on the evaluation result. The evaluation result includes: an anchor point sequence sorted according to the crack propagation rate and confidence level of the damage at each anchor point. The construction of a cultural relic knowledge graph based on the quadruple features of each type of damage includes: The ontology corresponding to the damage is determined based on the four-tuple characteristics of the damage. The ontology includes: cultural relics, components, and anchor points. The cultural relics are whole individuals, the components are stable structural units, and the anchor points are micro-area entities. The environmental parameters of each cultural relic are embedded into the ontology of the damage to obtain a semantic-physical coupled ontology. The type of the coupled ontology is determined by a preset mechanical constraint strategy. A cultural relic knowledge graph is constructed based on each coupled ontology and the type of each coupled ontology. The mechanically constrained strategy includes: By utilizing the single temporal uniqueness constraint and the metric closure constraint, the state of the damage at each time point in the anchor point corresponding to the coupled ontology is bound to the coupled ontology. The single temporal uniqueness constraint is used to ensure that any object has at most one normalized state at a given timestamp, and the metric closure constraint is used to ensure that the normalized state carries the quadruple feature and physical dimension label. By utilizing time monotonicity constraints, path additivity constraints, and residual upper bound constraints, the states of the coupled entity at adjacent or non-adjacent times are connected to obtain the state evolution information of the coupled entity. The residual upper bound is automatically generated by the PINN mechanical residual threshold. The external event triggering factors of the coupled ontology are determined using threshold constraints; The target evaluation model trained based on the cultural relic knowledge graph includes: The cultural relic knowledge graph is processed by a relation-aware graph convolutional network to obtain the message flow corresponding to each anchor point in the cultural relic knowledge graph, and the tensor information of each anchor point is obtained based on the message flow. The tensor information of each anchor point is calibrated at the equation level using a differentiable physics layer to obtain calibrated tensor information. The target prediction model is obtained by performing multi-step evolution prediction and uncertainty modeling on the calibrated tensor information using an event-gated time series algorithm.

2. The method for assessing the damage morphology of cultural relics according to claim 1, characterized in that, The collaborative preprocessing and comparative learning of multi-source, multi-scale data of each of the aforementioned cultural relics yields quadruple features of each type of damage in each of the cultural relics, including: Heterogeneous paired momentum contrast coding network is used to perform multimodal collaborative coding processing on multi-source multi-scale data to obtain encoded multi-source multi-scale data. The heterogeneous paired momentum contrast coding network consists of an image-material encoder and a geometric encoder. The image-material encoder is used to encode and analyze the two-dimensional data in the multi-source multi-scale data, and the geometric encoder is used to encode and analyze the three-dimensional data in the multi-source multi-scale data. Based on the encoded multi-source, multi-scale data, the damage characteristics of each damage are obtained by comparing and learning the momentum-updated teacher encoder with the cross-modal shared negative sample queue strategy. Confidence prediction is performed based on the damage characteristics of each damage to obtain the confidence level of each damage, and the quadruple features of each damage are obtained based on the damage characteristics and the confidence level of each damage.

3. The method for assessing the damage morphology of cultural relics according to claim 1, characterized in that, The types of coupled ontologies include: state class, process class, action class, and event class. The state class coupled ontologies are used to characterize the state of damage in the anchor point corresponding to the ontology at a preset time. The process class coupled ontologies are used to characterize the damage evolution process of damage in the anchor point corresponding to the ontology. The action class coupled ontologies are used to characterize the external influencing factors of damage in the anchor point corresponding to the ontology. The event class coupled ontologies are used to characterize the triggering event of damage in the anchor point corresponding to the ontology.

4. The method for assessing the damage morphology of cultural relics according to claim 1, characterized in that, The repair decision model outputs repair suggestion information based on the output results, including: Calculate the upper confidence boundary risk index of each anchor point, and sort the anchor points in descending order according to the upper confidence boundary risk index of each anchor point to obtain the disposal candidate set; For each anchor point in the candidate disposal set, the environmental parameters in the cultural relic knowledge graph are spliced ​​together to obtain each constrained anchor point; Decision analysis is performed sequentially on each constrained anchor point in the candidate disposal set to obtain repair suggestion information corresponding to each anchor point.

5. A device for assessing the damage morphology of cultural relics, characterized in that, include: The acquisition module is used to acquire multi-source, multi-scale data of cultural relics at different times. The multi-source, multi-scale data includes point cloud data, image data, and text data of the cultural relics. The text data includes material spectrum, material texture, and environmental parameters. The preprocessing module is used to perform collaborative preprocessing and comparative learning on the multi-source, multi-scale data of each of the cultural relics to obtain the quadruple features of each damage in each of the cultural relics. A construction module is used to construct a cultural relic knowledge graph based on the quadruple features of each of the aforementioned damages, wherein the cultural relic knowledge graph is time- and space-aligned; The training module is used to train a target evaluation model based on the cultural relic knowledge graph, obtain the evaluation result of the cultural relic to be evaluated based on the target evaluation model, and send the evaluation result to the restoration decision model. The decision module is used to output repair suggestions based on the assessment results from the repair decision model. The assessment results include an anchor point sequence sorted according to the crack propagation rate and confidence level of the damage at each anchor point. The building module is specifically used for: The ontology corresponding to the damage is determined based on the four-tuple characteristics of the damage. The ontology includes: cultural relics, components, and anchor points. The cultural relics are whole individuals, the components are stable structural units, and the anchor points are micro-area entities. The environmental parameters of each cultural relic are embedded into the ontology of the damage to obtain a semantic-physical coupled ontology. The type of the coupled ontology is determined by a preset mechanical constraint strategy. A cultural relic knowledge graph is constructed based on each coupled ontology and the type of each coupled ontology. The mechanically constrained strategy includes: By utilizing the single temporal uniqueness constraint and the metric closure constraint, the state of the damage at each time point in the anchor point corresponding to the coupled ontology is bound to the coupled ontology. The single temporal uniqueness constraint is used to ensure that any object has at most one normalized state at a given timestamp, and the metric closure constraint is used to ensure that the normalized state carries the quadruple feature and physical dimension label. By utilizing time monotonicity constraints, path additivity constraints, and residual upper bound constraints, the states of the coupled entity at adjacent or non-adjacent times are connected to obtain the state evolution information of the coupled entity. The residual upper bound is automatically generated by the PINN mechanical residual threshold. The external event triggering factors of the coupled ontology are determined using threshold constraints; The training module is specifically used for: The cultural relic knowledge graph is processed by a relation-aware graph convolutional network to obtain the message flow corresponding to each anchor point in the cultural relic knowledge graph, and the tensor information of each anchor point is obtained based on the message flow. The tensor information of each anchor point is calibrated at the equation level using a differentiable physics layer to obtain calibrated tensor information. The target prediction model is obtained by performing multi-step evolution prediction and uncertainty modeling on the calibrated tensor information using an event-gated time series algorithm.

6. An electronic device, characterized in that, It includes a memory and a processor, wherein the memory stores a computer program executable by the processor, and the processor executes the computer program to implement the steps of the method for assessing the damage morphology of cultural relics as described in any one of claims 1-4.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, performs the steps of the method for assessing the morphology of cultural relic damage as described in any one of claims 1-4.

Images