CN-122020199-A - Deep learning-based PC component BIM (binary association modeling) collaborative exchange method

CN122020199ACN 122020199 ACN122020199 ACN 122020199ACN-122020199-A

Abstract

The invention discloses a PC component BIM twinning collaborative exchange method based on deep learning, which aims to solve the problems that unstructured multi-mode data of a site or a factory is difficult to automatically identify the identity, the state and the quality of a PC component, accurately map with BIM twinning objects one by one and difficultly form incremental update which can be exchanged in a multiparty collaborative way, analyzing the BIM twin model to obtain the unique identification of the component and the prior characteristics such as the component type, the dimensional parameter, the spatial position, the component geometry and the like, merging the cross-modal characteristics under the guidance of the first test to generate a candidate matching relation, constructing an optimal transmission cost matrix containing characteristic differences and constraint penalties to obtain a soft matching matrix, adopting the matching confidence rejection and assignment solution to obtain one-to-one mapping, writing the identification result and the evidence data index into the twin object and carrying out component level difference generation evidentiary increment exchange data with the last version, and realizing the technical effects of high-reliability automatic updating and traceable collaborative exchange of the component level twin model.

Inventors

HU ZHIGANG
FENG ZHEN
ZHANG XUCHEN
FEI RAN

Assignees

中交投资南京有限公司

Dates

Publication Date: 20260512
Application Date: 20260129

Claims (10)

1. A PC component BIM twinning collaborative exchange method based on deep learning comprises the following steps: S1, unstructured data of a target scene are obtained, and preprocessing is carried out to obtain multi-mode input data; S2, carrying out instance recognition and feature extraction on PC components through a component recognition deep learning model by using multi-modal input data to obtain first component characteristics, component state information and component quality information of each identified PC component, extracting evidence data corresponding to each identified PC component from the multi-modal input data, S3, acquiring a BIM twin model corresponding to a target scene, extracting BIM components, extracting each component unique identification and BIM prior information, generating BIM prior characteristics according to the BIM prior information, S4, carrying out cross-modal fusion processing on the first component characteristics under the guidance of the BIM prior characteristics to obtain fusion component characteristics, generating candidate matching relation between the identified PC components and the BIM components according to the BIM prior information and combining preset constraint conditions, S5, constructing an optimal transmission cost matrix according to the fusion component characteristics, the BIM prior characteristics and the candidate matching relation, carrying out optimal transmission by jointly determining a characteristic difference item of the fusion component characteristics and the BIM prior characteristics and constraint penalty item corresponding to the preset constraint conditions to obtain a soft matching matrix, S6, calculating the unique identification characteristics of each identified component according to the soft matching matrix, carrying out one-to-one mapping on the corresponding to the twin components, generating a threshold value according to the BIM prior characteristics, and carrying out one-to-one mapping to the corresponding to the soft matching parameters, and carrying out the mapping to the unique matching information to obtain a threshold value, and giving a unique matching result to the best matching result to the identification by the PC component, and S8, carrying out differential calculation on the component unique identification granularity of the twin object update value and the to-be-rechecked mark and the twin data of the last version by using the twin data of the last version of the BIM twin model to obtain component-level increment update information, and generating increment exchange data for multiparty collaborative exchange.
2. The deep learning-based PC component BIM twinning collaborative exchange method according to claim 1, S1 includes: the unstructured data comprises image data, video data and point cloud data; Collecting the image data, the video data and the point cloud data in a target scene, and respectively recording a collection time stamp for each frame of the image data, each frame of the video data and each frame of the point cloud data; Establishing a unified time reference according to the acquisition time stamp, and performing time synchronization on the image data, the video data and the point cloud data to enable the frames of the image data, the video data and the point cloud data to form a corresponding relationship under the unified time reference; Performing coordinate calibration on the image data, the video data and the point cloud data according to a preset engineering coordinate system, determining a coordinate transformation relation between the point cloud coordinate system and the preset engineering coordinate system, and transforming the point cloud data to the preset engineering coordinate system; and respectively performing noise filtering processing on the image data, the video data and the point cloud data which finish time synchronization and coordinate calibration to obtain multi-mode input data, wherein the multi-mode input data is the image data, the video data and the point cloud data which finish time synchronization, coordinate calibration and noise filtering processing.
3. The deep learning-based PC component BIM twinning collaborative exchange method according to claim 1, S2 includes: The method comprises the steps of obtaining image data and video data, respectively carrying out PC component instance segmentation on the image data and the video data frame by using the multi-modal input data through a component recognition deep learning model to obtain component segmentation masks of each frame, carrying out PC component instance segmentation on the point cloud data to obtain point cloud component segmentation masks, establishing a corresponding relation between the component segmentation masks and the point cloud component segmentation masks according to a coordinate calibration result to obtain a multi-modal component segmentation result, carrying out component region feature convergence on features generated by the component recognition deep learning model according to the multi-modal component segmentation result to obtain first component features of each recognized PC component, generating component state information of each recognized PC component through a component state prediction network according to the first component features, and generating component quality information of each recognized PC component through a component quality prediction network according to the first component features; The component local image is an image segment which is cut out according to the corresponding region of the component segmentation mask in the image data, the component local video segment is a video segment which is cut out according to the corresponding region of the component segmentation mask in the video data and comprises continuous multiframes, and the component local point cloud segment is a point cloud segment which is cut out according to the corresponding point set of the point cloud component segmentation mask in the point cloud data.
4. The deep learning-based PC component BIM twinning collaborative exchange method according to claim 1, S3 includes: acquiring a BIM twin model corresponding to the target scene, and analyzing the BIM twin model to extract a plurality of BIM components; Reading a component unique identifier for each BIM component, and extracting component type information, size parameter information, spatial position information under a preset engineering coordinate system and component geometric information which are used as BIM prior information; And carrying out category coding on the component type information according to the BIM prior information, carrying out numerical normalization processing on the size parameter information and the space position information, generating geometric description parameters on the component geometric information, and generating the BIM prior feature for representing the BIM component.
5. The deep learning-based PC component BIM collaborative exchange method according to claim 1, S4 includes: According to the BIM prior information and combining with a preset constraint condition, generating a candidate matching relation between each identified PC component and each BIM component, wherein the candidate matching relation calculates feature similarity between the fusion component feature and the BIM prior feature, and the first K BIM components are selected from high to low according to the feature similarity in the BIM components meeting the preset constraint condition, K is a preset positive integer, and the candidate matching relation meets the preset constraint condition, and the preset constraint condition comprises at least one of component type constraint, size parameter constraint and spatial position constraint.
6. The deep learning-based PC component BIM collaborative exchange method according to claim 1, S5 includes: Calculating a feature difference term according to the fusion component features and the BIM prior features aiming at each BIM component in the candidate matching relation of each identified PC component and the corresponding identified PC component; Respectively calculating at least one of a component type constraint penalty value, a size parameter constraint penalty value and a space position constraint penalty value according to the preset constraint condition to obtain a constraint penalty item; the characteristic difference item and the constraint penalty item are weighted and summed to obtain an optimal transmission cost value, and an optimal transmission cost matrix is constructed according to the optimal transmission cost value; Setting matrix elements which do not belong to the candidate matching relationship as preset large cost values or shielding values; and executing optimal transmission solution on the optimal transmission cost matrix, wherein the optimal transmission solution adopts Sinkhorn iteration solution with entropy regularization terms to obtain a soft matching matrix which represents the matching probability between each identified PC component and each BIM component.
7. The deep learning-based PC component BIM twinning collaborative exchange method according to claim 1, S6 includes: determining a matching probability corresponding to each BIM component according to the soft matching matrix, and determining the maximum value of the matching probabilities as the matching confidence of the identified PC component; comparing the matching confidence coefficient with a preset confidence coefficient threshold value, and generating refused identification information for the identified PC component with the matching confidence coefficient lower than the preset confidence coefficient threshold value; rejecting the rows corresponding to the rejected identified PC components from the soft matching matrix according to the rejection identification information to obtain a matching probability submatrix for assignment solution; And executing assignment solving according to the matching probability submatrices, wherein the assignment solving adopts a Hungary algorithm or linear programming solving to generate a matching relation meeting one-to-one constraint to obtain a one-to-one mapping result, and the one-to-one mapping result is the corresponding relation between each unapproved identified PC component and the unique component identifier of the corresponding BIM component.
8. The deep learning-based PC component BIM collaborative exchange method according to claim 1, S7 includes: The method comprises the steps of locating a twin object corresponding to a unique component identifier in a BIM twin model according to a one-to-one mapping result, writing component state information into a component state field of the twin object for each unrecognized identified PC component, writing component quality information into a component quality field of the twin object, writing matching confidence into a matching confidence field of the twin object, generating a twin object update value, associating the refused identified PC component to a BIM space partition object pre-established in the BIM twin model according to refused identification information based on the spatial position of the refused identified PC component under a preset engineering coordinate system, and writing an update value containing a mark to be rechecked and an evidence data index corresponding to the refused identified PC component in the BIM space partition object, wherein the BIM space partition object comprises at least one of a floor object, a region partition object or a component container object.
9. The deep learning-based PC component BIM collaborative exchange method according to claim 1, S8 includes: Reading the twin data of the previous version of the BIM twin model, and aligning the twin object updated value and the updated value of the mark to be rechecked with the twin data of the previous version according to the unique identification of the component; the unique identifiers of the components after alignment are respectively compared with the field value before modification and the field value after modification, modification fields are determined, modification values corresponding to the modification fields are generated, and component level increment updating information is obtained; selecting evidence data corresponding to the change field from the evidence data according to the component level increment updating information, and calculating a hash check value for the evidence data; And generating incremental exchange data for multiparty collaborative exchange according to a preset exchange data format, wherein the incremental exchange data comprises a component unique identifier, a change field, a change value, a time stamp, a matching confidence degree, evidence data and the hash check value.
10. The deep learning based PC component BIM collaborative switching method according to claim 6, wherein the optimal transmission solution employs unbalanced optimal transmission or partially optimal transmission, such that the soft matching matrix allows for the presence of identified PC components or BIM components with unassigned quality to characterize at least one of newly added components, missed components, or duplicate identified components.

Description

Deep learning-based PC component BIM (binary association modeling) collaborative exchange method Technical Field The invention relates to the field of building informatization, in particular to a PC component BIM (building information modeling) twin collaborative exchange method based on deep learning. Background Production, transportation, yard management and field installation of precast concrete units in fabricated buildings typically rely on building information models for planning, positioning and quality traceability. Along with the development of a building information model and a digital twin technology, the industry is gradually managed by a manual standing book and a two-dimensional drawing, component level information management taking the building information model as a core is developed, and sensing means such as image acquisition, video acquisition, laser point cloud scanning and the like are introduced to acquire actual state data of a site or a factory. Meanwhile, deep learning is mature in aspects of target detection, instance segmentation and three-dimensional point cloud understanding, so that the method can automatically identify components from unstructured data and write back twin models, and in the aspect of multiparty collaboration, application requirements for cross-party synchronization based on model files or exchange data are gradually formed. The prior art has the defects when the unstructured sensing data is automatically converted into the component-level twin increment update which can be cooperatively exchanged, and mainly comprises the following steps: Firstly, the problems of time asynchronism, inconsistent coordinates, noise, shielding and the like exist among the multi-mode data, so that component instance identification and characteristic characterization are unstable, and component identity identification, state identification and quality judgment are difficult to be simultaneously considered; Secondly, the one-to-one mapping between the identified components and the building information model components usually depends on manual confirmation or greedy matching based on feature similarity, lacks a global optimal matching mechanism which utilizes prior constraints such as component types, dimensional parameters, spatial positions and the like, is easy to cause mismatching, reconfiguration and missed matching, and lacks rejection and rechecking treatment on low-confidence results; and the twin information updating and the collaborative exchange are mostly realized by adopting a whole model coverage type updating or coarse granularity synchronization mode, so that the semantic difference increment exchange of the granularity of the unique identification of the component is difficult to realize, and evidence data and a verification mechanism corresponding to the changed content are also lacking, so that the collaborative efficiency is low and the traceability is insufficient. Therefore, a method of bi-twinning collaborative exchange of a PC component, which solves the above-mentioned deficiencies of the prior art, is a problem that needs to be solved by those skilled in the art. Disclosure of Invention Aiming at the problems that in the prior art, image data, video data and point cloud data of a site or a factory are difficult to automatically identify component identities and states and quality thereof, and component-level incremental data which can be cooperatively exchanged are difficult to form by updating results and are difficult to accurately map one by one with a BIM twinning model component, the invention provides a technical scheme for carrying out time synchronization and coordinate calibration preprocessing on multi-modal data, carrying out component instance identification and feature extraction by using a deep learning model and obtaining evidence data, analyzing the prior feature of the BIM twinning model to generate the component, carrying out cross-modal fusion under the guidance of the prior experiment and generating a candidate matching relation based on constraint conditions, constructing an optimal transmission cost matrix containing a feature difference item and a constraint penalty item to obtain a soft matching matrix, carrying out rejection and one-to-one assignment solution based on the matching confidence, writing the states, the quality and the confidence to the twinning object and carrying out semantic difference of component unique identification granularity based on the last version to generate evidence incremental exchange data. The invention provides a deep learning-based PC component BIM (binary association exchange) collaborative exchange method, which comprises the following steps: S1, unstructured data of a target scene are obtained, and preprocessing is carried out to obtain multi-mode input data; S2, carrying out instance recognition and feature extraction on PC components through a component recognition deep learning model by using m