A Non-Interface-Based Secondary Collision Detection Method for BIM Models Based on Artificial Intelligence
By constructing a dual-branch fusion AI recognition network to automatically filter false collisions and classify true collisions, the problems of false collision misjudgment and large workload of review in native BIM software detection are solved, achieving efficient and accurate collision detection and report generation.
Patent Information
- Application Number
- CN202610911492.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2026-06-24
- Publication Date
- 2026-11-10
- Estimated Expiration
- 2046-06-24
AI Technical Summary
Existing BIM software's native collision detection cannot distinguish between compliant false collisions and real collisions that affect construction. Manual review is labor-intensive and complex, making it unsuitable for complex power engineering scenarios. Furthermore, the generation of collision reports is inconsistent, adding extra workload.
A non-interface-based secondary collision detection method for BIM models based on artificial intelligence is adopted. By constructing a dual-branch fusion AI recognition network, false collisions are automatically filtered and true collisions are classified, generating standardized reports. This includes constructing a dataset for the target engineering field, training the AI recognition network, image preprocessing, fusion decision-making, and automatic report generation.
It achieves a false collision filtering rate of ≥95% and a true collision recognition accuracy of ≥90%, significantly reducing the workload of manual review, shortening the review cycle, improving the efficiency and accuracy of design review, and generating standardized reports.
Smart Images

Figure CN122473657B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of BIM engineering design and artificial intelligence computer vision technology, and in particular to a non-interface-based secondary collision detection method for BIM models based on artificial intelligence. Background Technology
[0002] In multi-disciplinary collaborative design using BIM in power engineering, the native clash detection function built into mainstream BIM software such as PDMS and Revit is currently the core tool for clash detection in the engineering design phase. However, existing native clash detection and subsequent verification technologies have many unresolved core pain points: First, the native collision detection only determines collisions based on geometric space intersection rules, which cannot distinguish between compliant contact collisions and invalid collisions that actually affect construction. This leads to a large number of compliant false collisions in power engineering scenarios, such as contact between supports and pipes, contact between supports and structures, and side point contact between pipes, being misjudged. Secondly, native collision detection can only output the basic result of "whether there is a collision" and cannot automatically distinguish between real and false collisions or classify real collisions by type. Designers need to manually review, screen and classify thousands of collision results one by one. The collision result review work of complex power engineering projects takes 10-15 working days. The workload is huge and manual review has the problems of strong subjectivity and high error rate. Third, existing collision result verification technologies are mostly interface-based tools that can only filter out a small number of false collisions by using fixed rules such as preset component types and spacing thresholds. They cannot adapt to the complex and ever-changing collision scenarios in power engineering, and their filtering effect is limited. Fourth, after the existing technology completes the collision verification, it is still necessary to manually sort out the valid collision results, capture scene images, collect data and fill in the collision report, which adds a lot of manual workload. In addition, there are problems such as inconsistent report formats, data errors and omissions, and non-standard screenshots. It is impossible to achieve a fully automated closed loop from collision result verification to report delivery. Summary of the Invention
[0003] To address the shortcomings of existing technologies, this invention provides a non-interface-based secondary collision detection method for BIM models based on artificial intelligence. This method is specifically designed to perform secondary intelligent verification of the results output by the native collision detection of BIM software, achieving automated and accurate filtering of false collisions, intelligent classification of true collisions, and simultaneous automatic generation of standardized collision reports. This constructs a fully automated solution from collision verification to deliverables.
[0004] The technical solution adopted in this invention, a non-interface-based secondary collision detection method for BIM models based on artificial intelligence, is as follows: A non-interface-based secondary collision detection method for BIM models based on artificial intelligence includes the following steps: S1. Construct a dedicated dataset for secondary collision detection in the target engineering field using BIM, and train a dual-branch fusion AI recognition network based on the dataset. S2. Obtain the native collision detection result list from the BIM software. Based on the list, batch collect standardized view images of each collision point and perform image preprocessing. These standardized view images are images with clearly defined collision points, the BIM model angles are commonly used in production environments, the collision types are manually identifiable, and the image size conforms to the corresponding model training image size (YOLO typically uses...). ); S3. The preprocessed single collision point image is synchronously input into the dual-branch fusion AI recognition network. The network outputs the classification probability results of collision validity and collision type through the attribute classification branch, and outputs the positioning information of the collision core region and auxiliary classification probability results through the region verification branch. S4. Perform a fusion decision on the output results of the two branches to obtain the final secondary detection result of the collision point, automatically filter out the results judged as false collisions, and retain the association information judged as valid true collisions. S5. After traversing all native collision results to complete full detection, integrate the correlation information of all valid true collisions and automatically generate a standardized collision detection report.
[0005] A further improvement to the technical solution of this invention is that: the construction of a dedicated dataset for secondary collision detection in BIM in step S1 specifically includes: setting a standardized labeling system that includes false collision classes and true collision classes; collecting collision point samples from the native collision detection output of BIM software in actual projects in the target engineering field, collecting standardized view images of each collision point and labeling the true and false collision attributes, collision type and collision core area boundary box; cleaning and augmenting the labeled samples, constructing an effective labeled sample set and dividing it into a training set, a validation set and a test set according to a preset ratio.
[0006] A further improvement of the technical solution of the present invention is that: the dual-branch fusion AI recognition network includes a parallel-connected true / false collision classification branch, a collision region verification branch, and a fusion decision layer connected to the outputs of both branches simultaneously; wherein, the true / false collision classification branch is a classification network optimized based on the ResNet residual neural network, the collision region verification branch is a detection network optimized based on the YOLOv11 object detection network, and the fusion decision layer is used to perform fusion calculation on the output results of the two branches.
[0007] A further improvement of the technical solution of the present invention is that: the true and false collision classification branch optimizes the feature extraction layer of the ResNet network to meet the requirements of extracting subtle features of contact collisions, and uses the FocalLoss loss function to solve the problem of sample imbalance. The network input is the BIM model view image corresponding to the collision point, and the output is the classification probability results of each category under the standardized label system.
[0008] A further improvement of the technical solution of the present invention is that: the collision region verification branch optimizes the anchor box parameters and small target feature fusion layer of the YOLOv11 network for the localization requirements of the collision core contact area, and adopts the CIoU loss function as the bounding box regression loss function. The network input is the BIM model view image corresponding to the collision point, and the output is the bounding box coordinates, confidence and auxiliary classification probability results of the collision core area.
[0009] A further improvement of the technical solution of the present invention is that: the fusion decision layer adopts a weighted voting method for fusion calculation, sets the weight of the attribute classification branch output result as W1, and the weight of the region verification branch auxiliary classification result as W2, where W1+W2=1 and W1>W2, and outputs the final collision validity judgment result, collision type and collision core region coordinates.
[0010] A further improvement to the technical solution of this invention is that: the training of the dual-branch fusion AI recognition network in step S1 specifically includes: loading the general pre-trained weights of the ResNet network and the YOLOv11 network, and training them using transfer learning; adopting a two-stage training mode, in the first stage independently training the true and false collision classification branch and the collision region verification branch until the accuracy converges, and in the second stage fixing the parameters of the two branch backbone networks and jointly training the fusion decision layer until the false collision recognition accuracy reaches the preset index; using a test set to verify the model accuracy, and performing incremental training optimization for collision types whose accuracy does not meet the standard.
[0011] A further improvement of the technical solution of the present invention is that: in step S2, the native collision detection result list of the BIM software includes the ID, corresponding component, spatial location information and professional affiliation of each collision point; the image preprocessing specifically involves normalizing the size and standardizing the pixel values of the acquired BIM model view image to make it conform to the input requirements of the dual-branch fusion AI recognition network.
[0012] A further improvement of the technical solution of the present invention is that: in step S4, the associated information of the effective true collision includes the collision type, the coordinates of the collision core region, the corresponding standardized view image, and the professional information to which it belongs.
[0013] A further improvement of the technical solution of the present invention is that: in step S5, a standardized collision detection report is automatically generated, specifically: a report template that conforms to the enterprise delivery standard is preloaded, and the collision type, collision location, volume number and component number of the valid true collision are automatically filled into the corresponding positions of the report template. After the filling is completed, a standardized collision detection report in DOCX format is automatically exported.
[0014] The technological advancements achieved by this invention due to the adoption of the above technical solutions are as follows: This invention specifically performs secondary intelligent verification of the native collision detection results of BIM software. According to engineering measurements, the false collision filtering rate is ≥95% and the true collision recognition accuracy is ≥90%. It can directly reduce the workload of manual verification by more than 90%, and compress the collision verification cycle of complex projects from 10-15 working days to a few hours, greatly improving the efficiency and accuracy of BIM design review.
[0015] This invention adopts a dual-model fusion architecture, which performs specific optimizations on the subtle features of contact-type pseudo-collisions and the small target area of the collision core. It solves the problems of traditional rule-based filtering schemes being unable to adapt to complex engineering scenarios and having poor filtering effects, and has strong generalization ability.
[0016] This invention adopts a non-interface design, which does not require connecting to the native interface of BIM software or reading the underlying geometric data of the BIM model. It can directly perform secondary detection on the collision results of all mainstream BIM software such as Revit and PDMS, without changing the operating habits of designers, and the threshold for implementation is extremely low.
[0017] This invention can simultaneously filter out false collisions and accurately classify and locate the core area of real collisions, providing a precise basis for subsequent collision rectification and improving the precision of engineering design.
[0018] This invention enables the automatic generation of standardized collision reports, constructing a fully automated closed loop from importing native collision results, AI secondary detection, false collision filtering to standardized report delivery. The report generation accuracy is 100%, further reducing the manual workload of later delivery by more than 80%, and ensuring the standardization and uniformity of the delivered results. Attached Figure Description
[0019] Figure 1 This is a flowchart of a non-interface-based secondary collision detection method for BIM models based on artificial intelligence, according to the present invention. Figure 2 This is a schematic diagram of the label partitioning process in the dataset construction process of this invention; Figure 3 This is a schematic diagram of the architecture of the dual-branch fusion AI recognition network for secondary detection according to the present invention; Figure 4This is a data flow diagram of the secondary detection process of the present invention; Detailed Implementation
[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings. In the following description, descriptions of well-known structures and technologies are omitted to avoid unnecessarily obscuring the concept of this invention. Example 1
[0021] like Figure 1 As shown, this embodiment provides a non-interface-based secondary collision detection method for BIM models based on artificial intelligence. The overall process is as follows: "Dataset construction → Dual-branch network training → Collision result acquisition and image preprocessing → Dual-branch AI recognition → Fusion decision and result output → Automatic generation of standardized report". The specific implementation steps are as follows: 1. Construct a dedicated dataset for secondary collision detection in BIM for power engineering. 1) Establish a standardized labeling system. To address the verification needs of native collision detection results in power engineering, define four core false collision labels: supports and pipes, supports and structures, pipe side points, and other false collisions; define five core valid true collision labels: pipes-structure, pipes-building, equipment-structure, equipment-building, and other true collisions, forming nine standardized detection labels. The labeling process is as follows: Figure 2 As shown; 2) Collect collision point samples from the native collision detection output of BIM software in actual power engineering projects, collect BIM model view images from a standardized perspective for each collision point, and accurately label each image. The labeling content includes true and false collision attribute labels, collision type labels, and collision core area boundary box coordinate labels. 3) Clean the labeled samples and remove invalid samples that are blurry or incorrectly labeled; In view of the imbalance between false collision and true collision samples, data augmentation is performed by flipping, scaling, brightness adjustment and local occlusion simulation to expand the sample size and build a secondary detection dataset of no less than 2,000 sets of valid labeled samples, of which false collision samples account for no less than 50%, which is consistent with the distribution of actual engineering scenarios. 4) Divide the dataset into training set, validation set and test set in a ratio of 7:2:1, and use them for model training, accuracy verification and performance testing, respectively.
[0022] 2. Construct a dual-branch fusion AI recognition network architecture for secondary detection. Network architecture diagram as follows Figure 3 As shown.
[0023] 1) Construct a true / false collision classification branch based on ResNet residual neural network. The network input is the BIM model view image corresponding to the collision point, and the output is the classification probability result of 9 categories of labels. The core output is the true / false collision attribute determination result. In order to meet the core requirement of false collision recognition in secondary detection scenarios, the feature extraction layer of ResNet network is optimized to enhance the ability to extract subtle features of contact collision. FocalLoss loss function is used to solve the sample imbalance problem and improve the recognition accuracy of false collision. 2) Construct a collision region verification branch based on the YOLOv11 object detection network. The network input is the BIM model view image corresponding to the collision point, and the output is the bounding box coordinates, confidence score, and auxiliary classification probability results of the collision core region. For secondary detection scenarios, optimize the anchor box parameters and small target feature fusion layer of the YOLO network to accurately locate the collision core contact area and assist in verifying the effectiveness of the collision. The CIoU loss function is used as the bounding box regression loss function. 3) At the output of the two branches, a secondary detection fusion decision layer based on weighted voting method is constructed. The weight of the true and false collision classification branch is set as W1, and the weight of the auxiliary classification result of the collision region verification branch is set as W2, where W1+W2=1 and W1>W2. Through weighted fusion calculation, the final secondary detection result is output, including collision validity determination (true collision / false collision), collision type, and collision core region coordinates.
[0024] 3. Training and Optimization of a Dual-Branch Fusion AI Network for Secondary Detection 1) Load the pre-trained weights of the ResNet network on the ImageNet dataset and the pre-trained weights of the YOLOv11 network on the COCO dataset, and use transfer learning to train the model to reduce the risk of overfitting in small sample scenarios. 2) A two-stage training mode is adopted: In the first stage, the true and false collision classification branch and the collision region verification branch are trained independently. The network parameters of the two branches are optimized using training set data, and the hyperparameters are adjusted using validation set data until the accuracy of the two branches on the validation set converges. In the second stage, the fusion decision layer is jointly trained. The backbone network parameters of the two branches are fixed, and the weighted weights W1 and W2 are optimized until the false collision recognition accuracy of the fusion reaches the preset target. 3) Use test set data to verify the accuracy of the trained dual-model fusion network. The core verification metrics include false collision filtering rate, true collision recognition accuracy, and single collision point detection response time. For collision types whose recognition accuracy is not up to standard, supplement the corresponding samples to the dataset and perform incremental training optimization until the model reaches engineering-grade accuracy metrics.
[0025] 4. Execution of secondary collision detection for BIM models based on the trained model. The data flow diagram of the execution process is as follows: Figure 4 As shown.
[0026] 1) Obtain the list of native collision detection results from the BIM software to be reviewed. The list includes the ID, corresponding component, spatial location information, and professional affiliation of each collision point. 2) Based on the collision result list, BIM model view images of each collision point from a standardized perspective are collected in batches using automated scripts. The images are then preprocessed by size normalization and pixel value standardization to make them meet the model input requirements. 3) Input the preprocessed single collision point image into the trained true and false collision classification branch and collision region verification branch. The classification branch outputs the true and false collision attributes and the classification probability results of the collision type, while the verification branch outputs the collision core region coordinates, confidence score, and auxiliary classification probability results. 4) The fusion decision layer performs weighted fusion calculation on the output results of the two branches and outputs the final secondary detection result of the collision point; it automatically filters out results that are determined to be false collisions, and outputs the collision type, collision core area coordinates, corresponding view image, and professional information for results that are determined to be valid true collisions. 5) Traverse all original collision results, complete the secondary detection of all collision points, integrate and output the verified list of valid true collisions, and complete the entire secondary detection process.
[0027] 5. Standardized collision reports are automatically generated based on secondary detection results. 1) The program uses a pre-defined standard format for collision reports to obtain collision-related information from a single detection result; 2) Based on the list of valid true collisions output by the secondary detection, automatically fill the collision type, collision location, volume number, and component number into the report; 3) After all data is filled in, a standardized collision detection report in DOCX format will be automatically generated and exported. Example 2
[0028] This embodiment, without changing the dual-branch parallel architecture, input / output standards, and core detection logic of "true / false collision attribute classification + collision core region localization verification" in Embodiment 1, adapts and replaces the basic networks of the two branches with the same functionality. The true / false collision classification branch can be replaced with mainstream image classification networks such as the EfficientNet series, VisionTransformer (ViT), and SwinTransformer. To address the need for extracting subtle edge features in contact-type false collisions in power engineering, the convolutional modules, attention mechanisms, and feature extraction layers of the network are optimized to enhance the feature capture capability of small contact areas. The network output still matches the present invention. The system sets nine standardized detection label classification probability results, with the core output determining the true and false collision attributes. The collision region verification branch can be replaced with mainstream object detection networks such as YOLOv8 / YOLOv9, Faster R-CNN, and SSD. Targeting the scenario where the collision core contact area is mostly small targets and subtle regions, the system optimizes the network anchor box parameters, multi-scale feature fusion layer, and small target detection head. It also adapts to the use of bounding box regression loss functions such as CIoU / DIoU / GIoU to accurately locate the collision core contact area and outputs the bounding box coordinates, confidence score, and auxiliary classification probability results of the collision core area. It is fully compatible with the fusion decision layer in Example 1 and can achieve the same secondary detection effect. Example 3
[0029] This embodiment replaces and optimizes the fusion decision layer based on the fixed-weighted voting method in Embodiment 1. It eliminates the need to pre-fix the numerical proportions of the true / false collision classification branch weight W1 and the collision region verification branch weight W2. Instead, it employs an adaptive dynamic fusion decision scheme, using the confidence level of the two branch outputs as the core basis to construct dynamic weight calculation rules. When the confidence level of the classification result output by the true / false collision classification branch is greater than or equal to a preset engineering-level threshold, the weight proportion of the classification branch is automatically increased. Conversely, when the confidence level of the classification result output by the classification branch is lower than the preset threshold, the weight proportion of the collision region verification branch is automatically increased. Furthermore, it can be replaced with a fusion decision method based on DS evidence theory and Bayesian decision theory, performing uncertainty modeling and fusion on the outputs of the two branches to eliminate the risk of misjudgment from a single branch output. This results in a stable and accurate secondary detection result, fully retaining the core logic of the dual-branch result cross-verification in Embodiment 1, and adapting to the detection accuracy requirements of different engineering scenarios. Example 4
[0030] This embodiment, while fully retaining the core inventive concept and full-process logic of the present invention—"non-interface secondary intelligent verification of native collision detection results from BIM software, automated filtering of false collisions, accurate classification of true collisions, and automatic generation of standardized reports"—replaces the dual-branch fusion AI recognition network in Embodiment 1 with an end-to-end recognition scheme for a multimodal large model. It can use a closed-source commercial or open-source finely tuned multimodal large model adapted to engineering scenarios. A standardized Prompt engineering system matching the present invention's nine standardized label systems, engineering judgment rules, and mandatory output format is pre-constructed. The input process is completely consistent with Embodiment 1, and standardized BIM model views of collision points are collected in batches based on the native collision detection result list. The images, without needing to read the underlying geometric data of the BIM model or connect to the native interface of the BIM software, fully retain the non-interface core features. The preprocessed images and standardized prompts are synchronously input into the multimodal large model. Through the end-to-end visual understanding and scene reasoning capabilities of the large model, standardized results including collision validity determination, collision type, collision core area location, and engineering judgment basis are output. At the same time, the recognition accuracy of the large model can be optimized through scene-based fine-tuning and few-sample prompt learning. A collaborative reasoning scheme of "large model + lightweight small model" is constructed to adapt to the batch inspection needs of engineering. The output results are fully compatible with those of Example 1 and can be directly connected to the subsequent standardized report automatic generation process, achieving all the core objectives of this invention.
[0031] In the above embodiments, an AI-based, non-interface-based secondary collision detection method for BIM models is provided. This invention specifically performs secondary intelligent verification of the native collision detection results of BIM software. Real-world engineering measurements show a false collision filtering rate of ≥95% and a true collision recognition accuracy of ≥90%, directly reducing manual verification workload by over 90%. This compresses the collision verification cycle for complex projects from 10-15 working days to several hours, significantly improving the efficiency and accuracy of BIM design review. This invention employs a dual-model fusion architecture, specifically optimizing the subtle features of contact-type false collisions and the core small target areas of collisions. This solves the problem that traditional rule-based filtering schemes cannot adapt to complex engineering scenarios and have poor filtering effects, demonstrating strong generalization ability. Furthermore, this invention uses a non-interface-based design, eliminating the need to interface with native BIM software or read the underlying geometric data of the BIM model. It can directly support all mainstream BIM software such as Revit and PDMS. The software performs secondary collision detection without requiring changes to designers' operating habits, making its application extremely easy. While filtering false collisions, this invention simultaneously performs accurate classification and core collision location for true collisions, providing precise data for subsequent collision rectification and improving the precision of engineering design. Furthermore, this invention enables the automatic generation of standardized collision reports, constructing a fully automated closed loop from importing native collision results, AI secondary detection, false collision filtering to standardized report delivery. The report generation accuracy is 100%, further reducing manual workload in later delivery by over 80%, ensuring the standardization and uniformity of deliverables.
[0032] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the concept and scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the inventive concept should fall within the protection scope of the present invention. All technical contents for which protection is sought in this invention are fully described in the claims.
Claims
1. A non-interface-based secondary collision detection method for BIM models based on artificial intelligence, characterized in that, Includes the following steps: S1. Construct a dedicated dataset for secondary collision detection in the target engineering field using BIM, and train a dual-branch fusion AI recognition network based on the dataset. The dual-branch fusion AI recognition network includes a parallel-connected true / false collision classification branch, a collision region verification branch, and a fusion decision layer connected to the outputs of both branches simultaneously. The true / false collision classification branch is a classification network optimized based on the ResNet residual neural network, the collision region verification branch is a detection network optimized based on the YOLOv11 object detection network, and the fusion decision layer is used to perform fusion calculation on the outputs of the two branches. The true / false collision classification branch optimizes the feature extraction layer of the ResNet network to meet the requirements of extracting subtle features of contact collisions. It uses the FocalLoss loss function to solve the problem of sample imbalance. The network input is the BIM model view image corresponding to the collision point, and the output is the classification probability results of each category under the standardized label system. The collision region verification branch optimizes the anchor box parameters and small target feature fusion layer of the YOLOv11 network for the localization requirements of the collision core contact area. It adopts the CIoU loss function as the bounding box regression loss function. The network input is the BIM model view image corresponding to the collision point, and the output is the bounding box coordinates, confidence and auxiliary classification probability results of the collision core area. S2. Obtain the list of native collision detection results from the BIM software, and based on the list, batch collect standardized view images of each collision point and complete image preprocessing. S3. The preprocessed single collision point image is synchronously input into the dual-branch fusion AI recognition network. The network outputs the classification probability results of collision validity and collision type through the attribute classification branch, and outputs the positioning information of the collision core region and auxiliary classification probability results through the region verification branch. S4. Perform a fusion decision on the output results of the two branches to obtain the final secondary detection result of the collision point, automatically filter out the results judged as false collisions, and retain the association information judged as valid true collisions. S5. After traversing all native collision results to complete full detection, integrate the correlation information of all valid true collisions and automatically generate a standardized collision detection report.
2. The non-interface-based secondary collision detection method for BIM models based on artificial intelligence according to claim 1, characterized in that, The step S1 involves constructing a dedicated dataset for secondary collision detection in BIM, specifically including: setting a standardized labeling system that includes false collisions and true collisions; collecting collision point samples from the native collision detection output of BIM software in actual projects in the target engineering field; collecting standardized view images of each collision point and labeling them with true and false collision attributes, collision types, and collision core area bounding boxes; cleaning and augmenting the labeled samples to construct an effective labeled sample set and dividing it into training set, validation set, and test set according to a preset ratio.
3. The non-interface-based secondary collision detection method for BIM models based on artificial intelligence according to claim 1, characterized in that: The fusion decision layer uses a weighted voting method for fusion calculation. The weight of the attribute classification branch output result is set as W1, and the weight of the region verification branch auxiliary classification result is set as W2, where W1+W2=1 and W1>W2. The final collision validity judgment result, collision type and collision core region coordinates are output.
4. The non-interface-based secondary collision detection method for BIM models based on artificial intelligence according to claim 1, characterized in that, The training of the dual-branch fusion AI recognition network in step S1 specifically includes: loading the general pre-trained weights of the ResNet network and the YOLOv11 network, and training them using transfer learning; adopting a two-stage training mode, in the first stage independently training the true and false collision classification branch and the collision region verification branch until the accuracy converges, and in the second stage fixing the parameters of the two branch backbone networks and jointly training the fusion decision layer until the false collision recognition accuracy reaches the preset index; using a test set to verify the model accuracy, and incrementally training and optimizing for collision types whose accuracy does not meet the standard.
5. The non-interface-based secondary collision detection method for BIM models based on artificial intelligence according to claim 1, characterized in that: In step S2, the native collision detection result list of the BIM software includes the ID of each collision point, the corresponding component, the spatial location information and the discipline to which it belongs; the image preprocessing specifically involves normalizing the size and standardizing the pixel values of the acquired BIM model view image to make it conform to the input requirements of the dual-branch fusion AI recognition network.
6. The non-interface-based secondary collision detection method for BIM models based on artificial intelligence according to claim 1, characterized in that: In step S4, the associated information of a valid true collision includes the collision type, the coordinates of the collision core region, the corresponding standardized view image, and the professional information to which it belongs.
7. The non-interface-based secondary collision detection method for BIM models based on artificial intelligence according to claim 1, characterized in that, In step S5, a standardized collision detection report is automatically generated. Specifically, a report template that meets the enterprise's delivery standards is preloaded, and the collision type, collision location, volume number, and component number of the valid true collision are automatically filled into the corresponding positions of the report template. After filling, a standardized collision detection report in DOCX format is automatically exported.
Citation Information
Patent Citations
Building information model real-time collision detection method and system based on deep learning
CN120259741A
Three-dimensional surveying and mapping data fusion processing method and system based on BIM
CN122116077A
