A concealed engineering quality intelligent detection method of a fusion cloud data deep learning model

Abstract

The application relates to the technical field of building engineering quality detection, and particularly discloses a concealed engineering quality intelligent detection method fusing a cloud data deep learning model. The method first acquires multi-modal heterogeneous data through automatic equipment and performs strict time-space alignment; then extracts lightweight features by using an edge computing gateway and uploads the features to the cloud; in the cloud, a multi-modal cross-feature fusion network large model is used to output a detection result in combination with a cross-attention mechanism; finally, based on digital twinning technology, defects are mapped to a three-dimensional virtual model, and a rule engine is used for comprehensive health degree scoring and intelligent early warning. Through edge-cloud collaborative architecture and multi-modal data deep fusion, the application improves the precision and efficiency of concealed engineering defect recognition, realizes transparent and digital management and tamper-proof tracing of the whole life cycle of concealed parts, and effectively reduces construction quality risks.

Description

Technical Field

[0001] This invention belongs to the field of building engineering quality inspection technology, specifically involving an intelligent inspection method for hidden engineering quality that integrates cloud data deep learning models. Background Technology

[0002] Concealed works (such as steel reinforcement, underground waterproofing, and pile foundation) serve as the "skeleton" and "protective layer" of a building structure, and their quality directly determines the structural safety and service life of the building. However, because concealed works are covered by concrete pouring, earthwork backfilling, or decorative layers after completion, traditional quality inspection methods have significant limitations. 1. Limited and outdated testing methods: Traditional testing relies heavily on visual observation and manual measurement by supervisors, which is not only inefficient but also makes it difficult to conduct non-destructive backtracking after coverage, easily leading to the "permanent burial" of potential quality problems.

[0003] 2. Low integration of multi-source data: Existing non-destructive testing equipment (such as ground penetrating radar, ultrasonic, etc.) mostly analyzes data independently, lacking deep coupling between spatial and visual dimensions, resulting in a high false alarm and false alarm rate for complex defects (such as minor leaks in waterproofing layers, local omissions in the binding of densely reinforced steel bars).

[0004] 3. Insufficient cloud capabilities: The network environment at construction sites is complex, and the transmission of massive amounts of detection data is limited, making it difficult to implement high-precision deep learning models in real time at the construction site. Furthermore, the models cannot evolve themselves according to changes in different engineering environments.

[0005] 4. Poor virtual-real interaction and traceability: The inspection data is disconnected from the design model (BIM), and construction managers cannot intuitively and three-dimensionally grasp the actual status of hidden parts. Moreover, the inspection records are at risk of being tampered with, making it difficult to trace responsibility. Summary of the Invention

[0006] To overcome the above-mentioned technical problems, this invention provides an intelligent detection method for hidden engineering quality that integrates cloud data deep learning models.

[0007] The present invention adopts the following technical solution: A method for intelligent detection of hidden engineering quality that integrates cloud data and deep learning models includes the following steps: S1: Acquire multimodal heterogeneous acquisition data of the hidden engineering site through automated acquisition equipment, and perform strict time and space alignment processing on the multimodal heterogeneous acquisition data; S2: Using the edge computing gateway deployed at the construction site, the original data after spatiotemporal alignment is cleaned and denoised, and a lightweight deep learning model is called to perform preliminary feature extraction. The extracted lightweight feature data is then uploaded to the cloud data center. S3: Construct a large multimodal heterogeneous cross-feature fusion network model in a cloud data center. This model extracts high-level semantic features of different modalities through independent encoder branches and uses a cross-attention mechanism to fuse features, outputting the defect detection results of hidden engineering. S4: Based on building information modeling digital twin technology, the defect detection results output from the cloud are accurately mapped to the three-dimensional virtual digital twin of the hidden project using inverse coordinate transformation, and then visualized highlighting and inverse geometric reconstruction are performed. S5: Input the detection results into the built-in national standard rule engine, use fuzzy comprehensive evaluation theory to calculate the comprehensive health score of the hidden project, and trigger graded intelligent early warning and automatic control commands according to the severity of the defects.

[0008] Preferably, in step S1, the multimodal heterogeneous acquisition data includes high-resolution images acquired by a visible light camera, multispectral images acquired by a multispectral imager, three-dimensional point cloud data acquired by a three-dimensional laser scanner, and one-dimensional time series signals acquired by ground penetrating radar.

[0009] Preferably, in step S1, the time and space alignment process specifically includes: using the edge computing gateway as the master clock node to access the precise time protocol network, binding microsecond-level global timestamps to each modal data; calculating the pose matrix of the automated acquisition device through real-time dynamic differential positioning technology and inertial measurement unit, and using affine transformation to convert and unify the local data of each sensor into the global world coordinate system of the construction site.

[0010] Preferably, in step S2, the preliminary feature extraction of the edge computing gateway specifically includes: using a depthwise separable convolutional network to extract shallow edge and texture feature matrices of two-dimensional image data; using a simplified point cloud neural network to extract local geometric feature vectors of three-dimensional point cloud data; and uploading the feature matrix, feature vectors, and a small number of original data slices suspected of containing serious defects to the cloud after secure encryption.

[0011] Preferably, in step S3, the large model of the multimodal heterogeneous cross-feature fusion network includes three parallel independent feature extraction encoder branches: a two-dimensional visual feature depth encoder for processing image features, a three-dimensional spatial geometric encoder for processing point cloud features, and a one-dimensional physical signal time encoder for processing ground-penetrating radar time-series signals; the two-dimensional visual feature depth encoder adopts a visual transformer architecture based on a self-attention mechanism, and the one-dimensional physical signal time encoder adopts a long short-term memory network.

[0012] Preferably, in step S3, the cross-attention mechanism specifically includes: using the visual features processed by the encoder as the query vector, and the point cloud features and radar features as the key vector and value vector; by calculating the dot product similarity matrix between the query vector and the key vector, dynamically allocating the weights of different modal data at a specific spatial location to the defect indication, thereby outputting the fused global defect features.

[0013] Preferably, in step S4, the reverse geometric reconstruction specifically refers to: the cloud system replacing the corresponding ideal geometric surface in the design drawings in the building information model based on the actual acquired three-dimensional point cloud data, thereby generating an as-built digital twin containing the actual construction space error.

[0014] Preferably, the method further includes a federated learning model adaptive iteration step deployed between the cloud and the edge computing gateway: the edge computing gateway downloads the global initial deep learning model from the cloud, calculates the gradient update of the model weight parameters locally using the locally verified and labeled detection data, and uploads it in encryption; the cloud central server uses a secure aggregation algorithm to perform a weighted average of the gradient data of different edge nodes, updates the global model weights, and redistributes them.

[0015] Preferably, the adaptive iteration step of the federated learning model also has an adaptive hard sample mining mechanism: during cloud inference, when the model's classification confidence score for the detected target is within a preset uncertainty threshold range, the corresponding data is automatically marked as a hard sample and manual review is triggered. The results after review and confirmation are automatically added to the incremental learning training set to further optimize the model.

[0016] Preferably, this method incorporates blockchain anti-counterfeiting and traceability mechanisms throughout the entire process: after the edge computing gateway completes the collection of raw data and time synchronization, it immediately calculates the cryptographic hash value and packages the hash value with the device hardware identification code, operator biometric information and geographical coordinates, and stores it on the blockchain for evidence; the final detection report and inference keyframes output from the cloud are also written into the distributed consortium blockchain after hash calculation.

[0017] Compared with the prior art, the beneficial effects of the present invention are: 1. Multimodal deep fusion, achieving a leapfrog improvement in detection accuracy: By using a cross-attention mechanism to deeply fuse visible light texture, laser point cloud geometric information and ground-penetrating radar physical signals, the physical limits of a single sensor are broken, enabling accurate identification of deep defects behind the surface.

[0018] 2. Edge-cloud collaborative architecture balances real-time performance and computing power requirements: By utilizing edge computing gateways to perform preprocessing and feature extraction, network transmission pressure is greatly alleviated, enabling real-time early warning at the construction site, while leveraging the complex analysis capabilities of ultra-large-scale deep learning models in the cloud.

[0019] 3. Digital twin mapping enables "transparent" management of concealed works: Through reverse geometric reconstruction technology, the test results are presented in three dimensions, allowing managers to see the internal quality directly through the covering layer, realizing synchronous virtual and real monitoring of the entire construction process.

[0020] 4. Federated learning mechanism to ensure privacy while enabling continuous model evolution: The system can improve the global model accuracy through gradient aggregation without uploading original sensitive data, enabling the model to adapt to the detection needs of different regions and processes.

[0021] 5. Blockchain notarization ensures the absolute authenticity and authority of test data: The introduction of a hash-on-chain mechanism blocks the possibility of human tampering with data from the source, providing irrefutable technical support for the lifelong responsibility system for the quality of hidden projects. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the method flow of the present invention. Detailed Implementation

[0023] This embodiment abandons the traditional single visual detection method and proposes an intelligent detection scheme for hidden engineering based on "edge-cloud collaborative architecture", "multimodal heterogeneous data cross-attention fusion mechanism" and "cloud model continuous iteration technology based on federated learning".

[0024] This embodiment provides an intelligent detection method for the quality of concealed works that integrates cloud data and a deep learning model. This method relies on a system architecture that includes multi-source sensor acquisition nodes, an edge computing gateway, a cloud-based deep learning platform, and a building information model (BIM) digital twin terminal. Taking typical concealed works such as rebar tying and underground waterproofing coating as examples, this embodiment details the complete implementation steps and technical aspects of this detection method.

[0025] Phase 1: Intelligent Acquisition and Strict Spatiotemporal Alignment of Multimodal Heterogeneous Data Before concealed works are covered by concrete pouring or backfilling, the system initiates a multi-dimensional spatial and physical data acquisition process. This process relies on an automated acquisition matrix deployed at the construction site.

[0026] The acquisition matrix comprises a high-resolution visible light camera, a multispectral imager, a 3D laser scanner, and a ground-penetrating radar, mounted on a high-precision omnidirectional mobile robot or drone. The visible light camera captures surface texture, color, and obvious geometric defects. The multispectral imager is specifically designed to detect the uniformity of waterproofing material coatings and potential leaks, as waterproofing layers of different thicknesses and materials exhibit varying reflectivities at specific spectral frequencies. The 3D laser scanner emits high-frequency laser beams to acquire high-density 3D point cloud data of the concealed engineering surface, used for precise measurements of rebar spacing, protective layer thickness, and formwork flatness. The ground-penetrating radar emits high-frequency electromagnetic waves and receives the echoes, penetrating the surface to detect potential voids or foreign objects.

[0027] To ensure absolute consistency in time and space between data collected by different sensors, the system introduces a strict spatiotemporal alignment mechanism. For time synchronization, all acquisition devices are connected to a precise time protocol network, using an edge computing gateway as the master clock node to achieve microsecond-level timestamp binding. Every image frame, every radar echo, and every laser scan result is stamped with a unified global timestamp. For spatial alignment, the mobile robot integrates real-time dynamic differential positioning technology and an inertial measurement unit. The system calculates the robot's pose matrix in real time, transforming local data from the camera coordinate system, radar coordinate system, and laser scanner coordinate system into the global world coordinate system of the construction site. This process involves complex affine transformations and extrinsic parameter calibration to ensure perfect overlap of the physical positions of multimodal data in three-dimensional space.

[0028] Phase 2: Real-time preprocessing and lightweight feature extraction of the edge computing gateway Because the amount of multimodal data collected on-site in concealed engineering projects is extremely large, directly transmitting the entire amount to the cloud would lead to severe network latency and bandwidth overload. Therefore, this embodiment innovatively deploys an edge computing gateway at the construction site to perform pre-processing and lightweight feature extraction.

[0029] The edge computing gateway first cleans and denoises the received raw data. For visible light images, the system applies histogram equalization and adaptive illumination compensation algorithms to eliminate the impact of uneven lighting at the construction site. For 3D point cloud data, the system uses voxel filtering to downsample, removes redundant dense points, and uses statistical filtering algorithms to remove outliers caused by dust or equipment noise. For one-dimensional time-series signals from ground-penetrating radar, edge nodes undergo zero-point drift correction, background removal, and bandpass filtering to enhance the signal-to-noise ratio of the effective echo signal.

[0030] After data cleaning, the edge computing gateway uses a pre-programmed lightweight deep learning model for initial feature extraction. For 2D image data, a depthwise separable convolutional network is used on the edge side to extract shallow edge and texture feature matrices. For 3D point cloud data, the system uses a simplified point cloud neural network to extract local geometric feature vectors. These highly compressed feature matrices and vectors are only one percent the size of the original data. Subsequently, the edge computing gateway securely and encrypts these lightweight features, along with a small number of raw data slices suspected of containing serious defects, and uploads them to the cloud data center via 5G network security.

[0031] Phase 3: Construction of core cloud-based multimodal deep learning models and cross-attention fusion The cloud data center is the core of this detection method, deploying a distributed graphics processing unit cluster with massive computing resources. In the cloud, the system constructs a large-scale deep learning model called a multimodal heterogeneous cross-feature fusion network. This model includes three parallel, independent feature extraction encoder branches and a global cross-attention fusion module.

[0032] The first branch is a two-dimensional visual feature deep encoder, employing a visual transformer architecture based on a self-attention mechanism. This encoder receives shallow image features uploaded from the edge side, divides them into multiple image patches, and adds positional encoding. Through multiple multi-head self-attention layers, the model is able to capture long-distance dependencies between different areas of the hidden engineering surface, such as the contextual relationship between an untied rebar and the surrounding rusted rebar.

[0033] The second branch is a 3D spatial geometric encoder, specifically designed to process point cloud features. This branch employs a hierarchical feature learning structure, continuously sampling point sets and aggregating features within the local spatial domain to extract deep 3D topological information of the concealed works. Based on this, the system can accurately calculate the 3D spatial deviation between the three-dimensional network morphology of the reinforcing steel skeleton and the design drawings.

[0034] The third branch is a one-dimensional physical signal timing encoder that uses a long short-term memory network to process ground-penetrating radar characteristics. The long short-term memory network effectively processes the temporal attenuation characteristics of radar electromagnetic waves reflected in different media layers through its internal forget gate, input gate, and output gate mechanism, accurately characterizing the locations of abrupt changes in dielectric constant inside the concealed medium, thereby locating cavities at the bottom of concrete or water accumulation under the waterproof layer.

[0035] Cross-attention fusion mechanism for multimodal data Due to the physical limitations of single sensors—for example, cameras cannot see through standing water, and radar struggles to discern surface textures—the system, after extracting high-level semantic features from three modalities, did not employ a simple feature concatenation method. Instead, it introduced a multimodal cross-attention module. This module uses visual features as query vectors and point cloud features and radar features as key and value vectors, respectively. By calculating the dot product similarity matrix between the query vector and the key vector, the model can dynamically evaluate which modality's data is most critical for indicating defects at a specific spatial location. For example, when inspecting underground waterproofing layers, if visual features show an abnormal color at a certain location, the cross-attention mechanism automatically assigns higher weight to the radar features at that location to verify whether internal leakage exists. This mechanism achieves deep complementarity of multi-source physical information, significantly reducing the false alarm and false negative rates in the detection of hidden engineering defects.

[0036] Phase 4: Defect Space Mapping and Reverse Construction Based on Digital Twins The results output by the cloud-based deep learning model not only include the category and severity of defects, but also high-dimensional semantic segmentation masks and bounding boxes. To enable construction managers to intuitively understand the location of these defects in the actual project, this embodiment integrates building information modeling (BIM) digital twin technology.

[0037] The cloud system retrieves the building information model (BIM) from the design phase of the concealed project and creates a virtual digital twin that perfectly corresponds to the physical entity. Using the global world coordinate system parameters obtained in the first phase, the system executes a rigorous spatial registration algorithm. Defect points identified by the deep learning model in the image pixel coordinate system and the point cloud local coordinate system are precisely mapped to the three-dimensional spatial coordinates of the BIM model through an inverse coordinate transformation matrix.

[0038] Building upon this foundation, the system automatically generates dynamic highlight markers within the digital twin. For example, when a missing section of the rebar mesh is detected on a wall, the corresponding rebar node in the virtual model will flash a red warning light. Furthermore, the system performs reverse geometric reconstruction within the virtual model based on the detected 3D point cloud data. If the formwork for the cast-in-place foundation exhibits outward deformation, the system replaces the ideal geometric surface of the foundation in the design drawings with actual point cloud data, generating an as-built digital twin that incorporates the actual construction errors. This virtual-real mapping and reverse construction technology ensures that the internal state of concealed works remains clearly visible even after being covered, achieving transparent management of concealed works.

[0039] Phase 5: Cloud-based Data Management and Adaptive Model Iteration Based on Federated Learning Considering the data privacy protection needs of different construction companies and the significant differences in the environments of various construction sites, this embodiment innovatively introduces a federated learning framework in the cloud to achieve continuous optimization and generalization capability improvement of deep learning models.

[0040] The cloud platform does not directly collect original confidential drawings or full sets of original images containing commercially sensitive information from each construction site. Instead, the cloud platform maintains a global initial deep learning model. Edge computing nodes at each construction site connected to the system download this global model and, using the latest concealed works detection data generated locally and manually verified and annotated, perform fine-tuning training of the model locally. Edge nodes only calculate the gradient update of the model's weight parameters and upload the encrypted gradient data to the cloud central server.

[0041] The cloud-based central server receives gradient data from various geological conditions and construction techniques across the country. The system employs a secure aggregation algorithm to perform a weighted average of these gradients, thereby updating the weights of the global model in the cloud. The updated global model is then redistributed to each edge node. Through this federated learning mechanism, the model can implicitly learn the characteristics of waterproofing layer peeling in the humid environment of southern China, while also understanding the radar echo patterns of early frost heave in concrete in the frigid northern regions.

[0042] Furthermore, the system possesses adaptive hard sample mining capabilities. During routine cloud-based inference, if the model's classification confidence score for a certain detection target hovers within an uncertain threshold range for an extended period, the system will automatically mark that data segment as a hard sample and trigger a manual review process. The results confirmed by expert intervention will be added to the incremental learning training set, enabling the model to accumulate experience like a human engineer, continuously improving its accuracy in identifying complex, rare, and hidden engineering quality defects.

[0043] Phase 6: Multi-dimensional comprehensive quality assessment and intelligent early warning decision support The quality of concealed works is not a black-and-white issue, but requires a quantifiable evaluation system. After completing defect identification and spatial mapping, the cloud system will initiate a multi-dimensional comprehensive quality assessment process.

[0044] The system incorporates a vast rule engine knowledge base covering the unified national standards for construction quality acceptance. The structured results output by the detection algorithm are fed into this rule engine. For example, for reinforced concrete projects, the system automatically extracts the main reinforcement spacing deviation, the length of the stirrup reinforcement zone, and the distribution density of the protective layer spacers. The system then compares these extracted quantitative values ​​item by item with the preset allowable deviation thresholds in the rule engine.

[0045] To provide macro-level engineering quality control, the system employs fuzzy comprehensive evaluation theory to establish a mathematical model for assessing the quality of concealed works. The system assigns dynamic weights based on the degree of impact of different defects on structural safety. For example, missing main reinforcement bars in load-bearing walls receive a very high penalty weight, while minor surface rust spots receive a lower weight. Through matrix multiplication and normalization, the system automatically calculates the overall health score of the current batch of concealed works inspections.

[0046] Based on the health score and the severity of specific defects, the system triggers a tiered intelligent early warning mechanism. For major hidden defects that seriously endanger structural safety, such as large-area voids in the foundation or severe misalignment of the reinforcing steel in load-bearing columns, the cloud system will immediately trigger the highest-level red warning. This warning will not only push a strong alert pop-up to the mobile smart terminals of the on-site project manager and supervising engineer, including a 3D location screenshot of the defect and a radar characteristic waveform image, but will also automatically send a pause pouring command to the concrete mixing plant through the system interface, physically preventing the next covering process from proceeding.

[0047] For general quality defects, such as insufficient thickness of waterproof coating in certain areas, the system will generate a yellow alert and automatically produce an electronic quality rectification notice containing rectification suggestions, location coordinates, and before-and-after comparison requirements, which will be sent to the relevant construction team. After completing the local touch-up coating, the construction team can use a portable handheld data acquisition device to take directional re-photos of the rectified area. After the data is uploaded, the cloud model automatically performs a closed-loop re-inspection and comparison. The alert status will be automatically lifted only after the defect is confirmed to be eliminated, thus forming a complete closed loop for the quality control of concealed works.

[0048] Algorithm anti-counterfeiting and traceability mechanisms In the quality inspection of concealed works, the authenticity and tamper-proof nature of the data are crucial. To prevent construction companies from artificially modifying on-site photos to cover up quality problems, this embodiment introduces blockchain-based distributed ledger technology throughout the entire data flow lifecycle.

[0049] The edge computing gateway immediately calculates the cryptographic hash value of the data packet the moment it completes the raw data collection and timestamp binding. This hash value, along with the unique hardware identifier of the collection device, the operator's biometric information, and the current geographical coordinates, is packaged into a block and broadcast to the consortium blockchain nodes jointly maintained by the construction unit, supervision unit, construction unit, and quality supervision agency.

[0050] When the cloud-based deep learning model generates the final quality inspection report for concealed works, the report's results, the model version number used, and the keyframes of the derivation process are all hashed and stored on the blockchain. At any point in the future, whether during project completion and acceptance or when structural defects occur during operation requiring accountability, inspectors can retrieve historical inspection data of the relevant concealed works from the cloud database and recalculate the hash value for comparison with the records on the blockchain. Any pixel-level alteration of the original image, modification of the timestamp, or interception of radar echo data will cause a cascading change in the hash value, instantly detected by the blockchain's anti-counterfeiting mechanism. This deep technological trust mechanism completely solves the industry pain points of traditional concealed works image data being easily forged and difficult to trace, ensuring the absolute purity of the input data to the deep learning model and the legal validity of the output results.

[0051] Although embodiments of the present invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and variations can be made to the above embodiments without departing from the principles and spirit of the present invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A method for concealed engineering quality intelligent detection of a fusion cloud data deep learning model, characterized in that, Includes the following steps: S1: Acquire multimodal heterogeneous acquisition data of the hidden engineering site through automated acquisition equipment, and perform strict time and space alignment processing on the multimodal heterogeneous acquisition data; S2: Using the edge computing gateway deployed at the construction site, the original data after spatiotemporal alignment is cleaned and denoised, and a lightweight deep learning model is called to perform preliminary feature extraction. The extracted lightweight feature data is then uploaded to the cloud data center. S3: Construct a large multimodal heterogeneous cross-feature fusion network model in a cloud data center. This model extracts high-level semantic features of different modalities through independent encoder branches and uses a cross-attention mechanism to fuse features, outputting the defect detection results of hidden engineering. S4: Based on building information modeling digital twin technology, the defect detection results output from the cloud are accurately mapped to the three-dimensional virtual digital twin of the hidden project using inverse coordinate transformation, and then visualized highlighting and inverse geometric reconstruction are performed. S5: Input the detection results into the built-in national standard rule engine, use fuzzy comprehensive evaluation theory to calculate the comprehensive health score of the hidden project, and trigger graded intelligent early warning and automatic control commands according to the severity of the defects. 2.The method of claim 1, wherein In step S1, the multimodal heterogeneous acquisition data includes high-resolution images acquired by a visible light camera, multispectral images acquired by a multispectral imager, three-dimensional point cloud data acquired by a three-dimensional laser scanner, and one-dimensional time series signals acquired by ground penetrating radar.

3. The method for intelligent detection of hidden engineering quality based on a cloud data deep learning model according to claim 2, characterized in that, In step S1, the time and space alignment process specifically includes: using the edge computing gateway as the master clock node to access the precise time protocol network, binding microsecond-level global timestamps to each modal data; calculating the pose matrix of the automated acquisition device through real-time dynamic differential positioning technology and inertial measurement unit, and using affine transformation to convert and unify the local data of each sensor into the global world coordinate system of the construction site.

4. The intelligent detection method for hidden engineering quality based on a cloud data deep learning model according to claim 1, characterized in that, In step S2, the preliminary feature extraction of the edge computing gateway specifically includes: using a depthwise separable convolutional network to extract shallow edge and texture feature matrices of two-dimensional image data; using a simplified point cloud neural network to extract local geometric feature vectors of three-dimensional point cloud data; and uploading the feature matrix, feature vectors, and a small number of original data slices suspected of containing serious defects to the cloud after secure encryption.

5. The intelligent detection method for hidden engineering quality based on a cloud data deep learning model according to claim 1, characterized in that, In step S3, the large model of the multimodal heterogeneous cross-feature fusion network includes three parallel independent feature extraction encoder branches: a two-dimensional visual feature depth encoder for processing image features, a three-dimensional spatial geometric encoder for processing point cloud features, and a one-dimensional physical signal time encoder for processing ground-penetrating radar time signals; the two-dimensional visual feature depth encoder adopts a visual transformer architecture based on a self-attention mechanism, and the one-dimensional physical signal time encoder adopts a long short-term memory network.

6. The intelligent detection method for hidden engineering quality based on a cloud data deep learning model according to claim 5, characterized in that, In step S3, the cross-attention mechanism specifically includes: using the visual features processed by the encoder as the query vector, and the point cloud features and radar features as the key vector and value vector; by calculating the dot product similarity matrix between the query vector and the key vector, dynamically allocating the weights of different modal data at a specific spatial location to the defect indication, thereby outputting the fused global defect features.

7. The intelligent detection method for hidden engineering quality based on a cloud data deep learning model according to claim 1, characterized in that, In step S4, the reverse geometric reconstruction specifically refers to the cloud system replacing the corresponding ideal geometric surface in the design drawings in the building information model based on the actual acquired three-dimensional point cloud data, thereby generating an as-built digital twin containing the actual construction space error.

8. The intelligent detection method for hidden engineering quality based on a cloud data deep learning model according to claim 1, characterized in that, The method also includes a federated learning model adaptive iteration step deployed between the cloud and the edge computing gateway: the edge computing gateway downloads the global initial deep learning model from the cloud, calculates the gradient update of the model weight parameters locally using the locally verified and labeled detection data, and uploads it in encrypted form; The cloud-based central server uses a secure aggregation algorithm to perform a weighted average of the gradient data from different edge nodes, updates the global model weights, and then redistributes the data.

9. The intelligent detection method for hidden engineering quality based on a cloud data deep learning model according to claim 8, characterized in that, The adaptive iteration step of the federated learning model also has an adaptive hard sample mining mechanism: during cloud inference, when the model's classification confidence score for the detected target is within a preset uncertainty threshold range, the corresponding data is automatically marked as a hard sample and manual review is triggered. The results after review and confirmation are automatically added to the incremental learning training set to further optimize the model.

10. A method for intelligent detection of hidden engineering quality based on a cloud data deep learning model according to any one of claims 1 to 9, characterized in that, This method incorporates blockchain anti-counterfeiting and traceability mechanisms throughout the entire process: after the edge computing gateway completes the collection of raw data and time synchronization, it immediately calculates the cryptographic hash value and packages the hash value with the device hardware identification code, operator biometric information and geographical coordinates, and stores it on the blockchain for evidence; the final detection report and inference keyframes output from the cloud are also written into the distributed consortium blockchain after hash calculation.

Images