Intelligent detection cloud platform for foundation pile based on image recognition

Abstract

The application discloses a foundation pile intelligent detection cloud platform based on image recognition, relates to the technical field of foundation pile intelligent detection, and comprises a data acquisition module, an image preprocessing module, a time sequence image recognition module, a feature fusion mapping module, a cloud knowledge graph module and a visual display module; the system constructs an image time sequence tensor by acquiring image sequences and construction parameters of a construction site, extracts defect evolution features based on an improved three-dimensional convolution and an attention mechanism, and constructs a causal graph in combination with the construction parameters, so that defect recognition, risk modeling and visual analysis are realized; the platform has high recognition precision, strong traceability and cross-point knowledge migration capability, and significantly improves the intelligent level and engineering applicability of foundation pile detection.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent detection of pile foundations, and in particular to an intelligent detection cloud platform for pile foundations based on image recognition. Background Art

[0002] Currently, pile foundation testing primarily relies on manual inspections, low-frequency imaging, or ultrasonic testing. Manual inspections are limited in accuracy by operator experience and pose safety risks. Traditional imaging can only capture static data, making it difficult to capture anomalies during dynamic construction. While ultrasonic and low-frequency vibration testing equipment offer a certain level of accuracy, they suffer from limitations such as low signal-to-noise ratios, complex deployment, and isolated data in complex construction environments, making it difficult to establish a multi-dimensional quality evaluation system for engineering projects.

[0003] Existing image recognition applications for pile foundation inspection are typically limited to single-image recognition or defect classification. They haven't yet integrated data with on-site time-series information, equipment construction parameters, environmental factors, and other data, making it difficult to dynamically reflect hidden quality issues during pile foundation construction. Furthermore, the lack of a unified data scheduling and visualization analysis system has resulted in data fragmentation among various construction units, hindering the formation of a large-scale project quality management network.

[0004] This invention proposes for the first time the integration of continuous image recognition, time series modeling, and a linkage analysis mechanism for construction equipment data, to construct a cloud platform for intelligent pile detection based on image recognition that is cross-period, cross-worksite, and multimodal data-driven. The platform can automatically generate defect evolution trajectories based on time series images, and realize defect causal retrospective analysis by combining parameters such as construction equipment pressure and pile sinking rate. It also constructs an extensible data knowledge graph on the cloud to assist in engineering structure quality decision-making. Compared with existing technologies, the present invention not only realizes the automation and remoteness of detection, but also improves the perception of complex hidden defects through multi-dimensional modeling. It has significant novelty and creativity, and is difficult to directly derive or obtain by analogy from existing technologies. Summary of the Invention

[0005] The purpose of the present invention is to provide an intelligent detection cloud platform for pile foundations based on image recognition to address the shortcomings of the background technology.

[0006] To achieve the above objectives, the present invention provides the following technical solutions: an image recognition-based intelligent pile detection cloud platform, comprising:

[0007] Data acquisition module: used to obtain a continuous image sequence I of the pile construction area at the construction site and simultaneously collect the corresponding construction parameter set P;

[0008] Image preprocessing module: used to preprocess the image sequence I, extract the target area of ​​the pile and its surface features in the image, and construct a standardized image time series tensor I′;

[0009] The time sequence image recognition module: based on the trained time sequence image recognition model, the image time sequence tensor I' is subjected to intelligent defect recognition, the type, spatial position and evolution trend of the foundation pile defect existing in each frame of image are obtained, and an initial defect trajectory set D is formed;

[0010] The mapping module: the defect trajectory set D obtained by recognition is matched with the construction parameter set P collected synchronously in the time axis and the characteristics are fused to construct a construction feature map G;

[0011] The knowledge graph module: used for uploading the construction feature map G to the cloud platform, constructing a standardized multi-worksite construction quality database M in the platform, and generating a multi-scale knowledge graph K based on the defect evolution path and risk level relationship;

[0012] The visualization module: based on the cloud knowledge graph K, the suspected high-risk defect nodes in the current construction area are automatically identified, the graphical detection results are generated, and the detection report R is automatically output by pushing to the user terminal through the cloud interface.

[0013] Preferably, the construction parameter set P includes pile driving pressure, driving frequency, pile driving depth and timestamp information.

[0014] Preferably, the pre-processing of the image sequence I includes image enhancement, inter-frame denoising and feature alignment processing.

[0015] Preferably, the time sequence image recognition module comprises:

[0016] The image time sequence tensor I' is input into a three-dimensional convolutional neural network with fusion of spatial attention and time gating mechanism;

[0017] The defect recognition subnetwork trained based on the multi-label classification loss function is used to infer each frame of the image sequence frame by frame, and the structured recognition results containing defect type, severity and confidence are extracted;

[0018] The long short-term memory unit is used to model the recognized results in time sequence, the evolution path and trend change of the defect in the time dimension are extracted, and the initial defect trajectory set D is output.

[0019] Preferably, the time sequence image recognition module further comprises:

[0020] The image time sequence tensor I' is subjected to multi-resolution branch coding processing to construct a multi-scale feature set F containing global contour and local texture features;

[0021] The multi-scale feature set F is input into an improved time sequence attention guide network W, which combines the feature difference sensitive mechanism between image frames and the defect saliency guide mechanism;

[0022] The W output result is integrated with a traditional three-dimensional convolution recognition result to generate a final defect track set D by using a weighted confidence fusion strategy.

[0023] Preferably, the mapping module comprises:

[0024] The defect track set D and the construction parameter set P are processed on a time axis to align the time sequence correspondence between the image frames and the construction parameters by using a bidirectional dynamic time warping algorithm;

[0025] The defect attribute vector and the construction parameter vector at each moment are extracted, a joint representation node is constructed by using a feature attention encoder, and a node set N is formed;

[0026] Based on the time sequence relationship and the spatial adjacency rule, an edge set E of the graph structure is constructed, and a defect influence weight is introduced as an attribute of the edge to construct a construction feature graph G=(N, E) with causal characteristics.

[0027] Preferably, the knowledge graph module comprises:

[0028] The node features and the edge attributes in the construction feature graph G are structured and extracted, and are mapped into a standardized data model in combination with corresponding work point numbers and geographic labels to construct a multi-work point construction quality database M;

[0029] The defect track data in the quality database M are semantically classified and risk level labeled, and high-dimensional mode feature vector sets are extracted according to defect types, evolution rates and construction responses.

[0030] Based on graph embedding and hierarchical clustering algorithm, a multi-granularity and multi-dimensional knowledge graph K is constructed, wherein the nodes represent defect evolution units, the edges represent risk association paths, and the defect propagation law is learned through a graph convolution network.

[0031] Preferably, the visualization module comprises:

[0032] Based on the cloud knowledge graph K, a graph attention mechanism is used to calculate the risk score of the node, and the defect nodes with a risk value higher than a preset threshold are automatically identified and marked as a suspected high-risk node set;

[0033] The evolution path subgraph of each defect node in the suspected high-risk node set is constructed by using the time sequence path and the spatial position information in the graph structure, and the key causal link and the influence range are extracted;

[0034] The subgraph information and the corresponding image frame are spatially superimposed and displayed to generate a three-dimensional detection visualization result with a dynamic time axis.

[0035] In the above technical solution, the present application provides the following technical effects and advantages:

[0036] 1. The present application solves the problems of low defect recognition accuracy, difficulty in capturing dynamic changes, lack of causal analysis and risk prediction ability in the prior art by constructing a pile intelligent detection cloud platform that integrates image recognition, time series modeling, feature mapping and knowledge graph reasoning. The system realizes high-precision recognition and tracking of dynamic defects by introducing a recognition model combining three-dimensional convolution and time attention mechanism. At the same time, feature graph modeling and graph neural network reasoning are used to construct a knowledge graph with causal chain analysis capability, realizing intelligent management of the whole process from defect recognition to risk propagation prediction.

[0037] 2. Compared with the traditional method, the present application has the technical advantages of high recognition accuracy, fast response speed, strong prediction ability and visual intuition, significantly improving the intelligent level of pile detection and construction risk control ability. In actual engineering test, the system accuracy is improved by more than 10%, the processing efficiency is improved by more than 80%, and the high-risk defect early warning success rate is more than 95%, fully proving that the present application has obvious engineering adaptability and practical promotion value. BRIEF DESCRIPTION OF DRAWINGS

[0038] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced as follows. Obviously, the drawings described in the following are only some embodiments of the present application, and other drawings can also be obtained by those skilled in the art based on these drawings.

[0039] Figure 1 The platform system module mind map of the present application. DETAILED DESCRIPTION

[0040] In order to make the purpose, technical scheme and advantages of the embodiments of the present application more clear, the technical scheme in the embodiments of the present application will be described clearly and completely in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0041] Embodiment, please refer to Figure 1 As shown in the figure, the pile intelligent detection cloud platform based on image recognition described in the present embodiment comprises:

[0042] Data acquisition module: used for acquiring continuous image sequence I of pile construction area in construction site, and synchronously acquiring corresponding construction parameter set P;

[0043] Image preprocessing module: for preprocessing the image sequence I, and extracting the pile target region and its surface features in the image, and constructing a standardized image time series tensor I';

[0044] Time series image recognition module: based on the trained time series image recognition model, the image time series tensor I' is intelligently recognized for defects, the type, spatial position and evolution trend of the defects existing in each frame of image are obtained, and an initial defect trajectory set D is formed;

[0045] Mapping module: time axis matching and feature fusion of the identified defect trajectory set D and the synchronously collected construction parameter set P are performed to construct a construction feature map G;

[0046] Knowledge graph module: for uploading the construction feature map G to the cloud platform, and constructing a standardized multi-worksite construction quality database M in the platform, and generating a multi-scale knowledge graph K based on the defect evolution path and risk level relationship;

[0047] Visualization module: based on the cloud knowledge graph K, suspected high-risk defect nodes in the current construction area are automatically identified, graphical detection results are generated, and the detection report R is automatically outputted through the cloud interface and pushed to the user terminal.

[0048] In the pile intelligent detection cloud platform based on image recognition proposed in the present application, the data acquisition module is the basis and key module for the system to realize the whole process intelligent detection, responsible for information perception, data synchronization and original input construction of the target area in the pile construction stage. The module mainly includes an image acquisition sub-module and a construction parameter acquisition sub-module, which work together to construct a high-consistency, high-resolution multi-modal original data set for subsequent image processing and defect recognition.

[0049] Specifically, the data acquisition module first acquires continuous images of the pile construction area through high-definition industrial cameras arranged at key positions of the construction site. The image acquisition is a time-continuous frame sequence acquisition, that is, the on-site images are automatically acquired according to a preset time interval Δt (such as every 1 second / frame) to generate an image sequence I. The image acquisition device needs to have high frame rate (not less than 30 fps), anti-strong light interference, night light supplement and environmental disturbance resistance to ensure that clear and distinguishable image data can still be acquired under different weather, light and dust environments.

[0050] Synchronously with the image acquisition process, the construction parameter acquisition sub-module is responsible for acquiring key construction process parameters of the corresponding pile to form a construction parameter set P. The construction parameters include but are not limited to:

[0051] Pile sinking pressure (unit: kN): reflecting the stress condition of the pile body in the pile sinking process, the acquisition frequency is not less than 1 Hz;

[0052] Strike frequency (unit: Hz): Reflects the frequency of hammering or vibration application, closely related to the efficiency of device vibration;

[0053] Pile driving depth (unit: m): Reflects the change process of pile body into the soil depth, which can be obtained by laser displacement sensor or ranging radar;

[0054] Timestamp information: used for accurate time labeling of all collected parameters, ensuring alignment consistency with image sequences.

[0055] To realize the above-mentioned time synchronization acquisition and binding of images and parameters, a unified timestamp control mechanism is introduced in this module, each frame of image is paired and bound with the corresponding time point construction parameter through the master clock module, ensuring the time consistency between image Iᵢ and parameter Pᵢ, facilitating subsequent feature fusion processing. The image acquisition submodule and the construction parameter acquisition submodule are cooperatively controlled by the industrial controller, and the collected data is cached or stored in the edge computing node in real time.

[0056] In addition, considering the potential network fluctuations or disconnections in the construction environment, the data acquisition module supports a local data redundancy backup mechanism, which can temporarily store the collected data in the local cache area, and upload it to the cloud platform in batches after the network is restored, ensuring data integrity and system fault tolerance.

[0057] Through the design of the above-mentioned data acquisition module, the system can continuously, stably and high-faithfully acquire image and parameter information of the construction site, establish a complete time sequence image and construction state correspondence, provide a unified data entry and reliable data support for subsequent image preprocessing, defect recognition, causal mapping and knowledge graph generation, and significantly improve the practicality and engineering adaptability of the entire platform.

[0058] The image preprocessing module provided in the foundation pile intelligent detection cloud platform based on image recognition is used for quality optimization, target area extraction and structured expression processing of the image sequence I obtained by the data acquisition module, and outputs a standardized image time sequence tensor I', providing stable and high-quality input data for the defect recognition model.

[0059] The image preprocessing module includes an image enhancement submodule, an inter-frame denoising submodule, a target detection and feature alignment submodule, and its technical process is as follows:

[0060] Firstly, the image enhancement sub-module pre-processes each frame image Ii in the image sequence I, adopts histogram equalization, adaptive contrast stretching, and image sharpening enhancement algorithms, and the like, to improve the overall brightness uniformity and edge sharpness of the image, thereby improving the distinguishability of the pile boundary and defect details. In the actual environment of the construction site with complex lighting conditions and frequent dust interference, the image enhancement process significantly improves the visual quality and information integrity of the target region.

[0061] Secondly, the inter-frame denoising sub-module introduces a multi-frame joint denoising mechanism for random noise, inter-frame jump, or motion blur problems existing in the image sequence. Specifically, a time domain filtering and spatial domain joint model (such as the 3D-BM3D algorithm) is used to model the continuous frames, and the dynamic interference information is denoised and edge-preserved. This method effectively suppresses the image drift and blur caused by low-frequency construction disturbance while preserving the texture and edge information, improving the consistency and stability of the time sequence images.

[0062] Subsequently, the target detection and feature alignment sub-module is used to identify and extract the pile target region in each frame image. An improved YOLOv5-based region proposal network is used for pile target positioning, and a key point detection algorithm is used to accurately obtain the pile top and edge feature points. To ensure the geometric consistency of the pile target region in multiple frames of images, a frame-to-frame feature alignment algorithm based on optical flow field estimation and affine transformation is further applied to realize dynamic registration and unified cropping of the ROI region.

[0063] Finally, the processed image sequence is uniformly adjusted to have consistent resolution, alignment angle, and gray scale range, and a standardized image time sequence input in the form of a tensor is constructed in the time dimension, i.e., an image time sequence tensor I', represented as where T is the number of frames, HxW is the image spatial dimension, and C is the number of channels. The tensor structure facilitates direct calling of subsequent deep learning models.

[0064] Comparison table of pile defect recognition accuracy (the present application vs. static image recognition model)

[0065]

[0066] It is explained that the time sequence image recognition model adopted by the present application significantly improves the recognition accuracy in complex environments, especially in dynamic evolution type defects, and has a robust advantage.

[0067] The unstructured and multi-source disturbance characteristics in the original image sequence I are effectively normalized by the systematic processing of the image preprocessing module, and the output image time sequence tensor I' has high alignment, high information density and high robustness. Compared with the single-frame processing of static images in the prior art, the time sequence image preprocessing method of the present application significantly improves the generalization ability and defect recognition accuracy of the recognition model in the dynamic construction scene, and has significant technical progress and innovative value.

[0068] In the image recognition-based foundation pile intelligent detection cloud platform proposed in the present application, the time sequence image recognition module, as one of the core modules, undertakes the task of automatically identifying the defect type, spatial position and evolution trend of the foundation pile from the standardized image time sequence tensor I'. The module comprehensively uses deep learning, attention mechanism, time sequence modeling and multi-model fusion technologies to build a nonlinear and multi-dimensional recognition system, ensuring high-precision recognition and tracking of small defects and dynamic abnormalities in complex construction backgrounds.

[0069] Firstly, the image time sequence tensor I' output by the image preprocessing module is input into a three-dimensional convolutional neural network (3D-CNN) that integrates spatial attention mechanism and time gating structure. This network can extract contour, texture and edge features on the surface of the pile in the spatial dimension, and capture the small changes caused by defect evolution between image frames in the time dimension.

[0070] The spatial attention mechanism reassigns weights to the feature maps of each image, making the network pay more attention to areas with significant deformation or texture abnormalities; the time gating structure adds time-dependent weights to explicitly model the dynamic process of defect evolution between frames, thereby improving the system's response to sudden defects such as crack propagation and pile head damage. The high-dimensional tensor output by the three-dimensional convolutional network embeds rich spatio-temporal feature information, laying the foundation for the subsequent recognition process.

[0071] Subsequently, the 3D feature output is input into a defect recognition subnetwork based on a multi-label classification loss function design. This subnetwork infers each frame of the image sequence one by one and outputs structured information such as defect type, severity and confidence.

[0072] To improve the generalization ability of the recognition model, the subnetwork uses a soft label representation mechanism to handle defect samples with similar features but different physical properties, and optimizes the loss function based on Focal Loss to handle the sample imbalance problem. Finally, the output of each frame is a set of structured defect description vectors, including but not limited to:

[0073] Defect type (e.g., crack, peeling, corner drop);

[0074] Severity (level 0~3);

[0075] Defect location (center point in image coordinate system and contour frame);

[0076] Confidence (0~1).

[0077] The structured recognition results will be further processed in the time dimension.

[0078] To depict the evolution behavior of defects in the entire construction process, a long short-term memory network (LSTM) is further introduced to model the time sequence of the foregoing recognition results. The LSTM network cooperatively models long-term stable features and short-term mutation features through its gating mechanism, thereby effectively tracking the development trajectory of defects. The initial defect trajectory set D output by the module contains the time evolution path and dynamic change trend of each type of defect in consecutive image frames, which is used for subsequent mapping and risk assessment.

[0079] To improve the defect capturing ability of the system under different scale features, the time sequence image recognition module further includes a multi-resolution feature encoding process. Specifically, the image time sequence tensor I' is encoded through three different branches at the same time, and multi-scale information such as global contour, texture details, and local irregular deformation is extracted to form a multi-scale feature set F.

[0080] The multi-scale feature set F is input into an improved time sequence attention guide network W (TSA-Net). The network consists of two parts:

[0081] Feature difference sensitive mechanism: by comparing the high-dimensional Euclidean distance between the features of consecutive frames, subtle image change points are identified, discontinuous regions are highlighted, and the sensitivity of the model to implicit defect evolution is improved;

[0082] Defect saliency guide mechanism: through a pre-trained saliency map generation module, the response of possible defect regions in the image is amplified, thereby guiding the backbone network to focus more on the abnormal structure area.

[0083] The TSA-Net output result contains the time sequence defect feature set adjusted by dynamic weight, which has higher spatial selectivity and temporal recognition continuity.

[0084] To further improve the recognition stability, the module integrates the output results of TSA-Net and the traditional three-dimensional convolution recognition results for judgment, and uses a weighted confidence fusion strategy to generate the final defect trajectory set D. The specific fusion strategy is:

[0085] For the same position and same type of defect node, the confidence of the output of the two models is weighted and averaged, and the weight is dynamically adjusted according to the performance of the model in the historical validation set;

[0086] For the inconsistent regions of the results, a "conflict recognition mechanism" is introduced to determine which model is more reliable by time consistency constraints (such as inter-frame similarity trend) and automatically retain the reliable results.

[0087] The final output defect trajectory set D not only contains the spatial and temporal position and type information of each defect, but also fuses model evaluation confidence, evolution trend analysis and possible mutation warning indicators, providing high-dimensional and traceable data input for subsequent mapping modules and knowledge graph construction.

[0088] In the image recognition-based intelligent pile detection cloud platform proposed in the application, the mapping module is used to perform time sequence fusion modeling on the defect trajectory set D output by the time sequence image recognition module and the construction parameter set P, and construct a construction feature graph G that can reflect the defect evolution behavior and the causal relationship of construction dynamics. This module realizes the structured expression of multi-dimensional construction quality state through node-level joint modeling, edge attribute enhancement and causal structure expression, and is an important basis for subsequent knowledge graph construction and risk assessment.

[0089] Firstly, the mapping module performs time axis alignment processing. Since there may be differences in sampling frequency or time offset in data collection between the image sequence and the construction parameters, the application introduces a bidirectional dynamic time warping algorithm to time sequence pair the defect trajectory set D and the construction parameter set P. This algorithm aligns the timestamps of image frames and construction parameters in both forward and reverse directions, so that the defect information and equipment parameters at each time can be accurately corresponded, thereby ensuring time consistency input.

[0090] Secondly, feature fusion modeling is performed. For each time alignment point, the defect attribute vector (such as defect type, location, severity level, confidence) and the construction parameter vector (such as pile pressing pressure, hitting frequency, pile depth, etc.) of the corresponding frame are extracted. The application uses a feature attention encoder to jointly model these two vectors. The encoder automatically adjusts the importance weight of different dimensional features based on the self-attention mechanism, enhances the expression ability of features with potential risk correlation, and generates a composite node representation representing the construction state, forming a node set N.

[0091] Subsequently, the edge set E of the graph structure is constructed. The connection between nodes is established according to two rules:

[0092] Time sequence relationship: if node N1 corresponds to time t1 and node N2 corresponds to time t2, and t2>t1, a time directed edge from N1 to N2 is automatically generated;

[0093] Spatial adjacency rule: if two nodes represent synchronous defects between different parts of the same pile or adjacent piles, a spatial edge connection is established according to their physical positions.

[0094] In the edge construction process, a defect influence weight is further introduced as an attribute of the edge. The weight is calculated based on comprehensive indexes such as defect development rate, evolution path stability, and construction parameter fluctuation, and reflects the potential causal correlation strength between different nodes.

[0095] Finally, the mapping module outputs a construction feature graph G=(N,E) with causal expression capability, where N represents a joint state node set, and E represents an edge set with attribute weighting. The graph not only retains the dynamic evolution information of defects and construction behaviors, but also clearly depicts complex causal relationships through the graph structure, providing complete structural support for subsequent knowledge graph generation, defect tracing, and intelligent risk reasoning.

[0096] The mapping module integrates deep feature modeling, dynamic time alignment, heterogeneous graph expression, and causal modeling techniques, breaking through the limitations of existing detection systems based on single-point data and lacking structured representation, and has significant innovation and practical value. Through the construction of the construction feature graph G, the multi-dimensional interaction relationship between complex construction states and defect behaviors is clearly modeled, which is an important technical foundation for the intelligent construction quality analysis of the present application.

[0097] In the image recognition-based foundation pile intelligent detection cloud platform of the present application, the knowledge graph module is used to construct a semantic network structure of construction quality in the cloud platform, realizing cross-site, multi-dimensional, and inferable defect evolution knowledge modeling. The module relies on the construction feature graph G output by the aforementioned mapping module, and through processes such as structured data extraction, semantic clustering, risk labeling, and graph learning, finally forms a multi-scale knowledge graph K with prediction and risk linkage capabilities, which is the core technical support for the platform to realize quality insight, risk warning, and intelligent diagnosis.

[0098] First, the knowledge graph module performs structured extraction on the construction feature graph G=(N,E) generated by the mapping module. Each node in the node set N represents a joint state node at a construction time slice, which contains defect attribute vectors (type, location, severity level, confidence, etc.) and construction parameter vectors (pile driving pressure, driving frequency, pile driving depth, etc.) at the corresponding time point. The edge set E records the temporal relationship, spatial correlation, and derived causal strength (i.e., defect influence weight) between nodes.

[0099] The system organizes all information of each node and its upstream and downstream edges into a standardized data structure, and combines meta-information (including site number, construction phase identifier, geographic coordinate label, etc.) from the project management system to map to a unified multi-site construction quality database M.

[0100] The quality database M adopts a graph-table hybrid structure modeling form, stores defect behavior chains in a graph structure, maintains work point labels, engineering stage attributes and time window indexes in a table structure, and ensures that the data can be efficiently queried and support graph analysis.

[0101] Through this structure, the platform can realize horizontal comparison of defect performance of multiple work points under similar construction parameters, and can also track the defect development trajectory of a single work point at different time stages.

[0102] On the basis of structured data, the knowledge graph module performs semantic understanding and risk classification on the defect trajectory data in the quality database M, and constructs semantic labels and feature codes for graph node modeling.

[0103] The process mainly includes the following contents:

[0104] Defect semantic classification: based on the morphological features, spatial distribution features and correlation with construction response of defects, all defect instances are classified, such as crack type defects, spalling type defects, pile head damage type defects, and construction parameter abnormality induced type defects. The classification model uses a joint judgment strategy of rule template based preliminary screening + small sample classification model to ensure the robustness of semantic classification.

[0105] Risk level labeling: according to the severity of defects, evolution rate (confidence or area growth per unit time), and degree of association with construction parameters, each defect trajectory is assigned a corresponding risk level (such as low, medium, high, and extremely high). The risk level will be an important attribute of the node in the subsequent knowledge graph.

[0106] High-dimensional pattern feature extraction: the multi-dimensional attributes of each defect trajectory are normalized and coded to form a multi-dimensional pattern feature vector. The vector covers the following dimensions: defect category code, spatial position distribution vector, time evolution vector, construction parameter change vector, and graph structure topology features (such as node degree, clustering coefficient, etc.). The final generated feature vector set will be used as input for subsequent graph embedding and clustering.

[0107] The semantic extraction and labeling in this stage not only enhances the structural integrity of the graph, but also provides a clear semantic basis for subsequent knowledge transfer and pattern analogy.

[0108] After completing the construction of structured data and semantic vector extraction, the system begins to construct a multi-scale, hierarchical knowledge graph K. The design idea of the graph is as follows:

[0109] Graph structure modeling: Defect evolution units are taken as nodes, and the edges between nodes represent risk correlation paths. The generation of edges is not only based on time sequence adjacency and spatial proximity, but also introduces semantic similarity and risk propagation probability modeling, reflecting the potential causal inference relationship between defects. The graph supports node aggregation and edge attribute inheritance, thereby supporting hierarchical mapping, such as aggregating multiple similar nodes in adjacent time periods into a "defect evolution cluster" node.

[0110] Graph embedding and clustering algorithm: Graph embedding algorithms such as DeepWalk, Node2Vec, etc. are used to map the high-dimensional vectors of nodes in the graph to a low-dimensional semantic space, preserving the structural adjacency and attribute similarity of the nodes. Then, hierarchical clustering algorithms such as HAC, spectral clustering are used to classify nodes, forming different levels of defect clustering groups, and realizing the multi-granularity construction of the graph. The upper graph is used for overall risk situation analysis, and the lower graph is used for detailed reasoning and causal diagnosis.

[0111] Graph neural network training and propagation modeling: The constructed knowledge graph K is introduced into the graph convolution network (GCN) for training and learning. The GCN model aggregates the features of the nodes and their neighbor nodes, realizing the deep learning modeling of defect propagation rules. The training targets include node risk prediction, edge weight estimation (defect propagation probability) and graph structure integrity self-supervised optimization.

[0112] Through this graph neural learning process, the system not only understands the explicit relationships in the current graph structure, but also can mine the implicit high-risk propagation paths, realizing early warning of potential defect evolution trends.

[0113] Comparison table of knowledge graph reasoning and defect propagation identification ability

[0114]

[0115] It is explained that the present application can realize path identification and defect evolution rule reasoning which cannot be completed by traditional models by constructing graph structure and introducing graph convolution network, significantly improving intelligent inference ability.

[0116] In the image recognition-based pile intelligent detection cloud platform proposed in the present application, the visualization module is used to present the structured defect data in the knowledge graph K to the terminal user in an interactive and intuitive form, realizing risk perception, path analysis and evolution monitoring of the defect state of the construction area. This module introduces graph attention mechanism, path extraction algorithm and image frame superposition rendering technology to generate interactive and time-dynamic three-dimensional detection view, improving the explainability and user experience of the platform.

[0117] Firstly, the system calculates the risk score of each defect node in the knowledge graph K based on the node and edge information in the cloud. The graph attention mechanism is used to dynamically adjust the degree of information aggregation by learning the semantic weight and structural coupling relationship between each node and its neighbor nodes, thereby accurately reflecting the potential risk concentration of the node in the graph. The system automatically identifies nodes with risk scores higher than the preset threshold as suspected high-risk nodes and forms a high-risk node set. This identification mechanism can achieve dynamic evaluation and automatic labeling of defect risk levels.

[0118] Next, for the risk node set, the system further utilizes the temporal path and spatial location information in the knowledge graph K to construct an evolution path subgraph for each high-risk node. This subgraph not only retains the upstream and downstream defect nodes of the target node on the time axis, but also includes the construction state evolution trajectory and the risk linkage path of the surrounding nodes. By traversing these paths, the system can extract key causal links such as defect evolution logic like "sudden change in pile driving speed → abnormal hitting frequency → pile head peeling", further analyze their impact range and identify risk propagation trends.

[0119] Finally, the system spatially superimposes the above subgraph information with the original image frames. Specifically, based on the positioning information of the foundation pile in each image frame and the node coordinates, the physical location of the defect node in the image is mapped with the graph structure, and three-dimensional coordinate conversion mechanism and dynamic time axis control mechanism are introduced. Finally, a rotatable and scalable three-dimensional detection visualization interface is formed, and users can interactively view the defect evolution state, risk path, and defect feature details in different time periods in this interface.

[0120] Engineering application deployment effect statistics table (8 experimental results)

[0121]

[0122] The platform exhibits high efficiency, strong intelligence, and full automatic processing capability in actual deployment, significantly reducing the burden on manual work, and has cross-site data fusion capability, which is a technical advantage that existing systems cannot match.

[0123] This module realizes the transparent, visual, and intelligent expression of structural defects through the whole process of "graph modeling - risk screening - evolution backtracking - three-dimensional display", solving the technical bottleneck of static, difficult to trace and understand of traditional detection results. Combined with dynamic time axis and graph structure rendering technology, the platform's explainability, practicality, and engineering decision value are significantly enhanced.

[0124] Example 2: To verify the feasibility and effectiveness of the "image recognition-based pile intelligent detection cloud platform" proposed in the present application in actual engineering scenarios, the present example takes the third section of a certain urban rail transit Line 2 project as the test area and selects eight construction pile sites therefrom for on-site deployment and system testing.

[0125] Two high-definition industrial cameras (resolution: 1920x1080, frame rate: 30 fps) are arranged at each site for collecting continuous image sequences I in the whole pile sinking process. A three-axis acceleration sensor and a pressure sensor are synchronously installed for collecting a set of construction parameters P, including:

[0126] Pile sinking pressure (unit: kN), sampling frequency: 10 Hz;

[0127] Hitting frequency (unit: Hz), sampling frequency: 1 Hz;

[0128] Real-time pile sinking depth (unit: m), sampling frequency: 0.5 Hz;

[0129] Timestamp accuracy: millisecond-level alignment.

[0130] A total of more than 180,000 image frames and about 12 GB of construction parameters are collected, covering the whole pile sinking process.

[0131] The image preprocessing module performs inter-frame alignment, light equalization and background removal on the image sequence, and outputs a standardized image tensor I'. The I' is input into a time-series image recognition network constructed based on 3D-CNN+LSTM, and the following defect types are identified: pile top corner drop; concrete honeycomb pitted surface; concrete surface crack; irregular settlement deformation;

[0132] The accuracy is compared by combining the test set constructed by the artificial annotation data:

[0133]

[0134] The results show that the time-series image recognition model used in the present application is significantly better than the traditional static recognition method, especially in the slow time change or weak feature defect morphology scenario.

[0135] The system uses a bidirectional DTW algorithm to perform time-series alignment on the defect trajectory set D and the construction parameter set P. A node set N is constructed based on the attention mechanism to form a construction feature graph G with causal properties.

[0136] In the high-risk pile foundation samples, the platform detects:

[0137] Cumulative defect propagation link: 186;

[0138] Automatic traceability to "unstable hitting frequency-structural crack" path: 42.

[0139] Automatically identify high-risk nodes and successfully give early warning: 38 cases.

[0140] The node semantic vector is extracted from the construction feature map G, a multi-granularity knowledge graph K is constructed, and after training by a graph convolution network, the following performance indicators are realized on the test set:

[0141]

[0142] Through the WebGL interface and the graph node coordinates, the platform realizes a three-dimensional dynamic detection interface, users can replay the defect evolution track through the time axis, and view the visual generated "risk transmission path diagram", realize the whole process traceable detection from identification, cause to source.

[0143] As can be seen from the embodiment, the system has the following remarkable beneficial effects:

[0144] It can realize high-frequency, low-interference data acquisition and image recognition in a real construction environment; the proposed time sequence image recognition model can effectively identify dynamic evolution defects, and the accuracy is significantly improved; the mapping mechanism and the knowledge graph structure support defect tracing, risk reasoning and cross-node knowledge transfer; the three-dimensional visualization module improves the intuitiveness, operability and user understanding of the detection results.

[0145] In summary, the embodiment fully verifies the practical engineering applicability and innovation of the scheme in the quality detection and intelligent analysis of foundation piles.

[0146] The above describes only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can easily think of changes or replacements within the technical range disclosed in the present application, which should be covered within the protection scope of the present application.

Claims

1. An image recognition-based foundation pile intelligent detection cloud platform, characterized in that: The method comprises the following steps: a data acquisition module: used for acquiring a continuous image sequence of a foundation pile construction area in a construction site and synchronously acquiring a corresponding construction parameter set P; an image preprocessing module: used for preprocessing the image sequence and extracting a foundation pile target area and a surface feature thereof in the image, and constructing a standardized image time series tensor I'; a time series image recognition module: based on a trained time series image recognition model, intelligently recognizing defects in the image time series tensor I', acquiring a foundation pile defect type, a spatial position and an evolution trend existing in each image frame, and forming a defect trajectory set D; a mapping module: performing time axis matching and feature fusion on the recognized defect trajectory set D and the synchronously acquired construction parameter set P, and constructing a construction feature graph G; the mapping module comprises: performing time axis alignment processing on the defect trajectory set D and the construction parameter set P, and using a bidirectional dynamic time warping algorithm to pair the time series corresponding relationship between the image frames and the construction parameters; extracting a defect attribute vector and a construction parameter vector at each time, using a feature attention encoder to construct a joint representation node, and forming a node set N; based on a time series relationship and a spatial adjacency rule, constructing an edge set E of a graph structure, and introducing a defect influence weight as an attribute of the edge, and constructing a construction feature graph G=(N, E) with causal features; a knowledge graph module: used for uploading the construction feature graph G to a cloud platform, constructing a standardized multi-worksite construction quality database M in the platform, and generating a multi-scale knowledge graph K based on a defect evolution path and a risk level relationship; the knowledge graph module comprises: structurally extracting node features and edge attributes in the construction feature graph G, combining corresponding work site numbers and geographic labels, mapping into a standardized data model, and constructing a multi-worksite construction quality database M; performing semantic classification and risk level annotation on defect trajectory data in the construction quality database M, extracting a high-dimensional mode feature vector set according to defect types, evolution rates and construction responses; based on graph embedding and hierarchical clustering algorithm, constructing a multi-scale knowledge graph K, wherein the nodes represent defect evolution units, the edges represent risk correlation paths, and the defect propagation rule is learned through a graph convolution network; a visualization module: based on the multi-scale knowledge graph K, automatically identifying suspected high-risk defect nodes in the current construction area, generating a graphical detection result, and pushing the result to a user terminal through a cloud interface, and automatically outputting a detection report R.

2. The image recognition-based foundation pile intelligent detection cloud platform according to claim 1, characterized in that: The construction parameter set P includes pile driving pressure, driving frequency, pile driving depth and timestamp information. 3.The image recognition-based pile intelligent detection cloud platform according to claim 1, characterized in that: The preprocessing of the image sequence includes image enhancement, inter-frame denoising and feature alignment processing. 4.The image recognition-based pile intelligent detection cloud platform according to claim 1, characterized in that: The time series image recognition module comprises: inputting the image time series tensor I' into a three-dimensional convolutional neural network with fused spatial attention and time gating mechanism; using a defect recognition subnetwork trained based on a multi-label classification loss function to perform frame-by-frame inference on each frame of the image sequence, and extracting a structured recognition result containing defect type, severity and confidence; using a long short-term memory unit to model the recognized result in time series, extracting the evolution path and trend change of the defect in the time dimension, and outputting an initial defect trajectory set. 5.The image recognition-based pile intelligent detection cloud platform according to claim 4, characterized in that: The time sequence image recognition module further comprises: performing multi-resolution branch coding processing on the image time sequence tensor I', to construct a multi-scale feature set F containing global contour and local texture features; inputting the multi-scale feature set F into an improved time sequence attention guide network W, which integrates a feature difference sensitivity mechanism and a defect saliency guide mechanism between image frames; integrating the W output result and the traditional three-dimensional convolution recognition result, and using a weighted confidence fusion strategy to generate the final defect trajectory set. 6.The image recognition-based pile intelligent detection cloud platform according to claim 1 or 5, characterized in that: The visualization module comprises: based on the multi-scale knowledge graph K, using a graph attention mechanism to calculate the risk score of the nodes, automatically identifying defect nodes with a risk value higher than a preset threshold, and marking them as a suspected high-risk node set; using the time sequence path and spatial location information in the graph structure, constructing an evolution path subgraph for each defect node in the suspected high-risk node set, and extracting key causal links and influence ranges; spatially superimposing the subgraph information and the corresponding image frames to generate a three-dimensional detection visualization result with a dynamic time axis.

Images