Urban road and bridge intelligent diagnosis method and system based on digital twinning technology

Abstract

The application discloses a city road and bridge intelligent diagnosis method and system based on digital twinning technology, and relates to the field of municipal facility diagnosis.The method comprises the following steps: acquiring bridge digital twinning data and establishing real-time mapping of physical entities and digital models; constructing a bridge disease feature library and an expert knowledge base based on a knowledge graph, realizing structured storage of multi-source heterogeneous information and rule reasoning; adopting a multi-task deep learning fault mode recognition algorithm to output disease classification and spatial positioning; establishing a time-space joint driving adaptive reasoning engine, using a long short-term memory network and a graph neural network to realize trend prediction and spatial correlation analysis; generating disease development trends through multi-step iterative prediction and triggering early warning; and constructing a continuous learning mechanism containing experience playback and knowledge distillation to realize model closed-loop optimization.The application realizes intelligent diagnosis of the fusion of data driving and knowledge driving, and improves the intelligent level and prediction ability of bridge management and maintenance.

Description

Technical Field

[0001] This application belongs to the technical field of municipal facility diagnosis, specifically relating to an intelligent diagnosis method and system for urban roads and bridges based on digital twin technology. Background Technology

[0002] With the deepening of smart city construction, the safe operation and health management of urban roads and bridges, as important infrastructure, have become core issues in urban governance. Digital twin technology, by constructing a digital mapping of physical entities, enables real-time monitoring and simulation analysis of bridge structural conditions, providing a new technical means for bridge maintenance. Among these technologies, the digital twin model layer, as a key bridge connecting the physical entity and the digital space, can integrate multi-source sensing data to establish a three-dimensional visualization model of the bridge structure, supporting the assessment and prediction of structural health status.

[0003] Among them, intelligent bridge diagnosis methods based on digital twin technology are an important research direction in the current field of bridge maintenance. This method aims to improve the level of intelligent bridge maintenance by integrating digital twin modeling capabilities with intelligent analysis algorithms to achieve automatic identification, classification, and diagnosis of bridge defects. Specifically, the intelligent diagnosis method needs to complete a full analysis process from data acquisition and feature extraction to fault mode recognition, and provide maintenance decision recommendations based on the diagnostic results.

[0004] In the prior art, CN120296852B discloses a bridge structure analysis method based on digital twins. Its technical solution focuses on simulation and real-time structural analysis, achieving stress-strain response calculation and evaluation by establishing a finite element model of the bridge. CN112802021B discloses a bridge defect diagnosis method based on unmanned aerial vehicles (UAVs). Its technical solution focuses on using UAVs equipped with camera equipment to inspect the bridge's appearance, and using image processing technology to identify and locate the distribution of defects.

[0005] However, the aforementioned existing technologies all have obvious shortcomings: First, the technical solutions of both patents remain at the level of data-driven analysis and lack the ability to make intelligent diagnostic decisions based on the fusion of multi-source information; Second, the existing diagnostic methods mainly rely on preset thresholds or rule judgments, making it difficult to achieve adaptive identification and reasoning for complex diseases.

[0006] In summary, existing bridge diagnostic technologies are insufficient to meet the higher requirements for intelligent bridge management and maintenance in the context of smart cities. There is an urgent need to build an intelligent diagnostic engine with autonomous learning and reasoning capabilities to achieve a technological leap from data-driven analysis to intelligent decision-making and diagnosis. Summary of the Invention

[0007] The purpose of this invention is to provide an intelligent diagnostic method and system for urban roads and bridges based on digital twin technology, which can effectively solve the problems of existing technologies that are limited to data-driven analysis, lack intelligent diagnostic decision-making capabilities, cannot achieve adaptive identification and reasoning of complex defects, and lack the ability to predict future development trends.

[0008] To achieve the above objectives, the technical solution adopted by the present invention is as follows: an intelligent diagnostic method for urban roads and bridges based on digital twin technology, comprising the following specific steps: This includes acquiring digital twin data of bridges and establishing a real-time mapping relationship between the physical entity of the bridge and its digital model, and also includes: A bridge defect feature database and an expert knowledge base are constructed. The bridge defect feature database and the expert knowledge base are based on knowledge graph technology to perform structured representation and associated storage of multi-source heterogeneous information in the field of bridge defects. The deep learning-based fault mode recognition algorithm extracts general deep features from the bridge digital twin data and outputs the fault type classification results and the spatial location of the fault on the bridge structure. An adaptive inference engine is established, which includes a time series analysis module and a spatial correlation module. The time series analysis module uses a long short-term memory network to analyze and predict time series data, and the spatial correlation module uses a graph neural network to model the spatial topological relationship of bridge structures. The trend prediction module uses a multi-step iterative prediction strategy to generate disease development trend prediction results.

[0009] Furthermore, the construction of the bridge defect feature database and expert knowledge base includes: Define the core entity types in the field of bridge defects, including defect location entities, defect type entities, severity entities, cause analysis entities, and treatment plan entities; Establish relationships between various entities, including causal relationships, concurrent relationships, hierarchical relationships, positional relationships, and temporal relationships; The knowledge graph is stored using a graph database, and the storage architecture adopts an attribute graph model. Each entity node contains a unique identifier, entity type, attribute set, and associated edge information. Expert knowledge is transformed into production rules for representation and integration. Each production rule contains a condition part and a conclusion part. The condition part consists of one or more premise atomic formulas, supporting AND, OR, and NOT logical combinations. The conclusion part contains the reasoning conclusion and its confidence factor.

[0010] Furthermore, the process of transforming and fusing expert knowledge into production rule form also includes: The system is configured with two modes: forward chain reasoning and backward chain reasoning. Forward chain reasoning starts from known facts, gradually triggers production rules that meet the conditions, and arrives at new conclusions. Backward chain reasoning starts from the assumed conclusion and reversely retrieves the production rules and facts that support the conclusion. When multiple production rules are triggered simultaneously, the highest confidence factor among all triggered production rules is taken as the overall confidence level of the diagnostic conclusion.

[0011] Furthermore, the deep learning-based fault mode recognition algorithm includes: A shared feature representation layer is constructed for a multi-task learning framework. The shared feature representation layer uses a residual neural network as the backbone network, and the residual block uses a batch normalization layer and a ReLU activation function. A classification sub-task branch is constructed, which receives feature vectors from the shared feature representation layer and outputs disease type classification results through a fully connected layer and a Softmax layer. A localization sub-task branch is constructed. The localization sub-task branch receives feature vectors from the shared feature representation layer and generates a heat map of the disease location through upsampling and convolution operations. The resolution of the heat map corresponds to the original resolution of the input image or measurement signal. The coordinates of the disease center are determined by the heat map peak detection algorithm. Then, the two-dimensional image coordinates are mapped to three-dimensional spatial coordinates by combining the geometric model of the bridge structure. Design a joint loss function and configure weights for joint optimization. The joint loss function is a weighted sum of classification loss and localization loss.

[0012] Furthermore, the classification loss adopts the cross-entropy loss function, and the localization loss adopts either the mean squared error loss function or the IoU loss function; During joint training, the shared feature representation layer iteratively updates its network parameters through backpropagation and gradient descent under the joint supervision of the classification loss and the localization loss. When the value of the joint loss function converges to the point where it no longer decreases, it is considered that the optimal feature representation has been learned.

[0013] Furthermore, the deep learning-based fault mode recognition algorithm also includes: For vibration signals, time-frequency analysis is used to extract time-domain and frequency-domain features. The time-domain features include statistical features such as mean, variance, peak factor, waveform factor, impulse factor, and margin factor. The frequency-domain features are extracted through Fourier transform or wavelet transform. A visual Transformer architecture is used to extract local and global visual features from image data; An attention mechanism is used to achieve the fusion of multimodal features. The calculation formula for the attention mechanism is as follows: ; in, For querying the matrix, The key matrix, For value matrices, Key matrix Dimensions Key matrix The transpose of .

[0014] Furthermore, the graph structure is constructed based on the physical structure of the bridge, with graph nodes representing key components of the bridge structure and graph edges representing the spatial adjacency and mechanical transmission relationships between components; The trend prediction module uses a multi-step iterative prediction strategy to generate prediction results for the future time range. During the iteration process, the predicted value of the previous moment is included in the input sequence, while the oldest historical data point in the sequence is removed to keep the length of the input sequence fixed.

[0015] Furthermore, the digital twin model layer interacts with the intelligent diagnostic engine through a unified data interface. The physical simulation module establishes a bridge mechanics model based on the finite element method. The strain and displacement readings in the real-time monitoring data are converted and used as the boundary conditions of the finite element model, which are then loaded onto the corresponding nodes and elements of the finite element model.

[0016] Furthermore, it also includes building a continuous learning mechanism: An experience replay buffer is set up to store historical training samples. A circular queue structure is used for management. Each time the model is updated, historical samples are randomly selected from the experience replay buffer and mixed with newly collected samples to form a training batch. The mixing ratio of historical samples to new samples ranges from 1:1 to 1:3. Configure a priority experience replay mechanism, which assigns different sampling priorities based on the learning difficulty of the samples, and the learning difficulty is measured by the loss value of the samples; Configure the teacher and student models for knowledge distillation, using the diagnostic model before the update as the teacher model and the updated diagnostic model as the student model. Design a loss function and training strategy for knowledge distillation. The loss function consists of two parts: hard label loss and soft label loss. The hard label loss adopts the cross-entropy loss function, and the soft label loss adopts the KL divergence loss function. The soft labels are generated by inputting training samples into a teacher model, which then sets the distillation temperature parameter... The forward computation is performed, and the output, adjusted for distillation temperature parameters, serves as the class probability distribution as the soft label. Set to 4.0, during the distillation process, the distillation temperature parameter is used first. The student model is trained to learn the soft label distribution output by the teacher model, and then the temperature parameter is used. Make fine adjustments; A manual review mechanism and model update triggering conditions are established. Under normal operating conditions, the sampling ratio is 5% to 10%, which is increased to 20% to 30% under abnormal conditions. The criteria for judging abnormal conditions are: the false negative rate exceeds 5% in multiple consecutive batches, or the average prediction confidence of the model output is lower than 0.6. When the false negative rate exceeds 5% for two consecutive update cycles or the diagnostic accuracy is lower than 85% for two consecutive update cycles, the model update process is triggered.

[0017] A smart diagnostic system for urban roads and bridges based on digital twin technology includes: The data acquisition module is used to acquire bridge digital twin data and establish a real-time mapping relationship between the bridge physical entity and the digital model; The knowledge graph construction module is used to build a bridge defect feature database and an expert knowledge base, and to perform structured representation and associated storage of multi-source heterogeneous information in the field of bridge defects based on knowledge graph technology. The fault mode recognition module is used to extract general deep features from the bridge digital twin data based on the deep learning-based fault mode recognition algorithm, and output the classification results of the fault type and the spatial location of the fault on the bridge structure. The adaptive inference engine module includes a time series analysis module and a spatial correlation module. The time series analysis module uses a long short-term memory network to analyze and predict time series data, and the spatial correlation module uses a graph neural network to model the spatial topological relationships of bridge structures. The trend prediction and early warning module is used to generate disease development trend prediction results and trigger early warnings using a multi-step iterative prediction strategy.

[0018] In summary, this application includes at least one of the following beneficial technical effects: 1. This invention addresses the problem of existing technologies lacking intelligent diagnostic decision-making capabilities based on multi-source information fusion by constructing a bridge defect feature database and an expert knowledge base, structurally representing the diagnostic experience of domain experts, and performing collaborative reasoning with deep learning models.

[0019] 2. This invention designs a fault mode recognition algorithm based on deep learning, which adopts a multi-task learning framework and attention mechanism to achieve automatic classification and accurate location of complex diseases such as cracks, deformation, and corrosion, and solves the problem that existing technologies rely on preset thresholds or rules for judgment and cannot handle complex nonlinear disease modes.

[0020] 3. This invention establishes a time-series-space jointly driven adaptive inference engine, which jointly models time-series prediction and spatial inference to achieve accurate prediction and early warning of disease development trends. The prediction time span covers the preset prediction time span, providing a sufficient forward-looking decision-making time window for bridge maintenance.

[0021] 4. This invention constructs a continuous learning mechanism, achieving closed-loop optimization of the diagnostic model through experience replay and knowledge distillation techniques. The model can continuously improve diagnostic accuracy as data accumulates, solving the problem that existing technology models are static and fixed, unable to adapt to changes in scenarios.

[0022] 5. Achieve a deep integration architecture between the digital twin model layer and the intelligent diagnostic engine, supporting real-time linkage between physical entities and digital models. Diagnostic results can be intuitively displayed on the 3D model, improving the intelligence and visualization level of bridge maintenance. Attached Figure Description

[0023] Figure 1 A schematic diagram of the overall technical architecture of an intelligent diagnostic method for urban roads and bridges based on digital twin technology; Figure 2 A schematic diagram illustrating the core principle framework for the integration of knowledge graphs and deep learning; Figure 3 A flowchart outlining the main stages of data acquisition and multimodal perception for bridge digital twins; Figure 4 A schematic diagram of the multi-level interaction relationships and data flow of an adaptive inference engine jointly driven by temporal and spatial characteristics; Figure 5 A logical flowchart illustrating the continuous learning mechanism and the principle of model self-evolution. Figure 6 This is a schematic diagram of the architecture for the deep integration of the digital twin model layer and the intelligent diagnostic engine. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of this invention clearer, the following description is provided in conjunction with the appendix. Figure 1 To be continued Figure 6 The present invention will be further described in detail below with reference to specific embodiments.

[0025] This embodiment elaborates on a method for intelligent diagnosis of urban road bridges based on digital twin technology. The method comprises five core steps, utilizing the synergistic effect of knowledge graph technology, deep learning algorithms, adaptive inference engines, and continuous learning mechanisms to achieve intelligent diagnosis and prediction of bridge structures.

[0026] First, in step S1, a bridge defect feature database and an expert knowledge base are constructed. Step S1 is based on knowledge graph technology to perform structured representation and associated storage of multi-source heterogeneous information in the field of bridge defects, providing knowledge support for subsequent intelligent diagnosis. The specific implementation is carried out according to the following steps.

[0027] Step S101 defines the core entity types in the bridge defect domain. Defect location entities cover structural components such as bridge deck, main beams, secondary beams, piers, abutments, bearings, expansion joints, and drainage systems. Defect type entities include crack defects, deformation defects, corrosion defects, leakage defects, damage defects, and displacement defects.

[0028] Severity entities are classified into four levels according to industry standards: minor, moderate, severe, and critical. Cause analysis entities cover factors such as overload, material aging, construction defects, foundation settlement, vehicle impact, and environmental erosion. Treatment plans include measures such as routine maintenance, minor repairs, medium repairs, major repairs, reinforcement and renovation, and demolition and reconstruction.

[0029] Step S102 establishes the relationships between various entities. Causal relationships describe the logical connection between the causes and manifestations of defects, such as the causal relationship between overloaded transportation and bridge deck cracks. Concurrency relationships describe the correlation pattern of multiple defects occurring simultaneously, such as the possibility of bearing detachment and beam deformation occurring concurrently.

[0030] Hierarchical relationships describe the inclusion relationships between structural components, such as a main beam being contained within the superstructure. Locational relationships describe the spatial location of defects, such as a crack located on the bridge deck at a quarter-section from the support. Temporal relationships describe the time sequence of defect development and evolution, such as a minor crack evolving into a severe crack.

[0031] Step S103 involves storing the knowledge graph using a graph database. The storage architecture employs an attribute graph model, where each entity node contains a unique identifier, entity type, attribute set, and relational edge information. Entity attributes are differentiated according to different types. Disease type entities include attributes such as name, classification code, typical feature description, and detection method. Severity entities include attributes such as level number, quantification threshold range, and treatment time limit requirements. Relationship edges include attributes such as relationship type, weight coefficient, and confidence level to support multi-dimensional quantitative representation of relationships.

[0032] Step S104: Transform expert knowledge into production rule form for representation and integration. Each production rule includes a condition part and a conclusion part. The condition part consists of one or more premise atomic formulas, supporting AND, OR, and NOT logical combinations.

[0033] The conclusion section includes the reasoning conclusion and its confidence factor. The confidence factor ranges from 0 to 1 and is determined by domain experts based on engineering experience and the causal certainty between the premises and conclusions of the production rule. The higher the value, the more reliable the production rule is in engineering practice.

[0034] For example, if the width of the bridge deck crack is greater than 0.2 mm and the crack distribution density exceeds 3 cracks per square meter, and the average daily traffic volume of the bridge is greater than 1.2 times the design traffic volume, then the conclusion is that the load-bearing capacity of the bridge deck has decreased and a special inspection is recommended, with a confidence factor set at 0.85.

[0035] Step S103 persists entities and relationships in the form of a graph database, enabling the knowledge graph to efficiently read and write large-scale data and query complex relationships. Step S104 transforms the diagnostic experience of domain experts into computable production rules, upgrading the knowledge graph from a static fact store to a knowledge system with reasoning capabilities.

[0036] Step S105: Configure two modes for expert rule forward chain reasoning and backward chain reasoning. Forward chain reasoning starts from known facts, gradually triggers production rules that meet the conditions, and arrives at new conclusions. It is suitable for diagnostic deduction scenarios. Backward chain reasoning starts from the assumed conclusion and reversely retrieves the production rules and facts that support the conclusion. It is suitable for fault tracing scenarios.

[0037] When multiple production rules are triggered simultaneously, the highest confidence factor among all triggered production rules is taken as the overall confidence of the diagnostic conclusion, and the overall confidence and conclusion are output together.

[0038] For forward chain reasoning, the input information comes from real-time monitoring data collected by sensors in step S2 and generated after feature extraction in step S3, as well as static structural information stored in the graph database in step S103. The reasoning process matches the production rule conditions one by one, substituting the real-time data and static information into the premise atomic formulas of each production rule for logical judgment. When all the premise conditions of a production rule are satisfied, the production rule is triggered.

[0039] The inference output consists of the diagnostic conclusions and their confidence factors corresponding to the production rules that meet the conditions, which are presented to managers as auxiliary decision-making basis for subsequent steps. The input of backward chain inference is a description of the abnormal phenomenon to be traced. The inference process starts from the assumed conclusion, searches backward for the production rules that can derive the conclusion, and then recursively checks whether the facts corresponding to the preconditions of each production rule exist in the knowledge graph, until the root cause is found or further tracing is no longer possible.

[0040] In summary, step S1 completes the construction of the knowledge graph through entity definition, relationship establishment, graph database storage, production rule representation, and dual-mode reasoning configuration, providing domain knowledge constraints for subsequent deep learning models and supporting the integration and diagnosis of data-driven and knowledge-driven approaches.

[0041] The next step, S2, is to acquire digital twin data of the bridge. By deploying a multimodal sensor network, multidimensional data such as vibration, strain, displacement, and environment of the bridge are collected, and a real-time mapping relationship between the physical entity of the bridge and the digital model is established to provide a data foundation for subsequent intelligent diagnosis. This includes the following steps.

[0042] Step S201: Deploy accelerometers to collect dynamic response signals of the bridge. The accelerometers are capacitive or MEMS accelerometers, and the range is determined according to the bridge's design load level, usually within the range of ±2g to ±10g. The sampling frequency is not less than 1000Hz to meet the requirements of modal analysis.

[0043] Accelerometers are arranged along the longitudinal, transverse, and vertical directions of the bridge, with measuring points placed at key locations such as the mid-span, quarter-span, and above the supports. Each key measuring point has a triaxial accelerometer. Sensor signals are transmitted to a data acquisition unit via wired or wireless means. The data acquisition unit has real-time acquisition, buffering, and preprocessing functions. Before sampling begins, the sensors are calibrated and zero-point calibrated.

[0044] Step S202: Deploy strain sensors to monitor strain changes at key stress points. Fiber optic strain gauges or resistance strain gauges are selected as the strain sensors, with a measurement accuracy of no less than 0.1%FS and a resolution of no less than 1με. Strain sensors are deployed at locations with significant internal forces, such as the mid-span section of the main beam, sections near supports, and sections with abrupt structural changes. Multiple measuring points are arranged along the height of the section to obtain strain distribution curves.

[0045] Temperature-compensated strain sensors are paired with working strain sensors to eliminate the influence of temperature changes on strain measurements. The strain data acquisition frequency is dynamically adjusted according to the bridge's operating status; the acquisition frequency is 50Hz under normal operating conditions, and automatically increases to 200Hz when an abnormal event is triggered.

[0046] Step S203: Deploy displacement sensors to measure the static deformation of the bridge. The displacement sensors are either wire-type displacement sensors or GNSS displacement monitoring systems. The measurement range covers the bridge's design deflection, typically ±100mm to ±500mm, and the measurement accuracy is not less than 0.1mm.

[0047] Displacement sensors are installed at the abutments at both ends of the bridge, the tops of the piers, and the bearing installation locations to measure parameters such as the overall settlement, horizontal displacement, and bearing deformation of the bridge. For long-span bridges, total stations or GNSS instruments are used for automated displacement monitoring, enabling continuous observation around the clock.

[0048] Step S204: Deploy environmental sensors to collect environmental parameters such as temperature, humidity, and wind speed. The environmental sensors include a temperature sensor, humidity sensor, wind speed sensor, wind direction sensor, and precipitation sensor. The temperature sensor has a measurement range of -40°C to +80°C and an accuracy of ±0.5°C.

[0049] The humidity sensor measures from 0% to 100%RH with an accuracy of ±3%RH. The wind speed sensor measures from 0 to 60 m / s with an accuracy of ±0.3 m / s. The environmental sensors are installed at the windward locations on the bridge and in the clearance area beneath the bridge to avoid any obstruction of the measurement results by the bridge structure.

[0050] Step S204 supplements the ability to collect environmental parameters. Environmental factors influence the structural behavior of bridges. Temperature changes can cause beam expansion and contraction and stress redistribution, wind loads can induce vibration responses, and humidity is closely related to steel corrosion and concrete deterioration. After deploying environmental sensors, subsequent diagnostic analysis can correlate structural responses with environmental conditions, eliminating the interference of environmental factors in the assessment of structural condition.

[0051] Step S205 integrates the data collected by the multimodal sensors through a unified data interface. The data interface conforms to the IoT communication protocol standard and supports communication protocols such as Modbus, TCP / IP, and MQTT. After the raw data enters the data processing unit, it undergoes preprocessing operations such as noise reduction filtering, outlier removal, time synchronization, and format conversion in sequence.

[0052] Noise reduction filtering employs moving average filtering or low-pass filtering to eliminate high-frequency noise interference with the effective signal. Outlier removal follows the Laida criterion, marking measurements deviating from the mean by more than three standard deviations as outliers and removing them. Time synchronization uses a unified network time protocol server timestamp as a reference to align the data collected by each sensor.

[0053] The format conversion process unifies the data reported by different communication protocols into the standard JSON format. The preprocessed data is simultaneously pushed to both the real-time database and the historical database. The real-time database uses a time-series database to store the monitoring data for the most recent 72 hours, while the historical database uses a relational database to store all historical data.

[0054] Step S206 involves establishing a real-time mapping relationship between the physical entity of the bridge and its digital model using a digital twin platform. The digital twin platform constructs a three-dimensional geometric model of the bridge with an accuracy reaching LOD3 level, clearly representing the bridge's detailed structure. When establishing a mapping relationship between the three-dimensional geometric model and monitoring data, each sensor measurement point is assigned a unique identification number, and the measurement point number is bound one-to-one with the corresponding geometric coordinate point in the three-dimensional model.

[0055] After data acquisition, the spatial location of the measuring points in the 3D model is indexed by the measuring point number. Real-time data is presented intuitively on the 3D model through color mapping, animation, and other methods. The physical simulation module establishes a bridge mechanical model based on the finite element method. The inputs are the real-time monitoring data pushed in step S205 and the material constitutive parameters, including the elastic modulus, Poisson's ratio, density, etc. of each component.

[0056] The physical simulation module outputs the displacement and stress field distributions of each node of the bridge under various load conditions. During the simulation, the strain and displacement readings from the real-time monitoring data are converted and used as boundary conditions for the finite element model, which are then applied to the corresponding nodes and elements of the model, thereby achieving real-time simulation analysis of the bridge's working state.

[0057] In summary, step S2 establishes a real-time data channel between the physical bridge and the virtual model through sensor deployment, data integration, and digital twin mapping. The output multimodal monitoring data and real-time simulation results will serve as inputs to the fault mode recognition algorithm in step S3.

[0058] The next step, S3, involves designing a deep learning-based fault mode recognition algorithm, including the following steps: Step S301: Construct a shared feature representation layer for the multi-task learning framework. The shared feature representation layer uses a residual neural network as the backbone network, with a depth of 50 or 101 layers. Batch normalization layers and ReLU activation functions are used within the residual blocks. The shared feature representation layer extracts general deep features from the multimodal raw data, and the dimension of the output feature vector is set to 2048.

[0059] Step S302: Construct the classification sub-task branch. The classification sub-task branch receives a 2048-dimensional feature vector from the shared feature representation layer and outputs the disease type classification result through a fully connected layer and a Softmax layer. The disease types supported by the classification sub-task include: cracks, subdivided into network cracks, longitudinal cracks, transverse cracks, and diagonal cracks; deformation, subdivided into vertical deflection deformation, transverse offset deformation, and torsional deformation; corrosion, subdivided into steel reinforcement corrosion, concrete carbonation, and steel structure corrosion; and other disease types including leakage, damage, and abnormal vibration. The classification sub-task outputs the probability distribution of each category and selects the category with the highest probability as the final classification result.

[0060] Step S303: Construct a localization sub-task branch. This branch receives a 2048-dimensional feature vector from the shared feature representation layer and generates a heatmap of the lesion location through upsampling and convolution operations. The heatmap resolution corresponds to the original resolution of the input image or measurement signal, and the value of each pixel on the heatmap represents the probability of lesion presence at that location.

[0061] The localization subtask outputs the spatial location of the defect on the bridge structure. The center coordinates of the defect are determined by the peak detection algorithm of the heat map. Then, the two-dimensional image coordinates are mapped to three-dimensional spatial coordinates by combining the geometric model of the bridge structure.

[0062] Step S304: Design the joint loss function and configure weights for joint optimization. The classification loss for the classification subtask uses the cross-entropy loss function, and the localization loss for the localization subtask uses the mean squared error loss function or the IoU loss function. The joint loss function is a weighted sum of the classification loss and the localization loss, calculated as follows: ; in, For the joint loss function value, For classification loss weights, The cross-entropy loss function value is used for classification. To locate the loss weights, To locate the mean squared error loss function value or the IoU loss function value. Classification loss weights. Set to 1.0, location loss weight Set to 0.5.

[0063] Classification cross-entropy loss function The calculation formula is: ; in, This represents the total number of disease categories. For category indexing, For the first The true label for each category, with a value of 0 or 1. The model predicts the first The probability value for each category, ranging from 0 to 1.

[0064] Positioning mean square error loss function The calculation formula is: ; in, To determine the total number of dimensions at the output location of the positioning task, For dimensional indexing, For the first The true location coordinates of each dimension The model predicts the first Each dimension of the position coordinates.

[0065] Locating the IoU loss function The calculation formula is: ; in, For the predicted disease-affected areas, This represents the actual diseased area. To predict the area or number of voxels at the intersection of the predicted region and the real region, This is the area or number of voxels of the union between the predicted region and the ground truth region. The localization loss can be either the mean squared error loss function or the IoU loss function, chosen based on the specific training results.

[0066] During joint training, the shared feature representation layer updates its parameters under the joint supervision of classification loss and localization loss. In the joint loss function... Under joint supervision, the network parameters of the shared feature representation layer are iteratively updated through backpropagation and gradient descent. When the joint loss function value... When the feature representation converges to the point where it no longer decreases, it is considered to have learned the optimal feature representation, which is beneficial for both classification and localization tasks, thereby improving the overall recognition accuracy.

[0067] Step S305: Design differentiated feature extraction strategies for input data of different modes. For vibration signals, time-frequency analysis is used to extract time-domain and frequency-domain features. Time-domain features include statistical features such as mean, variance, peak factor, waveform factor, impulse factor, and margin factor. Frequency-domain features are extracted through Fourier transform or wavelet transform, including features such as dominant frequency, spectral centroid, spectral entropy, and power spectral density.

[0068] For image data, a visual Transformer architecture is used to extract local and global visual features. The visual Transformer divides the image into fixed-size image patches and models the dependencies between the image patches through a self-attention mechanism. The extracted visual features include local texture features and global semantic features.

[0069] Step S306: An attention mechanism is used to fuse multimodal features. The core calculation of the attention mechanism is scaled dot product attention, and its calculation formula is as follows: ; in, For querying the matrix, The key matrix, For value matrices, Key matrix Dimensions Key matrix The transpose of the matrix, For query matrix AND key matrix transpose Matrix multiplication between Key Dimension The square root of the product is used as a scaling factor to control the magnitude of the dot product result. The normalized exponential function maps the attention weights to a range of 0 to 1, with the sum of all weights being 1.

[0070] The specific computational process of the attention mechanism is as follows: query matrix transpose of the key matrix Perform a dot product operation to obtain the correlation score matrix between the modal features. The correlation score matrix Divide each element by the scaling factor This yields the scaled score matrix.

[0071] Apply row-by-row to the scaled score matrix The function is normalized to obtain the attention weight matrix. The attention weight matrix and the value matrix are combined. Multiplying yields the weighted aggregated fused feature vector. .

[0072] The attention mechanism automatically learns the contribution of different modal features. When the data quality of a certain modality is poor or missing, its corresponding attention weight is automatically reduced, thereby weakening the negative impact of that modality feature on the fusion result. The fused feature vector is set to 512 dimensions and is passed into the subsequent classification subtask branch and localization subtask branch.

[0073] Thus, step S3 completes the design of the deep learning-based fault mode recognition algorithm. From constructing the multi-task learning framework, setting the shared feature representation layer, implementing the classification and localization sub-tasks, to optimizing the joint loss function, implementing multimodal differentiated feature extraction strategies, and performing attention-based feature fusion, a fault mode recognition scheme is formed. The disease category and spatial location results output in step S3 will serve as input to the adaptive inference engine in step S4, supporting subsequent disease development trend prediction and early warning.

[0074] For step S4 in the method of this embodiment, an adaptive inference engine is established. Combining time series data analysis and spatial correlation, graph neural network technology is used to model the spatial topological relationship of the bridge structure to realize the prediction and early warning of the development trend of the disease. The following steps are followed.

[0075] Step S401: Construct the time series analysis module. The time series analysis module uses a Long Short-Term Memory (LSTM) network to analyze and predict time series data. The LSM network contains three gating units: an input gate, a forget gate, and an output gate. The input gate determines how much information from the current time step is written into the cell state. The forget gate determines how much of the cell state from the previous time step is retained.

[0076] The output gate determines how much of the current cell state is output. The synergistic effect of the three gating units enables the Long Short-Term Memory network to capture long-term dependencies in the monitored data, avoiding gradient vanishing.

[0077] Step S402: Configure the training strategy and operation mode of the time series analysis module. The input to the Long Short-Term Memory (LSTM) network is the time series of the monitoring data, and the output is the predicted value for the next time step. The number of hidden layer nodes is set to 128, the number of stacked layers is set to 2, and the hyperbolic tangent function is used as the activation function. The network adopts a bidirectional structure, containing LSM layers in both forward and backward propagation directions.

[0078] During the training phase, supervised learning is performed using historical monitoring data. The training data is constructed as follows: input segments are extracted from the time series of historical monitoring data using a sliding window of fixed length. The window length is the monitoring data of the past 24 time steps, and the supervision label corresponding to each input segment is the monitoring value of the next time step after the window.

[0079] During training, each iteration randomly selects a batch of samples from the training dataset. The input segments from these samples are fed into the network for forward computation, calculating the mean squared error loss between the predicted values ​​and the supervision labels. The network parameters are then updated via backpropagation. The Adam optimizer is used, with a learning rate set to 0.001. This iterative process is repeated until the loss function value no longer decreases on the validation set, at which point training stops.

[0080] Step S403: Construct a spatial association module. The spatial association module uses a graph neural network to model the spatial topological relationships of the bridge structure. The graph structure is constructed based on the physical structure of the bridge, and the graph nodes represent key components of the bridge structure, including the bridge deck, main beams, secondary beams, piers, abutments, supports, expansion joints, etc.

[0081] The edges in the graph represent the spatial adjacency and mechanical transmission relationships between components. Edges are established between adjacent components and between components with mechanical force transmission paths.

[0082] Step S404: Configure the node features, edge features, and aggregation function of the graph neural network. The node features of the graph neural network include the geometric and mechanical properties of the components. Geometric properties include component length, cross-sectional dimensions, area, moment of inertia, etc.

[0083] Mechanical properties include stress state, strain state, and damage index. The edge features of a graph neural network contain attributes of the relationships between components, including connection type, force transmission path, and stiffness ratio.

[0084] The aggregation function of the graph neural network is implemented using a graph attention mechanism. The graph attention mechanism assigns attention weights to each neighboring node. The attention weights are calculated by concatenating the feature vector of the current node with the feature vectors of its neighbors into a combined vector, which is then input into a single-layer fully connected network. The output is a scalar value as the unnormalized attention score. The weight parameters of the fully connected network are learned during the model training process.

[0085] After calculating the unnormalized attention scores for all neighbor nodes, the Softmax function is used to normalize these scores. The normalized attention weights are multiplied by the corresponding feature vectors of the neighbor nodes, and the weighted results of all neighbor nodes are summed to obtain the updated feature vector of the current node.

[0086] Graph attention mechanisms can adaptively learn the importance weights of different neighboring nodes. When the state of a neighboring node has a significant impact on the current node, the model assigns it a higher attention weight.

[0087] In bridge diagnostics, when a component exhibits an abnormal state, the model can automatically focus on that component and assess its impact on adjacent components. This has practical value for scenarios such as local damage propagation and stress redistribution in adjacent components caused by support failure.

[0088] Step S405: Construct the trend prediction module. The trend prediction module uses a multi-step iterative prediction strategy to generate prediction results for the future time range. First, it generates the predicted value for the next time step. The time step length is determined based on the sampling interval of the monitoring data, typically 1 hour or 1 day.

[0089] The predicted values ​​are then used as input for the next step, and combined with historical sequence data to iteratively generate prediction results for subsequent time steps. During the iteration process, the predicted values ​​from the previous time step are included in the input sequence, while the oldest historical data points in the sequence are removed, keeping the length of the input sequence constant.

[0090] Step S406: Configure the output and early warning triggering mechanism of the trend prediction module. The prediction time span covers the next 7 to 30 days. The trend prediction module output includes point prediction values ​​and confidence intervals. Point prediction values ​​are the predicted values ​​for each future time point. Confidence intervals are generated using Monte Carlo dropout or Bayesian inference methods to reflect the uncertainty of the prediction results.

[0091] The warning thresholds are determined based on the structural response limits specified in the bridge design code. 90% of the allowable value in the design code is used as the yellow warning trigger threshold, and 100% of the allowable value in the design code is used as the red warning trigger threshold. When the predicted value exceeds the yellow warning trigger threshold, a yellow warning message is sent to bridge management personnel.

[0092] When the predicted value of a point exceeds the red alert trigger threshold, a red alert is triggered, and an alert message is generated and pushed to bridge management personnel. The alert message includes the location of the exceeding measuring point, the predicted time of exceeding the limit, the predicted exceeding value, and the confidence interval.

[0093] In summary, step S4 captures the temporal evolution patterns of monitoring data through the time series analysis module, models the spatial topological dependencies of the bridge structure through the spatial correlation module, and generates multi-step prediction results with confidence intervals through the trend prediction module. These three modules work collaboratively to achieve a comprehensive assessment and early warning of bridge defect development trends. The prediction results and early warning information output from step S4 will be transmitted to subsequent steps for maintenance personnel to reference in decision-making and will be displayed in conjunction with the digital twin model layer.

[0094] Finally, step S5 involves constructing a continuous learning mechanism, which is implemented according to the following steps.

[0095] Step S501: Set up an experience replay buffer. This buffer stores historical training samples with a capacity of 10,000 samples, managed using a circular queue structure. New samples replace the oldest samples upon arrival. Each time the model is updated, historical samples are randomly selected from the replay buffer and mixed with newly collected samples to form a training batch. The mixing ratio of historical to new samples is set according to task requirements, ranging from 1:1 to 1:3.

[0096] Step S502: Configure the priority experience replay mechanism. The priority experience replay mechanism assigns different sampling priorities based on the learning difficulty of the samples. Learning difficulty is measured by the loss value of the samples; samples with larger loss values ​​have higher priorities. The samples in the buffer are divided into multiple priority levels according to their loss values, with higher priority levels assigned higher sampling probabilities. Uniform sampling is performed within the same level.

[0097] Step S503: Configure the teacher and student models for knowledge distillation. During knowledge distillation, the diagnostic model before the update serves as the teacher model, and the updated diagnostic model serves as the student model. The teacher and student models have the same network structure but different parameters. The parameters of the teacher model are relatively fixed, while the parameters of the student model are updated using the backpropagation algorithm.

[0098] Step S504: Design the loss function and training strategy for knowledge distillation. The loss function for knowledge distillation consists of two parts: hard label loss and soft label loss. The hard label loss uses the cross-entropy loss function to measure the difference between the student model's prediction and the true label.

[0099] The soft label loss uses the KL divergence loss function to measure the difference between the student model's predicted distribution on the soft labels generated by the teacher model and the teacher model's predicted distribution. The total loss function is a weighted sum of the hard label loss and the distillation loss, calculated as follows: ; in, The total loss function value for knowledge distillation. For hard labels, weight loss is applied. The value of the hard-label cross-entropy loss function. For distillation loss weighting, The values ​​represent the KL divergence loss function for soft-labeled samples. The weights represent the loss function for hard-labeled samples. Set to 0.7, distillation loss weight Set to 0.3.

[0100] The soft label generation method is as follows: the training samples are input into the teacher model, and the teacher model adjusts the distillation temperature parameter. The forward calculation is performed, and the output category probability distribution after adjusting the distillation temperature parameter is used as a soft label.

[0101] Distillation temperature parameters Setting it to 4.0 makes the prediction distribution of the teacher model smoother and enhances the transfer of knowledge about relationships between categories. During the distillation process, the distillation temperature parameter is used first. Train the student model to learn the soft label distribution output by the teacher model, and then use normal temperature parameters. Make minor adjustments.

[0102] Step S505: Define the evaluation index system for diagnostic results. The evaluation indexes include diagnostic accuracy and false negative rate.

[0103] Diagnostic accuracy Defined as the ratio of the number of correctly diagnosed samples to the total number of samples, the formula is: ; in, The number of samples that were correctly diagnosed. This represents the total number of samples.

[0104] False negative rate Defined as the ratio of the number of samples with actual disease but not detected to the number of samples with actual disease, the calculation formula is: ; in, This represents the number of samples that actually had defects but were not detected. This represents the actual number of samples with existing diseases.

[0105] Step S506: Establish a manual review mechanism and model update trigger conditions. The manual review mechanism performs sampling verification of the model diagnostic results, with the sampling ratio dynamically adjusted according to the operating status. Under normal operating conditions, the sampling ratio is 5% to 10%, increasing to 20% to 30% under abnormal conditions. The criteria for determining abnormal conditions are: a missed detection rate exceeding 5% in multiple consecutive batches, or an average prediction confidence level of the model output being lower than 0.6.

[0106] The manual review results include categories such as correct diagnosis, incorrect diagnosis, missed detection, and false positive. These results are fed back into the model training process to calculate the labels for incremental learning samples. The review results are also used to update evaluation metrics. The model update process is triggered when the missed detection rate exceeds 5% for two consecutive update cycles or the diagnostic accuracy falls below 85% for two consecutive update cycles.

[0107] Steps S505 and S506 complete the construction of the evaluation feedback module. The evaluation index system provides a measurement standard for the quantitative evaluation of model performance. The manual review mechanism introduces the judgment of domain experts into a closed loop, ensuring that the direction of continuous model optimization remains consistent with actual engineering needs.

[0108] Thus, step S5 preserves historical knowledge and focuses on difficult samples through the experience playback module, realizes the structured evolution and knowledge inheritance of the model through the knowledge distillation module, and establishes a closed-loop optimization mechanism of quantitative evaluation and manual review through the evaluation and feedback module. Together with steps S1 to S4, this forms a complete closed loop, enabling the entire intelligent diagnosis to continuously improve itself with data accumulation.

[0109] This invention also includes a deeply integrated architecture of a digital twin model layer and an intelligent diagnostic engine. The digital twin model layer provides a 3D visualization model and physical simulation capabilities for the bridge. The 3D visualization model is built based on BIM technology, achieving a LOD3 level accuracy and containing geometric, material, and texture information for each bridge component. The physical simulation module is implemented based on the finite element method and can simulate the mechanical response of the bridge under various load conditions.

[0110] The intelligent diagnostic engine uses knowledge graphs and deep learning for fault reasoning. The knowledge graph stores factual and rule-based knowledge in the bridge's fault domain, providing domain knowledge support for diagnostic reasoning; the deep learning model learns fault patterns from monitoring data, enabling automatic identification and classification of faults. The intelligent diagnostic engine receives real-time monitoring data pushed by the digital twin model layer and outputs diagnostic conclusions and early warning information.

[0111] The digital twin model layer interacts with the intelligent diagnostic engine through a unified data interface. This data interface defines standard data formats and communication protocols, supporting bidirectional transmission of monitoring data, model parameters, and diagnostic results. Real-time linkage between the physical entity and the digital model is achieved through a data synchronization mechanism. Data collected by sensors is preprocessed and then pushed to the digital twin model in real time. The geometric state and simulation results of the digital twin model reflect the current state of the bridge's physical entity in real time.

[0112] The diagnostic results are visualized on the 3D model. The visualization includes marking the location of defects, identifying defect types, coloring for severity, and indicating trends with arrows. Managers can intuitively understand the bridge's health status through the 3D model, and clicking on specific components allows them to view detailed diagnostic information and historical evolution curves.

[0113] To facilitate understanding, let's take a bridge spanning a river on a major urban road as an example. This bridge is a prestressed concrete continuous box girder bridge with a span arrangement of 50 meters + 80 meters + 50 meters, a bridge deck width of 28 meters, and six lanes in both directions. The bridge experiences heavy daily traffic, with frequent passage of heavy vehicles, and the maintenance unit places great emphasis on the structural safety of the bridge.

[0114] In step S1, a dedicated knowledge graph is constructed for this bridge. Entity types include: specific types of defects unique to this bridge type, such as diagonal cracks in the box girder web, transverse cracks in the bottom slab, and bearing detachment; typical causes of defects in this bridge type, such as prestress loss, shrinkage and creep, and temperature gradient effects; and commonly used treatment methods for this bridge type, such as external prestressing reinforcement, bonded steel plate reinforcement, and bearing replacement. Expert rules include: if the angle of the diagonal crack in the box girder web is greater than 45 degrees and the crack width is greater than 0.3 mm, the conclusion is that the crack is a structural crack and requires specialized testing and evaluation, with a confidence level of 0.90.

[0115] In step S2, a complete health monitoring system is deployed on the bridge. Accelerometers are placed at 24 monitoring points at the mid-span, quarter-span, and above the supports of each span, with each point arranged in three directions. Strain sensors are placed at 18 cross-sections at the mid-span and near the supports of each span, with 8 to 12 monitoring points per cross-section. Displacement sensors are placed at 8 monitoring points on the abutments and pier tops. Environmental sensors are placed at 4 monitoring points on both sides of the bridge. Sensor data is transmitted to the monitoring center in real time via a wired network, and the sampling frequency is dynamically adjusted according to the operational status.

[0116] In step S3, a fault mode recognition model is trained based on the bridge's historical inspection and monitoring data. The training data includes: approximately 5,000 damage records from the bridge's annual periodic inspection reports; and approximately 10,000 sets of vibration and strain data and their corresponding status labels accumulated by the health monitoring system. After the multi-task learning model is trained, its classification accuracy on the test set is approximately 92%, and its positioning accuracy is approximately 95%.

[0117] In step S4, the adaptive inference engine performs real-time analysis of the bridge's monitoring data. The time-series analysis module predicts the strain development trend of each measuring point on the bridge over the next 7 days, while the spatial correlation module analyzes the state correlation between various components. On a certain day, the time-series analysis module predicts that the mid-span strain of a certain span will exceed the warning threshold in 3 days, while the spatial correlation module discovers that adjacent supports are in a state of detachment. The correlation analysis confirms that there is a risk of overall deflection in this span. A warning is issued 3 days in advance, giving the maintenance unit time to respond.

[0118] In step S5, the continuous learning mechanism continuously optimizes the diagnostic model. The model update cycle is set to once a month. During the update, historical samples are extracted from the experience replay buffer and mixed with the newly collected data from the previous month for training. The knowledge distillation process ensures that the model's performance on historical data does not degrade. Through continuous learning, the diagnostic accuracy increases from approximately 90% initially to approximately 95%, and the false negative rate decreases from approximately 8% to approximately 4%.

[0119] The deep integration of the digital twin model layer and the intelligent diagnostic engine provides an intuitive technical means for the bridge's management and maintenance. Managers can view the health status of each bridge component in real time through a 3D visualization platform, with diagnostic results intuitively labeled on the 3D model. Early warning information is pushed to on-site maintenance personnel via mobile devices, guiding precise maintenance operations.

[0120] On the other hand, the intelligent diagnostic system for urban roads and bridges based on digital twin technology disclosed in this invention includes: The data acquisition module is used to acquire bridge digital twin data and establish a real-time mapping relationship between the bridge physical entity and the digital model; The knowledge graph construction module is used to build a bridge defect feature database and an expert knowledge base, and to perform structured representation and associated storage of multi-source heterogeneous information in the field of bridge defects based on knowledge graph technology. The fault mode recognition module is used to extract general deep features from bridge digital twin data using a deep learning-based fault mode recognition algorithm, and output the classification results of fault types and the spatial location of faults on the bridge structure. The adaptive inference engine module includes a time series analysis module and a spatial correlation module. The time series analysis module uses a long short-term memory network to analyze and predict time series data, while the spatial correlation module uses a graph neural network to model the spatial topological relationships of bridge structures. The trend prediction and early warning module is used to generate disease development trend prediction results and trigger early warnings using a multi-step iterative prediction strategy.

[0121] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention. Therefore, the embodiments should be regarded as exemplary and non-limiting in all respects.

[0122] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for intelligent diagnosis of urban road bridges based on digital twin technology, comprising acquiring bridge digital twin data and establishing a real-time mapping relationship between the bridge physical entity and the digital model, characterized in that: Also includes: A bridge defect feature database and an expert knowledge base are constructed. The bridge defect feature database and the expert knowledge base are based on knowledge graph technology to perform structured representation and associated storage of multi-source heterogeneous information in the field of bridge defects. The deep learning-based fault mode recognition algorithm extracts general deep features from the bridge digital twin data and outputs the fault type classification results and the spatial location of the fault on the bridge structure. An adaptive inference engine is established, which includes a time series analysis module and a spatial correlation module. The time series analysis module uses a long short-term memory network to analyze and predict time series data, and the spatial correlation module uses a graph neural network to model the spatial topological relationship of bridge structures. The trend prediction module uses a multi-step iterative prediction strategy to generate disease development trend prediction results.

2. The method according to claim 1, characterized in that: The construction of the bridge defect feature database and expert knowledge base includes: Define the core entity types in the field of bridge defects, including defect location entities, defect type entities, severity entities, cause analysis entities, and treatment plan entities; Establish relationships between various entities, including causal relationships, concurrent relationships, hierarchical relationships, positional relationships, and temporal relationships; The knowledge graph is stored using a graph database, and the storage architecture adopts an attribute graph model. Each entity node contains a unique identifier, entity type, attribute set, and associated edge information. Expert knowledge is transformed into production rules for representation and integration. Each production rule contains a condition part and a conclusion part. The condition part consists of one or more premise atomic formulas, supporting AND, OR, and NOT logical combinations. The conclusion part contains the reasoning conclusion and its confidence factor.

3. The method according to claim 2, characterized in that: The process of transforming and integrating expert knowledge into production rule form also includes: The system is configured with two modes: forward chain reasoning and backward chain reasoning. Forward chain reasoning starts from known facts, gradually triggers production rules that meet the conditions, and arrives at new conclusions. Backward chain reasoning starts from the assumed conclusion and reversely retrieves the production rules and facts that support the conclusion. When multiple production rules are triggered simultaneously, the highest confidence factor among all triggered production rules is taken as the overall confidence level of the diagnostic conclusion.

4. The method according to claim 1, characterized in that: The deep learning-based fault mode recognition algorithm includes: A shared feature representation layer is constructed for a multi-task learning framework. The shared feature representation layer uses a residual neural network as the backbone network, and the residual block uses a batch normalization layer and a ReLU activation function. A classification sub-task branch is constructed, which receives feature vectors from the shared feature representation layer and outputs disease type classification results through a fully connected layer and a Softmax layer. A localization sub-task branch is constructed. The localization sub-task branch receives feature vectors from the shared feature representation layer and generates a heat map of the disease location through upsampling and convolution operations. The resolution of the heat map corresponds to the original resolution of the input image or measurement signal. The coordinates of the disease center are determined by the heat map peak detection algorithm. Then, the two-dimensional image coordinates are mapped to three-dimensional spatial coordinates by combining the geometric model of the bridge structure. Design a joint loss function and configure weights for joint optimization. The joint loss function is a weighted sum of classification loss and localization loss.

5. The method according to claim 4, characterized in that: The classification loss uses the cross-entropy loss function, and the localization loss uses either the mean squared error loss function or the IoU loss function. During joint training, the shared feature representation layer iteratively updates its network parameters through backpropagation and gradient descent under the joint supervision of the classification loss and the localization loss. When the value of the joint loss function converges to the point where it no longer decreases, it is considered that the optimal feature representation has been learned.

6. The method according to claim 1, characterized in that: The deep learning-based fault mode recognition algorithm also includes: For vibration signals, time-frequency analysis is used to extract time-domain and frequency-domain features. The time-domain features include statistical features such as mean, variance, peak factor, waveform factor, impulse factor, and margin factor. The frequency-domain features are extracted through Fourier transform or wavelet transform. A visual Transformer architecture is used to extract local and global visual features from image data; An attention mechanism is used to achieve the fusion of multimodal features. The calculation formula for the attention mechanism is as follows: ； in, For querying the matrix, The key matrix, For value matrices, Key matrix Dimensions Key matrix The transpose of .

7. The method according to claim 1, characterized in that: The graph structure is constructed based on the physical structure of the bridge. Graph nodes represent key components of the bridge structure, and graph edges represent the spatial adjacency and mechanical transmission relationships between components. The trend prediction module uses a multi-step iterative prediction strategy to generate prediction results for the future time range. During the iteration process, the predicted value of the previous moment is included in the input sequence, while the oldest historical data point in the sequence is removed to keep the length of the input sequence fixed.

8. The method according to claim 1, characterized in that: The digital twin model layer interacts with the intelligent diagnostic engine through a unified data interface. The physical simulation module establishes a bridge mechanics model based on the finite element method. The strain and displacement readings in the real-time monitoring data are converted and used as the boundary conditions of the finite element model, which are then applied to the corresponding nodes and elements of the finite element model.

9. The method according to claim 1, characterized in that: This also includes building a continuous learning mechanism: An experience replay buffer is set up to store historical training samples. A circular queue structure is used for management. Each time the model is updated, historical samples are randomly selected from the experience replay buffer and mixed with newly collected samples to form a training batch. The mixing ratio of historical samples to new samples ranges from 1:1 to 1:

3. Configure a priority experience replay mechanism, which assigns different sampling priorities based on the learning difficulty of the samples, and the learning difficulty is measured by the loss value of the samples; Configure the teacher and student models for knowledge distillation, using the diagnostic model before the update as the teacher model and the updated diagnostic model as the student model. Design a loss function and training strategy for knowledge distillation. The loss function consists of two parts: hard label loss and soft label loss. The hard label loss adopts the cross-entropy loss function, and the soft label loss adopts the KL divergence loss function. The soft labels are generated by inputting training samples into a teacher model, which then sets the distillation temperature parameter... The forward computation is performed, and the output, adjusted for distillation temperature parameters, serves as the class probability distribution as the soft label. Set to 4.0, during the distillation process, the distillation temperature parameter is used first. The student model is trained to learn the soft label distribution output by the teacher model, and then the temperature parameter is used. Make fine adjustments; A manual review mechanism and model update triggering conditions are established. Under normal operating conditions, the sampling ratio is 5% to 10%, which is increased to 20% to 30% under abnormal conditions. The criteria for judging abnormal conditions are: the false negative rate exceeds 5% in multiple consecutive batches, or the average prediction confidence of the model output is lower than 0.

6. When the false negative rate exceeds 5% for two consecutive update cycles or the diagnostic accuracy is lower than 85% for two consecutive update cycles, the model update process is triggered.

10. An intelligent diagnostic system for urban roads and bridges based on digital twin technology, characterized in that: include: The data acquisition module is used to acquire bridge digital twin data and establish a real-time mapping relationship between the bridge physical entity and the digital model; The knowledge graph construction module is used to build a bridge defect feature database and an expert knowledge base, and to perform structured representation and associated storage of multi-source heterogeneous information in the field of bridge defects based on knowledge graph technology. The fault mode recognition module is used to extract general deep features from the bridge digital twin data based on the deep learning-based fault mode recognition algorithm, and output the classification results of the fault type and the spatial location of the fault on the bridge structure. The adaptive inference engine module includes a time series analysis module and a spatial correlation module. The time series analysis module uses a long short-term memory network to analyze and predict time series data, and the spatial correlation module uses a graph neural network to model the spatial topological relationship of bridge structures. The trend prediction and early warning module is used to generate disease development trend prediction results and trigger early warnings using a multi-step iterative prediction strategy.

Images