Bridge technology condition prediction method based on fusion of graph neural network and time series modeling

Abstract

The present application relates to a bridge technical condition prediction method based on the fusion of graph neural network and time series modeling, belongs to the field of traffic infrastructure maintenance and artificial intelligence application technology, and solves the problems of insufficient utilization of space-time characteristics, lack of engineering logic constraints and rough prediction granularity of existing methods. The present application first collects historical technical condition data and pre-processes to obtain structured space-time sequence data; a double-branch feature extraction network extracts spatial correlation feature vectors and time series evolution feature vectors in parallel; a degradation trend prior feature vector is fused through a gating mechanism to generate a comprehensive feature vector; a double-branch continuous ordinal prediction head is used to output the initial continuous prediction score of the target bridge in the prediction year; finally, a time series consistency post-processing algorithm is used for logical constraint correction to generate the bridge technical condition grade prediction result. The present application effectively utilizes space-time characteristics and has strong engineering logic interpretability, and can realize continuous and accurate bridge technical condition prediction.

Description

Technical Field

[0001] This invention relates to the field of transportation infrastructure maintenance and artificial intelligence application technology, specifically to a method for predicting the technical condition of bridges based on the fusion of graph neural networks and temporal modeling. Background Technology

[0002] As vital links in transportation networks, the accurate prediction of bridge technical condition is crucial for developing scientific maintenance plans and ensuring traffic safety. Currently, bridge technical condition assessment mainly relies on regular manual inspections, which suffers from high subjectivity, high costs, and long inspection cycles. With the accumulation of bridge health monitoring data, data-driven prediction methods are gradually becoming a research hotspot. Several techniques have already been applied to the field of bridge technical condition prediction, such as:

[0003] For example, the paper "Prediction of Bridge Technical Condition Based on Regression Analysis" published in the Journal of Beijing University of Civil Engineering and Architecture proposed to use regression analysis to fit a model of bridge technical condition decline, and then predict the development trend of bridge technical condition.

[0004] For example, the paper "Research on the Prediction Method of Development Trend of Technical Condition Level of In-Service Bridges" published in the journal "Highway Transportation Technology (Applied Technology Edition)" established a development trend model of the technical condition level of various bridges with the operating time by performing linear regression on historical data of technical condition level assessment of highway bridges.

[0005] For example, the paper "Research on Grey Markov Chain Model for Predicting Bridge Technical Condition" published in the Journal of Wuhan University of Technology (Transportation Science and Engineering Edition) combines the grey model GM(1,1) with Markov chain to consider the overall changes and local fluctuations in the technical condition of bridges.

[0006] For example, the paper "Study on the Degradation Law of Beam Bridge Based on Markov Process" published in the Journal of Jiamusi University (Natural Science Edition) uses Markov theory and nonlinear programming genetic algorithm to study the time-varying law of preventive maintenance technology of concrete highway beam bridge.

[0007] For example, the Chinese patent with publication number CN113569908A, "A method for predicting bridge technical condition and defects based on deep learning", proposes a method for predicting bridge technical condition and defects using a CPSO-BP neural network model. By optimizing the neural network structure and parameters, it can accurately predict the location, type and degree of bridge defects.

[0008] For example, the Chinese patent "Method for Predicting Service Performance Degradation of Reinforced Concrete Bridges" with publication number CN111160528A proposes to use data from inspection reports over the years to train and tune multiple LSTM neural networks to accurately predict the technical condition score of reinforced concrete bridges in the next few years.

[0009] For example, the Chinese patent with publication number CN114036258A, "A method for rapid identification of bridge technical condition levels based on natural language processing", organizes the health status information of bridges into text descriptions, classifies them into technical condition levels, and uses machine learning algorithms to build a training model to achieve rapid identification of bridge technical condition levels.

[0010] For example, the paper "Technical Condition Assessment of Small and Medium Span Bridges Based on Machine Learning" published in the Journal of Chang'an University (Natural Science Edition) proposes an intelligent assessment method for bridge technical condition based on machine learning. By constructing a bridge condition database and machine learning algorithms, it achieves an accurate assessment of the technical condition of bridges.

[0011] The shortcomings of existing technologies are mainly reflected in:

[0012] (1) Lack of deep spatiotemporal integration: Most existing methods are difficult to simultaneously take into account the "spatial group effect" and "temporal evolution law" of bridges;

[0013] (2) Ignoring the logic of engineering intervention: Existing models are usually a "black box" and do not have the ability to distinguish between "natural degradation" and "artificial modification". In actual data, the bridge grade suddenly improves because of major repairs or reinforcements. If the model cannot identify the modification sequence, it will lead to confusion in the prediction logic.

[0014] (3) Discretization of prediction results: Bridge rating levels are usually discrete integers from 1 to 5 categories. Direct classification prediction will lose the ordinal nature between the levels. Direct regression prediction is difficult to align with the threshold. Moreover, existing methods lack the ability to output continuous scores, making it difficult to reflect the subtle state of "two-class preference" or "two-class bias". Summary of the Invention

[0015] The purpose of this invention is to provide a bridge technical condition prediction method based on the fusion of graph neural networks and temporal modeling. This method utilizes deep learning techniques, particularly graph neural networks (GNNs) and the Transformer architecture, to predict the technical condition rating of bridges. By constructing a dual-branch spatiotemporal feature extraction framework and combining physical degradation priors with temporal logic for reconstruction, this method addresses the problems of insufficient utilization of spatiotemporal features, lack of engineering logic constraints, and coarse prediction granularity in existing technologies.

[0016] To achieve the above objectives, the present invention adopts the following technical solution:

[0017] A bridge technical condition prediction method based on the fusion of graph neural networks and temporal modeling includes the following steps:

[0018] Step 1: Collect historical technical condition data for multiple bridges, including bridge static attributes, historical rating sequence, rating date, renovation completion date, and renovation history. Preprocess the collected data to obtain structured spatiotemporal sequence data.

[0019] Step 2: Input the structured spatiotemporal sequence data into the constructed dual-branch feature extraction network to extract the spatial correlation feature vector and temporal evolution feature vector of the bridge in parallel;

[0020] In the spatial branch, a bridge spatial association map is constructed based on the similarity of bridge static features and geolocation coding. A graph attention network is used to aggregate neighborhood information to extract spatial association feature vectors that reflect the structural and geographical dependencies between bridges.

[0021] In the temporal branch, the historical rating sequence, mask sequence, time span and time interval are embedded and encoded in multiple dimensions, and the Transformer encoder with cross-temporal attention mechanism is used to extract the temporal evolution feature vector that reflects the degradation law of the bridge.

[0022] Step 3: Concatenate the spatial correlation feature vector with the temporal evolution feature vector to obtain a concatenated feature vector, and then perform gated fusion with the degradation trend prior feature vector calculated based on physical laws to generate a comprehensive feature vector;

[0023] Step 4: Input the comprehensive feature vector into the dual-branch continuous ordinal prediction head, output the ordinal prediction value of ordinal regression and the scalar score value of continuous regression respectively, and obtain the initial continuous prediction score of the target bridge in the prediction year by weighted fusion of the two.

[0024] Step 5: Use the time-series consistency post-processing algorithm to perform logical constraint correction on the initial continuous prediction scores to generate the final bridge technical condition level prediction result.

[0025] Furthermore, the preprocessing includes missing value imputation, numerical feature normalization, historical degradation rate calculation, and date-based modification time series analysis, as well as generating a mask sequence based on the historical rating level sequence and calculating the time span and time interval.

[0026] Furthermore, the date-based modification time series analysis process includes the following steps:

[0027] The completion date and evaluation date of the target bridge's reconstruction are analyzed, accurate to the month and day.

[0028] By comparing the month and day data of the renovation completion date with the month and day data of the evaluation date, and combining the year information, renovation effect labels are generated. :

[0029] If the renovation completion year is earlier than or equal to the predicted year, and the renovation completion date is earlier than the assessment date, then mark it as such. This indicates that the effects of the renovation will take effect within the current year.

[0030] If the renovation completion year is earlier than or equal to the predicted year, but the renovation completion date is later than the assessment date, then mark it as such. This indicates that the effects of the renovation will take effect the following year;

[0031] If there is no valid modification history, mark it. .

[0032] Furthermore, the process of constructing the spatial relationship diagram of the bridge beams includes:

[0033] Calculate the cosine similarity of the static feature vectors of all bridges, and combine it with the similarity of geolocation encoding to select the bridge with the highest similarity. Establish edge connections between neighboring nodes and add self-loop edges.

[0034] Furthermore, the formula for the graph attention network is:

[0035] (3)

[0036] in, This represents the extracted spatial correlation feature vector; This represents the Sigmoid activation function; For the target node The set of neighboring nodes; For shared linear transformation matrices; The normalized attention coefficient is calculated using the following formula:

[0037] (4)

[0038] in, , , They are nodes ,node and nodes The initial static feature vector; For attention vectors, Represents the attention vector The transpose operation; Indicates a splicing operation; This is the LeakyReLU activation function.

[0039] Furthermore, the multi-dimensional embedding encoding in step 2 uses the Time2Vec method to encode the time span, with the following formula:

[0040] (5)

[0041] in, This is a time-coded vector; The time span relative to the target year; For encoding dimensions; and Each is for the first The learnable frequency and phase parameters of each feature.

[0042] Furthermore, the gating fusion process in step 3 includes:

[0043] Step 3.1: Calculate the prior feature vector of degradation trend based on the age and historical degradation rate of the target bridge. The formula is as follows:

[0044] (6)

[0045] in, The basic time-varying degradation factor is calculated using an exponential degradation model:

[0046] (7)

[0047] in, The current age of the bridge; Normalized years; and These are learnable physical parameters, representing the basic degradation magnitude and degradation acceleration rate, respectively; Indicates a splicing operation; Material impact factor; Structural influencing factors;

[0048] Step 3.2: Calculate the gating weights :

[0049] (8)

[0050] in, This indicates the concatenation of feature vectors; This represents the learnable weight matrix of the gated fusion network; Represents the learnable bias vector of the gated fusion network; This represents the Sigmoid activation function; Indicates a splicing operation;

[0051] Step 3.3: Calculate the comprehensive feature vector Its formula is:

[0052] (9)

[0053] in, This indicates element-wise multiplication.

[0054] Furthermore, in step 4, the ordinal regression branch maps the comprehensive feature vector to a Logits vector. And calculate ordinal predicted values ​​based on the expected rank method. :

[0055] (10)

[0056] in, Logits vector The first in One component; For the Sigmoid function, Total number of grades;

[0057] Initial continuous prediction score The calculation formula is:

[0058] (11)

[0059] in, This is the scalar score output by the continuous regression branch; Hyperparameters for balancing the weights of the two branches; This is a truncation function.

[0060] Furthermore, the timing consistency post-processing algorithm satisfies the following constraint logic:

[0061] (13)

[0062] in, This is the final prediction result for the bridge's technical condition level; This is the last time in history that a rating has been assigned; This is the initial continuous prediction score output from step 4; For historical degradation rate, It indicates a trend of degradation.

[0063] Furthermore, following step 5, the following steps are also included:

[0064] Step 6: Evaluate the model performance of the bridge technical condition level prediction results using evaluation indicators, including the quadratic weighted Kappa coefficient, mean absolute error, and root mean square error.

[0065] Compared with the prior art, the present invention has the following beneficial effects:

[0066] (1) Achieve multi-dimensional fusion modeling and significantly improve prediction accuracy. Existing methods are usually limited to a single dimension or only use the time series of a single bridge, or only rely on the similarity of static attributes. However, this invention uses a dual-branch feature extraction network architecture to aggregate the features of bridge groups with geographical proximity and structural similarity using a graph attention network. At the same time, it uses a Transformer encoder to capture long-distance temporal dependencies and further introduces physical degradation priors (based on the exponential decay law of bridge age and historical slope). This fusion mechanism not only makes up for the problem of sparse data of a single bridge by utilizing the common laws of the bridge group, but also constrains the overfitting tendency of the neural network through physical priors.

[0067] (2) Precise modeling of the impact of maintenance behavior on state evolution. Addressing the shortcomings of existing technologies that typically use "year" to roughly label renovation records, leading to the model's inability to distinguish between "renovations at the beginning of the year taking effect that year" and "renovations at the end of the year taking effect the following year," this invention designs a date-based renovation time-series analysis algorithm. This algorithm refines the time granularity to "month / day," dynamically generating renovation effect labels by comparing the order of completion and evaluation dates. This enables the model to accurately identify the true reasons for bridge grade improvement, effectively eliminating noise caused by misaligned data labeling time sequences. This allows the model to correctly learn two distinct state evolution modes: "natural degradation" and "human intervention," avoiding the erroneous generalization of grade recovery caused by engineering renovations as natural restoration, thereby greatly improving the model's predictive reliability in scenarios with frequent maintenance.

[0068] (3) Achieving continuous and refined assessment of bridge health status. Traditional bridge assessments only output discrete integers from 1 to 5, which are insufficient to reflect the subtle state of a bridge at a "critical point" (e.g., about to fall from category 2 to category 3). This invention innovatively designs a dual-branch continuous ordinal prediction head, combining the probability distribution of ordinal regression with the numerical constraints of scalar regression to output an initial continuous prediction score with decimals. This design not only retains the category attribute of the assessment level but also provides a higher resolution health score (e.g., a prediction value of 2.8 indicates that although the bridge is nominally category 2, it is already very close to category 3). This refined output can provide "early warnings" for maintenance departments, solving the problem of predictive jumps and instability at the category boundaries in traditional classification models, and helping management departments to intervene in preventive maintenance earlier.

[0069] (4) Ensure that the prediction results conform to the laws of physical evolution and enhance the credibility of the model. Deep learning models sometimes produce outputs that violate physical laws (e.g., the bridge grade automatically improves without any reinforcement measures). This invention uses a time-consistency post-processing algorithm to apply monotonicity constraints to the initial continuous prediction scores based on historical degradation rate and modification effect labels, ensuring that the final prediction results are logically rigorous and interpretable;

[0070] (5) Efficiently adapts to non-uniform and non-standardized historical inspection data. Addressing the issue that bridge inspection data often has inconsistent time intervals (e.g., some are inspected annually, others every three years), this invention employs Time2Vec time encoding and interval embedding technology to explicitly encode the absolute time span and relative time interval into the network. This allows the model to no longer rely on fixed time steps but to understand the degradation differences caused by different inspection intervals (e.g., the degradation at a 5-year interval should be greater than at a 1-year interval). This enables the invention to directly utilize historical, non-standardized inspection data without extensive interpolation or discarding, maximizing the utilization rate of historical data. Attached Figure Description

[0071] To more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope of protection. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort. In the drawings:

[0072] Figure 1 This is a flowchart of a bridge technical condition prediction method based on the fusion of graph neural networks and temporal modeling in an embodiment of the present invention;

[0073] Figure 2 This is a schematic diagram of the network architecture of the dual-branch feature extraction network and the dual-branch continuous ordinal prediction head in this invention;

[0074] Figure 3 This is a logical diagram illustrating the date-based modification timing analysis.

[0075] Figure 4 The flowchart is for the timing consistency post-processing algorithm.

[0076] Figure 5 This is a comparison chart of the predicted and actual values ​​of a bridge at different time points in an embodiment of the present invention. Detailed Implementation

[0077] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention. It should be understood that the specific embodiments described are merely illustrative of the present invention and are not intended to limit the present invention.

[0078] This invention proposes a bridge technical condition prediction method based on the fusion of graph neural networks and temporal modeling. This method primarily involves refined preprocessing of multi-source heterogeneous detection data, particularly employing a retrofit temporal analysis algorithm based on the comparison of completion and assessment dates (month and day) to accurately identify the effective nodes of engineering interventions. Building upon this, a dual-branch feature extraction network framework is constructed, and a graph attention network is used in parallel to aggregate spatial group association features of bridges. A Transformer network combined with Time2Vec encoding is used to capture the long-term temporal evolution patterns of individual bridges, and a gating mechanism is used to fuse prior features of physical degradation. Finally, continuous ordinal prediction and temporal consistency post-processing are performed, using a dual-branch continuous ordinal prediction head to output a refined score. Engineering logic constraints are applied to the prediction results based on the retrofit effect label and historical degradation rate. This method, combining deep spatiotemporal fusion with strong physical rule constraints, effectively solves the problems of existing technologies' inability to distinguish between natural degradation and human intervention, and the lack of interpretability in prediction results. It significantly improves the accuracy and engineering practicality of predictions, providing strong technical support for scientific maintenance decisions for bridges.

[0079] Specifically, such as Figure 1 As shown in the figure, this embodiment provides a bridge technical condition prediction method based on the fusion of graph neural network and time series modeling. The method mainly includes the following steps 1 to 5.

[0080] Step 1 (S100): Data Acquisition and Structured Preprocessing. Historical technical condition data for multiple bridges is collected, including static attribute data, historical rating sequence, rating date, reconstruction completion date, and reconstruction history. The static attribute data includes static attribute category features and static numerical features, including geometric structural features and engineering physical mapping coefficients. The collected data is preprocessed, including missing value imputation, numerical feature normalization, historical degradation rate calculation, date-based reconstruction time series analysis, and generation of a mask sequence based on the historical rating sequence, as well as calculation of time span and time interval. This yields structured spatiotemporal sequence data containing bridge static feature vectors and multidimensional historical state time series.

[0081] For the collected multi-source heterogeneous detection data, i.e., historical technical status data, this embodiment adopts a hierarchical missing value imputation strategy and a robust numerical normalization method. The specific processing procedure is as follows:

[0082] First, hierarchical missing value imputation based on feature attributes is performed. For static attribute category features of bridges (such as superstructure material name, type of land cover crossed, design load level, etc.), if missing, they are uniformly filled with a specific "unknown" identifier so that they can be coded as independent categories later. For bridge technical condition rating levels in the historical rating sequence, if missing, they are filled with the default intermediate level value. For geometric structural features with physical significance (such as span ratio, width-to-span ratio) and engineering inference features (such as estimated number of spans, estimated average span, etc.), an imputation strategy combining statistical median and domain experience values ​​is adopted. Specifically, the median of the feature in the global dataset is first calculated. If the median exists, it is used for imputation; if the median cannot be calculated due to sample sparsity, preset domain experience default values ​​(e.g., the default value for span ratio is set to 0.5, and the default value for width-to-span ratio is set to 0.1) are used for imputation. For mapping coefficients involving engineering physical properties (such as material degradation coefficient, load factor, seismic strength, etc.), when specific values ​​cannot be matched, preset constants based on civil engineering codes are directly used as a fallback.

[0083] Secondly, to eliminate the impact of extreme outliers on the stability of subsequent model training, quantile-based tail reduction is performed on the absolute geometric dimensions of the bridge (including total bridge length, total span length, maximum single-span length, and clear bridge deck width). Let a certain feature vector be... Calculate its quantiles and quantiles Values ​​exceeding this range are forcibly truncated to boundary values, resulting in a processed feature vector. Represented as:

[0084] (1)

[0085] in, Indicates feature direction The elements in.

[0086] Subsequently, robust standardization is performed on all numerical static features that have undergone the above processing. Considering that bridge data may still contain long-tailed data with non-normal distributions, this invention employs a robust scaling method based on the median and interquartile range to normalize the numerical features of the data. For feature vectors... Each element in Its normalized value The calculation formula is:

[0087] (2)

[0088] in, For feature vectors The median of and They are the feature vectors The upper quartile (i.e. quantiles) and lower quartiles (i.e.) (Quantiles). This method leverages the robustness of centering and scaling statistics to effectively reduce the pulling effect of outliers on the center and scale of the feature distribution, ensuring that the feature distribution input to the neural network has better numerical stability.

[0089] The calculation of historical degradation rate specifically involves: extracting the historical assessment grade sequence and the corresponding assessment year sequence for each bridge; using the least squares method to linearly fit the trend of grade change over years; and calculating the slope of the fitted line, which is the historical degradation rate. This feature quantifies the average rate at which the bridge's technical condition deteriorates over time within a historical period, and is used to assist subsequent models in judging future degradation inertia.

[0090] Finally, date-based modification time-series analysis is performed, that is, the modification completion date and evaluation date are precisely analyzed to generate modification effect labels reflecting the effectiveness of the engineering intervention, such as... Figure 3 As shown, the process includes:

[0091] The completion date and evaluation date of the target bridge's reconstruction are analyzed, accurate to the month and day.

[0092] By comparing the month and day data of the renovation completion date with the month and day data of the evaluation date, and combining the year information, renovation effect labels are generated. If the renovation completion year is earlier than or equal to the predicted year, and the renovation completion date is earlier than or equal to the assessment date, then mark it as such. This indicates that the renovation effect takes effect in the current year; if the renovation completion year is earlier than or equal to the predicted year, but the renovation completion date is later than the assessment date, then it is marked as such. This indicates that the modification effect will take effect the following year; if there is no valid modification history, it will be marked as such. .

[0093] Finally, a mask sequence is generated based on the historical rating sequence, and the time span and time interval are calculated. The mask sequence has the same length as the historical rating sequence and is used to identify whether there is real detection data at each time point. If it exists, it is marked as 1, and if it is missing, it is marked as 0. The time span is the time difference (in years) between each rating date and the target prediction year. The time interval is the time interval between two adjacent ratings (in years).

[0094] Step 2 (S200): Two-branch spatiotemporal feature extraction. Construct a two-branch feature extraction network and input the structured spatiotemporal sequence data obtained in Step 1 into the two-branch feature extraction network to extract the spatial correlation feature vector and temporal evolution feature vector of the bridge in parallel.

[0095] like Figure 2 As shown, the prediction model of the present invention includes a dual-branch feature extraction network, a gated fusion network, and a dual-branch continuous ordinal prediction head, wherein the dual-branch feature extraction network includes parallel spatial branches and temporal branches.

[0096] In the spatial branch, a spatial association graph of bridges is constructed based on the similarity of bridge static feature vectors and geolocation encoding. A graph attention network (GAT) is used to aggregate neighborhood information to extract spatial association feature vectors that reflect the structural and geographical dependencies between bridges.

[0097] Specifically, the process of constructing the bridge spatial association map includes: calculating the cosine similarity of the static feature vectors of all bridges, and combining this with the similarity of geolocation codes to select the bridges with the highest similarity. By establishing edges connecting neighboring nodes and adding self-loop edges, a spatial association graph of the bridge is constructed. For example, for the target bridge... Select the one with the highest similarity By establishing edge connections between neighboring nodes and adding self-loop edges, a bridge spatial association graph is obtained.

[0098] The formula for graph attention networks is:

[0099] (3)

[0100] in, Indicates the updated target node The feature vector (i.e., the extracted spatial correlation feature vector); This represents the Sigmoid activation function; For the target node The set of neighboring nodes (including the target node) itself); The normalized attention coefficients represent the neighboring nodes. For the target node Weighted contribution; For shared linear transformation matrices; For neighboring nodes The initial static feature vector. Attention coefficients. Automatically learn neighboring bridges (such as bridges made of the same material and located in the same county) for the target bridge. The degradation impact weights are used to capture regional environmental erosion patterns.

[0101] The spatial correlation feature vector extraction process employs a multi-head attention mechanism, with attention coefficients... The calculation formula is:

[0102] (4)

[0103] in, , , They are nodes ,node and nodes The initial static feature vector; For attention vectors, Represents the attention vector The transpose operation; Indicates a splicing operation; This is the LeakyReLU activation function, used to solve the gradient vanishing problem in the negative interval.

[0104] In the time-series branch, the historical rating sequence (such as...) The mask sequence, time span, and time interval are embedded and encoded in multiple dimensions, and a Transformer encoder with a cross-time attention mechanism is used to extract temporal evolution feature vectors that reflect the degradation pattern of bridges.

[0105] The multi-dimensional embedding encoding employs the Time2Vec method to periodically encode the time span, enabling the model to distinguish the degradation differences between "continuous annual detection" and "detection at three-year intervals." The formula for periodically encoding the time span using the Time2Vec method is as follows:

[0106] (5)

[0107] in, This is a time-coded vector; The time span relative to the target year; For encoding dimensions; and Each is for the first Learnable frequency and phase parameters of each feature are used to simultaneously capture degenerate linear trends and periodic patterns.

[0108] A 3-layer Transformer encoder (with residual connections) is employed, utilizing a self-attention mechanism to capture long-term dependencies in the historical rating sequence, and outputting a temporal evolution feature vector. .

[0109] Step 3 (S300): Physical Prior Fusion. The spatial correlation feature vector obtained in Step 2 is concatenated with the temporal evolution feature vector to obtain a concatenated feature vector. This concatenated feature vector is then gated and fused with the degradation trend prior feature vector calculated based on physical laws to generate a comprehensive feature vector.

[0110] Step 3.1: First, based on the target bridge Bridge age (e.g., 25 years) and historical degradation rate (e.g.) (Level / Year), calculate the prior feature vector of degradation trend. It is composed of three physical components, and the calculation formula is as follows:

[0111] (6)

[0112] in, The basic time-varying degradation factor is calculated using an exponential degradation model:

[0113] (7)

[0114] in, The current age of the bridge; The normalized lifespan is 50 years in this embodiment. and These are the learnable physical parameters (initial values ​​are set to 0.3 and 0.15 respectively), representing the basic degradation magnitude and degradation acceleration rate, respectively; This indicates the splicing operation. Formula (7) explicitly simulates the physical law that the structural performance decreases nonlinearly and rapidly with increasing service time. As a material influence factor, discrete material type IDs are mapped to high-dimensional feature vectors through an embedding layer. These vectors are then passed through a linear projection layer and a sigmoid activation function to further map them to... The scalar coefficients in the interval reflect the differences in the ability of different materials (such as steel and concrete) to resist environmental erosion. The structural influence factor is obtained through adaptive learning of independent embedding layers and linear projection layers, reflecting the impact of different structural forms (such as beam bridges and arch bridges) on the degradation rate.

[0115] Step 3.2: Concatenate the feature vectors Prior feature vectors of degradation trend After concatenation, the data is input into the gated fusion network to calculate the gate weights. The calculation formula is as follows:

[0116] (8)

[0117] in, This represents the learnable weight matrix of the gated fusion network; Represents the learnable bias vector of the gated fusion network; This represents the Sigmoid activation function; This indicates a splicing operation.

[0118] If the data features exhibit gradual degradation that conforms to physical laws, the gating weights... Towards equilibrium;

[0119] If data features fluctuate drastically due to noise, the gating mechanism will adaptively adjust the weights and use physical priors to smooth the features.

[0120] Step 3.3: Calculate the comprehensive feature vector Its formula is:

[0121] (9)

[0122] in, This is the concatenated feature vector obtained by concatenating the spatial correlation feature vector and the temporal evolution feature vector; It is a priori feature vector of degradation trend that includes bridge age, the last historical grade, and the historical degradation slope; This indicates element-wise multiplication.

[0123] Step 4 (S400): Continuous Ordinal Prediction. The comprehensive feature vector obtained in Step 3 is input into the dual-branch continuous ordinal prediction head. The dual-branch continuous ordinal prediction head outputs the ordinal prediction value of the ordinal regression and the scalar score value of the continuous regression, respectively. The two are weighted and fused to obtain the initial continuous prediction score of the target bridge in the prediction year. The initial continuous prediction score retains the category attribute of the level and provides a continuous health score.

[0124] The dual-branch continuous ordinal prediction head includes a weighted fusion layer and parallel ordinal regression and continuous regression branches, which integrate feature vectors. Enter the ordinal regression branch and the continuous regression branch respectively.

[0125] The ordinal regression branch consists of a single-layer fully connected network used to synthesize the feature vectors. Mapped to Logits, denoted as the Logits vector. Ordinal predicted values ​​are calculated based on the expected rank method. The calculation formula is:

[0126] (10)

[0127] in, Logits vector The first in Each component indicates a technical condition level not exceeding [a certain value]. The linear prediction of the cumulative probability; This indicates the total number of grades in the bridge's technical condition assessment. ); This is the Sigmoid activation function.

[0128] The continuous regression branch employs a multilayer perceptron (MLP) structure to synthesize feature vectors. The input to the first fully connected layer is mapped to a 64-dimensional latent space. After ReLU nonlinear activation, a Dropout layer is used to prevent overfitting, and finally, the scalar score is directly output through the second fully connected layer. Its value range is [1, ... The real number ].

[0129] Ordinal predictions of the ordinal regression branch output by the weighted fusion layer Scalar scores from the continuous regression branch output Weighted fusion is performed to obtain the initial continuous prediction score of the target bridge in the prediction year. The calculation formula is as follows:

[0130] (11)

[0131] in, Hyperparameters for balancing the weights of the two branches; This is a truncation function. Assume... ,but This score retains high-precision continuity information, reflecting that the target bridge is in a critical state of "end of Category 2, about to enter Category 3".

[0132] In this invention, the prediction model is optimized using a joint loss function during training, wherein the joint loss function is:

[0133] (12)

[0134] in, This represents the mean squared error loss of the continuous regression branch. This is a binary cross-entropy loss with a Focal term for the ordinal regression branch, used to address the sample imbalance problem in ordinal regression. These are the weighting coefficients for the regression loss.

[0135] Step 5 (S500): Time-series consistency post-processing. Based on the modification effect labels and historical degradation rates calculated in Step 1, the initial continuous prediction scores are logically constrained and corrected using the time-series consistency post-processing algorithm to generate the final bridge technical condition level prediction result. This step applies logical constraints to the initial continuous prediction scores (e.g., no improvement without modification) to ensure that the prediction results conform to the objective laws of civil engineering.

[0136] See Figure 4 The timing consistency post-processing algorithm satisfies the following constraints:

[0137] (13)

[0138] in, The corrected prediction values ​​represent the final predicted bridge technical condition level, with level 1 being the best and level 5 being the worst. This is the last time in history that a rating has been assigned; This is the initial continuous prediction score output from step 4; For historical degradation rate, It indicates a trend of degradation.

[0139] The above constraint logic means that when there is a valid history of modification, the predicted level value is allowed to decrease (i.e., the level is improved), and the minimum value is taken; when there is no valid history of modification and there is a trend of degradation, the predicted value of the technical condition level is restricted from being less than the historical assessment level (i.e., a non-incremental constraint is applied, and the predicted level cannot be improved); in other cases, the model prediction value is maintained.

[0140] Step 6 (S600): Model Performance Evaluation. Based on the structured spatiotemporal sequence data obtained in Step 1, a validation set (or test set) is created. Evaluation metrics are used to compare and analyze the predicted bridge technical condition level output in Step 5 with the actual assessment level. This step uses Quadratically Weighted Kappa (QWK), Mean Absolute Error (MAE), and Root Mean Square Error (RMSE) as core evaluation metrics to quantitatively assess the model's performance in capturing degradation trends and predicting accuracy. If the evaluation metrics meet the preset accuracy requirements or reach convergence, the currently trained network parameters are saved as the final "Bridge Technical Condition Degradation Prediction Model" for subsequent prediction of the technical condition of bridges to be predicted, as well as subsequent engineering deployment and decision support. Otherwise, the model parameters are adjusted based on the feedback error gradient, and the training steps are repeated until the evaluation metrics meet the preset accuracy requirements or reach convergence.

[0141] The formula for calculating QWK is:

[0142] (14)

[0143] in, These are the values ​​of the double-weighted Kappa coefficients; The observed confusion matrix; The expected matrix is ​​based on random consistency. It is a quadratic weight matrix, and its calculation method is as follows: ; and These represent the numerical values ​​of the sample's actual technical condition level and the model's predicted level (i.e., the predicted result of the bridge's technical condition level), respectively. This indicates the overall grade of the bridge's technical condition assessment. .

[0144] The formula for calculating MAE is:

[0145] (15)

[0146] in, This represents the value of the mean absolute error; This indicates the total number of test samples (or validation samples) used in the model performance evaluation; Indicates the first The ground truth level of a bridge sample, which is the standard level confirmed by human assessment or historical records; The model represents the first The predicted technical condition grade of each bridge sample. Indicates the first The absolute value of the prediction error for each sample.

[0147] The formula for calculating RMSE is:

[0148] (16)

[0149] in, This represents the root mean square error.

[0150] To verify the effectiveness of the prediction method proposed in this invention, experiments were conducted on a real bridge detection dataset using the aforementioned evaluation metrics. The specific experimental setup and results are as follows:

[0151] Using historical technical condition data from 2847 bridges as the raw data, a total of 4489 samples were collected for validation. The results were compared with those of random forest and XGBoost regression methods to verify the effectiveness and superiority of the prediction method of this invention. The experimental results are shown in Table 1.

[0152] Table 1 Experimental Results

[0153]

[0154] like Figure 5 The figure illustrates the predictive effect of the method of this invention on a bridge before and after its renovation in 2022. The solid blue line represents the actual assessment level (i.e., the true value) determined manually, the dashed red line represents the predicted value (i.e., the predicted value) of the bridge's technical condition level output by this invention, and the red shaded area represents the confidence interval of the prediction result. It is worth noting that the predicted value output by this invention is a continuous floating-point number. For example, in 2024, the actual assessment level was Class 2, while the predicted value of this invention was approximately 2.3. This not only accurately corresponds to the Class 2 level range (1.5-2.5), but also more keenly captures the slight physical degradation that the bridge had undergone compared to 2022 (predicted value 2.0). This sub-level refined predictive capability can provide maintenance departments with earlier risk warnings than traditional integer ratings, demonstrating the technical advantages of this invention in temporal continuous modeling.

[0155] In summary, this invention provides a bridge technical condition prediction method based on the fusion of graph neural networks and temporal modeling, along with physical rule constraints. This method encompasses key aspects such as refined preprocessing and temporal analysis of multi-source heterogeneous data, construction of a dual-branch spatiotemporal feature extraction framework, gating fusion of physical degradation priors, establishment of a continuous ordinal prediction model, and post-processing constraints for temporal consistency. Ultimately, it achieves accurate assessment of bridge technical condition in complex maintenance scenarios. Through collaborative modeling using graph attention networks and Transformers, the spatial grouping effect and long-term temporal dependence of bridges are fully explored. Combined with monthly and daily-accurate temporal analysis of reconstruction timelines and physical degradation patterns, it effectively solves the problem of data-driven models struggling to distinguish between natural degradation and human intervention. Finally, the use of continuous ordinal regression output and post-processing logic constraints significantly improves the granularity and engineering rationality of the prediction results, providing maintenance departments with a more forward-looking condition assessment. Compared with traditional methods, this invention has advantages such as high utilization of spatiotemporal features, strong interpretability of engineering logic, and excellent generalization ability. This invention, through the deep integration of data intelligence and physical common sense, effectively overcomes the prediction bias under limited detection data, realizes the scientific quantification of bridge health status, and provides a reliable scientific basis for bridge life cycle management and preventive maintenance decisions.

[0156] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0157] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A bridge technical condition prediction method based on the fusion of graph neural networks and temporal modeling, characterized in that, Includes the following steps: Step 1: Collect historical technical condition data for multiple bridges, including bridge static attributes, historical rating sequence, rating date, renovation completion date, and renovation history. Preprocess the collected data to obtain structured spatiotemporal sequence data. Step 2: Input the structured spatiotemporal sequence data into the constructed dual-branch feature extraction network to extract the spatial correlation feature vector and temporal evolution feature vector of the bridge in parallel; In the spatial branch, a bridge spatial association map is constructed based on the similarity of bridge static features and geolocation coding. A graph attention network is used to aggregate neighborhood information to extract spatial association feature vectors that reflect the structural and geographical dependencies between bridges. In the temporal branch, the historical rating sequence, mask sequence, time span and time interval are embedded and encoded in multiple dimensions, and the Transformer encoder with cross-temporal attention mechanism is used to extract the temporal evolution feature vector that reflects the degradation law of the bridge. Step 3: Concatenate the spatial correlation feature vector with the temporal evolution feature vector to obtain a concatenated feature vector, and then perform gated fusion with the degradation trend prior feature vector calculated based on physical laws to generate a comprehensive feature vector; Step 4: Input the comprehensive feature vector into the dual-branch continuous ordinal prediction head, output the ordinal prediction value of ordinal regression and the scalar score value of continuous regression respectively, and obtain the initial continuous prediction score of the target bridge in the prediction year by weighted fusion of the two. Step 5: Use the time-series consistency post-processing algorithm to perform logical constraint correction on the initial continuous prediction scores to generate the final bridge technical condition level prediction result.

2. The bridge technical condition prediction method based on the fusion of graph neural network and temporal modeling according to claim 1, characterized in that, The preprocessing includes missing value imputation, numerical feature normalization, historical degradation rate calculation, and date-based modification time series analysis, as well as generating a mask sequence based on the historical rating sequence and calculating the time span and time interval.

3. The bridge technical condition prediction method based on the fusion of graph neural network and temporal modeling according to claim 2, characterized in that, The date-based modification time series analysis process includes the following steps: The completion date and evaluation date of the target bridge's reconstruction are analyzed, accurate to the month and day. By comparing the month and day data of the renovation completion date with the month and day data of the evaluation date, and combining the year information, renovation effect labels are generated. : If the renovation completion year is earlier than or equal to the predicted year, and the renovation completion date is earlier than the assessment date, then mark it as such. This indicates that the effects of the renovation will take effect within the current year. If the renovation completion year is earlier than or equal to the predicted year, but the renovation completion date is later than the assessment date, then mark it as such. This indicates that the effects of the renovation will take effect the following year; If there is no valid modification history, mark it. .

4. The bridge technical condition prediction method based on the fusion of graph neural network and temporal modeling according to any one of claims 1 to 3, characterized in that, The process of constructing the spatial relationship diagram of the bridge beams includes: Calculate the cosine similarity of the static feature vectors of all bridges, and combine it with the similarity of geolocation encoding to select the bridge with the highest similarity. Establish edge connections between neighboring nodes and add self-loop edges.

5. The bridge technical condition prediction method based on the fusion of graph neural network and temporal modeling according to any one of claims 1 to 3, characterized in that, The formula for the graph attention network is: （3） in, This represents the extracted spatial correlation feature vector; This represents the Sigmoid activation function; For the target node The set of neighboring nodes; For shared linear transformation matrices; The normalized attention coefficient is calculated using the following formula: （4） in, , , They are nodes ,node and nodes The initial static feature vector; For attention vectors, Represents the attention vector The transpose operation; Indicates a splicing operation; This is the LeakyReLU activation function.

6. The bridge technical condition prediction method based on the fusion of graph neural network and temporal modeling according to any one of claims 1 to 3, characterized in that, The multi-dimensional embedding encoding in step 2 uses the Time2Vec method to encode the time span, and its formula is as follows: （5） in, This is a time-coded vector; The time span relative to the target year; For encoding dimensions; and Each is for the first The learnable frequency and phase parameters of each feature.

7. The bridge technical condition prediction method based on the fusion of graph neural network and temporal modeling according to any one of claims 1 to 3, characterized in that, Step 3, the gating fusion process includes: Step 3.1: Calculate the prior feature vector of degradation trend based on the age and historical degradation rate of the target bridge. The formula is as follows: （6） in, The basic time-varying degradation factor is calculated using an exponential degradation model: （7） in, The current age of the bridge; Normalized years; and These are learnable physical parameters, representing the basic degradation magnitude and degradation acceleration rate, respectively; Indicates a splicing operation; Material impact factor; Structural influencing factors; Step 3.2: Calculate the gating weights : （8） in, This indicates the concatenation of feature vectors; This represents the learnable weight matrix of the gated fusion network; Represents the learnable bias vector of the gated fusion network; This represents the Sigmoid activation function; Indicates a splicing operation; Step 3.3: Calculate the comprehensive feature vector Its formula is: （9） in, This indicates element-wise multiplication.

8. The bridge technical condition prediction method based on the fusion of graph neural network and temporal modeling according to any one of claims 1 to 3, characterized in that, In step 4, the ordinal regression branch maps the composite feature vector to a Logits vector. And calculate ordinal predicted values ​​based on the expected rank method. : （10） in, Logits vector The first in One component; For the Sigmoid function, Total number of grades; Initial continuous prediction score The calculation formula is: （11） in, This is the scalar score output by the continuous regression branch; Hyperparameters for balancing the weights of the two branches; This is a truncation function.

9. The bridge technical condition prediction method based on the fusion of graph neural network and temporal modeling according to any one of claims 1 to 3, characterized in that, The timing consistency post-processing algorithm satisfies the following constraints: （13） in, This is the final prediction result for the bridge's technical condition level; This is the last time in history that a rating has been assigned; This is the initial continuous prediction score output from step 4; For historical degradation rate, It indicates a trend of degradation.

10. The bridge technical condition prediction method based on the fusion of graph neural network and temporal modeling according to any one of claims 1 to 3, characterized in that, Following step 5, the following is also included: Step 6: Evaluate the model performance of the bridge technical condition level prediction results using evaluation indicators, including the quadratic weighted Kappa coefficient, mean absolute error, and root mean square error.

Images