Road infrastructure service performance big data integration analysis method

Abstract

The invention discloses a road infrastructure service performance big data integration analysis method, and belongs to the technical field of road data analysis. Data fusion is realized through a federated learning architecture, digital twin bodies constructed by multi-source data are fused, and an accurate data basis is provided for subsequent scene deduction and performance prediction through low-delay data transmission and a state correction algorithm; through combination of deep reinforcement learning and a random process, an adaptive random process model is adopted for different types of uncertainty factors, probability distribution of a future scene can be quantified, and through matching combination of a scene deduction result and a digital twinborn performance degradation model and entropy item design of a reward function, the probability distribution of the future scene can be quantified. The limitation that the decision depends on single optimistic and pessimistic scenes is avoided; the dynamic interaction effect of damage factors is accurately captured through a space-time attention mechanism, the method can adapt to road performance prediction under different climate and traffic conditions, and the generalization ability of the model can be effectively enhanced.

Description

Technical Field

[0001] This invention relates to the field of road data analysis technology, and more specifically to a big data integrated analysis method for road infrastructure service performance. Background Technology

[0002] Big data integration and analysis of road infrastructure service performance refers to the process of systematically collecting, integrating, processing and intelligently analyzing the service status, performance and degradation patterns of road infrastructure throughout its entire life cycle (design, construction, operation, maintenance and decommissioning) using big data technology. Its core objective is to achieve accurate assessment, prediction and decision support for the health status, safety level, service capacity and future performance evolution trend of road infrastructure through the integration and in-depth mining of multi-source heterogeneous data.

[0003] Road infrastructure has a lifespan of several decades, historical data has a limited time span, and there are uncertainties in future traffic flow, load levels, climate change, etc. Existing prediction models are mostly based on static fitting of historical data, lacking the ability to dynamically extrapolate future scenarios, and are difficult to effectively support long-term operation and maintenance decisions. Summary of the Invention

[0004] The purpose of this invention is to provide a big data integration and analysis method for road infrastructure service performance, which can solve the technical problems mentioned in the background art.

[0005] The objective of this invention can be achieved through the following technical solutions: A big data integration and analysis method for road infrastructure service performance includes: Construct a digital twin of road infrastructure by integrating historical monitoring data, real-time sensor data and environmental simulation data through multi-source heterogeneous data fusion technology to establish a dynamic mapping relationship between physical entities and digital models; Based on the constructed digital twin, a scenario inference model is trained using deep reinforcement learning algorithms. By introducing stochastic process theory, the model simulates the uncertain evolution of future traffic flow, load level, and climate change, and outputs multi-scenario evolution results. By utilizing the multi-scenario evolution results, the dynamic interaction effects of key damage factors are captured through a spatiotemporal attention mechanism, enabling accurate prediction of the long-term performance of road infrastructure.

[0006] Furthermore, historical monitoring data, real-time sensor data, and environmental simulation data are acquired through a distributed acquisition network; the collected data are standardized, and a horizontal federated learning architecture is adopted to achieve privacy and security fusion across data sources.

[0007] Furthermore, based on the fused feature vectors, a digital twin is constructed that includes a three-dimensional geometric model, a physical and mechanical model, and a behavioral evolution model.

[0008] Furthermore, to address the scene deduction requirements in the continuous action space, an Actor-Critic dual network architecture is constructed.

[0009] Furthermore, for different types of uncertainty factors, we use adapted stochastic process models for refined simulation, including simulation of stochastic fluctuations in traffic flow, simulation of climate change uncertainty, and simulation of extreme load events.

[0010] Furthermore, the initial state data of road infrastructure, regional baseline climate data, and baseline traffic data are input into the digital twin, and the initial state vector of the road is output. The current road state is input into the Actor network, and the Actor network outputs deterministic evolutionary actions. ;in, This represents the annual growth rate of traffic flow. The annual growth rate of the proportion of heavy-duty vehicles; This refers to the annual increase in temperature. This represents the annual increase in precipitation. Gaussian noise is superimposed on deterministic actions to generate the final action. , It has a zero-mean Gaussian noise distribution.

[0011] Furthermore, when performing road condition evolution calculations, the input data includes the current road condition, generated temperature, precipitation, annual growth rate of traffic flow, and annual growth rate of the proportion of heavy vehicles. Generate updated current road status; The road states obtained from different road scenario simulations are sorted and combined to obtain multi-scenario evolution results.

[0012] Furthermore, by using the road evolution results across multiple scenarios as input data, the system outputs a road smoothness prediction curve for the next 10 years and an attention weight analysis report of key damage factors.

[0013] Furthermore, a dual-branch attention structure is designed to capture the spatial interaction effect and temporal evolution effect of damage factors respectively. For the spatial attention branch, the contribution weights of six state features to performance degradation are calculated to highlight the core damage factors. For the time attention branch, calculate the weight of the impact of the state at different time steps on future performance, highlighting the period of accelerated decay.

[0014] Furthermore, the road performance degradation model, which has been trained and validated, is used to predict the long-term performance of roads, and an attention weight analysis report of key damage factors is output.

[0015] Compared to existing solutions, the beneficial effects achieved by this invention are: This invention achieves data fusion through a federated learning architecture, eliminating the need to transmit raw data, thus avoiding the risk of data leakage and complying with data security regulations. The digital twin constructed by fusing multi-source data can effectively reduce geometric errors and mechanical response prediction errors, accurately restoring the physical characteristics of road infrastructure. Through low-latency data transmission and state correction algorithms, real-time synchronization between the physical entity and the digital model is achieved, providing a precise data foundation for subsequent scenario simulation and performance prediction.

[0016] This invention combines deep reinforcement learning with stochastic processes to achieve dynamic projection of future traffic, loads, and the environment, thus solving the problem that traditional models can only statically fit historical data. It employs adaptive stochastic process models for different types of uncertainties, quantifying the probability distribution of future scenarios and providing a risk quantification basis for long-term operation and maintenance decisions. By matching and combining the scenario projection results with a digital twin performance degradation model, it effectively improves simulation accuracy compared to traditional simulation schemes. Through the design of the entropy term in the reward function, the generated scenarios cover more than 95% of possible future states, avoiding the limitations of relying on a single optimistic or pessimistic scenario for decision-making.

[0017] This invention accurately captures the dynamic interaction effects of damage factors through a spatiotemporal attention mechanism, which can effectively reduce errors and improve accuracy compared to traditional LSTM models. The model, trained on 1,000 sets of multi-scenario evolution data, can adapt to road performance prediction under different climate and traffic conditions, such as rainy road sections in the south and heavy-load road sections in the north, which can effectively enhance the model's generalization ability. The output attention weights clearly identify key damage factors and the period of accelerated decay, providing reliable data support for targeted maintenance. Attached Figure Description

[0018] The invention will now be further described with reference to the accompanying drawings.

[0019] Figure 1 This is a flowchart of a big data integration and analysis method for road infrastructure service performance according to the present invention. Detailed Implementation

[0020] 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 embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] like Figure 1As shown, this invention is a big data integration and analysis method for road infrastructure service performance, comprising: Constructing a digital twin of road infrastructure involves integrating historical monitoring data, real-time sensor data, and environmental simulation data through multi-source heterogeneous data fusion technology to establish a dynamic mapping relationship between physical entities and digital models. Specific steps include: Historical monitoring data, real-time sensor data, and environmental simulation data are acquired through a distributed data acquisition network. Among them, historical monitoring data includes static data such as road construction archives and annual inspection reports, such as pavement settlement values, crack widths, and structural strength. Real-time sensing data, including dynamic data such as strain data, displacement data, and temperature and humidity data collected by IoT sensors deployed inside the road structure; IoT sensors, specifically including but not limited to strain gauges, displacement gauges, and temperature and humidity sensors; Environmental simulation data, including environmental data such as precipitation and freeze-thaw cycles obtained from meteorological and hydrological stations; Standardization processing is performed on the collected data, including but not limited to using the K-nearest neighbor algorithm to fill missing data, using the 3σ principle to identify and remove outlier data, and mapping all data to the [0,1] interval. These are all existing conventional technical solutions, and the specific implementation steps will not be elaborated here. When using a horizontal federated learning architecture to achieve privacy and security fusion across data sources, each data holder trains a feature extraction model locally. The feature extraction model adopts a convolutional neural network (CNN) structure, with preprocessed single-source data as input and feature vectors of that data source as output. Data holders include, for example, transportation departments and maintenance units. Training the feature extraction model is an existing conventional technical solution, and the specific implementation steps are not detailed here. Each data holder uploads the parameters of its local model, such as weights and biases, to the federated learning server. The federated learning server then aggregates the model parameters using a weighted average method, involving the following expression: ;in, These are the aggregated global model parameters; For the first The number of samples held by each data holder; Assign a number to the data holder. =1, 2, 3, ... , The total number of data holders; This represents the total number of samples; For the first Local model parameters for each data holder; The federated learning server distributes global model parameters to each data holder, and each data holder updates its local model and continues training, repeating the above process until the model converges. The trained global model is used to fuse features from standardized multi-source data, and the fused unified feature vector is output. , The dimension of the fused features; Based on the fused feature vectors, a digital twin is constructed that includes a 3D geometric model, a physical and mechanical model, and a behavioral evolution model; specifically: When constructing the three-dimensional geometric model, BIM (Building Information Modeling) technology is used to construct a 1:1 scale three-dimensional model of the road based on road design drawings and laser scanning data, which includes geometric information of structures such as pavement, roadbed, bridges, and tunnels; When constructing the physical and mechanical model, a mechanical model of the road structure is established using finite element analysis software to simulate the stress-strain response of the road under different loads and environmental conditions. The core formula is Hooke's Law, and the relevant expressions are: ;in, For stress; The elastic modulus of the material; In response to the situation; When constructing the behavior evolution model, a road performance degradation model is trained based on a Long Short-Term Memory (LSTM) network. The input is the fused feature vector F, and the output is the evolution trend of road performance indicators, such as road surface smoothness and structural strength. The construction steps of the road performance degradation model include: The fused feature vectors are divided into training set, validation set and test set, with data proportions of 70%, 20% and 10%, respectively. Input features are selected from factors that are strongly correlated with road performance degradation, including structural features, load features, environmental features and historical performance features. Structural characteristics, specifically including static parameters such as pavement thickness, subgrade compaction, and concrete strength; Load characteristics, specifically including dynamic parameters such as average daily traffic volume, proportion of heavy vehicles, and axle load distribution; Environmental characteristics, specifically including climate parameters such as average annual temperature, annual precipitation, and number of freeze-thaw cycles; Historical performance characteristics, specifically including test data on road surface smoothness, structural strength, etc. over the years; The output features are selected from core performance indicators, such as road surface smoothness, structural strength, and crack rate. The sliding window method is used to construct time series samples with a window size of 5 years, that is, the input features of the first 5 years are used to predict the performance indicators of the 6th year. The network structure includes an input layer, LSTM layer 1, LSTM layer 2, LSTM layer 3, a fully connected layer, and an output layer. The number of neurons in LSTM layer 1, LSTM layer 2, LSTM layer 3, fully connected layer and output layer are 64, 32, 16, 8 and 3, respectively. The activation function for LSTM layer 1, LSTM layer 2, LSTM layer 3, and the fully connected layer is ReLU, while the activation function for the output layer is Linear. The dropout rates for LSTM layer 1, LSTM layer 2, and LSTM layer 3 are all 0.2. The network uses the Adam optimizer with a learning rate of 0.001 and a loss function of mean squared error (MSE). The relevant expression is as follows: ;in, These are the actual performance metrics; These are the model's predicted values; For sample number, =1, 2, 3, ... , The number of samples; Training the LSTM network using the training set and validating the trained model using the test set are both existing conventional techniques, and the specific implementation steps will not be elaborated here. The trained and validated model was set as a road performance degradation model and deployed to the digital twin platform; When establishing a real-time dynamic mapping between physical entities and digital twins, a low-latency data transmission channel is built, and the MQTT protocol is used to synchronize real-time sensing data to the digital twin platform, with the transmission latency controlled within 100ms. Design a state correction algorithm to adjust the model's predicted values ​​based on real-time sensor data. The relevant expressions are: ;in, for The model state at time t. for The model state at time t; For correction factors, 0 < <1, dynamically adjusted according to the degree of data fluctuation, with a value range of 0.2-1.0; for Real-time sensing data of physical entities at any given moment; for The predicted value of the time-matter model; A two-way interactive mechanism is constructed to support the simulation of maintenance measures, such as road repair and structural reinforcement, on the digital twin, and the simulation results are fed back to the operation and maintenance decisions of the physical entity.

[0022] In this embodiment of the invention, data fusion is achieved through a federated learning architecture, eliminating the need to transmit original data, thus avoiding the risk of data leakage and complying with data security regulations. The digital twin constructed by fusing multi-source data can effectively reduce geometric errors and mechanical response prediction errors, and can accurately restore the physical characteristics of road infrastructure. Through low-latency data transmission and state correction algorithms, real-time synchronization between the physical entity and the digital model is achieved, providing an accurate data foundation for subsequent scenario simulation and performance prediction.

[0023] Based on the constructed digital twin, a scenario simulation model is trained using deep reinforcement learning algorithms. By introducing stochastic process theory, the model simulates the uncertain evolution of future traffic flow, load levels, and climate change, and outputs multi-scenario evolution results. Specific steps include: Based on the output data of the constructed digital twin, the state space S, action space A, and reward function R of deep reinforcement learning are defined; specifically: The state space S integrates a multi-dimensional state vector that incorporates current road performance, traffic load, and environmental impact. The specific expression is as follows: Among them, IRI is the current road surface smoothness, with a value range of [0,10], reflecting the road surface driving quality. The smaller the value, the higher the road surface driving quality. L is the current road structure deflection value, with a value range of [0,2], reflecting the overall strength of the roadbed and pavement. The smaller the value, the stronger the overall strength of the roadbed and pavement. T is the average temperature of the past 5 years, with a value range of [−20,40], reflecting the regional climate baseline. P is the average annual precipitation of the past 5 years, with a value range of [0,2000], reflecting the hydrological environment characteristics. F is the current average daily traffic flow, with a value range of [0,100000], reflecting the traffic load intensity. V is the current proportion of heavy vehicles, with a value range of [0,100], reflecting the degree of load influence. The larger the value, the higher the degree of load influence. Action space A defines the evolution range of future uncertainties. It is a continuous action space that simulates the dynamic changes of traffic, load, and environment. The specific expression is: Wherein, ΔF is the annual growth rate of average daily traffic flow over the next 10 years, ranging from [−5, 15], with negative growth corresponding to traffic diversion scenarios; ΔV is the annual growth rate of the proportion of heavy vehicles over the next 10 years, ranging from [−2, 8], with negative growth corresponding to industrial restructuring scenarios; ΔT is the annual change in average temperature over the next 10 years, ranging from [−1, 2], reflecting climate change trends; and ΔP is the annual change rate of precipitation over the next 10 years, ranging from [−10, 20], reflecting changes in precipitation patterns. The reward function R focuses on the matching degree between scenario extrapolation and the performance degradation of the digital twin, while also taking into account the diversity of scenarios. The relevant expressions are as follows: ;in, The flatness is matched with a weighting coefficient, with a default value of 0.6; This is the flatness error penalty coefficient, with a default value of 0.5; To predict the flatness under simulated scenarios; This is the baseline flatness degradation value; The weighting coefficient for structural strength is set, with a default value of 0.3; This is the strength error penalty coefficient, with a default value of 0.8; To predict the deflection value under the simulation scenario; The baseline deflection value is the decay value; This is the diversity penalty weighting coefficient, with a default value of 0.1; The entropy value of the action space. , Let g be the probability distribution of the g-th action, where g is the action number, and g = 1, 2, 3, 4; actions include traffic growth rate, heavy vehicle proportion growth rate, etc. The DDPG algorithm is used to handle scene deduction requirements in continuous action spaces, and an Actor-Critic dual-network architecture is constructed; specifically: An Actor network, also known as a policy network, takes a state space S as input and outputs the optimal action sequence A. Its structure includes an input layer, hidden layer 1, hidden layer 2, and an output layer. The number of neurons in hidden layer 1, hidden layer 2, and the output layer are 128, 64, and 4, respectively. The activation functions for hidden layer 1 and hidden layer 2 are ReLU, and the activation function for the output layer is Tanh. The relevant expressions are: ;in, The value of the loss function for the Actor network; For the expectation operation on the experience replay pool, E denotes mathematical expectation. Representing state From the experience replay pool Random sampling is used to represent the average value calculation for all states in the replay pool; The state-action value function corresponds to the evaluation value output by the Critic network, representing the state... Next action The total long-term rewards that can be obtained; For experience replay pool; The policy function of the Actor network, with input state. The corresponding action 'a' is output, which is the core output of the Actor network; The Critic network, or value network, takes a state space S and an action space A as inputs and outputs a state-action value assessment. The network structure includes an input layer, hidden layer 1, hidden layer 2, and an output layer. The number of neurons in hidden layer 1, hidden layer 2, and the output layer are 128, 64, and 1, respectively. The activation functions for hidden layer 1 and hidden layer 2 are ReLU, and the activation function for the output layer is Linear. The relevant expressions are: ;in, The value of the loss function for the Critic network; For the expectation operation on the experience replay pool, E denotes mathematical expectation. Indicates from the experience replay pool The random sampling of state-action-reward-new state samples in the replay pool is used to calculate the average error over all samples in the replay pool. This is an instant reward value; Discount factor; The next state value of the target Critic network; This is the state vector for the next time step; The policy function for the target Actor network; For the current Critic network's value prediction; For different types of uncertainties, adapted stochastic process models are used for refined simulations, including simulations of stochastic fluctuations in traffic flow, uncertainties in climate change, and extreme load events; specifically: Traffic flow stochastic fluctuation simulation uses Markov chains to simulate traffic flow state transitions, dividing traffic flow into three states: low, medium, and high, corresponding to <20,000 vehicles / day, 20,000-50,000 vehicles / day, and >50,000 vehicles / day, respectively. The transition matrix is ​​as follows: ;in, Let be the transition matrix, and let the elements in the transition matrix be... , indicating from state Transition to state The probability of a low-flow state being maintained is 0.8, and the probability of transitioning to a high-flow state is 0.05. Climate change uncertainty simulation uses a Gaussian process to simulate the continuous uncertainty of temperature and precipitation. The mean function has a linear trend, and the covariance function uses a squared exponential kernel. The relevant expressions are as follows: ; ;in, The output value of the mean function; It is a time variable; The intercept of the mean function; The slope of the mean function; The value is the covariance function value; For any two time points; The variance of the signal; For length scale; For noise variance; Let Kronecker function be used when hour Otherwise, it is 0; It should be noted that the mean function is used to characterize the long-term linear trend of climate change and serves as the benchmark prediction for future climate values. This ensures that climate change simulations conform to the long-term trends reported by the IPCC and avoids generating extreme scenarios that violate scientific principles. The covariance function is used to characterize the correlation and random fluctuations of climate variables over time. By using signal variance and length scale, it simulates the interannual fluctuations and temporal correlations of climate, providing a more realistic uncertainty input for predicting road performance degradation. The extreme load event simulation uses a Poisson process to simulate the occurrence of extreme heavy vehicle loads, with event intensity... Set to twice a year, with the time interval following an exponential distribution, the relevant expression is: ;in, Let be the probability density function of the exponential distribution, representing the time interval between two extreme load events. The probability density value; The time interval between two extreme events; event intensity This represents the average number of extreme load events occurring per unit time. The constructed digital twin will be input into the initial state data of road infrastructure, regional baseline climate data, and baseline traffic data; The initial state data of road infrastructure includes, but is not limited to, smoothness and structural deflection values. Regional baseline climate data, including but not limited to annual average temperature and annual precipitation; Baseline traffic data, including but not limited to average daily traffic volume and the proportion of heavy vehicles; Output the initial state vector of the road ; Current road status Input to the Actor network, the Actor network outputs deterministic evolutionary actions. ;in, This represents the annual growth rate of traffic flow. The annual growth rate of the proportion of heavy-duty vehicles; This refers to the annual increase in temperature. This represents the annual increase in precipitation. To avoid scene homogenization, Gaussian noise is superimposed on deterministic actions to generate the final action. , The noise distribution is zero-mean Gaussian, and the noise standard deviation is 0.1 to ensure a moderate degree of randomness in the actions; The sampled motion parameters are cropped to within the preset constraint boundaries, for example... If it exceeds 10%, it will be forcibly set to 10% to ensure that the scenario meets the engineering rationality; When performing climate state simulations, the annual temperature increase output by the Actor network is used. Annual increase in precipitation As input data for climate state simulation; Time steps are generated based on the mean function and covariance function. temperature Precipitation Includes the mean and 95% confidence interval, for example =15.2℃±0.8℃; When performing traffic state simulation, the annual growth rate of traffic flow is based on the output of the Actor network. Annual growth rate of the proportion of heavy-duty vehicles Through formula Update daily average traffic volume ,formula Update the proportion of heavy-duty vehicles ; And, using event intensity A Poisson process with a value of 2 generates time steps. Records of extreme heavy vehicle incidents, including but not limited to the number of occurrences, travel time, and load intensity ≥100 tons; When performing road state evolution calculations using a pavement performance degradation model, the input data is the current road state. The generated temperature Precipitation Annual growth rate of traffic flow Annual growth rate of the proportion of heavy-duty vehicles ; Generate updated current road state ; The road states obtained from different road scenario simulations are sorted and combined to obtain multi-scenario evolution results.

[0024] In this embodiment of the invention, by combining deep reinforcement learning with stochastic processes, dynamic projections of future traffic, loads, and the environment are achieved, which can solve the problem that traditional models can only statically fit historical data. Adaptive stochastic process models are used for different types of uncertainty factors, which can quantify the probability distribution of future scenarios and provide a risk quantification basis for long-term operation and maintenance decisions. By matching and combining the scenario projection results with the digital twin performance degradation model, the simulation accuracy can be effectively improved compared with traditional simulation schemes. Through the design of the entropy term of the reward function, the generated scenarios cover more than 95% of possible future states, avoiding the limitations of decision-making relying on a single optimistic or pessimistic scenario.

[0025] By leveraging the multi-scenario evolution results, a spatiotemporal attention mechanism is used to capture the dynamic interaction effects of key damage factors, enabling accurate prediction of the long-term performance of road infrastructure. Specific steps include: The output multi-scenario evolution results are used as input data for the road performance degradation model, with each group containing time steps of [missing data]. state sequence Among them, the multi-scenario evolution results of roads can specifically correspond to 1000 sets of the next 10 years; the time step can be 10, that is... ; When outputting the road smoothness prediction curve for the next 10 years and the attention weight analysis report of key damage factors, the multi-scenario time series data is converted into input-friendly structured samples. The first 5 years' state sequence for each scenario is used as input features, and the IRI values ​​for the next 5 years are used as prediction labels. The training sample format is as follows: ;in, These are conditional characteristics, including flatness, deflection value, temperature, precipitation, traffic flow, and the proportion of heavy vehicles. ; This represents the true IRI value for the next 5 years. Time step Specifically, the first [number] after the road is put into use Year; Min-Max normalization is performed on all features to eliminate dimensional differences. This is a conventional technical solution, and the specific implementation steps will not be elaborated here. When dividing the dataset, the 1000 scenarios were divided into training set, validation set and test set in an 8:1:1 ratio for model training, optimization and performance evaluation. When designing a dual-branch attention structure to capture the spatial interaction and temporal evolution effects of damage factors, the contribution weights of six state features to performance degradation are calculated for the spatial attention branch, highlighting core damage factors such as traffic flow and smoothness. The relevant expressions are as follows: ;in, This is a spatial attention weight vector, where each element corresponds to the weight of a state feature; The activation function is Softmax normalized. This is the spatial attention weight matrix, a trainable matrix with dimensions of 6×30, where 6 corresponds to 6 features and 30 corresponds to the dimension of the concatenated feature vector. For time series feature concatenation, the feature matrix X of 5 time steps is concatenated row by row into a 30×1 one-dimensional vector; This is a spatial attention bias vector, a trainable vector with dimensions of 6×1, where each element corresponds to a basic bias value of a feature. For the time-attention branch, the weights of the impact of the state at different time steps on future performance are calculated, highlighting the period of accelerated degradation, such as the 3rd to 5th year of road use. The relevant expression is as follows: ;in, This is the temporal attention weight vector; The temporal attention weight matrix is ​​a trainable row vector with a dimension of 1×64, used to learn the association strength between hidden features and future decay. This is a time attention bias vector, a trainable vector with dimensions of 1×5, where each element corresponds to the base bias value for a time step. This is the output matrix of the LSTM hidden layer; The spatial and temporal attention weights are element-wise multiplied to obtain the spatiotemporal fusion attention weights, which are then used to weight the original features: ;in, is the weighted feature sequence; ⊙ is the element-wise product, which multiplies the spatial weights by the temporal weights of the corresponding time steps; For matrix multiplication under the broadcast mechanism, the fusion weights are multiplied by the original feature matrix to obtain the weighted feature sequence; When using a validated road performance degradation model to predict long-term road performance and outputting an attention weight analysis report of key damage factors, the current initial state of the road is input. ; After the road performance degradation model outputs the IRI prediction value every 5 years, it uses the last state of that period as the input for the next round and iteratively outputs the IRI change curve for the next 10-20 years. In addition, it simultaneously outputs heatmaps of spatial attention weights and temporal attention weights, for example: Spatial attention weights specifically include: traffic flow F (0.3) > flatness IRI (0.25) > temperature T (0.15) > precipitation P (0.1) > structural deflection L (0.1) > heavy vehicle proportion V (0.1); The time-based attention weighting specifically includes: the weighting of heavy road usage in the 3rd to 5th year reaches 60%, indicating that this stage is a period of accelerated performance degradation.

[0026] In this embodiment of the invention, the dynamic interaction effect of damage factors is accurately captured through a spatiotemporal attention mechanism. For example, the synergistic effect of traffic flow and high temperature on the accelerated damage to road surface smoothness can effectively reduce errors and improve accuracy compared to the traditional LSTM model. The model trained on 1,000 sets of multi-scenario evolution data can adapt to road performance prediction under different climate and traffic conditions, such as rainy road sections in the south and heavy-load road sections in the north, which can effectively enhance the model's generalization ability. The output attention weights clearly identify key damage factors and the period of accelerated decay, such as "traffic flow in the 3rd to 5th year is the core driving factor of performance decay", providing data support for targeted maintenance, such as strengthening road maintenance and optimizing traffic management during this period.

[0027] In the several embodiments provided by this invention, it should be understood that the disclosed methods can be implemented in other ways. For example, the embodiments of the invention described above are merely illustrative; for instance, the division of modules is only a logical functional division, and there may be other division methods in actual implementation.

[0028] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0029] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module. The integrated module can be implemented in hardware or in the form of hardware plus software functional modules.

[0030] 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 essential characteristics of the present invention.

[0031] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for big data integration and analysis of road infrastructure service performance, characterized in that, include: Construct a digital twin of road infrastructure by integrating historical monitoring data, real-time sensor data and environmental simulation data through multi-source heterogeneous data fusion technology to establish a dynamic mapping relationship between physical entities and digital models; Based on the constructed digital twin, a scenario inference model is trained using deep reinforcement learning algorithms. By introducing stochastic process theory, the model simulates the uncertain evolution of future traffic flow, load level, and climate change, and outputs multi-scenario evolution results. By utilizing the multi-scenario evolution results, the dynamic interaction effects of key damage factors are captured through a spatiotemporal attention mechanism, enabling accurate prediction of the long-term performance of road infrastructure.

2. The method for big data integration and analysis of road infrastructure service performance according to claim 1, characterized in that, Historical monitoring data, real-time sensor data, and environmental simulation data are acquired through a distributed acquisition network; the collected data are standardized and a horizontal federated learning architecture is used to achieve privacy and security fusion across data sources.

3. The method for big data integration and analysis of road infrastructure service performance according to claim 2, characterized in that, Based on the fused feature vectors, a digital twin is constructed that includes a three-dimensional geometric model, a physical and mechanical model, and a behavioral evolution model.

4. The method for big data integration and analysis of road infrastructure service performance according to claim 3, characterized in that, To handle the scene deduction requirements of continuous action space, an Actor-Critic dual network architecture is constructed.

5. The method for big data integration and analysis of road infrastructure service performance according to claim 4, characterized in that, For different types of uncertainties, we use adapted stochastic process models for refined simulation, including simulation of stochastic fluctuations in traffic flow, simulation of uncertainties in climate change, and simulation of extreme load events.

6. The method for big data integration and analysis of road infrastructure service performance according to claim 5, characterized in that, Input the initial state data of road infrastructure, regional baseline climate data, and baseline traffic data into the digital twin, and output the initial state vector of the road. The current road state is input into the Actor network, and the Actor network outputs deterministic evolutionary actions. ;in, This represents the annual growth rate of traffic flow. This represents the annual growth rate of the proportion of heavy-duty vehicles. This refers to the annual increase in temperature. This represents the annual increase in precipitation. Gaussian noise is superimposed on deterministic actions to generate the final action. , It has a zero-mean Gaussian noise distribution.

7. The method for big data integration and analysis of road infrastructure service performance according to claim 6, characterized in that, When performing road condition evolution calculations, the input data are the current road condition, generated temperature, precipitation, annual growth rate of traffic flow, and annual growth rate of the proportion of heavy vehicles; Generate updated current road status; The road states obtained from different road scenario simulations are sorted and combined to obtain multi-scenario evolution results.

8. The method for big data integration and analysis of road infrastructure service performance according to claim 7, characterized in that, Using the road evolution results in multiple scenarios as input data, the system outputs a road smoothness prediction curve for the next 10 years and an attention weight analysis report of key damage factors.

9. The method for big data integration and analysis of road infrastructure service performance according to claim 8, characterized in that, When designing a dual-branch attention structure to capture the spatial interaction effect and temporal evolution effect of damage factors, the contribution weights of six state features to performance degradation are calculated for the spatial attention branch to highlight the core damage factors. For the time attention branch, calculate the weight of the impact of the state at different time steps on future performance, highlighting the period of accelerated decay.

10. The method for big data integration and analysis of road infrastructure service performance according to claim 9, characterized in that, The road performance degradation model, which has been trained and validated, is used to predict the long-term performance of roads and outputs an attention weight analysis report of key damage factors.

Images