Bridge asphalt pavement monitoring method and system based on big data

Abstract

The application relates to the technical field of data processing, in particular to a bridge asphalt pavement monitoring method and system based on big data, which comprises the following steps: acquiring a data set, preprocessing the data set, and dividing the data set into a training set and a verification set; respectively extracting time domain features, frequency domain features and time-frequency features; performing feature combination to generate a high-dimensional feature vector; defining a Particle particle class and initializing a particle swarm, defining a fitness function, performing iterative optimization, and outputting optimal hyperparameters; constructing a CNN model, training the model, and verifying the model based on the verification set; inputting to-be-measured data into the trained model, performing forward propagation calculation, outputting probability distribution values of three states, and determining a final prediction result according to the output probability distribution values. The application fundamentally improves the intelligent level, prediction accuracy and scientific nature of maintenance decision of the bridge asphalt pavement through big data technology driving, and provides efficient and accurate technical support for safe operation of the bridge.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a method and system for monitoring bridge asphalt pavement based on big data. Background Technology

[0002] As a crucial node in highway transportation networks, bridges bear the combined effects of vehicle loads, environmental erosion, and the bridge's own structural vibrations on their asphalt pavement, making it a core weak link in ensuring safe operation throughout the bridge's entire life cycle. Compared to ordinary highway pavements, bridge asphalt pavements face a more complex service environment: the vertical vibration and torsional deformation of the bridge's main beams significantly amplify the interlayer shear stress; the bridge deck experiences large temperature gradients (surface temperatures can reach over 60°C in summer, and freezing is likely in winter); stress concentration is prone to occur near expansion joints; water can easily seep in after the waterproofing layer fails, leading to loosening and voids in the base layer; and the development rate of defects such as rutting under heavy traffic, fatigue cracking, and water damage is much higher than that of ordinary pavements. Statistics show that approximately 30% of bridge pavements in service develop early defects within their design service life, directly threatening traffic safety and significantly increasing maintenance costs.

[0003] Traditional bridge asphalt pavement monitoring requires a full-chain assessment of "apparent defects - internal damage - structural response - long-term performance" to support the shift from post-maintenance to pre-prevention. However, the unique characteristics of bridge asphalt pavements significantly increase the difficulty of monitoring: First, data sources are fragmented, and heterogeneous data such as traffic load, ambient temperature and humidity, bridge vibration, pavement dynamic response, and historical maintenance records are difficult to integrate effectively; second, the causes of defects are complex, with multi-field coupling nonlinear effects of temperature, load, and vibration, making it difficult for traditional shallow feature extraction to capture potential damage; third, there is a strong correlation between bridge dynamic characteristics (such as modal frequencies and damping) and pavement damage, making ordinary highway monitoring methods unsuitable; and fourth, the processing of massive real-time data and the linkage with decision-making are insufficient, making it difficult to achieve risk classification and early warning and intelligent maintenance decision-making.

[0004] The urgent needs of bridge maintenance necessitate the construction of an efficient, accurate, and real-time intelligent monitoring system. Big data technology, through multi-source heterogeneous data fusion, deep learning mining, and distributed computing on cloud platforms, provides a technological foundation for solving these challenges. This application proposes a big data-based bridge asphalt pavement monitoring method to achieve comprehensive, real-time, and intelligent monitoring of dedicated bridge pavements. Summary of the Invention

[0005] This application provides a bridge asphalt pavement monitoring method and system based on big data. By using big data technology, it overcomes all the shortcomings of existing monitoring methods and realizes intelligent, precise, real-time and integrated decision-making for bridge asphalt pavement monitoring, fundamentally improving bridge operation safety and maintenance efficiency.

[0006] To address the aforementioned technical problems, in a first aspect, embodiments of this application provide a bridge asphalt pavement monitoring method and system based on big data, comprising the following steps: First, acquiring a dataset, preprocessing the dataset, and dividing the preprocessed dataset into a training set and a validation set; then, dividing the time series data in the training set into a fixed sliding window, and extracting time-domain features, frequency-domain features, and time-frequency features within each time window; next, combining the extracted time-domain features, frequency-domain features, and time-frequency features to generate a high-dimensional feature vector; then, defining a particle class and initializing the particle swarm, defining a fitness function, performing iterative optimization, and outputting the optimal hyperparameters; based on the optimal hyperparameters, constructing a CNN model, training the model, and validating the model based on the validation set to obtain a trained model; finally, inputting the test data into the trained model, performing forward propagation calculation, outputting the probability distribution values ​​of three states—healthy, slightly damaged, and severely damaged—through the Softmax activation function, and determining the final prediction result based on the output probability distribution values.

[0007] In some exemplary embodiments, the dataset is preprocessed, including: grouping by sensor type and sorting the data in each group in chronological order; if data is missing, linear interpolation is used to fill the missing data with the average of the two valid data points before and after it to ensure the continuity of the time series; the data is cleaned, and then the mean and standard deviation of all data for each sensor are calculated and normalized to output clean data.

[0008] In some exemplary embodiments, data cleaning includes: calculating the mean and standard deviation of the complete dataset for each sensor, identifying outliers by applying the three-times-standard-deviation principle, and replacing outliers with adjacent normal values ​​to eliminate noise interference; for temperature-sensitive strain data, further combining synchronous readings from temperature sensors to establish a linear regression model between strain values ​​and temperature, subtracting the strain component caused by temperature from the measured strain value to obtain the true structural stress and strain, thus eliminating the interference of ambient temperature on the monitoring results.

[0009] In some exemplary embodiments, extracting frequency domain features includes: performing a fast Fourier transform on the data sequence within the same window to convert the time-domain signal into a frequency-domain signal and extracting frequency domain features; extracting time-frequency features includes: using wavelet packet transform to perform a three-level decomposition on the window signal to obtain time-frequency features of different frequency bands, further capturing local variation information of the signal in time and frequency; frequency domain features include spectral energy, main frequency components, and peak power spectral density; time domain features include mean, standard deviation, peak value, root mean square value, skewness, and kurtosis.

[0010] In some exemplary embodiments, the Particle class includes a position array, a velocity array, an optimal position array, and an optimal fitness value; wherein, the position array represents the current hyperparameter combination, the velocity array represents the direction of parameter change, the optimal position array is the recorded historical best hyperparameter combination for an individual, and the optimal fitness value is the recorded corresponding minimum loss value.

[0011] In some exemplary embodiments, initializing the particle swarm includes: setting the number of particles to n, and randomly generating an initial position and initial velocity for each particle within its hyperparameter search space; defining a fitness function includes: the fitness function takes the hyperparameter vector as input, calls the model training function to complete training on the training set, and returns the loss value on the validation set, where a smaller loss value indicates a better optimization effect of the set of hyperparameters.

[0012] In some exemplary embodiments, during each iteration, the optimization steps include: calculating the new velocity of each particle according to the standard velocity update formula of the particle swarm optimization algorithm; superimposing the new velocity onto the current position to obtain the new position, and performing boundary checks on each dimension of the new position, truncating it to the boundary value if it exceeds a preset boundary; calling the fitness function to evaluate the loss value of the new position, and updating its individual optimal position and optimal fitness value if the loss value of the new position is less than the current individual optimal loss value of the particle; traversing all particles to find the particle with the smallest current individual optimal loss, and updating its global optimal position and global optimal fitness value if its loss value is less than the global optimal loss value.

[0013] In some exemplary embodiments, constructing a CNN model includes: constructing a convolutional module: adding a first convolutional layer, using particle swarm optimization (PSO) to optimize the number and size of convolutional kernels, setting a stride of 1, and introducing a ReLU activation function to enhance the model's non-linear expressive power; adding a max pooling layer with a pooling window size of 2 and a stride of 2 to downsample the feature map, reducing the number of parameters while retaining key feature information; adding a second convolutional layer, using the optimal parameters optimized by PSO to further extract higher-level abstract features, and then connecting it again to a ReLU activation function and a max pooling layer; finally, adding an output layer, mapping the features to three dimensions through a fully connected layer, and using a Softmax activation function to convert the output into a probability distribution; wherein the three dimensions correspond to the three categories of bridge asphalt pavement health status: healthy, slightly damaged, and severely damaged.

[0014] In some exemplary embodiments, after adding a second convolutional layer, the multidimensional feature map is flattened into a one-dimensional vector and enters a fully connected layer. This fully connected layer uses the number of neurons optimized by PSO for feature mapping and randomly discards some neurons according to the Dropout rate optimized by PSO to effectively prevent model overfitting.

[0015] Secondly, this application also provides a bridge asphalt pavement monitoring method system based on big data. This system is used to implement the bridge asphalt pavement monitoring method based on big data as described in the above embodiments. The system includes a data preprocessing module, a feature extraction module, a feature combination module, an iterative optimization module, a model training module, and a prediction module connected in sequence. The data preprocessing module acquires a dataset, preprocesses the dataset, and divides the preprocessed dataset into a training set and a validation set. The feature extraction module divides the time series data in the training set into fixed sliding windows, extracting time-domain features, frequency-domain features, and time-frequency features within each time window. The feature combination module... The module is used to combine extracted time-domain features, frequency-domain features, and time-frequency features to generate high-dimensional feature vectors; the iterative optimization module is used to define particle classes and initialize the particle swarm, define the fitness function, perform iterative optimization, and output the optimal hyperparameters; the model training module is used to build a CNN model based on the optimal hyperparameters, train the model, and validate the model based on the validation set to obtain the trained model; the prediction module is used to input the test data into the trained model, perform forward propagation calculation, output the probability distribution values ​​of three states—healthy, slightly damaged, and severely damaged—through the Softmax activation function, and determine the final prediction result based on the output probability distribution values.

[0016] The technical solution provided in this application has at least the following advantages: This application provides a method and system for monitoring bridge asphalt pavement based on big data. The method includes the following steps: First, acquire a dataset, preprocess the dataset, and divide the preprocessed dataset into a training set and a validation set; then, divide the time series data in the training set into a fixed sliding window, and extract time-domain features, frequency-domain features, and time-frequency features in each time window; next, combine the extracted time-domain features, frequency-domain features, and time-frequency features to generate a high-dimensional feature vector; then, define a particle class and initialize a particle swarm, define a fitness function, perform iterative optimization, and output the optimal hyperparameters; based on the optimal hyperparameters, construct a CNN model, train the model, and validate the model based on the validation set to obtain a trained model; finally, input the test data into the trained model, perform forward propagation calculation, and output the probability distribution values ​​of three states—healthy, slightly damaged, and severely damaged—through the Softmax activation function, and determine the final prediction result based on the output probability distribution values.

[0017] This application provides a bridge asphalt pavement monitoring method and system based on big data, which completely overcomes all the shortcomings of existing technologies and achieves the following core objectives: (1) Construct an automated, multi-source real-time data acquisition and intelligent analysis platform to solve the problems of low monitoring efficiency, poor real-time performance and reliance on manual labor, and achieve continuous monitoring that is all-weather, objective and efficient.

[0018] (2) Establish a multi-source heterogeneous data fusion system to break down data silos and fully integrate data such as traffic, environment, bridge vibration and road surface response to solve the problems of single data and insufficient fusion.

[0019] (3) Introduce big data analysis and machine learning deep mining algorithms to achieve nonlinear coupling feature extraction and high-precision prediction, and solve the problems of shallow data processing and low prediction accuracy.

[0020] (4) Design a special collaborative monitoring model for special working conditions of bridges to solve the problems of poor adaptability and weak special monitoring capabilities; at the same time, the monitoring results are directly converted into risk classification early warning and intelligent maintenance decision support to realize the monitoring-evaluation-decision closed loop and solve the problems of fragmented results and inability to provide real-time early warning.

[0021] (5) Based on the big data cloud platform, realize a low-cost and scalable distributed architecture to solve the problems of high deployment cost and poor scalability.

[0022] In summary, this application, driven by big data technology, fundamentally improves the intelligence level of bridge asphalt pavement monitoring, the accuracy of prediction, and the scientific nature of maintenance decisions, providing efficient and precise technical support for the safe operation of bridges. Attached Figure Description

[0023] One or more embodiments are illustrated by way of example with reference to the accompanying drawings. These illustrations do not constitute a limitation on the embodiments, and unless otherwise stated, the figures in the drawings are not to be limited by scale.

[0024] Figure 1 This is a flowchart illustrating a bridge asphalt pavement monitoring method based on big data, provided as an embodiment of this application.

[0025] Figure 2 A flowchart illustrating data preprocessing provided for embodiments of this application.

[0026] Figure 3 A flowchart for feature extraction provided in an embodiment of this application.

[0027] Figure 4 The flowchart illustrates the iterative process and monitoring model optimization provided in the embodiments of this application.

[0028] Figure 5 A flowchart of the training process provided for an embodiment of this application.

[0029] Figure 6 This is a schematic diagram of the structure of a bridge asphalt pavement monitoring system based on big data, provided in an embodiment of this application. Detailed Implementation

[0030] As can be seen from the background technology, existing bridge asphalt pavement monitoring technologies have many shortcomings, such as relying mainly on manual inspection, single sensor embedded monitoring, traditional image / radar non-destructive testing, and partial application of bridge structural health monitoring systems.

[0031] The existing technical solutions most similar to this application mainly focus on the full-cycle performance evaluation of asphalt pavement, sensor-embedded structural monitoring, image / radar non-destructive testing, and bridge structural health monitoring systems. The following describes in detail the four most representative implementation schemes and their core technical characteristics: The first implementation scheme proposes a multi-index-based full-cycle performance monitoring and evaluation method for asphalt pavements. This scheme constructs a pavement performance evaluation model by collecting data on pavement structural material parameters, climate and traffic conditions, pavement age, and maintenance level. It then combines external observation indicators such as smoothness, skid resistance, surface deflection, and pavement damage degree with structural dynamic mechanical indicators such as flexural tensile stress at the bottom of the surface layer, vertical compressive stress at the top of the base layer, and internal shear stress of the pavement for comprehensive evaluation. Data acquisition mainly relies on falling weight deflectometers, laser smoothness meters, manual inspections, or fixed sensors. Data processing methods include traditional statistical models or simple finite element simulations, ultimately outputting a comprehensive pavement health status index. This scheme achieves multi-index fusion from materials to structure, but the data sources are still mainly discrete point measurements and intermittent detection. It lacks a real-time multi-source heterogeneous big data fusion platform, insufficiently explores the vibration-temperature coupling effect of bridges, and uses shallow linear or semi-empirical models for prediction. This makes it difficult to achieve early warning of dynamic damage specific to bridge decks, and the monitoring results are disconnected from the maintenance decision-making system.

[0032] The second implementation scheme proposes an embedded fiber optic sensor or distributed fiber optic monitoring method for asphalt pavement structural performance. This type of scheme (such as the embedded or slotted application of fiber optic strain sensors in asphalt pavements) involves encapsulating the fiber optic sensor and embedding it in the surface or base layer during pavement construction to collect parameters such as temperature, strain, and vibration in real time. The structural response is then inferred using the Bragg wavelength drift formula, enabling long-term monitoring of pavement internal strain. Some schemes combine infrared differential thermal imaging or ground-penetrating radar to assist in detecting permeability and interlayer voids. This technology is widely used in the pavement paving of newly built or reconstructed bridges, offering advantages such as high precision, resistance to electromagnetic interference, and good durability. However, sensor deployment costs are high, the survival rate is greatly affected by construction temperature and compaction, the monitoring range is limited to the deployment points, and data processing remains at the level of single-parameter threshold judgment. It lacks the introduction of a big data platform for multi-source fusion and deep machine learning mining, resulting in insufficient correlation analysis between the overall bridge vibration modes and pavement damage. Furthermore, its scalability is poor, making it difficult to adapt to the unified management of large-scale bridge groups.

[0033] The third approach involves a related technology that proposes a non-destructive testing scheme for bridge deck pavement defects based on image recognition, 3D laser scanning, or GPR (Geometric Pavement Recognition). This scheme utilizes drones equipped with high-definition cameras combined with deep learning to automatically identify surface defects such as cracks, potholes, and ruts. Alternatively, it employs vehicle-mounted / airborne 3D laser point cloud scanners, millimeter-wave radar, and GPR to achieve continuous scanning of pavement thickness, internal voids, and interlayer damage (e.g., generating GPR degradation index maps under the ASTM D-6087 standard). Some advanced solutions integrate BIM models to achieve three-dimensional visualization and localization of defects. Typical applications include steel bridge deck pavement defect identification or "air-water-ground" three-dimensional inspection of urban bridges. While these solutions offer high detection efficiency and enable contactless operation, they rely on limited data (primarily images or radar reflections) and lack integration with real-time traffic load, environmental data, and bridge structural responses. AI identification primarily focuses on extracting superficial features and fails to predict nonlinear coupling damage. Furthermore, special bridge conditions (such as expansion joint interference and vibration noise) lead to a high misjudgment rate, and monitoring results cannot be directly converted into risk warnings and intelligent maintenance decisions. Consequently, system deployment remains isolated, limiting scalability.

[0034] The fourth implementation scheme: Another related technology proposes a bridge structural health monitoring system that combines road surface data. This type of system constructs a comprehensive bridge structural health monitoring platform using accelerometers, strain gauges, GNSS displacement monitoring, etc., collecting data such as vibration, cable force, and displacement, and using finite element model updates or Bayesian methods to assess the structural state. Some schemes introduce big data analysis to locate damage and provide safety warnings from massive monitoring data, and attempt to establish preliminary correlations with road surface inspection data. In recent years, emerging digital twin schemes have further integrated sensor data with physical models to achieve real-time simulation. However, existing structural health monitoring schemes mainly focus on the main bridge structure, with insufficient attention paid to dedicated monitoring of bridge deck asphalt pavement: road surface data is mostly collected as auxiliary or discrete data, and no dedicated fusion model has been built for the unique disease mechanisms of bridge deck pavement (such as waterproof layer failure, interlayer shear); big data applications remain at the storage and simple statistical level, lacking machine learning-driven deep feature mining and real-time correlation analysis of multi-source heterogeneous data (traffic-environment-vibration-road surface response); at the same time, the system is disconnected from maintenance decision-making, making it difficult to achieve an integrated closed loop from monitoring to intelligent early warning.

[0035] In summary, although existing technologies have made progress in single dimensions (such as sensor embedding, image recognition, and structural monitoring), they generally suffer from common problems such as low monitoring efficiency, single data sources, shallow processing, poor adaptability to bridge pavements, disconnect between monitoring and decision-making, and insufficient system scalability. These issues make it difficult to meet the all-weather, precise, and intelligent monitoring needs of bridge asphalt pavements in complex service environments.

[0036] Therefore, existing bridge asphalt pavement monitoring technologies mainly rely on manual inspections, single-sensor embedded monitoring, traditional image / radar non-destructive testing, and partial applications of bridge structural health monitoring systems, which have the following prominent drawbacks: (1) The monitoring efficiency is low, the real-time performance is poor, and it relies heavily on subjective human judgment. Traditional methods mostly use intermittent manual inspections or fixed-point sensors for periodic sampling, which cannot achieve all-weather, continuous, and automated data collection. This results in delayed disease detection, is easily affected by the experience of the inspectors, weather, and traffic control, has large subjective errors, and has a long overall monitoring cycle and low efficiency.

[0037] (2) The data source is singular and cannot effectively integrate multi-source heterogeneous data. Existing solutions rely on only a single type of data (such as road surface images, local strain or vibration), which makes it difficult to simultaneously integrate massive heterogeneous data such as traffic load, environmental temperature and humidity, bridge dynamic response, and historical maintenance records, forming data silos and failing to capture the complex disease mechanism of temperature-load-vibration multi-field coupling.

[0038] (3) The data processing is superficial, the feature mining is insufficient, the prediction accuracy is low and the foresight is lacking. Traditional statistical models or simple finite element simulations only perform shallow feature extraction, which cannot deeply explore nonlinear correlations. The identification rate of potential internal damage (such as interlayer voids, water damage, fatigue cracks) is low, the prediction model accuracy is poor, and it is difficult to predict long-term performance trends.

[0039] (4) Difficult to adapt to special working conditions of bridges and weak special monitoring capabilities. For example, when ordinary road surface monitoring methods are directly applied to bridges, the unique service environment of bridges, such as amplified vibration, large temperature gradient, stress concentration of expansion joints, and failure of waterproof layer, is ignored, resulting in a high rate of missed detection and high rate of misjudgment, and it is impossible to accurately assess the special defects of bridge pavement.

[0040] (5) The monitoring results are severely disconnected from maintenance decisions and real-time early warning is impossible. Most existing technologies only output isolated detection data and lack closed-loop linkage with intelligent maintenance decisions. They cannot achieve risk classification early warning and pre-emptive prevention, and remain at the post-event remedial stage, resulting in insufficient timeliness and scientific rigor in maintenance.

[0041] This application addresses these shortcomings by directly resolving the following technical problems one by one: (1) Solving the technical problems of low monitoring efficiency, poor real-time performance, and reliance on subjective human judgment in existing technologies. This application realizes automated, multi-source real-time data acquisition and intelligent analysis through a big data platform, completely changing the traditional manual inspection or intermittent detection mode, achieving all-weather, continuous, and efficient road condition monitoring, and significantly improving the timeliness and objectivity of monitoring. Solving the technical problems of existing technologies having a single data source and being unable to integrate multi-source heterogeneous data. This application constructs a multi-source data fusion system based on big data, which can simultaneously integrate massive heterogeneous data such as traffic load, ambient temperature / humidity, bridge structure vibration, vehicle dynamic response, and historical maintenance records, overcoming the limitations of traditional methods that rely only on a single sensor or surface image, and realizing comprehensive and correlated road defect diagnosis.

[0042] (2) Solve the technical problems of shallow data processing, insufficient feature mining, and low prediction accuracy in existing technologies. This application adopts big data analysis and machine learning algorithms to deeply mine nonlinear correlations (such as temperature-load coupled damage, vibration-asphalt aging interaction effects), realize accurate identification and long-term performance prediction from surface defects to potential internal damage, and significantly improve prediction accuracy and foresight.

[0043] (3) Solving the technical problem that existing technologies are difficult to conduct special monitoring for special working conditions of bridges. This application is specifically adapted to the unique stress environment of bridge asphalt pavement (bridge vibration, large temperature gradient, impact of expansion joints, risk of waterproof layer failure, etc.), and solves the problem of missed detection or misjudgment caused by the neglect of bridge dynamic characteristics in traditional highway pavement monitoring methods through a big data-driven bridge-pavement collaborative monitoring model, so as to achieve high-precision condition assessment for bridge deck.

[0044] (4) Solve the technical problem of the disconnect between existing monitoring results and maintenance decisions, and the inability to provide real-time early warning. This application directly transforms big data analysis results into structural health assessment, risk classification and early warning, and intelligent maintenance decision support, overcoming the drawbacks of the traditional method of disconnect between "detection-assessment-decision", realizing the transformation from post-remediation to pre-prevention, and improving the overall safety and durability of bridge asphalt pavement.

[0045] In summary, this application comprehensively overcomes all the shortcomings of existing monitoring methods through big data technology, realizing intelligent, precise, real-time, and integrated decision-making for bridge asphalt pavement monitoring, fundamentally improving bridge operation safety and maintenance efficiency.

[0046] The embodiments of this application will now be described in detail with reference to the accompanying drawings. However, those skilled in the art will understand that many technical details have been provided in the embodiments of this application to facilitate a better understanding of the application. However, the technical solutions claimed in this application can be implemented even without these technical details and various variations and modifications based on the following embodiments.

[0047] See Figure 1 This application provides a method and system for monitoring bridge asphalt pavement based on big data, including the following steps: Step S1: Obtain the dataset, preprocess the dataset, and divide the preprocessed dataset into a training set and a validation set.

[0048] Step S2: Divide the time series data in the training set into fixed sliding windows, and extract time-domain features, frequency-domain features, and time-frequency features in each time window.

[0049] Step S3: Combine the extracted time-domain features, frequency-domain features, and time-frequency features to generate a high-dimensional feature vector.

[0050] Step S4: Define the Particle class and initialize the particle swarm, define the fitness function, perform iterative optimization, and output the optimal hyperparameters.

[0051] Step S5: Based on the optimal hyperparameters, construct a CNN model, train the model, and validate the model based on the validation set to obtain the trained model.

[0052] Step S6: Input the test data into the trained model, perform forward propagation calculation, and output the probability distribution values ​​of three states: healthy, slightly damaged, and severely damaged through the Softmax activation function. Determine the final prediction result based on the output probability distribution values.

[0053] It should be noted that the preprocessed dataset was randomly divided into training and validation sets in an 8:2 ratio, and a stratified sampling strategy was used to ensure that the distribution of samples of each category remained consistent between the training and validation sets. Before inputting the test data into the trained model, data preprocessing (including missing value imputation, outlier removal, temperature correction, and Z-score normalization) and feature extraction (time domain, frequency domain, and time-frequency domain features) were performed strictly according to the same steps as in the training phase, generating a high-dimensional feature vector consistent with the training sample format. Subsequently, this feature vector was input into the loaded model, and forward propagation was performed to output the probability distribution values ​​of three states: healthy, slightly damaged, and severely damaged.

[0054] Step S1 mainly involves data preprocessing, and its flowchart is as follows: Figure 2As shown. In some embodiments, step S1 preprocesses the dataset, including: grouping by sensor type and sorting the data in each group in chronological order; if data is missing, linear interpolation is used to fill the missing data with the average of the two valid data points before and after it to ensure the continuity of the time series; the data is cleaned, and then the mean and standard deviation of all data for each sensor are calculated and normalized to output clean data.

[0055] In some exemplary embodiments, data cleaning includes: calculating the mean and standard deviation of the complete dataset for each sensor; applying the three-standard-deviation principle (3σ criterion) to identify outliers; and replacing outliers with adjacent normal values—that is, data points exceeding three standard deviations above or below the mean are considered outliers and replaced with adjacent normal values ​​to eliminate noise interference; for temperature-sensitive strain data, further combining synchronous readings from temperature sensors to establish a linear regression model between strain values ​​and temperature, subtracting the temperature-induced strain component from the measured strain value to obtain the true structural stress and strain, thus eliminating the interference of ambient temperature on the monitoring results.

[0056] After completing the above data cleaning, the mean and standard deviation of all data from each sensor are calculated, and Z-score standardization is performed to convert all data into a normal distribution with a mean of 0 and a standard deviation of 1, eliminating the influence of differences in the dimensions of different sensors. Finally, the preprocessed cleaned data is completely written into the HBase database, while retaining all fields of the original data for subsequent traceability and auditing (e.g., ...). Figure 2 (As shown). This process forms a complete and logically rigorous mechanism for cleaning multi-source heterogeneous data of bridge asphalt pavement, providing high-quality and highly consistent input data for subsequent big data fusion and intelligent analysis.

[0057] Step S2 is mainly used for feature extraction. After extracting time-domain features, frequency-domain features, and time-frequency features respectively, step S3 is executed to combine the features. The flowchart is as follows: Figure 3As shown, the preprocessed time series data is divided into fixed sliding windows. Within each time window, time-domain features are extracted, including statistics such as mean, standard deviation, peak value, root mean square value, skewness, and kurtosis. Next, a Fast Fourier Transform (FFT) is performed on the data sequence within the same window to convert the time-domain signal to a frequency-domain signal, extracting frequency-domain features such as spectral energy, main frequency components, and peak power spectral density. Then, wavelet packet transform is used to perform a three-level decomposition of the window signal to obtain time-frequency domain features of different frequency bands, further capturing local variations in time and frequency. Afterward, the time-domain, frequency-domain, and time-frequency-domain features extracted from each window are merged to construct a high-dimensional feature vector, forming a complete sample. Simultaneously, key metadata such as the sensor number, installation location, window start time, and end time are recorded for each window. Finally, the feature vectors and related metadata for all time windows are completely written into the HBase database, forming a structured feature dataset that can be used for subsequent model training and analysis. This feature extraction process systematically transforms the original time-series signal into a feature vector containing rich time-domain, frequency-domain, and time-frequency-domain information, providing a high-quality input foundation for in-depth mining of multi-source data on bridge asphalt pavement and identification of nonlinear coupled damage (e.g., Figure 3 (As shown).

[0058] Step S4 mainly involves the iterative process and monitoring model optimization, and its flowchart is as follows: Figure 4 As shown. Specifically, during iterative optimization, firstly, a Particle class is defined, which contains a position array (representing the current hyperparameter combination), a velocity array (representing the direction of parameter change), an optimal position array (recording the individual's historical best hyperparameter combination), and an optimal fitness value (recording the corresponding minimum loss value). Then, the particle swarm is initialized, with the number of particles set to n. For each particle, an initial position (within preset boundaries for each dimension) and an initial velocity (randomly generated within a defined velocity range) are randomly generated within its hyperparameter search space. Next, a fitness function is defined. This function takes the hyperparameter vector as input, calls the model training function to complete training on the training set, and returns the loss value on the validation set. The smaller the loss value, the better the optimization effect of that set of hyperparameters.

[0059] In each iteration, optimization is performed as follows: First, the new velocity of each particle is calculated according to the standard velocity update formula of the particle swarm optimization algorithm; then, the new velocity is superimposed on the current position to obtain the new position, and boundary checks are performed on each dimension of the new position. If it exceeds the preset boundary, it is truncated to the boundary value; subsequently, the fitness function is called to evaluate the loss value of the new position. If the loss value of the new position is less than the current individual optimal loss value of the particle, its individual optimal position and optimal fitness value are updated; finally, all particles are traversed to find the particle with the smallest current individual optimal loss. If its loss value is less than the global optimal loss value, the global optimal position and global optimal fitness value are updated. The entire optimization process is performed n iterations, and the current global optimal loss value is output every m iterations to monitor the optimization progress in real time. Through the above iterative processes of particle initialization, fitness evaluation, velocity and position updates, and individual and global optimal updates, efficient global search and optimization of model hyperparameters are finally achieved. This particle swarm optimization process provides efficient and stable automatic optimization capabilities for key hyperparameters (such as learning rate, number of neural network layers, window size, etc.) of bridge asphalt pavement monitoring models, effectively improving the performance of subsequent damage identification and prediction models.

[0060] Step S5 mainly involves model building and training, and its flowchart is as follows: Figure 5As shown, the input layer first receives the high-dimensional feature vector after feature extraction and transforms it into an input shape suitable for a one-dimensional convolutional neural network through a reshape operation. Then, convolutional modules are constructed sequentially: a first convolutional layer is added, using particle swarm optimization (PSO) to optimize the number and size of convolutional kernels, setting a stride of 1, and introducing a ReLU activation function to enhance the model's non-linear expressive power; next, a max-pooling layer is added with a pooling window size of 2 and a stride of 2 to downsample the feature map, reducing the number of parameters while retaining key feature information. Then, a second convolutional layer is added, also using optimal parameters optimized by PSO, to further extract higher-level abstract features, followed by another ReLU activation function and max-pooling layer. After the convolutional modules, the multi-dimensional feature map is flattened into a one-dimensional vector and enters a fully connected layer. This fully connected layer uses the number of neurons optimized by PSO for feature mapping and randomly discards some neurons according to the Dropout rate optimized by PSO to effectively prevent overfitting. Finally, an output layer is added, which maps the features to three dimensions (corresponding to the three categories of bridge asphalt pavement health status: healthy, slightly damaged, and severely damaged) through a fully connected layer, and uses the Softmax activation function to convert the output into a probability distribution. Before model training, the feature dataset is randomly divided into training and validation sets in an 8:2 ratio, and a stratified sampling strategy is used to ensure that the distribution of samples of each category is consistent between the training and validation sets. During training, iterative optimization is performed on the training set, calculating cross-entropy loss and gradient, and updating network parameters through the backpropagation algorithm until the model converges.

[0061] Finally, step S6 is executed to load the trained CNN model parameters from the file, input the test data into the trained CNN model, and output the final result. For newly acquired real-time data from bridge asphalt pavement sensors, data preprocessing and feature extraction are performed strictly according to the same steps as in the training phase to generate a high-dimensional feature vector consistent with the training sample format. Subsequently, this feature vector is input into the loaded model, and forward propagation calculation is performed. The probability distribution values ​​of three states—healthy, slightly damaged, and severely damaged—are output through the Softmax activation function. Finally, the final prediction result is determined based on the output probability values ​​(usually the category with the highest probability is used as the prediction label), while retaining the specific probability values ​​of the three categories to provide a quantitative basis for subsequent risk grading assessment.

[0062] See Figure 6This application also provides a bridge asphalt pavement monitoring method system based on big data. This system is used to implement the bridge asphalt pavement monitoring method based on big data as described in the above embodiments. The system includes a data preprocessing module 101, a feature extraction module 102, a feature combination module 103, an iterative optimization module 104, a model training module 105, and a prediction module 106 connected in sequence. The data preprocessing module 101 is used to acquire a dataset, preprocess the dataset, and divide the preprocessed dataset into a training set and a validation set. The feature extraction module 102 is used to divide the time series data in the training set according to a fixed sliding window, and extract time-domain features, frequency-domain features, and time-frequency features within each time window. The feature combination module 103 is used to combine the extracted time-domain features, frequency-domain features, and time-frequency features to generate a high-dimensional feature vector; the iterative optimization module 104 is used to define the particle class and initialize the particle swarm, define the fitness function, perform iterative optimization, and output the optimal hyperparameters; the model training module 105 is used to build a CNN model based on the optimal hyperparameters, train the model, and validate the model based on the validation set to obtain the trained model; the prediction module 106 is used to input the test data into the trained model, perform forward propagation calculation, output the probability distribution values ​​of three states—healthy, slightly damaged, and severely damaged—through the Softmax activation function, and determine the final prediction result based on the output probability distribution values.

[0063] In summary, the core technical solution described in this application is as follows: Through real-time acquisition of multi-source heterogeneous data from bridges and pavements, a big data fusion platform, in-depth nonlinear coupling damage mining, dedicated risk grading assessment and intelligent maintenance decision-making closed loop, and a distributed cloud platform architecture, integrated real-time and intelligent monitoring and decision-making for all elements of bridge asphalt pavements are achieved. This solution comprehensively addresses all the shortcomings of existing technologies in terms of efficiency, fusion capability, prediction accuracy, adaptability, decision-making closed loop, and scalability, thereby achieving the purpose of the invention.

[0064] This application explicitly states that it is not limited to the only specific implementation methods described in the above embodiments. Any technical means that are equivalent or similar to the technical solutions of the above embodiments (including but not limited to the substitution of some structures, devices, method steps, or equivalent replacements of the complete technical solution), as long as they can achieve the same inventive purpose (i.e., multi-source data fusion of bridge asphalt pavement, accurate identification of coupled damage, and integrated intelligent early warning and maintenance decision-making), fall within the protection scope of this application. Several alternative solutions are detailed below. These alternative solutions can be directly obtained or slightly modified by those skilled in the art without creative effort, aiming to expand the protection scope of this patent and prevent others from circumventing patent protection through simple substitution or circumvention.

[0065] The alternatives to some of the structures, devices, or method steps in this application are as follows: (1) Replacement of real-time acquisition of multi-source heterogeneous data.

[0066] The embodiments described above employ a specific combination of vibration sensors, temperature / humidity sensors, vehicle dynamic response sensors, and road surface strain sensors. Alternative solutions include: replacing some vibration sensors with accelerometers, tilt sensors, and displacement monitors; replacing wired sensor deployment with distributed fiber optic cables and wireless IoT nodes; or adding UAV / vehicle-mounted mobile laser scanners and millimeter-wave radar as supplementary data acquisition methods. As long as the simultaneous real-time acquisition of all elements—bridge structural response, bridge deck pavement response, traffic load, and environmental parameters—is achieved, and specific calibrations are performed for bridge vibration amplification, temperature gradient, and the impact of expansion joints, the same basic data acquisition objective can be accomplished, falling within the scope of this application.

[0067] (2) Replacement of big data fusion platform.

[0068] The embodiments described above employ a cloud / edge collaborative platform combined with specific temporal alignment, spatial mapping, and multimodal correlation fusion algorithms. Alternative solutions include: using a pure edge computing architecture (where all data is fused at local edge nodes) or a purely cloud-based centralized fusion approach; the fusion algorithm can be replaced with a multimodal fusion model such as Kalman filtering and graph neural networks; and spatial mapping can be replaced with a BIM digital twin model or a GIS geographic information system. As long as deep fusion of multi-source heterogeneous data on traffic, environment, vibration, and pavement response is achieved, and bridge-pavement coupling features (vibration-shear, temperature-aging, etc.) are extracted, the same fusion objective can be achieved.

[0069] (3) Replacement of nonlinear coupling damage depth mining and prediction model.

[0070] The embodiments described above employ machine learning / deep learning for nonlinear mining and long-term performance prediction. Alternative solutions include: using a support vector machine + random forest combined model, Bayesian networks, or a physics-data hybrid driven finite element + neural network model; the prediction module can be replaced with time series prediction or a reinforcement learning dynamic optimization model. As long as the goal of "accurate identification of superficial defects to internal damage (interlayer voids, water damage, fatigue) and forward-looking prediction of long-term decay trends" is achieved, the same high-precision mining and prediction objective can be accomplished.

[0071] (4) Replacement of risk classification assessment and intelligent maintenance decision-making closed-loop mechanism.

[0072] The embodiments described above in this application construct a dedicated health assessment indicator system and map it to maintenance decisions. Alternative solutions include using fuzzy comprehensive evaluation, analytic hierarchy process (AHP), or expert system rule bases to replace machine learning assessments; the decision output can be directly interfaced with a third-party maintenance management system, or a maintenance report can be automatically generated using knowledge graphs and natural language processing. As long as a closed-loop linkage of "monitoring-assessment-risk grading and early warning-recommendation of preventative / remedial maintenance measures" is formed, the same integrated decision-making objective can be achieved.

[0073] In addition to the aforementioned partial substitutions, the same inventive objective can also be achieved through the following complete alternative technical solutions: (1) Alternative solution based on digital twin + physical simulation.

[0074] Using a bridge-road digital twin model as the core, multi-source sensor data is injected in real time. Real-time finite element analysis combined with a proxy model replaces pure data-driven data mining, achieving coupled damage simulation and prediction. The decision-making part uses a model predictive control algorithm to directly output maintenance strategies. This solution also completes a closed loop of real-time monitoring, fusion, prediction, and decision-making, and has greater advantages in scenarios with limited computing resources.

[0075] (2) Alternative solutions based on 5G + edge intelligence + satellite remote sensing.

[0076] By leveraging 5G low-latency networks and satellite remote sensing to supplement bridge displacement data, a lightweight neural network is deployed at the edge for initial local data mining, while the cloud performs global fusion and decision-making. This solution is suitable for remote bridge clusters, achieving comprehensive monitoring and intelligent decision-making at a lower deployment cost.

[0077] (3) Cross-bridge group alternative based on federated learning + privacy computing.

[0078] Data from multiple bridges is not uploaded centrally; instead, models are trained locally using federated learning, and model parameters are shared to achieve cross-bridge knowledge transfer and global prediction. This solution can still achieve the same goal under scenarios with strict data privacy regulations, while enhancing system scalability.

[0079] (4) Lightweight alternative based on knowledge graph + rule engine.

[0080] This solution constructs a knowledge graph of bridge asphalt pavement defects based on the fusion of big data, and uses a rule engine instead of deep learning for risk assessment and decision-making reasoning. It requires fewer computational resources, is suitable for monitoring small and medium-sized bridges, and achieves the same goals of accurate identification, early warning, and closed-loop decision-making.

[0081] All the above-described alternatives (partial or complete) can achieve the purpose of this application (i.e., intelligent, precise, real-time, and integrated decision-making for bridge asphalt pavement monitoring) with effects no less than the core solutions of the embodiments described above, and can be obtained by those skilled in the art without creative effort through the above substitutions or combinations. Therefore, any technical solution that achieves the same function using any of the above-described alternatives or any combination thereof falls within the protection scope of this application, thereby effectively preventing others from circumventing patent protection by simply replacing sensor types, algorithm frameworks, platform architectures, or overall technical paths.

[0082] In summary, the core of this application lies in leveraging big data technology to achieve an intelligent upgrade of bridge asphalt pavement monitoring. This application provides a complete technical solution for multi-source data acquisition, fusion, analysis, and decision-making. These key points demonstrate the substantial differences and inventive contributions of this application compared to existing technologies.

[0083] (1) Real-time acquisition of multi-source heterogeneous data and bridge-road collaborative perception mechanism.

[0084] This application utilizes multiple sensors (including vibration, temperature / humidity, vehicle dynamic response, and pavement strain) deployed in the bridge structure, pavement layer, and traffic environment, combined with traffic load data and historical maintenance records, to achieve real-time synchronous acquisition of all elements. Furthermore, it designs specialized sensing deployment and calibration methods for special conditions such as bridge vibration amplification effects, temperature gradients, and the impact of expansion joints. This is the unique aspect of the data foundation of this application, distinguishing it from existing single-sensor or ordinary highway pavement monitoring schemes.

[0085] (2) Multi-source heterogeneous data fusion and bridge-specific feature engineering method based on big data platform.

[0086] This application constructs a big data fusion platform and employs a specific fusion algorithm to deeply integrate traffic load, environmental parameters, bridge dynamic response (modal frequency, damping, etc.) with pavement response data, and extracts bridge-pavement coupling features. This solves the data silo problem in existing technologies.

[0087] (3) Big data-driven nonlinear coupling damage in-depth mining and prediction model.

[0088] This application introduces machine learning / deep learning algorithms to perform nonlinear correlation mining on multi-source fusion data, achieving accurate identification from superficial defects (cracks, rutting) to internal potential damage (interlayer voids, water damage, fatigue), and establishing a long-term performance degradation prediction model to support forward-looking risk trend analysis. Breaking through the limitations of existing shallow statistical or single finite element models, high-precision coupled prediction is the technical highlight of this application.

[0089] (4) Closed-loop mechanism for risk classification assessment and intelligent maintenance decision-making for bridge asphalt pavement. Based on the fusion and data mining results, this application establishes a dedicated health assessment index system for bridge decks, enabling real-time early warning of disease risk levels. The monitoring output is directly mapped to intelligent maintenance decision-making suggestions (timing and recommended preventative maintenance measures), forming an integrated closed loop of monitoring, assessment, early warning, and decision-making. This addresses the existing problem of the disconnect between monitoring and decision-making.

[0090] Compared with existing technologies, the big data-based bridge asphalt pavement monitoring method provided in this application has significant advantages. While existing technologies introduce big data storage and simple statistical analysis, and attempt to preliminarily correlate bridge vibration, displacement, and other structural response data with a small amount of pavement detection data, they still primarily focus on monitoring the main bridge structure, with pavement data as an auxiliary component. Data fusion remains at a superficial level, lacking a dedicated bridge-pavement coupling model, and failing to achieve a closed loop for monitoring and maintenance decisions. This application adopts a complete technical solution of "real-time acquisition of multi-source heterogeneous data - big data fusion platform - deep learning nonlinear mining - dedicated risk assessment and intelligent decision-making closed loop," which has the following significant advantages compared with the aforementioned best existing technologies: (1) The real-time performance and efficiency of monitoring are greatly improved, completely solving the bottlenecks of reliance on manual labor and intermittent detection. The best existing technology still relies on discrete sensor sampling and manual assisted inspection, resulting in long data acquisition cycles and limited coverage. This application achieves real-time synchronous acquisition of all elements through multi-source sensors in the bridge structure, bridge deck pavement layer and traffic environment, and relies on the big data cloud platform for edge-cloud collaborative processing, realizing all-weather, continuous and automated monitoring. After adopting the technical solution of this application, the time for disease detection is shortened from several hours / day to minutes, the monitoring efficiency is more than 3 times higher than the best existing technology, subjective judgment errors are eliminated, and the timely guarantee capability of bridge asphalt pavement operation safety is significantly improved.

[0091] (2) The ability to deeply fuse multi-source heterogeneous data is significantly enhanced, enabling the extraction of bridge-pavement-specific coupling features. The best existing technology only simply correlates bridge health monitoring data with a small number of pavement indicators, without constructing a dedicated fusion model for special working conditions such as bridge vibration amplification, temperature gradient, and expansion joint effects, resulting in prominent data silos. This application constructs a big data fusion platform based on temporal alignment, spatial mapping, and multimodal correlation, which deeply fuses massive heterogeneous data such as traffic load, ambient temperature and humidity, bridge modal frequency / damping, pavement strain, and historical maintenance records, and extracts bridge-specific coupling features such as vibration-shear stress and temperature-aging interaction. After adopting the solution of this application, the data utilization rate is improved by more than 80% compared with the best existing technology, which can accurately capture complex disease mechanisms under multi-field coupling and solve the technical shortcomings of existing technologies that cannot adapt to the unique service environment of bridge asphalt pavement.

[0092] (3) The accuracy of damage identification and performance prediction is greatly improved, and it has the ability to analyze long-term trends. The best existing technology uses shallow statistical models or finite element updates, and feature mining remains at the surface level. It is difficult to handle nonlinear coupling effects, and the identification rate and long-term prediction accuracy of internal damage (such as interlayer voids and water damage) are low. This application introduces machine learning / deep learning algorithms to perform nonlinear correlation deep mining on the fused data, and establishes an accurate identification model from apparent defects to potential internal damage and a long-term performance degradation prediction model. After adopting the technical solution of this application, the defect prediction accuracy is improved by more than 40% compared with the best existing technology. It can warn of potential risks 1-2 maintenance cycles in advance, realize the fundamental transformation from "post-event remediation" to "pre-event prevention", and significantly extend the service life of bridge asphalt pavement.

[0093] (4) Monitoring results are directly transformed into intelligent maintenance decisions, forming a complete closed-loop management mechanism. Current best practices output isolated health indices or early warning information, leaving monitoring and maintenance decisions disconnected and unable to provide risk grading and recommended measures. This application, based on fusion and data mining results, constructs a dedicated health assessment index system for bridge asphalt pavements, enabling real-time early warning of disease risk grading and directly mapping monitoring outputs to intelligent maintenance decision recommendations. After adopting this solution, the scientific accuracy of maintenance decisions is more than twice that of current best practices, achieving an integrated closed loop of "monitoring-assessment-early warning-decision," reducing maintenance costs by 25%-35% compared to traditional methods, and significantly improving the precision and economy of bridge maintenance. Based on the above technical solutions, this application provides a method and system for monitoring bridge asphalt pavement based on big data. The method includes the following steps: First, acquire a dataset, preprocess the dataset, and divide the preprocessed dataset into a training set and a validation set; then, divide the time series data in the training set into a fixed sliding window, and extract time-domain features, frequency-domain features, and time-frequency features in each time window; next, combine the extracted time-domain features, frequency-domain features, and time-frequency features to generate a high-dimensional feature vector; then, define a particle class and initialize a particle swarm, define a fitness function, perform iterative optimization, and output the optimal hyperparameters; based on the optimal hyperparameters, construct a CNN model, train the model, and validate the model based on the validation set to obtain a trained model; finally, input the test data into the trained model, perform forward propagation calculation, output the probability distribution values ​​of three states—healthy, slightly damaged, and severely damaged—through the Softmax activation function, and determine the final prediction result based on the output probability distribution values.

[0094] The purpose of this application is to provide a big data-based bridge asphalt pavement monitoring method that completely overcomes all the shortcomings of existing technologies and achieves the following core objectives: Constructing an automated, multi-source real-time data acquisition and intelligent analysis platform to solve the problems of low monitoring efficiency, poor real-time performance, and reliance on manual labor, achieving all-weather, objective, and efficient continuous monitoring; Establishing a multi-source heterogeneous data fusion system to break down data silos and comprehensively integrate data from traffic, environment, bridge vibration, and pavement response, solving the problems of single data and insufficient fusion; Introducing big data analysis and machine learning deep mining algorithms to achieve nonlinear coupled feature extraction and high-precision prediction, solving the problems of shallow data processing and low prediction accuracy; Designing a dedicated collaborative monitoring model for special bridge conditions to solve the problems of poor adaptability and weak specialized monitoring capabilities; Directly transforming monitoring results into risk classification early warning and intelligent maintenance decision support, realizing a closed loop of monitoring-assessment-decision, solving the problems of fragmented results and inability to provide real-time early warning; Implementing a low-cost, scalable distributed architecture based on a big data cloud platform to solve the problems of high deployment costs and poor scalability.

[0095] In summary, the purpose of this application is to fundamentally improve the intelligence level of bridge asphalt pavement monitoring, the accuracy of prediction, and the scientific nature of maintenance decisions through big data technology, so as to provide efficient and accurate technical support for the safe operation of bridges.

[0096] Those skilled in the art will understand that the above-described embodiments are specific examples of implementing this application, and in practical applications, various changes in form and detail may be made without departing from the spirit and scope of this application. Any person skilled in the art can make their own modifications and alterations without departing from the spirit and scope of this application; therefore, the scope of protection of this application should be determined by the scope defined in the claims.

Claims

1. A method for monitoring bridge asphalt pavement based on big data, characterized in that, Includes the following steps: Obtain the dataset, preprocess the dataset, and divide the preprocessed dataset into training and validation sets; The time series data in the training set are divided into fixed sliding windows, and time-domain features, frequency-domain features, and time-frequency features are extracted in each time window. The extracted time-domain features, frequency-domain features, and time-frequency features are combined to generate a high-dimensional feature vector; Define the Particle class and initialize the particle swarm, define the fitness function, perform iterative optimization, and output the optimal hyperparameters; Based on the optimal hyperparameters, a CNN model is constructed, the model is trained, and the model is validated based on the validation set to obtain the trained model. The test data is input into the trained model, forward propagation is performed, and the probability distribution values ​​of three states—healthy, slightly damaged, and severely damaged—are output through the Softmax activation function. The final prediction result is determined based on the output probability distribution values.

2. The bridge asphalt pavement monitoring method based on big data according to claim 1, characterized in that, Preprocessing of the dataset includes: Group the data by sensor type and sort the data in each group in chronological order. If data is missing, linear interpolation is used to fill the gap by averaging the two valid data points before and after the missing data, thus ensuring the continuity of the time series. The data is cleaned, and then the mean and standard deviation of all data from each sensor are calculated, normalized, and clean data is output.

3. The bridge asphalt pavement monitoring method based on big data according to claim 2, characterized in that, Data cleaning includes: For the complete dataset of each sensor, the mean and standard deviation are calculated. Outliers are identified by applying the three-standard-deviation principle and replaced with adjacent normal values ​​to eliminate noise interference. For temperature-sensitive strain data, a linear regression model between strain value and temperature is established by combining synchronous readings from temperature sensors. The strain component caused by temperature is subtracted from the measured strain value to obtain the true structural stress and strain, thus eliminating the interference of ambient temperature on the monitoring results.

4. The bridge asphalt pavement monitoring method based on big data according to claim 1, characterized in that, Extracting frequency domain features includes: performing a fast Fourier transform on the data sequence within the same window to convert the time-domain signal into a frequency-domain signal and extracting frequency domain features; Extracting time-frequency features includes: Wavelet packet transform is used to decompose the window signal into three levels to obtain the time-frequency characteristics of different frequency bands, and further capture the local variation information of the signal in time and frequency. The frequency domain characteristics include spectral energy, main frequency components, and power spectral density peaks; The time-domain features include mean, standard deviation, peak value, root mean square value, skewness, and kurtosis.

5. The bridge asphalt pavement monitoring method based on big data according to claim 1, characterized in that, The Particle class includes a position array, a velocity array, an optimal position array, and an optimal fitness value; among these... The position array represents the current hyperparameter combination, the velocity array represents the direction of parameter change, the optimal position array is the recorded individual's historical best hyperparameter combination, and the optimal fitness value is the recorded corresponding minimum loss value.

6. The bridge asphalt pavement monitoring method based on big data according to claim 1, characterized in that, Initialize the particle swarm, including: setting the number of particles to n, and randomly generating the initial position and initial velocity for each particle within its hyperparameter search space; Define the fitness function, which takes the hyperparameter vector as input, calls the model training function to complete training on the training set, and returns the loss value on the validation set. The smaller the loss value, the better the optimization effect of the hyperparameter set.

7. The bridge asphalt pavement monitoring method based on big data according to claim 1, characterized in that, In each iteration, the optimization steps include: Calculate the new velocity of each particle according to the standard velocity update formula of the particle swarm optimization algorithm; The new velocity is superimposed on the current position to obtain the new position, and boundary checks are performed on each dimension of the new position. If it exceeds the preset boundary, it is truncated to the boundary value. Call the fitness function to evaluate the loss value at the new position. If the loss value at the new position is less than the current best loss value of the particle, then update its best position and best fitness value. Iterate through all particles and find the particle with the smallest current individual optimal loss. If its loss value is less than the global optimal loss value, then update the global optimal position and the global optimal fitness value.

8. The bridge asphalt pavement monitoring method based on big data according to claim 1, characterized in that, Building a CNN model includes: Constructing the convolutional module: Add the first convolutional layer, use the particle swarm optimization algorithm to optimize the number and size of convolutional kernels, set the stride to 1, and introduce the ReLU activation function to enhance the non-linear expressive power of the model; Add a max pooling layer with a pooling window size of 2 and a stride of 2 to downsample the feature map, thereby reducing the number of parameters while retaining key feature information; A second convolutional layer is added, using the optimal parameters after particle swarm optimization, to further extract higher-level abstract features, and then ReLU activation function and max pooling layer are applied again; Finally, an output layer is added, which maps the features to three dimensions through a fully connected layer, and uses the Softmax activation function to convert the output into a probability distribution; where the three dimensions correspond to the three categories of the health status of bridge asphalt pavement: healthy, slightly damaged, and severely damaged.

9. The bridge asphalt pavement monitoring method based on big data according to claim 8, characterized in that, After adding the second convolutional layer, the multidimensional feature map is flattened into a one-dimensional vector and enters the fully connected layer. This fully connected layer uses the number of neurons optimized by PSO for feature mapping and randomly discards some neurons according to the Dropout rate optimized by PSO to effectively prevent the model from overfitting.

10. A bridge asphalt pavement monitoring method system based on big data, the system being used to implement the bridge asphalt pavement monitoring method based on big data as described in any one of claims 1 to 9, characterized in that, The system comprises a data preprocessing module, a feature extraction module, a feature combination module, an iterative optimization module, a model training module, and a prediction module, connected in sequence; among them, The data preprocessing module is used to acquire the dataset, preprocess the dataset, and divide the preprocessed dataset into a training set and a validation set. The feature extraction module is used to divide the time series data in the training set into fixed sliding windows, and extract time-domain features, frequency-domain features and time-frequency features in each time window respectively; The feature combination module is used to combine the extracted time-domain features, frequency-domain features, and time-frequency features to generate a high-dimensional feature vector; The iterative optimization module is used to define the Particle class and initialize the particle swarm, define the fitness function, perform iterative optimization, and output the optimal hyperparameters. The model training module is used to construct a CNN model based on the optimal hyperparameters, train the model, and validate the model based on the validation set to obtain the trained model. The prediction module is used to input the data to be tested into the trained model, perform forward propagation calculation, output the probability distribution values ​​of three states—healthy, slightly damaged, and severely damaged—through the Softmax activation function, and determine the final prediction result based on the output probability distribution values.

Images