Highway maintenance demand prediction system based on big data analysis

Abstract

The present application belongs to the technical field of data processing, and relates to a highway maintenance demand prediction system based on big data analysis. The system obtains a road surface image sequence and historical maintenance text through a visual perception terminal and a business system interface. A convolutional neural network is used by a feature extraction processor to generate an apparent disease morphology feature vector, and a pre-trained language model is used to generate a process semantic feature vector. A cross-modal cross-attention module in the fusion prediction processor calculates the morphology vector as a query item and the semantic vector as a key-value item to generate a dynamic adjustment coefficient of disease development rate. A road structure mechanics attenuation model solves the mechanics equation based on the dynamic adjustment coefficient and outputs a remaining service life prediction curve. The present scheme establishes a directional correlation between disease morphology and construction process at the feature level, eliminates the causal loss caused by direct splicing, and makes the prediction curve correspond to the real deterioration trajectory of the road structure.

Description

Technical Field

[0001] This invention belongs to the field of data processing technology and relates to a highway maintenance demand prediction system based on big data analysis. Background Technology

[0002] Existing highway maintenance demand forecasting solutions generally employ a multimodal joint prediction architecture when processing multi-source heterogeneous data. In practice, existing systems acquire pavement appearance images through vision devices and obtain historical maintenance construction log text through business systems. The system uses convolutional networks to perform convolution operations on the images, extracting numerical feature vectors representing the geometric dimensions and texture distribution of defects. Simultaneously, the system uses a pre-trained language model to encode the construction log text, converting material properties and construction steps in the text into numerical feature vectors of the same dimension. Subsequently, the existing system directly concatenates the image numerical feature vectors and the text numerical feature vectors along the vector dimension, combining them into a high-dimensional long vector containing information from all modalities. Finally, the system inputs this high-dimensional long vector into a time series prediction network or regression model, outputting pavement performance degradation trajectories and maintenance time node prediction results based on the numerical change patterns of historical data.

[0003] The aforementioned existing technical solutions, when processing multimodal data, directly concatenate image numerical features with text numerical features and input them into the prediction network, thus disrupting the original causal correspondence between different modal data. Since the concatenated values ​​cannot reflect the actual intervention of specific construction techniques on specific disease morphologies, when encountering the same apparent disease but with different material ratios or processes used in historical construction, the model cannot distinguish the differences in their structural evolution paths, leading to a deviation between the output maintenance cycle prediction and the actual structural attenuation state of the pavement. Summary of the Invention

[0004] The purpose of this invention is to provide a highway maintenance demand prediction system based on big data analysis, which can effectively solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: The highway maintenance demand prediction system based on big data analysis includes a visual perception terminal, a business system interface, a feature extraction processor, a fusion prediction processor, and a maintenance triggering server. The visual perception terminal is used to acquire a sequence of road surface appearance images; The business system interface is used to obtain historical maintenance construction log text; The feature extraction processor uses a convolutional neural network to extract features from the pavement appearance image sequence to generate appearance defect morphology feature vectors, and uses a pre-trained language model to extract features from historical maintenance construction log text to generate semantic feature vectors of material ratio and process procedures. The fusion prediction processor has a built-in cross-modal cross-attention module. It uses the apparent disease morphology feature vector as the query term and the semantic feature vector of material ratio and process as the key term to perform cross-attention calculation to generate a dynamic adjustment coefficient for the disease development rate. The fusion prediction processor also has a built-in pavement structure mechanical attenuation model. The dynamic adjustment coefficient is input into the pavement structure mechanical attenuation model and the mechanical equation is solved to output the remaining service life prediction curve. The maintenance trigger server compares the remaining service life prediction curve with the preset service life threshold and generates a maintenance demand trigger signal when the remaining service life is lower than the preset service life threshold.

[0006] Preferably, the convolutional neural network in the feature extraction processor includes a multi-scale feature extraction layer and a feature concatenation layer; The multi-scale feature extraction layer contains three parallel convolutional branches with progressively increasing kernel sizes. These three branches perform convolution operations on the road surface image sequence to output fine-grained texture feature maps, medium-scale morphological feature maps, and macroscopic topological structure feature maps, respectively. The feature splicing layer concatenates fine-grained texture feature maps, medium-scale morphological feature maps, and macroscopic topological structure feature maps in the channel dimension to generate a preliminary fused feature matrix. The feature extraction processor performs global average pooling on the preliminary fused feature matrix and maps it through a fully connected layer to generate an apparent lesion morphological feature vector.

[0007] Preferably, the pre-trained language model in the feature extraction processor includes a named entity recognition layer and a syntactic dependency parsing layer; The named entity recognition layer performs sequence annotation on historical maintenance construction log text to extract material attribute entity words and process action entity words; The syntactic dependency parsing layer constructs a syntactic tree from the historical maintenance construction log text and extracts the dependency relationship edges between material attribute entities and process action entities. The pre-trained language model constructs a heterogeneous graph structure using material attribute entities and process action entities as nodes and dependency relationship edges as connection paths. The feature extraction processor aggregates and updates the heterogeneous graph structure through a graph attention network to generate semantic feature vectors of material ratio and process.

[0008] Preferably, the cross-modal attention module includes a query mapping matrix, a key-value mapping matrix, and an activation function layer; The query mapping matrix performs a linear transformation on the apparent morphological feature vector of the disease to generate the target query matrix; The key-value mapping matrix performs a linear transformation on the semantic feature vectors of material proportions and process steps to generate the source key-value matrix; The cross-modal cross-attention module calculates the dot product of the target query matrix and the transpose of the source key matrix to obtain the cross-modal interaction similarity matrix. The activation function layer performs nonlinear mapping on the cross-modal interaction similarity matrix and calculates it through a normalized exponential function to output the attention weight distribution sequence corresponding to each process step. The fusion prediction processor multiplies the attention weight distribution sequence with the disease baseline development rate to generate the dynamic adjustment coefficient of the disease development rate.

[0009] Preferably, the pavement structure mechanical attenuation model includes a pavement three-dimensional finite element mesh construction component and a stress-strain solver; The road surface 3D finite element mesh construction component discretizes the highway structure into multiple eight-node hexahedral elements based on the road structure layer parameters and assigns material constitutive model parameters to each eight-node hexahedral element; The stress-strain solver receives the dynamic adjustment coefficient and superimposes it as the attenuation multiplier of the material constitutive model parameter into the current material modulus. Under the set axle load boundary conditions, the stress-strain solver performs iterative equilibrium calculations on the eight-node hexahedral element containing the attenuation multiplier to obtain the tensile stress distribution state at the bottom of the pavement structure layer. The fusion prediction processor calculates the cumulative fatigue damage variable based on the tensile stress distribution state at the bottom of the pavement structure layer and fits it to generate the remaining service life prediction curve.

[0010] Preferably, the maintenance trigger server is equipped with a sliding time window comparator and a status register; The sliding time window comparator retrieves multiple historical prediction values ​​of the remaining service life prediction curve before the current time node and within a preset window period. The status register determines whether all the historical prediction values ​​are lower than the preset service life threshold and whether the current prediction value of the remaining service life prediction curve at the current time node is lower than the preset service life threshold. When the result of the status register is yes, the maintenance triggering server generates a maintenance demand trigger signal and sends the maintenance demand trigger signal along with the timestamp of the current time node to the external maintenance scheduling terminal.

[0011] Preferably, all three convolutional branches in the multi-scale feature extraction layer contain cascaded dilated spatial pyramid pooling modules; The dilated spatial pyramid pooling module contains four parallel dilated convolutional layers with different dilation rates. Each of these four layers extracts multi-receptive field features from the feature maps input to its respective convolutional branch and outputs four sub-feature maps with different receptive fields. The dilated spatial pyramid pooling module then performs element-wise addition on the four sub-feature maps to output an enhanced fine-grained texture feature map, an enhanced medium-scale morphological feature map, and an enhanced macroscopic topological structure feature map. The feature concatenation layer receives the enhanced fine-grained texture feature map, the enhanced medium-scale morphological feature map, and the enhanced macroscopic topological structure feature map and performs a cascade operation along the channel dimension.

[0012] Preferably, the graph attention network includes an edge feature mapping layer and a node feature update layer; The edge feature mapping layer extracts relation type embedding vectors from the dependency edges in the heterogeneous graph structure and performs a linear transformation on the relation type embedding vectors to generate the edge weight coefficient matrix; The node feature update layer uses material attribute entity words and process action entity words as the initial node features. The node feature update layer calculates the attention coefficient between adjacent node features based on the edge weight coefficient matrix. The graph attention network performs a weighted summation and aggregation operation on the adjacent node features based on the attention coefficient and performs a residual connection between the aggregation result and the initial node features to generate updated heterogeneous graph node features. The feature extraction processor performs average pooling on the updated heterogeneous graph node features to generate semantic feature vectors of material ratio and process.

[0013] Preferably, the cross-modal cross-attention module also includes a text relevance mask generator; The text relevance mask generator reads the process segment identifiers in the historical maintenance construction log text and generates a Boolean mask matrix that matches the dimension of the source key value matrix based on the process segment identifiers. In the Boolean mask matrix, the positions corresponding to the process segments that correspond to the currently calculated disease type are set to true values ​​and the remaining positions are set to false values. Before performing nonlinear mapping at the activation function layer, the cross-modal cross-attention module performs an element-wise logical AND operation on the cross-modal interaction similarity matrix and the Boolean mask matrix to generate a filtered similarity matrix. The activation function layer performs nonlinear mapping on the filtered similarity matrix and calculates it using a normalized exponential function to output the filtered attention weight distribution sequence.

[0014] Preferably, the stress-strain solver includes a modulus update iteration step and a convergence judgment step; The modulus update iteration step reads the historical material modulus of the eight-node hexahedral element in the previous iteration step and multiplies the historical material modulus by the dynamic adjustment coefficient to obtain the updated material modulus of the current iteration step. The convergence judgment step recalculates the tensile stress at the bottom of the pavement structure layer based on the updated material modulus and obtains the current tensile stress calculation value of the current iteration step. The convergence judgment step calculates the absolute value of the difference between the current tensile stress calculation value and the historical tensile stress calculation value of the previous iteration step. When the absolute value of the difference is greater than the preset convergence tolerance, the modulus update iteration step is triggered to perform the next round of calculation. When the absolute value of the difference is less than or equal to the preset convergence tolerance, the final tensile stress distribution state is output to the fusion prediction processor.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This scheme generates apparent defect morphology feature vectors and semantic feature vectors for material ratios and construction processes using a feature extraction processor. A cross-modal attention module is incorporated into the fusion prediction processor. The apparent defect morphology feature vectors are used as query terms, and the semantic feature vectors as key terms to calculate dynamic adjustment coefficients. These dynamic adjustment coefficients are then input into the pavement structure mechanical attenuation model to solve for and output a remaining service life prediction curve. This approach establishes a directional correlation between defect morphology and construction process at the feature level, avoiding direct splicing of features from different modalities. This ensures that the dynamic adjustment coefficients input into the mechanical attenuation model reflect the suppression of specific defect morphology evolution by specific construction processes, eliminating interference from differences in construction quality. Consequently, the output remaining service life prediction curve corresponds to the actual performance degradation trajectory of the pavement structure itself.

[0016] 2. By setting three parallel convolutional branches and a feature splicing layer in the convolutional neural network to extract multi-scale features, and combining the multi-receptive field extraction and element-wise addition operation of the hollow space pyramid pooling module, the texture and topological structure information of defects of different sizes can be preserved. By constructing a heterogeneous graph structure through a named entity recognition layer and a syntactic dependency parsing layer, and combining it with a graph attention network to calculate attention coefficients based on the edge weight coefficient matrix for weighted summation and residual connection, the logical relationship between materials and processes can be extracted. By introducing a text relevance mask generator in the cross-attention calculation, a Boolean mask matrix is ​​generated based on paragraph identifiers to perform logical AND operation, filtering out interference from irrelevant process features. By setting a modulus update iteration step and a convergence judgment step in the stress-strain solver, and controlling the iteration process based on the comparison relationship between the absolute value of the tensile stress difference and the preset convergence tolerance, the numerical stability of the mechanical equation solution process is ensured, thereby improving the accuracy of the final generated maintenance demand trigger signal. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the overall execution process of the highway maintenance demand prediction system based on big data analysis of the present invention. Figure 2 This is a flowchart of the feature extraction process of the convolutional neural network in the feature extraction processor of the present invention; Figure 3 This is a flowchart of the feature extraction process of the pre-trained language model in the feature extraction processor of the present invention; Figure 4 This is a flowchart illustrating the computation of the cross-modal cross-attention module in the fusion prediction processor of the present invention. Figure 5 This is a flowchart of the solution process for the pavement structure mechanical attenuation model of the present invention; Figure 6 This is a flowchart illustrating the signal generation and judgment process of the maintenance trigger server of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0019] Please refer to Figure 1 This embodiment provides a highway maintenance demand prediction system based on big data analysis, including a visual perception terminal, a business system interface, a feature extraction processor, a fusion prediction processor, and a maintenance trigger server. The visual perception terminal is deployed on an adjustable mounting platform of a highway inspection vehicle. The attitude parameters of the adjustable mounting platform can be adjusted through a matching control unit. The visual perception terminal adopts a data acquisition architecture combining continuous scanning and fixed-frame shooting. It continuously scans the road surface along the travel direction of the inspection vehicle to generate a continuous panoramic image of the road surface along the road mileage. Simultaneously, it takes fixed-focus shots of the road surface at fixed time intervals to generate frame images containing local areas of the road surface. The panoramic images and frame images together form a sequence of road surface images. The visual perception terminal performs standardized preprocessing operations on the acquired road surface image sequence. First, it uses an intrinsic parameter matrix to perform distortion correction on the original image, eliminating geometric distortion caused by lens imaging. Then, it performs grayscale normalization on the image, mapping the image pixel values ​​to a fixed numerical range to eliminate differences in pixel value distribution under different lighting conditions. Finally, it performs region of interest cropping on the image, removing non-road surface areas and retaining image data of the effective road surface area, providing standardized input for subsequent feature extraction.

[0020] Table 1. Specifications for Preprocessing Road Surface Appearance Image Sequences The table above is used to standardize the input data for the road surface image sequence output by the visual perception terminal, ensuring that the image data collected under different inspection batches and environmental conditions have consistent input dimensions and data distribution, and avoiding the impact of fluctuations in input data on the stability of subsequent feature extraction.

[0021] The business system interface establishes an encrypted communication connection with the external maintenance business database through a standardized application programming interface (API). Data transmission is performed using Transmission Control Protocol (TCP) and Secure Sockets Layer (SSL). The business system interface sends data retrieval requests to the maintenance business database at preset time intervals. These requests include parameters such as the road number of the target road segment, mileage range, and data collection time range. Upon receiving the retrieval request, the maintenance business database returns the historical maintenance construction log text for the corresponding road segment and time range to the business system interface. The business system interface performs format cleaning on the received historical maintenance construction log text, removing garbled characters, special symbols, and blank paragraphs. It also uniformly replaces format control characters, converting the unstructured log text into standardized paragraph-style text data. Simultaneously, it adds metadata tags for each construction log text, including the corresponding road segment number, construction time, and construction mileage range, establishing a mapping relationship between the log text and the corresponding road segment. This ensures that the semantic features extracted subsequently can accurately match the apparent defects of the corresponding road segment.

[0022] The feature extraction processor receives road surface image sequences output from the visual perception terminal and historical maintenance construction log text output from the business system interface, and performs feature extraction processing on each. The feature extraction processor uses a convolutional neural network (CNN) to extract features from the road surface image sequences, generating surface defect morphology feature vectors. The CNN consists of an input layer, multiple convolutional layers, pooling layers, and fully connected layers. The input layer receives the preprocessed road surface image sequence, converting each frame into a fixed-dimensional tensor. The convolutional layers perform multiple convolution operations on the input image tensors, extracting edge, texture, shape, and size-related features of road defects. The pooling layers downsample the feature maps output from the convolutional layers, reducing the spatial dimension of the feature maps and retaining key feature information. After multiple convolutional and pooling processes, a fixed-dimensional feature map is generated. The feature extraction processor performs global average pooling on the generated feature map, converting the two-dimensional feature map into a one-dimensional feature vector. This vector is then linearly mapped by the fully connected layers to generate the surface defect morphology feature vector.

[0023] The mathematical expression for convolution is: in, To output the pixel value of the feature map at coordinates (x, y), The input feature map contains the pixel values ​​at the corresponding coordinates. The height of the convolution kernel. The width of the convolution kernel. For the convolution kernel in coordinates The weight value at that location, This is the bias term for the convolution operation.

[0024] The feature extraction processor uses a pre-trained language model to extract features from historical maintenance construction log text, generating semantic feature vectors for material ratios and technological processes. The pre-trained language model employs a Transformer-based architecture. First, it segments the input standardized construction log text into multiple sub-word units, adding positional encoding to each unit to mark its location within the text sequence. Then, the segmented sub-word units are converted into word embedding vectors and input into the encoder layer of the pre-trained language model. The encoder layer uses a multi-head self-attention mechanism to calculate the relationships between words in the text sequence, extracting semantic information related to material ratios and technological processes. After processing by multiple encoder layers, a semantic encoding matrix for the text sequence is generated. The feature extraction processor then performs mean pooling on the semantic encoding matrix to generate fixed-dimensional semantic feature vectors for material ratios and technological processes.

[0025] The fusion prediction processor incorporates a cross-modal cross-attention module and a pavement structure mechanical attenuation model. The cross-modal cross-attention module receives the apparent disease morphology feature vector and the material ratio and process semantic feature vector output by the feature extraction processor. It uses the apparent disease morphology feature vector as the query term and the material ratio and process semantic feature vector as the key term, performs cross-attention calculation, and generates a dynamic adjustment coefficient for the disease development rate.

[0026] The basic mathematical expression for cross-attention calculation is: Where Q is the query matrix, generated by linear transformation of the apparent disease morphology feature vector; K is the key matrix; and V is the value matrix, both generated by linear transformation of the semantic feature vectors of material ratio and process steps. Let be the dimension of the key matrix. This is a normalized exponential function, and the output is a weighted sum of the attention weights and the value matrix.

[0027] The attention weight distribution output by the cross-attention calculation reflects the degree of influence of different material ratios and process steps on the current apparent morphology of the disease. The feature extraction processor multiplies the attention weight distribution by a preset baseline disease development rate to generate a dynamic adjustment coefficient for the disease development rate. The corresponding mathematical expression is: in, This represents the dynamic adjustment coefficient for the disease development rate, where N is the number of process steps. The attention weight corresponding to the i-th process step. Let be the baseline rate of disease development corresponding to the i-th process step.

[0028] The pavement structure mechanical attenuation model receives a dynamic adjustment coefficient, which is then input into the model to solve the mechanical equations, outputting a remaining service life prediction curve. The pavement structure mechanical attenuation model is constructed based on the theory of elastic layered systems, dividing the pavement structure into multiple continuous structural layers. Each structural layer corresponds to a set elastic modulus, Poisson's ratio, and thickness parameter. The dynamic adjustment coefficient is used as an attenuation factor for the material's elastic modulus, superimposed on the elastic modulus of each structural layer to obtain the material modulus of the structural layer in the current state. The mathematical expression corresponding to the attenuation law of the material's elastic modulus over time is: in, Let be the elastic modulus of the structural layer material at time t. The initial elastic modulus of the structural layer material. t is the dynamic adjustment coefficient for the rate of disease development, and t is the service time of the pavement structure.

[0029] Under the given vehicle axle load boundary conditions, the pavement structure mechanical attenuation model solves the mechanical equilibrium equations of the elastic layered system, calculates the tensile stress value at the bottom of the pavement structure layer, and then calculates the cumulative fatigue damage of the pavement structure based on the tensile stress value at the bottom of the layer. The calculation of the cumulative fatigue damage is based on the linear fatigue cumulative damage theory, and the corresponding mathematical expression is: Where D is the cumulative fatigue damage variable of the pavement structure, and M is the number of axle load levels. The actual number of times the i-th level axle load is applied. The fatigue life of the pavement structure under the i-th level axle load is the maximum number of loads required for the tensile stress at the bottom of the structural layer to reach the fatigue limit.

[0030] When the cumulative fatigue damage variable D reaches the critical damage threshold, the pavement structure reaches its service life limit. The fusion prediction processor uses the calculation results of the cumulative fatigue damage variable at different time points to fit and generate a prediction curve for the remaining service life. The horizontal axis of the curve is the service time, and the vertical axis is the value of the remaining service life.

[0031] The maintenance triggering server receives the remaining service life prediction curve output by the fusion prediction processor, retrieves a preset service life threshold, and compares the value of the remaining service life prediction curve with the preset service life threshold. When the value of the remaining service life prediction curve is lower than the preset service life threshold, a maintenance demand trigger signal is generated. The preset service life threshold is set according to the highway's grade, design service life, and traffic flow parameters. The maintenance triggering server reads the predicted value of the remaining service life prediction curve at the current time node in real time and compares this value with the preset service life threshold for the corresponding road segment point by point. When the predicted value is less than the preset service life threshold, a maintenance demand trigger signal is generated. The trigger signal contains the road number, mileage range, current remaining service life value, and predicted maintenance time node information for the corresponding road segment, and can be directly sent to an external maintenance dispatch terminal.

[0032] In this embodiment, a sequence of pavement surface images is acquired through a visual perception terminal, and historical maintenance construction log text is obtained through a business system interface. A feature extraction processor generates morphological feature vectors of apparent defects and semantic feature vectors of material ratios and process procedures, respectively. A dynamic adjustment coefficient for defect development rate is generated through the cross-modal cross-attention module of the fusion prediction processor. The remaining service life prediction curve is obtained by solving the pavement structural mechanics attenuation model. Finally, a maintenance demand trigger signal is generated through a maintenance trigger server, realizing the prediction of highway maintenance demand. A directional correlation between defect morphology and construction process is established at the feature level, avoiding the destruction of causal relationship caused by direct splicing of different modal features, and ensuring the consistency between the prediction results and the actual pavement structural attenuation state.

[0033] In a preferred embodiment, reference Figure 2 The convolutional neural network in the feature extraction processor includes a multi-scale feature extraction layer and a feature concatenation layer. The multi-scale feature extraction layer receives a preprocessed sequence of road surface images and performs multi-scale feature extraction processing on the image sequence. The multi-scale feature extraction layer contains three parallel convolutional branches with progressively increasing kernel sizes, each corresponding to a different receptive field, used to extract different scale features of road surface defects from the images. The first convolutional branch uses a small kernel to extract fine-grained texture features from the road surface images, including edge textures of road cracks, pitting, and loose micro-texture information, outputting a fine-grained texture feature map. The second convolutional branch uses a medium-sized kernel to extract medium-scale morphological features from the road surface images, including the length and width of cracks, the outline size of potholes, and other geometric morphological information of defects, outputting a medium-scale morphological feature map. The third convolutional branch uses a large kernel to extract macroscopic topological features from the road surface images, including the overall distribution of ruts, the overall topological structure of crack networks, and the overall deformation information of road settlement, outputting a macroscopic topological feature map.

[0034] Each convolutional branch contains cascaded convolutional layers, batch normalization layers, and linear rectified activation function layers. The convolutional layers perform convolution operations on the input image tensors. The batch normalization layers perform batch normalization on the feature maps output by the convolutional layers, normalizing the numerical distribution of the feature maps to an interval with a mean of 0 and a variance of 1, eliminating distribution differences between different batches of data. The linear rectified activation function layers perform nonlinear mapping on the normalized feature maps, introducing nonlinear transformation capabilities and improving the network's ability to fit complex disease features.

[0035] Each convolutional branch is followed by a cascaded dilated spatial pyramid pooling module. This module contains four parallel dilated convolutional layers with different dilation rates. Without changing the kernel size, different dilation rates expand the receptive field of the convolutional layers while preserving the spatial resolution of the feature maps. The mathematical expression for dilated convolution is: in, The dilated convolution outputs the pixel value of the feature map at coordinates (x, y). Let r be the pixel value of the input feature map at the corresponding coordinates, and r be the dilation rate of the dilated convolution. , These are the height and width of the convolution kernel, respectively. These are the weights of the convolution kernel. This is a bias term.

[0036] Four parallel dilated convolutional layers with different dilation rates extract multi-receptive field features from the feature maps input to their respective convolutional branches, outputting four sub-feature maps with different receptive fields. A dilated spatial pyramid pooling module performs element-wise addition on these four sub-feature maps, fusing feature information from different receptive fields to output an enhanced feature map. The element-wise addition operation requires that the spatial dimension and channel dimension of the four sub-feature maps be completely consistent. The addition operation sums the pixel values ​​at the same coordinate positions, achieving the fusion of features from different receptive fields. This preserves the fine-grained details corresponding to a small dilation rate while incorporating the large-scale contextual features corresponding to a large dilation rate, improving the feature map's ability to express lesions of different sizes. The first convolutional branch outputs an enhanced fine-grained texture feature map, the second convolutional branch outputs an enhanced medium-scale morphological feature map, and the third convolutional branch outputs an enhanced macroscopic topological structure feature map.

[0037] Table 2. Parameter Configuration Table for Multi-Scale Convolutional Branching and Hollow Space Pyramid Pooling Module The table above is used to clarify the core parameters of the three parallel convolution branches and the dilated spatial pyramid pooling module, ensuring that different branches can extract disease features at the corresponding scale, while ensuring that the output feature maps have a consistent spatial dimension to meet the requirements of subsequent channel cascading operations.

[0038] The feature concatenation layer receives enhanced fine-grained texture feature maps, enhanced medium-scale morphological feature maps, and enhanced macroscopic topological structure feature maps from the outputs of three convolutional branches. It then performs a concatenation operation on these three feature maps along the channel dimension to generate a preliminary fused feature matrix. The mathematical expression for the feature concatenation operation is: in, To initially fuse the feature matrices, This is the enhanced fine-grained texture feature map. This is an enhanced medium-scale morphological feature map. This is the enhanced macroscopic topological feature map. This is a cascade operation function. This indicates that the cascading operation is performed at the channel level.

[0039] The feature extraction processor performs global average pooling on the preliminary fused feature matrix. Global average pooling averages the pixel values ​​at all spatial locations for each channel of the preliminary fused feature matrix, converting the two-dimensional feature map of each channel into a single numerical value, generating a one-dimensional pooled feature vector. The dimension of the pooled feature vector is the same as the number of channels in the preliminary fused feature matrix. The pooled feature vector is then input into a fully connected layer. The fully connected layer performs linear transformations and non-linear mappings on the pooled feature vector, mapping it to a preset fixed dimension to generate an apparent lesion morphology feature vector. The fully connected layer contains two linear transformation layers. The first linear transformation layer maps the pooled feature vector to a high-dimensional space, and the second linear transformation layer maps the high-dimensional feature vector to the preset fixed dimension. A Dropout layer is placed between the two linear transformation layers. The Dropout layer randomly discards a certain proportion of neurons during training to prevent overfitting and improve the generalization ability of the feature vector.

[0040] In this embodiment, three parallel convolutional branches with progressively increasing kernel sizes are set up to extract fine-grained texture features, medium-scale morphological features, and macroscopic topological features from the pavement appearance image, respectively. Combined with the cascaded hollow spatial pyramid pooling module in each convolutional branch, the extraction and fusion of multiple receptive field features are achieved, enhancing the expressive ability of the feature map for pavement defects of different sizes and types. Through the cascaded operation of the feature splicing layer in the channel dimension, the initial fusion of multi-scale features is achieved. The final generated appearance defect morphological feature vector can comprehensively and accurately reflect the morphology, size, and distribution information of pavement defects, providing accurate query input for subsequent cross-modal cross-attention calculation.

[0041] In another preferred embodiment, reference Figure 3 The pre-trained language model in the feature extraction processor includes a named entity recognition layer and a syntactic dependency parsing layer. The pre-trained language model receives standardized historical maintenance and construction log text output from the business system interface and performs semantic feature extraction on the text. The named entity recognition layer uses a sequence labeling model based on conditional random fields (CRFs) to perform sequence labeling on the input historical maintenance and construction log text, extracting material attribute entities and process action entities from the text. The named entity recognition layer first performs character-level encoding on the input text, converting each character into a corresponding character embedding vector, which is then input to a bidirectional long short-term memory (LSTM) network layer. The LSM network layer encodes the character embedding vectors, extracts the contextual features of the text sequence, and outputs a contextual feature vector corresponding to each character. Subsequently, the contextual feature vector is input to a CRF layer, which jointly decodes the label of each character, outputting the entity label corresponding to each character. Entity labels include start and intermediate labels for material attribute entities, start and intermediate labels for process action entities, and non-entity labels. Based on the output entity tags, continuous material attribute entity words and process action entity words are extracted from the text sequence. Material attribute entity words include various words related to maintenance materials, and process action entity words include various words related to maintenance construction processes.

[0042] The syntactic dependency parsing layer employs a transfer-based dependency parsing model to construct a syntactic tree from the historical maintenance construction log text, extracting dependency relationship edges between material attribute entities and process action entities. The layer first performs word segmentation and part-of-speech tagging on the input text, labeling each word with its corresponding part of speech. Then, based on a transfer system, it performs dependency parsing on the text sequence, identifying dependency relationships between words and constructing the corresponding syntactic tree structure. In the syntactic tree structure, each word corresponds to a tree node, and the dependency relationships between words correspond to connection edges between tree nodes. The syntactic dependency parsing layer extracts the dependency relationship edges between nodes corresponding to material attribute entities and nodes corresponding to process action entities from the syntactic tree structure, recording the type of the dependency relationship edge and the information of the two connected nodes.

[0043] The pre-trained language model uses extracted material attribute entities and process action entities as nodes, and extracted dependency relationship edges as connection paths between nodes to construct a heterogeneous graph structure. The heterogeneous graph structure contains two types of nodes: material attribute nodes and process action nodes. The initial features of each node are the word embedding vectors of the corresponding entity words, generated by the word embedding layer of the pre-trained language model. The heterogeneous graph structure also contains multiple types of edges, each corresponding to a different dependency relationship type. The features of each edge are the relationship type embedding vectors of the corresponding dependency relationship. Each material attribute node is only connected to process action nodes that have a dependency relationship with it. Connections between process action nodes can be established based on temporal dependency relationships in the text, reflecting the sequence of construction processes and ensuring that the heterogeneous graph structure can fully reflect the logical relationships between materials and processes, and between processes themselves, in the construction log text.

[0044] Table 3. Definition of Text Entity Types and Dependency Relationships in Historical Maintenance Construction Logs The table above is used to standardize the output of the named entity recognition layer and the syntactic dependency parsing layer, ensuring that the extracted entity words and dependency edges have a unified type definition, providing standardized input data for the construction of heterogeneous graph structures, and ensuring the stability of the graph attention network processing and the accuracy of feature extraction.

[0045] The feature extraction processor aggregates and updates the heterogeneous graph structure through a graph attention network, generating semantic feature vectors for material proportions and process steps. The graph attention network includes an edge feature mapping layer and a node feature update layer. The edge feature mapping layer extracts the corresponding relation type embedding vector for each dependency edge in the heterogeneous graph structure, performs a linear transformation on the relation type embedding vector, and generates an edge weight coefficient matrix. The corresponding mathematical expression is: in, This is the edge weight coefficient matrix. Embed the relation type vector corresponding to the dependency edge. It is a linear transformation function containing a trainable weight matrix and bias terms.

[0046] The node feature update layer uses the word embedding vectors corresponding to material attribute entity words and process action entity words as the initial features of the nodes. It calculates the attention coefficients between features of adjacent nodes based on the edge weight coefficient matrix. The corresponding mathematical expression is: in, Let be the attention coefficient between node i and node j. For attention weight vectors, Transform the weight matrix for node features. Let be the initial feature vector of node i. Let be the initial feature vector of node j. Let be the edge weight coefficient matrix between node i and node j. This is a vector concatenation operation. It is a linear rectified activation function with leakage.

[0047] The calculated attention coefficients are normalized, and the corresponding mathematical expression is: in, The normalized attention coefficient. Let i be the set of all neighboring nodes of node i. It is a natural exponential function.

[0048] The graph attention network uses normalized attention coefficients to perform a weighted summation and aggregation operation on the features of neighboring nodes, generating aggregated features for each node. These aggregated features are then residually connected to the initial node features to generate updated heterogeneous graph node features. The corresponding mathematical expression is: in, Let i be the updated feature vector of node i. This is a non-linear activation function, and the meanings of the other parameters are the same as those in the aforementioned formula.

[0049] After aggregation and update processing by a multi-layer graph attention network, all node features in the heterogeneous graph structure are updated. The updated node features integrate information from adjacent nodes and edge relationships, accurately reflecting the logical connections between material properties and process actions. The feature extraction processor performs average pooling on all node features of the updated heterogeneous graph, averaging the feature vectors of all nodes along their corresponding dimensions to generate semantic feature vectors of material proportions and process actions with fixed dimensions. Average pooling transforms the node-level features of the heterogeneous graph into graph-level global features, fully preserving the overall semantic information of material proportions and process actions in the construction log text, while eliminating feature dimension differences caused by text sequences of different lengths. This ensures that the generated semantic feature vectors have fixed dimensions, meeting the input requirements for subsequent cross-modal cross-attention calculations.

[0050] In this embodiment, material attribute entities and process action entities are extracted from historical maintenance construction log text through a named entity recognition layer. Dependency relationship edges between entities are extracted through a syntactic dependency parsing layer. A heterogeneous graph structure containing material nodes and process nodes is constructed. The node features of the heterogeneous graph structure are aggregated and updated through a graph attention network. The attention coefficient is calculated by combining the edge weight coefficient matrix. This achieves deep extraction of the association information between materials and processes. The resulting material ratio and process semantic feature vector can accurately reflect the logical relationship between materials and processes in the construction log text, providing accurate key-value input for subsequent cross-modal cross-attention calculation.

[0051] In yet another preferred embodiment, reference is made to Figures 4 to 6 The cross-modal attention module built into the fusion prediction processor includes a query mapping matrix, a key-value mapping matrix, and an activation function layer. This module receives the apparent morphological feature vectors and the semantic feature vectors of material ratios and process steps output by the feature extraction processor and performs cross-attention calculations. The query mapping matrix is ​​a trainable linear transformation weight matrix that performs a linear transformation on the apparent morphological feature vectors to generate the target query matrix Q. The target query matrix has a fixed dimension of a two-dimensional tensor, where the first dimension is the length of the query sequence and the second dimension is the dimension of the feature vectors. The key-value mapping matrix includes a key mapping matrix and a value mapping matrix, both of which are trainable linear transformation weight matrices. The key mapping matrix performs a linear transformation on the semantic feature vectors of material ratios and process steps to generate the source key matrix K. The value mapping matrix performs a linear transformation on the semantic feature vectors of material ratios and process steps to generate the source value matrix V. The dimensions of the source key matrix K and the source value matrix V are consistent to ensure that the dimensions of the target query matrix and the source key matrix match, satisfying the requirements of subsequent dot product operations.

[0052] The cross-modal cross-attention module calculates the dot product of the target query matrix and the transpose of the source key matrix to obtain the cross-modal interaction similarity matrix. The dot product result reflects the degree of similarity between the apparent morphological features of the disease and the semantic features of the material ratio and process steps. The higher the similarity, the larger the corresponding dot product value, indicating a higher correlation between the process step and the current disease morphology. The mathematical expression for calculating the cross-modal interaction similarity matrix is: Where S is the cross-modal interaction similarity matrix, Q is the target query matrix, K is the source key matrix, and d is the dimension of the key matrix. This is a scaling factor used to prevent the dot product result from becoming too large, which could cause the gradient of the subsequent normalized exponential function to vanish.

[0053] The cross-modal cross-attention module also includes a text relevance mask generator. This generator reads process segment identifiers from historical maintenance construction log text and generates a Boolean mask matrix that matches the dimensions of the source key matrix. When the business system interface preprocesses the historical maintenance construction log text, it adds a corresponding segment identifier to each process segment. This identifier includes information such as process type, applicable disease type, and construction time. The text relevance mask generator reads the applicable disease type information from the process segment identifier and matches it with the currently calculated apparent disease type to generate a Boolean mask matrix. The dimensions of the Boolean mask matrix are identical to those of the cross-modal interaction similarity matrix. Positions in the matrix corresponding to process segments that correspond to the currently calculated disease type are set to true values, while other positions are set to false values. True values ​​are represented by 1, and false values ​​by 0.

[0054] Before performing nonlinear mapping at the activation function layer, the cross-modal cross-attention module performs an element-wise logical AND operation on the cross-modal interaction similarity matrix and the Boolean mask matrix to generate a filtered similarity matrix. The rule for the element-wise logical AND operation is that the value at the corresponding position in the similarity matrix is ​​retained only if the value at the corresponding position in the mask matrix is ​​1; otherwise, the value at the corresponding position in the similarity matrix is ​​set to negative infinity. The value of negative infinity will be mapped to 0 in the subsequent normalized exponential function calculation, thereby filtering out irrelevant process features and eliminating interference from process features unrelated to the current disease type in the calculation of attention weights.

[0055] The activation function layer performs a nonlinear mapping on the filtered similarity matrix, followed by calculation using a normalized exponential function to output the filtered attention weight distribution sequence. The activation function layer employs a linear rectified activation function to nonlinearly transform the values ​​in the filtered similarity matrix, mapping values ​​less than 0 to 0 and retaining values ​​greater than 0, thus enhancing the model's nonlinear expressive power. The normalized exponential function normalizes the nonlinearly mapped similarity matrix, mapping the values ​​in the matrix to the [0,1] interval, with the sum of all values ​​being 1. This generates the attention weight distribution sequence corresponding to each process step. A larger attention weight indicates a higher degree of influence of the corresponding process step on the current disease morphology. The fusion prediction processor multiplies the attention weight distribution sequence by the baseline disease development rate to generate a dynamic adjustment coefficient for the disease development rate. The baseline development rate of pavement defects is a pre-calibrated value based on a large amount of pavement performance monitoring data. Different process procedures correspond to different baseline development rates of defects. By multiplying the attention weight distribution sequence with the baseline development rate of defects by weight, the generated dynamic adjustment coefficient can comprehensively reflect the actual impact of historical construction processes on the current defect development rate, thus realizing the directional correlation between defect morphology and construction process.

[0056] The fusion prediction processor's built-in pavement structure mechanical attenuation model includes a 3D finite element mesh construction component and a stress-strain solver. The 3D finite element mesh construction component discretizes the highway structure into multiple eight-node hexahedral elements based on pavement structure layer parameters. These layer parameters include the thickness, material type, and design parameters of the surface layer, base layer, subbase layer, and subgrade. The eight-node hexahedral elements are linear 3D solid elements, each containing 8 nodes, each node having 3 translational degrees of freedom, capable of simulating the 3D deformation state of the pavement structure. The 3D finite element mesh construction component assigns corresponding material constitutive model parameters to each eight-node hexahedral element. The material constitutive model adopts a linear elastic constitutive model, with parameters including elastic modulus, Poisson's ratio, and density. Elements corresponding to different structural layers are assigned the corresponding material constitutive model parameters for that structural layer, ensuring that the finite element model can reflect the actual mechanical properties of the pavement structure.

[0057] Table 4. Parameter Settings for Finite Element Calculation of Road Structure Mechanical Attenuation Model The table above is used to standardize the input parameters of the three-dimensional finite element mesh construction component and the stress-strain solver for road surfaces, ensuring that the finite element model can simulate the actual mechanical properties of the road surface structure, guaranteeing the accuracy of the mechanical equilibrium equation solution results, and providing reliable mechanical calculation data support for fitting the remaining service life prediction curve.

[0058] The stress-strain solver receives the dynamic adjustment coefficients output by the cross-modal cross-attention module. These coefficients are used as attenuation multipliers for the material constitutive model parameters and added to the current material modulus. Under the set axle load boundary conditions, iterative equilibrium calculations are performed on the eight-node hexahedral elements containing the attenuation multipliers to obtain the tensile stress distribution at the bottom of the pavement structure layer. The stress-strain solver includes a modulus update iteration step and a convergence judgment step. The modulus update iteration step reads the historical material modulus of the eight-node hexahedral elements from the previous iteration step, multiplies the historical material modulus by the dynamic adjustment coefficients, and obtains the updated material modulus for the current iteration step. The corresponding mathematical expression is: in, For the updated material modulus in the nth iteration step, The historical material modulus of the (n-1)th iteration step. This is the dynamic adjustment coefficient for the disease development rate.

[0059] The convergence judgment step updates the material modulus, reassembles the overall stiffness matrix of the finite element model, solves the mechanical equilibrium equations under the set axle load boundary conditions, calculates the tensile stress at the bottom of the pavement structure layer, and obtains the current tensile stress calculation value for the current iteration step. The axle load boundary conditions are set according to the standard axle load parameters in the highway design specifications. The axle load is applied to the corresponding element nodes on the surface of the pavement layer. At the same time, fixed displacement constraints are set on the bottom boundary of the model, and horizontal displacement constraints are set on the lateral boundary of the model to ensure that the boundary conditions conform to the actual stress state of the pavement structure.

[0060] The convergence judgment step calculates the absolute value of the difference between the current tensile stress calculation value and the historical tensile stress calculation value from the previous iteration step. This absolute value is compared to a preset convergence tolerance, which is set based on the required calculation accuracy. When the absolute value of the difference is greater than the preset convergence tolerance, a modulus update iteration step is triggered to perform the next round of calculation, updating the material modulus and resolving the mechanical equilibrium equations. When the absolute value of the difference is less than or equal to the preset convergence tolerance, the iterative calculation is considered converged, and the final tensile stress distribution state at the bottom of the pavement structure layer is output to the fusion prediction processor. The mathematical expression for convergence judgment is: in, This represents the current tensile stress calculated in the nth iteration step. The historical tensile stress calculation value for the (n-1)th iteration step. To preset the convergence tolerance, This is the absolute value operator.

[0061] The fusion prediction processor calculates the cumulative fatigue damage variable based on the tensile stress distribution at the bottom of the pavement structure layers, and then fits this variable to generate a remaining service life prediction curve. The calculation of the cumulative fatigue damage variable is based on linear fatigue cumulative damage theory, combined with the fatigue life equation of the pavement structure. The fatigue life equation is determined according to the formulas in the highway asphalt pavement design specifications. Based on the magnitude of the tensile stress at the bottom of the layers, the fatigue life of the pavement structure under the corresponding axle load is calculated. Combined with design traffic volume parameters, the cumulative fatigue damage variable at different time points is calculated. When the cumulative fatigue damage variable reaches the critical damage threshold, the pavement structure reaches its service life limit. The fusion prediction processor uses time as the x-axis and remaining service life as the y-axis, and performs polynomial fitting on the remaining service life calculation results at different time points to generate a continuous remaining service life prediction curve. This curve reflects the change in the remaining service life of the pavement structure over time, providing data support for determining maintenance needs.

[0062] The maintenance triggering server is equipped with a sliding time window comparator and a status register. It receives the remaining service life prediction curve output by the fusion prediction processor and uses the sliding time window comparator and status register to determine the triggering of maintenance needs. The sliding time window comparator retrieves multiple historical prediction values ​​from the remaining service life prediction curve before the current time point, with a duration equal to a preset window period. The preset window period is set according to the highway's grade and monitoring frequency, and the number of historical prediction values ​​matches the number of monitoring sessions within the preset window period.

[0063] The status register receives multiple historical predicted values ​​from the sliding time window comparator and the current predicted value of the remaining service life prediction curve at the current time point. It first determines whether all historical predicted values ​​are below a preset service life threshold, and then determines whether the current predicted value is also below the preset service life threshold. When both conditions are met, the status register outputs a valid trigger signal. Upon receiving the valid trigger signal, the maintenance trigger server generates a maintenance demand trigger signal and sends it along with the timestamp of the current time point, the road number of the corresponding road segment, the mileage interval, the remaining service life prediction value, and the defect type information to the external maintenance dispatch terminal. When either condition is not met, the status register outputs an invalid signal, and the maintenance trigger server does not generate a maintenance demand trigger signal, continuing to monitor the changes in the remaining service life prediction curve in real time. By setting the sliding time window comparator and the status register, false triggering caused by fluctuations in a single predicted value can be avoided. This ensures that a maintenance demand trigger signal is only generated when the remaining service life is consistently below the preset service life threshold within the preset window period, improving the accuracy and reliability of maintenance demand triggering.

[0064] In this embodiment, by setting a query mapping matrix, a key-value mapping matrix, and a text relevance mask generator in the cross-modal cross-attention module, a directional cross-association of apparent disease morphological features and construction process semantic features is achieved, filtering out interference from irrelevant process features. The generated dynamic adjustment coefficient can reflect the actual impact of historical construction processes on the disease development rate. Through the pavement three-dimensional finite element mesh construction component and stress-strain solver, combined with an iterative modulus update and convergence judgment mechanism, the pavement structural mechanical response is solved, ensuring the accuracy of the remaining service life prediction curve. Through the sliding time window comparator and status register in the maintenance trigger server, false triggering caused by fluctuations in predicted values ​​is avoided, improving the reliability of maintenance demand trigger signals, and ultimately realizing the prediction of highway maintenance needs.

Claims

1. A highway maintenance demand forecasting system based on big data analysis, characterized in that, This includes a visual perception terminal, a business system interface, a feature extraction processor, a fusion prediction processor, and a maintenance trigger server. The visual perception terminal is used to acquire a sequence of road surface appearance images; The business system interface is used to obtain historical maintenance construction log text; The feature extraction processor uses a convolutional neural network to extract features from the pavement appearance image sequence to generate appearance defect morphology feature vectors, and uses a pre-trained language model to extract features from historical maintenance construction log text to generate semantic feature vectors of material ratio and process procedures. The fusion prediction processor has a built-in cross-modal cross-attention module. It uses the apparent disease morphology feature vector as the query term and the semantic feature vector of material ratio and process as the key term to perform cross-attention calculation to generate a dynamic adjustment coefficient for the disease development rate. The fusion prediction processor also has a built-in pavement structure mechanical attenuation model. The dynamic adjustment coefficient is input into the pavement structure mechanical attenuation model and the mechanical equation is solved to output the remaining service life prediction curve. The maintenance trigger server compares the remaining service life prediction curve with the preset service life threshold and generates a maintenance demand trigger signal when the remaining service life is lower than the preset service life threshold.

2. The highway maintenance demand prediction system based on big data analysis according to claim 1, characterized in that, The convolutional neural network in the feature extraction processor includes a multi-scale feature extraction layer and a feature concatenation layer; The multi-scale feature extraction layer contains three parallel convolutional branches with progressively increasing kernel sizes. These three branches perform convolution operations on the road surface image sequence to output fine-grained texture feature maps, medium-scale morphological feature maps, and macroscopic topological structure feature maps, respectively. The feature splicing layer concatenates fine-grained texture feature maps, medium-scale morphological feature maps, and macroscopic topological structure feature maps in the channel dimension to generate a preliminary fused feature matrix. The feature extraction processor performs global average pooling on the preliminary fused feature matrix and maps it through a fully connected layer to generate an apparent lesion morphological feature vector.

3. The highway maintenance demand prediction system based on big data analysis according to claim 1, characterized in that, The pre-trained language model in the feature extraction processor includes a named entity recognition layer and a syntactic dependency parsing layer; The named entity recognition layer performs sequence annotation on historical maintenance construction log text to extract material attribute entity words and process action entity words; The syntactic dependency parsing layer constructs a syntactic tree from the historical maintenance construction log text and extracts the dependency relationship edges between material attribute entities and process action entities. The pre-trained language model constructs a heterogeneous graph structure using material attribute entities and process action entities as nodes and dependency relationship edges as connection paths. The feature extraction processor aggregates and updates the heterogeneous graph structure through a graph attention network to generate semantic feature vectors of material ratio and process.

4. The highway maintenance demand prediction system based on big data analysis according to claim 1, characterized in that, The cross-modal attention module includes a query mapping matrix, a key-value mapping matrix, and an activation function layer; The query mapping matrix performs a linear transformation on the apparent morphological feature vector of the disease to generate the target query matrix; The key-value mapping matrix performs a linear transformation on the semantic feature vectors of material proportions and process steps to generate the source key-value matrix; The cross-modal cross-attention module calculates the dot product of the target query matrix and the transpose of the source key matrix to obtain the cross-modal interaction similarity matrix. The activation function layer performs nonlinear mapping on the cross-modal interaction similarity matrix and calculates it through a normalized exponential function to output the attention weight distribution sequence corresponding to each process step. The fusion prediction processor multiplies the attention weight distribution sequence with the disease baseline development rate to generate the dynamic adjustment coefficient of the disease development rate.

5. The highway maintenance demand prediction system based on big data analysis according to claim 1, characterized in that, The pavement structure mechanical attenuation model includes a pavement three-dimensional finite element mesh construction component and a stress-strain solver; The road surface 3D finite element mesh construction component discretizes the highway structure into multiple eight-node hexahedral elements based on the road structure layer parameters and assigns material constitutive model parameters to each eight-node hexahedral element; The stress-strain solver receives the dynamic adjustment coefficient and superimposes it as the attenuation multiplier of the material constitutive model parameter into the current material modulus. Under the set axle load boundary conditions, the stress-strain solver performs iterative equilibrium calculations on the eight-node hexahedral element containing the attenuation multiplier to obtain the tensile stress distribution state at the bottom of the pavement structure layer. The fusion prediction processor calculates the cumulative fatigue damage variable based on the tensile stress distribution state at the bottom of the pavement structure layer and fits it to generate the remaining service life prediction curve.

6. The highway maintenance demand prediction system based on big data analysis according to claim 1, characterized in that, The maintenance trigger server is equipped with a sliding time window comparator and a status register. The sliding time window comparator retrieves multiple historical prediction values ​​of the remaining service life prediction curve before the current time node and within a preset window period. The status register determines whether all the historical prediction values ​​are lower than the preset service life threshold and whether the current prediction value of the remaining service life prediction curve at the current time node is lower than the preset service life threshold. When the result of the status register is yes, the maintenance triggering server generates a maintenance demand trigger signal and sends the maintenance demand trigger signal along with the timestamp of the current time node to the external maintenance scheduling terminal.

7. The highway maintenance demand prediction system based on big data analysis according to claim 2, characterized in that, The three convolutional branches in the multi-scale feature extraction layer all contain cascaded dilated spatial pyramid pooling modules; The dilated spatial pyramid pooling module contains four parallel dilated convolutional layers with different dilation rates. Each of these four layers extracts multi-receptive field features from the feature maps input to its respective convolutional branch and outputs four sub-feature maps with different receptive fields. The dilated spatial pyramid pooling module then performs element-wise addition on the four sub-feature maps to output an enhanced fine-grained texture feature map, an enhanced medium-scale morphological feature map, and an enhanced macroscopic topological structure feature map. The feature concatenation layer receives the enhanced fine-grained texture feature map, the enhanced medium-scale morphological feature map, and the enhanced macroscopic topological structure feature map and performs a cascade operation along the channel dimension.

8. The highway maintenance demand prediction system based on big data analysis according to claim 3, characterized in that, Graph attention networks consist of an edge feature mapping layer and a node feature update layer; The edge feature mapping layer extracts relation type embedding vectors from the dependency edges in the heterogeneous graph structure and performs a linear transformation on the relation type embedding vectors to generate the edge weight coefficient matrix; The node feature update layer uses material attribute entity words and process action entity words as the initial node features. The node feature update layer calculates the attention coefficient between adjacent node features based on the edge weight coefficient matrix. The graph attention network performs a weighted summation and aggregation operation on the adjacent node features based on the attention coefficient and performs a residual connection between the aggregation result and the initial node features to generate updated heterogeneous graph node features. The feature extraction processor performs average pooling on the updated heterogeneous graph node features to generate semantic feature vectors of material ratio and process.

9. The highway maintenance demand prediction system based on big data analysis according to claim 4, characterized in that, The cross-modal attention module also includes a text relevance mask generator; The text relevance mask generator reads the process segment identifiers in the historical maintenance construction log text and generates a Boolean mask matrix that matches the dimension of the source key value matrix based on the process segment identifiers. In the Boolean mask matrix, the positions corresponding to the process segments that correspond to the currently calculated disease type are set to true values ​​and the remaining positions are set to false values. Before performing nonlinear mapping at the activation function layer, the cross-modal cross-attention module performs an element-wise logical AND operation on the cross-modal interaction similarity matrix and the Boolean mask matrix to generate a filtered similarity matrix. The activation function layer performs nonlinear mapping on the filtered similarity matrix and calculates it using a normalized exponential function to output the filtered attention weight distribution sequence.

10. The highway maintenance demand prediction system based on big data analysis according to claim 5, characterized in that, The stress-strain solver includes a modulus update iteration step and a convergence judgment step. The modulus update iteration step reads the historical material modulus of the eight-node hexahedral element in the previous iteration step and multiplies the historical material modulus by the dynamic adjustment coefficient to obtain the updated material modulus of the current iteration step. The convergence judgment step recalculates the tensile stress at the bottom of the pavement structure layer based on the updated material modulus and obtains the current tensile stress calculation value of the current iteration step. The convergence judgment step calculates the absolute value of the difference between the current tensile stress calculation value and the historical tensile stress calculation value of the previous iteration step. When the absolute value of the difference is greater than the preset convergence tolerance, the modulus update iteration step is triggered to perform the next round of calculation. When the absolute value of the difference is less than or equal to the preset convergence tolerance, the final tensile stress distribution state is output to the fusion prediction processor.

Images