Highway intensive maintenance intelligent decision-making system based on multi-source collaborative awareness

Abstract

The invention discloses a highway intensive maintenance intelligent decision-making system based on multi-source collaborative awareness, and belongs to the technical field of highway maintenance management. The system comprises a multi-source heterogeneous sensing data acquisition module, a multi-level data fusion processing module, a disease development situation prediction module, an intensive operation scheme optimization module and a cross-department cooperative scheduling management module, and multi-source data is acquired through cooperative sensing of an unmanned aerial vehicle, a vehicle-mounted laser radar and an Internet of Things sensor; after adaptive Kalman filtering cleaning and space-time correlation fusion processing, the disease development trend is predicted based on a multi-factor coupling LSTM network, a one-time enclosing multi-process synchronization optimization operation scheme is generated based on a multi-target constraint genetic algorithm, and closed-loop feedback management is realized through a cross-department data interface. According to the method, the problems of multi-source data splitting, insufficient decision intelligence, low efficiency of cross-department collaboration and lack of disease prediction in the prior art are solved, and the maintenance efficiency and the intensification level of the high-flow highway can be effectively improved.

Description

Technical Field

[0001] This invention belongs to the field of highway maintenance and management technology, specifically involving a highway intensive maintenance intelligent decision-making system based on multi-source collaborative perception. This system integrates air-ground integrated monitoring technology, deep learning algorithms and cross-departmental collaborative mechanisms, and is suitable for daily maintenance and management scenarios of highways with high traffic volume and high density of operation. Background Technology

[0002] With the continuous growth of highway operating mileage and the surge in traffic flow, the demand for highway maintenance under high-load operation is becoming increasingly urgent. Taking a typical high-traffic section as an example, its daily cross-sectional flow can reach hundreds of thousands of vehicles, with a high proportion of heavy-duty vehicles, and frequent problems such as pavement defects and damage to ancillary facilities. The requirements for intensive maintenance place extremely high demands on the efficiency and coordination of maintenance operations.

[0003] Currently, there are several technical solutions in the field of highway maintenance and management. The first category is single-mode defect monitoring technology, such as using drones for road inspection. According to the research "Rapid Detection of Highway Pavement Defects Based on UAV Imagery and YOLOv8" published in *Highway Transportation Technology* by Wang et al. in 2024, this solution can achieve rapid inspection over a large area, but it suffers from insufficient defect location accuracy and the inability to link with real-time traffic data. The second category is vehicle-mounted detection equipment. According to research published in *Highway Transportation Technology*, Volume 39, Issue 5, 2022, vehicle-mounted lidar has high detection accuracy in highway pavement defect detection, but its inspection range is limited, making it difficult to adapt to the needs of routine monitoring across the entire road section. The third category is intelligent maintenance decision-making platforms. These platforms can achieve information management of maintenance records, but lack intelligent operation planning functions based on real-time operating conditions.

[0004] In recent years, the international academic community has made significant progress in the field of pavement distress detection and maintenance decision-making. According to a 2024 study published in *Scientific Reports*, researchers proposed the SMG-YOLOv8 model for multi-scenario asphalt pavement distress identification. By integrating spatial-to-depth modules and multi-scale convolutional attention mechanisms, it achieved an accuracy improvement of 8.2% to 12.5% ​​compared to baseline models. According to a 2024 study published in *Advanced Engineering Informatics*, researchers established a multi-output LSTM pavement performance prediction model based on the Guizhou Province expressway network. Using a Bayesian optimization algorithm to optimize hyperparameters, the average coefficient of determination for the four prediction indicators reached 0.823. According to a 2024 study published in *Structural Safety*, researchers proposed a three-objective optimization framework based on a bidirectional LSTM deep neural network model and a genetic algorithm for pavement maintenance and repair decision-making.

[0005] However, existing technologies still have the following shortcomings: First, multi-source monitoring data are fragmented, with data from drone inspections, vehicle-mounted inspections, and traffic flow monitoring not being effectively integrated, resulting in a lack of comprehensiveness in disease assessment; Second, the level of intelligence in maintenance decision-making is insufficient, with work plans relying on experience-based judgment, making it difficult to achieve the intensive goal of simultaneous multi-process control in a single enclosure; Third, the cross-departmental collaboration mechanism is imperfect, with data interfaces between maintenance, traffic police, and operation support units not being connected, leading to significant delays in collaborative response; Fourth, the ability to predict disease development trends is lacking, as existing technologies can only achieve static identification of diseases and cannot support the formulation of preventative maintenance decisions. Summary of the Invention

[0006] To address the technical problems of fragmented multi-source data, insufficient intelligent decision-making, inefficient cross-departmental collaboration, and lack of disease prediction in existing technologies, this invention provides an intelligent decision-making system for intensive highway maintenance based on multi-source collaborative perception.

[0007] The technical problems to be solved by this invention include: First, constructing a multi-source collaborative sensing data acquisition network to integrate data from UAV inspections, vehicle-mounted LiDAR detection, and IoT sensor monitoring, thereby improving the comprehensiveness and positioning accuracy of disease identification; Second, developing an intelligent matching algorithm based on historical maintenance data and real-time operating conditions to automatically generate a multi-process simultaneous operation plan for a single enclosure, thereby improving the intelligence level and efficiency of maintenance decision-making; Third, building a cross-departmental collaborative platform to connect the data interfaces of maintenance, traffic police, and operation support units, achieving closed-loop management and shortening the collaborative response delay; Fourth, integrating multi-dimensional data such as traffic flow and meteorology to construct a disease development trend prediction model, providing a scientific basis for the formulation of preventive maintenance plans.

[0008] The technical solution adopted in this invention is a highway intensive maintenance intelligent decision-making system based on multi-source collaborative perception, which includes a multi-source heterogeneous perception data acquisition module, a multi-level data fusion processing module, a disease development trend prediction module, an intensive operation plan optimization module, and a cross-departmental collaborative scheduling and management module.

[0009] The multi-source heterogeneous sensing data acquisition module is used to acquire visible light image data and infrared thermal image data of the road surface through the UAV inspection unit, acquire three-dimensional point cloud data and GPS positioning data of the road surface through the vehicle-mounted lidar unit, and acquire road surface temperature data, humidity data and traffic flow data through the Internet of Things sensor unit, and send the acquired data to the multi-level data fusion processing module.

[0010] The multi-level data fusion processing module receives data sent by the multi-source heterogeneous sensing data acquisition module. It performs noise filtering and outlier removal on the received data through an adaptive Kalman filter data cleaning unit. It extracts disease feature vectors from the cleaned image data and point cloud data through a multi-scale feature extraction unit. It then uses a spatiotemporal correlation fusion unit to perform spatiotemporal correlation binding between the disease feature vectors and the corresponding environmental data based on the positioning data, generates a fused disease information package, and sends it to the disease development trend prediction module.

[0011] The disease development trend prediction module receives and integrates disease information packages, historical maintenance data, traffic flow data, and meteorological data. It encodes the historical maintenance data using a time-series feature encoder and uses a multi-factor coupled prediction network to predict the development level and expansion range of the disease within a preset prediction time domain based on multi-dimensional input features. The module then generates disease development prediction results and sends them to the intensive operation scheme optimization module.

[0012] The intensive operation plan optimization module receives the disease development prediction results, the preset enclosure time window and construction resource constraint data. Based on the multi-objective constraint genetic optimization algorithm, it jointly optimizes the construction sequence of multiple processes, equipment scheduling paths and personnel configuration within the preset enclosure time window, generates an optimized operation plan for simultaneous multi-process enclosure, and sends it to the cross-departmental collaborative scheduling management module.

[0013] The cross-departmental collaborative scheduling and management module is used to receive optimized work plans, push optimized work plans to the traffic police system and construction teams through cross-departmental data interfaces, receive traffic control response data and construction progress feedback data, calculate closed-loop feedback parameters based on the deviation between construction progress feedback data and optimized work plans, send closed-loop feedback parameters to the disease development trend prediction module to update prediction model parameters, and send them to the intensive work plan optimization module to dynamically adjust optimization constraints.

[0014] Preferably, the drone inspection unit is equipped with a high-definition camera with a resolution of no less than 1920×1080 and an infrared thermal imager with a temperature measurement accuracy of no less than ±0.5℃. Preferably, the vehicle-mounted LiDAR unit integrates a LiDAR with a ranging accuracy of no less than ±2cm and a GPS positioning module with a positioning accuracy of no less than ±1m. Preferably, the IoT sensor unit includes a sensor array deployed every 300 to 500 meters along the highway.

[0015] Preferably, the adaptive Kalman filter data cleaning unit adaptively adjusts the noise covariance matrix based on the residual between sensor measurements and predicted values. Preferably, the multi-scale feature extraction unit uses a spatial pyramid pooling structure to extract multi-scale features. Preferably, the spatiotemporal correlation fusion unit uses an attention-weighted mechanism for feature-level fusion.

[0016] Preferably, the temporal feature encoder employs a bidirectional long short-term memory network structure. Preferably, the multi-factor coupled prediction network includes a feature embedding layer, a multi-head self-attention layer, and a prediction output layer. Preferably, the preset prediction time domain length is 1 month to 6 months.

[0017] Preferably, the multi-objective constrained genetic optimization algorithm aims to minimize the total containment time, minimize the equipment idle distance, and maximize the parallelism of processes. Preferably, the multi-objective constrained genetic optimization algorithm employs an elite retention strategy and adaptive crossover mutation probability.

[0018] Preferably, the cross-departmental collaborative scheduling and management module also includes an electronic fence early warning unit and a progress deviation monitoring unit.

[0019] The beneficial effects of this invention include: First, multi-source data fusion improves the accuracy of disease identification and positioning. Through the collaborative perception of drones, vehicle-mounted LiDAR, and IoT sensors, combined with multi-level data fusion processing, the disease identification coverage can reach over 95%, and the positioning accuracy can be improved to ±1m. Second, intelligent decision-making improves maintenance efficiency and intensification. Based on deep learning prediction models and genetic algorithm optimization models, a multi-process synchronous plan for one-time enclosure can be automatically generated, reducing the number of enclosure operations per road section by more than 50% and improving the efficiency of multi-process collaborative operations by more than 30%. Third, cross-departmental collaboration enables closed-loop management. By connecting data interfaces of multiple departments, the collaborative response latency can be shortened to within 1 hour. Fourth, disease prediction supports preventive maintenance. The prediction model can predict the development trend of diseases in advance, with a prediction accuracy of over 85%, which can effectively reduce maintenance costs. Attached Figure Description

[0020] Figure 1 This is the overall architecture diagram of the intelligent decision-making system for intensive highway maintenance based on multi-source collaborative perception, which is based on the present invention.

[0021] Figure 2 This is a structural diagram of the multi-source heterogeneous sensing data acquisition module of the present invention.

[0022] Figure 3 This is a structural diagram of the multi-level data fusion processing module of the present invention.

[0023] Figure 4 This is a structural diagram of the disease development trend prediction module of the present invention.

[0024] Figure 5 This is a structural diagram of the intensive operation scheme optimization module of the present invention. Detailed Implementation

[0025] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. The following embodiments use high-traffic highway maintenance as an application scenario to fully describe the technical solution of the present invention.

[0026] like Figure 1 As shown, the intelligent decision-making system for intensive highway maintenance based on multi-source collaborative sensing provided by this invention adopts a layered architecture design, including a multi-source heterogeneous sensing data acquisition module 1, a multi-level data fusion processing module 2, a disease development trend prediction module 3, an intensive operation plan optimization module 4, and a cross-departmental collaborative scheduling and management module 5. These five modules form a deeply coupled closed-loop collaborative system. The output of the multi-source heterogeneous sensing data acquisition module 1 serves as the input of the multi-level data fusion processing module 2, the output of the multi-level data fusion processing module 2 serves as the input of the disease development trend prediction module 3, the output of the disease development trend prediction module 3 serves as the input of the intensive operation plan optimization module 4, and the output of the intensive operation plan optimization module 4 serves as the input of the cross-departmental collaborative scheduling and management module 5. Simultaneously, the cross-departmental collaborative scheduling and management module 5 transmits the closed-loop feedback parameters back to the disease development trend prediction module 3 and the intensive operation plan optimization module 4, realizing dynamic adjustment and continuous optimization of system parameters.

[0027] Data flow between modules follows these principles: the output data of the preceding module is transmitted to the following module via a standardized interface; the following module then executes corresponding processing logic based on the received data and generates new output data. This closed-loop feedback mechanism enables the system to dynamically adjust prediction model parameters and optimize constraints based on actual execution conditions, thereby improving the system's adaptability and decision-making accuracy.

[0028] like Figure 2 As shown, the multi-source heterogeneous sensing data acquisition module 1 includes a drone inspection unit 11, a vehicle-mounted lidar unit 12, and an Internet of Things sensor unit 13. These three units work together to complete the multi-source sensing data acquisition task required for highway maintenance.

[0029] The design of the UAV inspection unit 11 fully considers the coverage and detection accuracy requirements of highway inspection. In one embodiment of the invention, this unit uses an industrial-grade UAV equipped with a high-definition camera and an infrared thermal imager. The high-definition camera has a resolution of 1920×1080 pixels and a frame rate of 30fps, which can clearly capture image features of visible defects such as road surface cracks and potholes. The infrared thermal imager has a temperature measurement range of -20℃ to 150℃, a temperature measurement accuracy of ±0.5℃, and a resolution of 640×480 pixels, which can detect differences in road surface temperature distribution and assist in identifying hidden defects such as road surface voids and water accumulation. The UAV's inspection route is set to fully cover the main highway and interchanges, with the flight altitude controlled within the range of 50 meters to 80 meters and the flight speed set between 30 km / h and 50 km / h. The selection of these parameters is based on the following: a flight altitude of 50 to 80 meters ensures a large ground coverage area while maintaining image resolution, with each image covering a road surface area of ​​approximately 60 meters × 40 meters; a flight speed of 30 km / h to 50 km / h improves inspection efficiency while maintaining image clarity, allowing a single inspection to cover a road segment of 50 km to 100 km. The drone is equipped with a 4G / 5G communication module, enabling real-time uploading of collected image data to the multi-level data fusion processing module 2, with data transmission latency controlled within 200 ms.

[0030] The purpose of the vehicle-mounted LiDAR unit 12 is to accurately verify and detect suspected defect areas discovered by drone inspections. In one embodiment of the invention, this unit integrates a LiDAR, a GPS positioning module, and a high-definition camera on a maintenance inspection vehicle. The LiDAR has a ranging range of 0.1 meters to 100 meters, a ranging accuracy of ±2 cm, an angular resolution of 0.2°, and a scanning frequency of 10 Hz to 20 Hz, enabling it to acquire high-precision three-dimensional point cloud data of the road surface. The GPS positioning module supports BeiDou + GPS dual-mode positioning with a positioning accuracy of ±1 m, providing accurate spatial location reference for the point cloud data. The high-definition camera is configured identically to the drone inspection unit 11, used to acquire visible light images synchronized with the point cloud data. The inspection vehicle travels at a low speed along the road, with the speed controlled between 5 km / h and 10 km / h, focusing on scanning suspected defect areas. The selection of these parameters is based on the following: a ranging accuracy of ±2cm allows for accurate measurement of geometric parameters such as depth and width of the damage; a vehicle speed of 5km / h to 10km / h ensures the density and integrity of the point cloud data, with 500 to 1000 point cloud data points per square meter of road surface. The vehicle-mounted LiDAR unit 12 is also equipped with a 4G / 5G communication module, which uploads the collected point cloud data, GPS positioning data, and image data to the multi-level data fusion processing module 2.

[0031] The IoT sensor unit 13 is designed to achieve real-time monitoring of environmental parameters and traffic flow along highways. In one embodiment of the invention, the unit includes sensor arrays deployed every 300 to 500 meters along the highway. Each array includes a road surface temperature sensor, a humidity sensor, and a traffic flow detector. The road surface temperature sensor has a measurement range of -40°C to 85°C and a measurement accuracy of ±0.5°C, enabling real-time monitoring of road surface temperature changes. The humidity sensor has a measurement range of 0 to 100%RH and a measurement accuracy of ±3%RH, enabling monitoring of road surface humidity. The traffic flow detector uses geomagnetic induction or video detection technology, with a detection accuracy of no less than 95%, and can statistically analyze traffic flow, vehicle speed, and vehicle type distribution data. In critical road sections such as bridges and tunnels, vibration sensors are additionally configured in the sensor array, with a measurement range of 0 to 20g and a measurement accuracy of ±0.001g, used to monitor structural stability. All sensors use a power supply mode combining solar power and lithium battery backup, and upload data via a 4G module at a data upload frequency of 1 to 5 times per minute. The selection of a sensor deployment interval of 300 to 500 meters is based on the fact that this interval can control deployment costs while ensuring monitoring coverage, and that 2 to 3 sets of sensor arrays can be deployed per kilometer of road section to achieve effective monitoring of environmental parameters.

[0032] The three units of the multi-source heterogeneous sensing data acquisition module 1 each focus on different data types, forming a complementary relationship. The UAV inspection unit 11 focuses on large-scale rapid inspection and visible defect identification; the vehicle-mounted LiDAR unit 12 focuses on the precise measurement of defect geometric parameters; and the IoT sensor unit 13 focuses on real-time monitoring of environmental parameters and traffic flow. The fusion of these three types of data can provide comprehensive data support for subsequent defect analysis and maintenance decisions.

[0033] like Figure 3 As shown, the multi-level data fusion processing module 2 includes an adaptive Kalman filter data cleaning unit 21, a multi-scale feature extraction unit 22, and a spatiotemporal correlation fusion unit 23. These three units sequentially perform data cleaning, feature extraction, and data fusion tasks according to the data processing flow.

[0034] The adaptive Kalman filter data cleaning unit 21 is designed to perform noise filtering and outlier removal on the raw data uploaded by the multi-source heterogeneous sensing data acquisition module 1, thereby improving data quality. Traditional Kalman filtering assumes that the covariance matrix of process noise and observation noise is a known constant. However, in actual highway maintenance scenarios, sensor noise characteristics change with environmental conditions. Therefore, this invention proposes an adaptive Kalman filter algorithm, the core innovation of which lies in adaptively adjusting the noise covariance matrix based on measurement residuals.

[0035] The specific implementation of the adaptive Kalman filter data cleaning algorithm is as follows. Let the system state equation and observation equation be: , , in, For the first The system state vector at time t is a vector containing sensor measurements and their rates of change. This is the state transition matrix; The process noise has the following covariance matrix: ; For the first The observation vector at time; The observation matrix; The covariance matrix of the observation noise is: .

[0036] The adaptive mechanism of this invention dynamically estimates the noise covariance matrix based on the statistical properties of the measurement residual sequence. The measurement residual is defined as: , in, This is a one-step prediction based on the estimate from the previous time step. The observation noise covariance matrix is ​​estimated using the sliding window method: , in, The length of the sliding window is set to 20 in this embodiment; This is the one-step prediction error covariance matrix. Simultaneously, based on the innovation sequence, the noise covariance matrix of the estimation process is: , in, Kalman gain; The new information covariance matrix; Let be the posterior error covariance matrix. Through the aforementioned adaptive mechanism, the algorithm can dynamically adjust the filtering parameters according to the actual noise characteristics of the sensor, thereby improving its adaptability to environmental changes while ensuring filtering accuracy.

[0037] In terms of outlier removal, this invention employs an outlier detection method based on Mahalanobis distance. The Mahalanobis distance of each observation relative to the predicted value is calculated as follows: , When the Mahalanobis distance exceeds a preset anomaly detection threshold, the observation is determined to be an anomaly and discarded. In this embodiment, the anomaly detection threshold is set to 3.0, corresponding to a 99.7% confidence interval.

[0038] The purpose of the multi-scale feature extraction unit 22 is to extract feature vectors of road defects from cleaned visible light image data and 3D point cloud data. Considering the multi-scale characteristics of road defects, this invention adopts a spatial pyramid pooling structure to achieve multi-scale feature extraction.

[0039] The specific implementation of the multi-scale spatial pyramid feature extraction algorithm for visible light image data is as follows. First, a convolutional neural network is used to extract the primary feature map of the image. Let the input image be... The feature map is obtained after multiple convolution operations. ,in and The height and width of the feature map, Let be the number of channels. Then, multi-scale spatial pyramid pooling is performed on the feature map. Let the number of pyramid layers be . , No. The pooling grid size of the layer is Then the first The pooling output of the layer is: , in, Indicates the first Layer A set of pixels within a pooling region; This represents the number of pixels within the region. In this embodiment, the pyramid layer count is set to 4, and the pooling mesh sizes are respectively... , , and Finally, the pooling results from each layer are flattened and concatenated to obtain multi-scale image feature vectors. ,in .

[0040] For 3D point cloud data, a voxelization method is used to extract geometric features. First, the point cloud space is divided into a regular voxel grid, with each voxel size set to 5cm × 5cm × 2cm. Then, for each non-empty voxel, the statistical features of its internal point cloud are calculated, including the number of points, center coordinates, and eigenvalues ​​of the covariance matrix. Finally, a 3D convolutional neural network is used to extract features from the voxelized point cloud data, obtaining the point cloud feature vector. .

[0041] The disease feature vector output by the multi-scale feature extraction unit 22 is a concatenation of the image feature vector and the point cloud feature vector. In this embodiment, , The total feature dimension is 768.

[0042] The purpose of the spatiotemporal correlation fusion unit 23 is to bind the disease feature vector with the corresponding environmental data in a spatiotemporal correlation based on GPS positioning data to generate a fused disease information package. This invention proposes a spatiotemporal correlation attention fusion algorithm, the core innovation of which lies in using an attention mechanism to adaptively determine the fusion weights of each data source.

[0043] The specific implementation of the spatiotemporal correlation attention fusion algorithm is as follows. Let the disease feature vector be... The road surface temperature data is Humidity data is Traffic flow data is First, the environmental data is mapped to an embedding vector with the same dimension as the disease feature vector: , , , in, For learnable embedding weight vectors; This is the bias vector. Then, the attention weights for each data source are calculated: , in, To query the weight vector. , and The calculation method is similar. Finally, weighted fusion is performed based on attention weights: , Fusion of disease feature vectors Together with GPS positioning data, timestamps, and other metadata, the data constitutes a fused disease information package, which is sent to the disease development trend prediction module 3. The data structure of the fused disease information package includes: a unique disease identifier, a fused feature vector, GPS coordinates, detection timestamps, preliminary disease classification labels, and confidence levels.

[0044] like Figure 4 As shown, the disease development trend prediction module 3 includes a time-series feature encoder 31 and a multi-factor coupled prediction network 32, which are used to predict the development trend of diseases based on historical maintenance data and multi-dimensional input features.

[0045] The temporal feature encoder 31 is designed to encode the temporal features of disease evolution sequences in historical maintenance data. Considering the temporal dependence and bidirectional correlation of disease evolution, this invention adopts a bidirectional long short-term memory (Bi-LSTM) network structure. In one embodiment of this invention, the temporal feature encoder 31 contains two layers of Bi-LSTM, each layer has 256 hidden units, and the dropout ratio is 0.2. Let the evolution sequence of a certain disease in the historical maintenance data be... ,in For the first The feature vector of the disease state at time step [time]. The hidden state update of the forward LSTM is: , The hidden state of the inverse LSTM is updated as follows: , The final temporal coding feature is the concatenation of the forward and backward hidden states: , The output of the temporal feature encoder 31 is the temporal encoded feature at the end of the sequence. .

[0046] The purpose of the multi-factor coupled prediction network 32 is to integrate multi-factor data such as disease characteristics, traffic flow, and meteorology to predict the development level and expansion range of diseases. This invention proposes a multi-factor coupled LSTM disease prediction algorithm, the core innovation of which lies in using a multi-head self-attention mechanism to achieve deep interaction of multi-factor features.

[0047] The multi-factor coupled prediction network 32 includes a feature embedding layer, a multi-head self-attention layer, and a prediction output layer. The feature embedding layer maps traffic flow data and meteorological data into embedding vectors of the same dimension as the disease feature vectors. Let the feature vector in the fused disease information package be... Traffic flow data sequence is The meteorological data series is ,in The length of the historical data. The traffic flow embedding vector is calculated as follows: , The calculation of the meteorological data embedding vector is as follows: , in, For embedding weight matrix; This is the bias vector.

[0048] Multi-head self-attention layers are used to capture the interactions between multi-factor features. Temporally encoded features are then used. , Integration of disease characteristics Traffic flow embedding vector and meteorological data embedding vector Concatenate into the input sequence The calculation of multi-head self-attention is as follows: , , , in, For query, key, and value projection matrices; This is the scaling factor; The number of attention heads is set to 8 in this embodiment; This is for outputting the projection matrix.

[0049] The prediction output layer maps the output of the multi-head self-attention layer to the disease development prediction result. The prediction output includes two parts: the probability distribution of disease development level and the estimate of disease expansion range. The disease development level is divided into 5 levels, corresponding to stable disease status, slight deterioration, moderate deterioration, severe deterioration, and urgent treatment. The level probability distribution is calculated as follows: , in, This is the output of the multi-head self-attention layer; This is the classification weight matrix; It is the bias vector; This represents the probability distribution for each level.

[0050] The extent of disease expansion is estimated using a regression method. Let the current disease area be... The predicted expansion rate is Then predict the time domain The area of ​​the diseased tissue after treatment is: , Expansion rate The calculation is as follows: , in, This is the regression weight vector; It is a bias scalar. The sigmoid function limits the spread rate to between 0 and 1.

[0051] The training of module 3 for predicting disease development trends employs a multi-task learning strategy, with the total loss function being a weighted sum of classification and regression losses: , in, Cross-entropy is used for classification loss; The mean squared error regression loss; and The loss weights are set to 0.6 and 0.4 respectively in this embodiment. The model is trained using the Adam optimizer, with an initial learning rate of 0.001, a batch size of 32, and 100 training epochs.

[0052] like Figure 5 As shown, the intensive operation scheme optimization module 4 is used to generate an optimized operation scheme for simultaneous multi-process enclosure based on the disease development prediction results, the preset enclosure time window, and construction resource constraint data. This invention proposes a multi-objective constrained genetic optimization algorithm, whose core innovation lies in simultaneously optimizing three objectives: total enclosure time, equipment idle distance, and process parallelism.

[0053] The specific implementation of the multi-objective constrained genetic optimization algorithm is as follows. First, define the decision variables, objective function, and constraints of the optimization problem. Let the set of diseases to be processed be... Each disease The set of procedures to be performed is The available construction equipment is as follows Decision variables include: process start time. Process execution equipment Assignment of construction workers .

[0054] The three optimization objectives are defined as follows. The first objective is to minimize the total containment time: , in, For process The execution time. The second objective is to minimize the device's idle distance: , in, For equipment The sequence of processes performed; For the first in the sequence The location of each process step; Let be the distance function between the two points. The third objective is to maximize the parallelism of the processes: , This indicator represents the average utilization rate of the equipment; a higher value indicates a higher degree of parallelism in the processes.

[0055] The constraints include: a preset enclosure time window constraint, i.e. and ,in and The available construction time window boundaries reported by the traffic police department; construction safety distance constraints, that is, a safe distance must be maintained between different work processes on the same road section. Equipment operation capacity constraints, meaning that each piece of equipment can only perform one process at a time; personnel skill matching constraints, meaning that personnel performing specific processes must possess the corresponding skill qualifications.

[0056] The genetic algorithm's encoding scheme employs a multi-chromosome structure, with each chromosome containing a process scheduling chromosome, an equipment allocation chromosome, and a personnel allocation chromosome. The process scheduling chromosome uses priority-based encoding, with each gene representing the execution priority of the corresponding process; the equipment allocation chromosome and the personnel allocation chromosome use integer encoding, with each gene representing the equipment number or personnel number assigned to the corresponding process.

[0057] The genetic algorithm uses a tournament selection method with a tournament size of 3. Different strategies are employed for crossover on different chromosomes: partial mapping crossover (PMX) is used for work scheduling chromosomes, while uniform crossover is used for equipment allocation and personnel allocation chromosomes. The mutation operation uses an adaptive mutation probability strategy. , in, The initial mutation probability is set to 0.05 in this embodiment; The average fitness of the population; This represents the optimal fitness for the population. This strategy automatically increases the mutation probability when population diversity decreases, preventing premature convergence.

[0058] The genetic algorithm employs an elite retention strategy, preserving the top 10% of individuals with the best fitness in each generation. Multi-objective optimization uses a non-dominated sorting method (NSGA-II), ranking and selecting individuals based on dominance and crowding distance. The population size is set to 100, the number of iterations to 200 generations, and the preset iteration convergence threshold is set to terminate iteration when there is no significant improvement at the Pareto front for 20 consecutive generations.

[0059] Before performing optimization, the intensive operation scheme optimization module 4 first performs spatial clustering of the diseases to be maintained. The DBSCAN clustering algorithm is used to group diseases with similar geographical locations into the same cluster. The neighborhood radius is set to 500 meters, and the minimum sample size is set to 2. For diseases within the same cluster, a directed acyclic graph (DAG) is constructed based on process dependencies. Nodes in the graph represent processes, and directed edges represent dependencies between processes. Topological sorting based on the DAG determines the feasible order of process execution and generates the initial population accordingly.

[0060] The optimized work plan output includes: the start and end times of each process, the equipment and personnel allocated to each process, the equipment scheduling routes, and the coordinates of the safety protection boundaries of the construction area. The plan is sent to the cross-departmental collaborative scheduling management module 5 in structured data format.

[0061] The cross-departmental collaborative scheduling and management module 5 includes a cross-departmental data interface, an electronic fence early warning unit, and a progress deviation monitoring unit, which are used to realize the cross-departmental push of optimized work plans, construction process monitoring, and closed-loop feedback.

[0062] The cross-departmental data interface is designed to enable bidirectional data interaction with the traffic police traffic control system and the operation support monitoring system. In one embodiment of this invention, the data interface adopts the RESTful API specification, supports HTTP / HTTPS protocols, and uses JSON format for data transmission. To ensure data transmission security, the interface incorporates TLS 1.3 encryption and OAuth 2.0 authentication mechanisms. The interface response time is controlled within 100ms, and it supports processing more than 100 concurrent requests per second.

[0063] The main functions of the cross-departmental data interface include: pushing optimized work plans to the traffic police system and receiving traffic control response data; pushing work tasks to construction teams and receiving construction progress feedback data; and pushing maintenance information to the operation and support system and receiving road condition monitoring data. Data push adopts a message queue mechanism, supporting reliable message transmission and breakpoint resumption.

[0064] The purpose of the electronic fence early warning unit is to delineate the boundary of the construction safety zone based on GPS positioning data, enabling real-time monitoring and early warning of the construction area. The boundary of the electronic fence is determined according to the coordinates of the construction area in the optimized operation plan, and the boundary is extended outward by a preset safety buffer distance, which is set to 50 meters in this embodiment. The electronic fence early warning unit receives GPS positioning data of work vehicles and personnel in real time to determine whether they are within the safety zone. When a non-work vehicle or personnel is detected entering the construction safety zone, an early warning process is triggered: first, an alarm message is sent to the traffic police system through a cross-departmental data interface, requesting enhanced traffic guidance; simultaneously, a vibration alert is pushed to the mobile devices of personnel at the construction site; finally, an early warning record is generated on the management platform for subsequent analysis.

[0065] The purpose of the schedule deviation monitoring unit is to monitor the construction progress in real time and trigger dynamic adjustments to the plan when the schedule deviation exceeds a threshold. The formula for calculating schedule deviation is: , in, For the first The actual progress of each process; To the planned completion schedule; This represents the percentage of schedule deviation. In this embodiment, the deviation is set to 20% when it exceeds a preset schedule deviation threshold, triggering a dynamic adjustment process.

[0066] The closed-loop feedback mechanism is one of the core innovations of this invention. The cross-departmental collaborative scheduling and management module 5 calculates closed-loop feedback parameters based on construction progress feedback data and actual execution results, including prediction model correction coefficients and optimization constraint correction values.

[0067] The calculation method for the prediction model correction coefficient is as follows. Let the prediction development level of a certain disease by the disease development trend prediction module 3 be... The actual observed development level is The prediction error is then... After accumulating prediction errors for multiple diseases, the correction coefficient for the prediction model is calculated: , This correction coefficient is used to adjust the output bias of the multi-factor coupled prediction network, making the prediction results closer to the actual situation.

[0068] The calculation method for the optimization constraint correction value is as follows. Let the planned execution time of a certain process in the optimized work plan be... The actual execution time is The time deviation is After calculating the time deviations of multiple processes, the time constraint parameters of the multi-objective constrained genetic optimization algorithm are updated: , in, This represents the average time deviation. Similarly, resource constraint parameters are updated based on actual resource consumption.

[0069] Closed-loop feedback parameters are sent to the disease development trend prediction module 3 and the intensive operation scheme optimization module 4 via a cross-departmental data interface. After receiving the prediction model correction coefficients, the disease development trend prediction module 3 performs online fine-tuning of the multi-factor coupled prediction network, updating the network weights and bias parameters. After receiving the optimization constraint correction values, the intensive operation scheme optimization module 4 dynamically adjusts the constraints of subsequent optimization tasks to improve the executability of the scheme.

[0070] The system of this invention employs a hybrid storage architecture to achieve efficient management of multi-source heterogeneous data. In one embodiment of this invention, the database design follows these principles: structured data and unstructured data are stored separately; hot data and cold data are managed in layers; and read / write separation improves concurrency performance.

[0071] Structured data storage employs a relational database; in this embodiment, MySQL 8.0 is selected. The main data tables include: a basic disease information table, storing fields such as unique disease identifier, GPS coordinates, first detection time, disease type, and severity level; a maintenance record table, storing fields such as maintenance task number, execution time, construction sequence, construction personnel, construction equipment, and completion status; a traffic control rules table, storing fields such as control period, control section, control type, and implementing department; and an equipment parameter table, storing fields such as equipment number, equipment type, operational capability parameters, and current status. The database incorporates built-in data validation rules to constrain GPS coordinate range, timestamp format, and the validity of enumerated values, ensuring data integrity and consistency.

[0072] Unstructured data storage employs a distributed file system; in this embodiment, MongoDB 5.0 is selected. The main datasets include: an inspection image dataset, storing visible light images and infrared thermal images captured by the UAV, using GridFS to store large files; a point cloud dataset, storing 3D point cloud data collected by the vehicle-mounted LiDAR, using a compressed storage format to reduce storage space; and a feature vector dataset, storing high-dimensional feature vectors output by the multi-scale feature extraction unit, supporting vector similarity retrieval.

[0073] Time-series data storage employs a dedicated time-series database; in this embodiment, InfluxDB is selected. The main time-series data includes: sensor monitoring data, stored as timestamp-indexed continuous monitoring values ​​such as temperature, humidity, and traffic flow; and system performance metrics, storing performance indicators such as processing latency, throughput, and error rate for each module. The time-series data retention strategy is set as follows: raw data is retained for 30 days, hourly aggregated data for 1 year, and daily aggregated data for 5 years.

[0074] The data synchronization and consistency guarantee mechanism is as follows: Asynchronous synchronization between data sources is achieved through message queues; in this embodiment, Kafka is chosen as the message queue. When data is written, it first enters the Kafka message queue, and then the consumer program writes it to the corresponding database. For critical data requiring strong consistency, a distributed transaction protocol is used to ensure the atomicity of writes to multiple databases. Data version control uses an optimistic locking mechanism; each record includes a version number field, and version verification is performed during updates to avoid concurrent write conflicts.

[0075] The system of this invention employs a multi-layered protection mechanism for data security and system security. In one embodiment of this invention, the security mechanism includes the following aspects.

[0076] Regarding data transmission security, all cross-network data transmissions are encrypted using the TLS 1.3 protocol, with a key length of at least 256 bits. Sensitive data, such as GPS coordinates and personnel information, is anonymized before transmission. The data interface employs a two-way authentication mechanism, with the client and server mutually verifying the validity of digital certificates.

[0077] Regarding data storage security, static data is encrypted using the AES-256 algorithm. Database access follows the principle of least privilege, with different functional modules using different database accounts and only necessary table access permissions granted. Important data is backed up regularly, with backup data stored in an off-site disaster recovery center. Backups are performed daily incrementally and weekly fully.

[0078] For identity authentication and access control, the system employs a multi-factor authentication mechanism, including combinations of authentication methods such as username and password, SMS verification code, and digital certificate. The role-based access control model defines various roles, including maintenance administrators, construction workers, traffic police, and operations personnel, each with different functional permissions and data access scopes. Session management uses a JWT token mechanism, with a token validity period of 2 hours, supporting both active cancellation and automatic cancellation upon timeout.

[0079] In terms of auditing and monitoring, all system operations are recorded in audit logs, including information such as operation time, user, operation type, target, and result. Audit logs are stored in a tamper-proof manner and cannot be modified after being written. The security monitoring system detects abnormal access behavior in real time, such as frequent login failures and abnormal data access patterns, triggering alarms and automatically taking protective measures.

[0080] In a complete embodiment of the present invention, the system deployment architecture adopts an edge-cloud collaborative mode to achieve optimized configuration of computing resources and efficient execution of data processing.

[0081] The edge computing layer is deployed along the highway using edge computing nodes. In this embodiment, industrial-grade edge servers are employed, configured with Intel Core i7 processors, 16GB of RAM, and 512GB of SSD storage. Edge nodes are deployed at a density of one node every 50km, responsible for performing computational tasks such as data acquisition, data cleaning, and preliminary feature extraction. Edge nodes connect to sensor devices via wired Ethernet or wireless networks, with data acquisition latency controlled within 50ms. Edge nodes are interconnected via a dedicated fiber optic network, supporting localized data processing and distributed collaboration.

[0082] The cloud computing layer is deployed in the maintenance management center's server cluster, which adopts a containerized deployment scheme in this embodiment, based on the Kubernetes cluster management platform. The server cluster configuration includes: 8 compute nodes, each equipped with an Intel Xeon E5-2680v4 CPU, 64GB of memory, and 2TB of SSD storage; 2 GPU nodes, each equipped with an NVIDIA Tesla V100 graphics card, for accelerating deep learning model inference; and 4 storage nodes, forming a distributed storage cluster with a total storage capacity of 100TB. The cloud is responsible for executing computationally intensive tasks such as multi-factor coupled prediction network inference, genetic optimization algorithm solving, and cross-departmental data interaction.

[0083] Data synchronization between the edge and the cloud employs an incremental transmission mechanism. Edge nodes only upload the changed data, using differential coding and compression algorithms to reduce the amount of data transmitted, achieving a compression rate of over 70%. Data transmission priority is divided into three levels: high priority is used for alarm information and emergency events, using an instant transmission method; medium priority is used for disease detection results and status updates, using a near real-time transmission method with a delay of no more than 5 minutes; low priority is used for raw images and point cloud data, using a batch transmission method, executed during network idle periods.

[0084] The system's software architecture adopts a microservice design pattern, with each functional module encapsulated as an independent microservice container and orchestrated and managed using Kubernetes. Microservices communicate synchronously via the gRPC protocol and asynchronously via Kafka message queues. Service registration and discovery use Consul, load balancing uses Nginx, and the API gateway uses Kong. The system supports horizontal scaling, dynamically adjusting the number of container instances based on business load; the automatic scaling strategy is based on CPU utilization and request response latency metrics.

[0085] The system of this invention provides comprehensive operation and maintenance functions, supporting long-term stable operation and continuous optimization. In one embodiment of this invention, the operation and maintenance functions include the following aspects.

[0086] In terms of system monitoring and alarms, the monitoring system collects operational status indicators for each module, including basic indicators such as CPU utilization, memory usage, network traffic, and disk I / O, as well as business indicators such as data processing latency, model inference time, and solution generation time. Monitoring data is collected every minute and stored in a time-series database for historical querying. Alarm rules employ multi-level threshold settings, divided into three levels: alert, warning, and critical. Different levels use different notification methods and processing procedures.

[0087] Regarding model updates and iterations, the deep learning model in the disease development prediction module supports online updates. Once sufficient new sample data has been accumulated, the system automatically triggers the model retraining process. After training, the new model is validated in the testing environment, and after successful validation, it is smoothly switched to the production environment via blue-green deployment. Model version management records the training data, model parameters, and evaluation metrics for each version, and supports version rollback.

[0088] Regarding parameter tuning and performance optimization, the system provides a parameter tuning interface, allowing administrators to adjust key parameters based on actual operating conditions, such as the window length of the adaptive Kalman filter, the population size and iteration count of the genetic algorithm, and the threshold for monitoring progress deviations. After parameter adjustment, the system automatically records performance comparison data before and after the adjustment, providing a reference for subsequent optimization. Performance optimization adopts a continuous improvement model, regularly analyzing system bottlenecks and developing optimization plans.

[0089] In practical applications, the system of this invention has been piloted on a high-traffic highway. The pilot section is 100km long, with an average daily cross-sectional flow of approximately 500,000 vehicles, and heavy-load vehicles accounting for about 35%, making it a typical high-load operating highway. The pilot deployment lasted for 6 months, and the operational data after deployment fully verified the technical effectiveness of the system.

[0090] In terms of disease identification and localization, after system deployment, the disease identification coverage increased from 85% before the pilot to 97%, an improvement of 14%. Disease localization accuracy improved from ±5m to ±1m, a five-fold increase. The number of identifiable disease types expanded from 8 to 12, with newly identified types including pavement voids, network cracks, subsidence, and swells. The false negative rate decreased from 15% to 3%, and the false positive rate decreased from 8% to 2%. These improvements are attributed to the application of multi-source data fusion and multi-scale feature extraction technologies.

[0091] In terms of maintenance decision-making efficiency, the time for generating maintenance decisions has been reduced from 2 to 3 days for manual decisions to within 2 hours for automatic system generation, representing an efficiency improvement of over 20 times. The number of maintenance closures on a single road segment has decreased by 58%, effectively reducing the impact on traffic. The efficiency of multi-process collaborative operations has increased by 42%, allowing more maintenance processes to be completed within the same closure time. Equipment utilization has increased from 65% to 85%, reducing equipment idle and empty-running costs. These achievements are attributed to the application of a multi-objective constrained genetic optimization algorithm.

[0092] In terms of cross-departmental collaboration, the response time was reduced from 4 hours before the pilot program to 45 minutes, a reduction of 81%. The time for generating traffic control plans was shortened from 30 minutes for manual formulation to 5 minutes for automatic system generation. The real-time nature of construction progress information was improved from hourly to minute-level, supporting more precise progress monitoring and dynamic adjustments. The electronic fence early warning function effectively prevented three safety risk incidents of non-operational vehicles mistakenly entering the construction area. These achievements are attributed to the application of cross-departmental data interfaces and closed-loop feedback mechanisms.

[0093] In terms of disease prediction, the accuracy rate of disease development prediction reached 87%, with a disease severity prediction accuracy rate of 89% and a disease expansion range prediction error of less than 15%. The proportion of preventative maintenance increased from 15% to 35%, effectively avoiding high-cost repairs caused by disease expansion. Maintenance work scheduled in advance based on prediction results accounted for 40%, reducing the frequency and cost of emergency maintenance. The total annual maintenance cost decreased by 18% compared to before the pilot program, while maintenance quality indicators remained stable. These results are attributed to the application of a multi-factor coupled LSTM prediction algorithm and a closed-loop feedback mechanism.

[0094] In summary, the intelligent decision-making system for intensive highway maintenance based on multi-source collaborative perception provided by this invention effectively solves the technical problems of fragmented multi-source data, insufficient intelligent decision-making, inefficient cross-departmental collaboration, and lack of disease prediction in existing technologies by organically combining multi-source data fusion, deep learning prediction, genetic algorithm optimization, and cross-departmental collaborative management. It realizes intensive planning and efficient collaborative implementation of highway maintenance operations, and has significant technological advancements and broad application and promotion value.

[0095] The embodiments of the present invention are not limited to the specific embodiments described above. Those skilled in the art can make various equivalent changes or substitutions based on the technical solutions of the present invention, and all such changes or substitutions should be included within the protection scope of the present invention.

Claims

1. A highway intensive maintenance intelligent decision-making system based on multi-source collaborative sensing, characterized in that, include: The multi-source heterogeneous sensing data acquisition module is used to acquire visible light image data and infrared thermal image data of the road surface through the UAV inspection unit, acquire three-dimensional point cloud data and GPS positioning data of the road surface through the vehicle-mounted lidar unit, and acquire road surface temperature data, humidity data and traffic flow data through the Internet of Things sensor unit. The module then sends the visible light image data, the infrared thermal image data, the three-dimensional point cloud data, the GPS positioning data, the road surface temperature data, the humidity data and the traffic flow data to the multi-level data fusion processing module. The multi-level data fusion processing module is used to receive data sent by the multi-source heterogeneous sensing data acquisition module, perform noise filtering and outlier removal on the received data through an adaptive Kalman filter data cleaning unit, extract disease feature vectors from the cleaned visible light image data and three-dimensional point cloud data through a multi-scale feature extraction unit, and bind the disease feature vectors with the corresponding road surface temperature data and humidity data in a spatiotemporal correlation fusion unit based on the GPS positioning data to generate a fused disease information package and send it to the disease development trend prediction module. The disease development trend prediction module is used to receive the fused disease information package, historical maintenance data, traffic flow data and meteorological data, encode the historical maintenance data with time-series features through a time-series feature encoder, and predict the development level and expansion range of the disease within a preset prediction time domain through a multi-factor coupled prediction network based on the encoded time-series features, the disease feature vector in the fused disease information package, the traffic flow data and the meteorological data, generate the disease development prediction result and send it to the intensive operation scheme optimization module; The intensive operation plan optimization module is used to receive the disease development prediction results, the preset enclosure time window and construction resource constraint data, and based on the multi-objective constraint genetic optimization algorithm, jointly optimize the construction sequence of multiple processes, equipment scheduling paths and personnel configuration within the preset enclosure time window, generate an optimized operation plan for simultaneous multi-process enclosure and send it to the cross-departmental collaborative scheduling management module. The cross-departmental collaborative scheduling and management module is used to receive the optimized operation plan, push the optimized operation plan to the traffic police system and construction team through the cross-departmental data interface, receive traffic control response data and construction progress feedback data, calculate closed-loop feedback parameters based on the deviation between the construction progress feedback data and the optimized operation plan, send the closed-loop feedback parameters to the disease development trend prediction module to update the prediction model parameters, and send them to the intensive operation plan optimization module to dynamically adjust the optimization constraints.

2. The system according to claim 1, characterized in that, In the multi-source heterogeneous sensing data acquisition module: the UAV inspection unit is equipped with a high-definition camera with a resolution of not less than 1920×1080 and an infrared thermal imager with a temperature measurement accuracy of not less than ±0.5℃, with a flight altitude of 50 to 80 meters and a flight speed of 30 km / h to 50 km / h; the vehicle-mounted lidar unit integrates a lidar with a ranging accuracy of not less than ±2 cm and a GPS positioning module with a positioning accuracy of not less than ±1 m; the IoT sensor unit includes road surface temperature sensors, humidity sensors and traffic flow detectors deployed every 300 to 500 meters along the highway, with a data upload frequency of 1 to 5 times per minute.

3. The system according to claim 1, characterized in that, In the multi-level data fusion processing module: the adaptive Kalman filter data cleaning unit adaptively adjusts the process noise covariance matrix and the observation noise covariance matrix based on the residual between the sensor measurement value and the predicted value; the multi-scale feature extraction unit uses a spatial pyramid pooling structure to extract multi-scale texture features and edge features from the visible light image data, and uses a point cloud voxelization method to extract geometric shape features and depth features from the three-dimensional point cloud data. The spatiotemporal correlation fusion unit employs an attention weighting mechanism to adaptively allocate weights and fuse feature vectors from different sensors at the feature level.

4. The system according to claim 1, characterized in that, The multi-level data fusion processing module is also used to: perform weighted fusion of the disease feature vector based on a preset fusion weight coefficient, wherein the preset fusion weight coefficient is dynamically determined according to the signal-to-noise ratio of each sensor data; and classify the diseases in the fused disease information package into levels according to a preset disease severity threshold, wherein the levels include three levels: minor disease, moderate disease, and severe disease.

5. The system according to claim 1, characterized in that, In the disease development trend prediction module: the time-series feature encoder adopts a bidirectional long short-term memory network structure to perform forward and reverse time-series feature encoding on the disease evolution sequence in the historical maintenance data; The multi-factor coupled prediction network includes a feature embedding layer, a multi-head self-attention layer, and a prediction output layer. The feature embedding layer maps the traffic flow data and the meteorological data into embedding vectors of the same dimension as the disease feature vectors. The multi-head self-attention layer performs cross-attention calculation on the embedded multi-factor features. The prediction output layer outputs the probability distribution of disease development level within a preset prediction time domain.

6. The system according to claim 1, characterized in that, In the intensive operation scheme optimization module: the multi-objective constrained genetic optimization algorithm takes minimizing the total enclosure time, minimizing the equipment idle running distance, and maximizing the parallelism of the process as optimization objectives; the constraints of the multi-objective constrained genetic optimization algorithm include the preset enclosure time window constraint, construction safety distance constraint, equipment operation capability constraint, and personnel skill matching constraint; the multi-objective constrained genetic optimization algorithm adopts an elite retention strategy and adaptive crossover mutation probability.

7. The system according to claim 1, characterized in that, The disease development trend prediction module is also used to: update the network weights and bias parameters of the multi-factor coupled prediction network according to the closed-loop feedback parameters; the preset prediction time domain length is 1 month to 6 months, and the prediction output frequency is updated once a week.

8. The system according to claim 1, characterized in that, The intensive operation scheme optimization module is also used for: spatially clustering the diseases to be maintained according to their geographical location to generate disease clusters; constructing a directed acyclic graph of processes for diseases within the same disease cluster according to process dependencies; determining a set of feasible schemes for parallel execution of processes based on the directed acyclic graph of processes and the preset enclosure time window; and searching for Pareto optimal solutions in the set of feasible schemes using the multi-objective constrained genetic optimization algorithm.

9. The system according to claim 1, characterized in that, The cross-departmental collaborative scheduling and management module also includes: an electronic fence early warning unit, used to delineate the boundary of the construction safety area based on the GPS positioning data, and generate early warning information when non-operational vehicles or personnel are detected entering the construction safety area; and a progress deviation monitoring unit, used to calculate the deviation value between the construction progress feedback data and the planned progress in the optimized operation plan, and trigger dynamic adjustment of the plan when the deviation value exceeds a preset progress deviation threshold.

10. The system according to claim 1, characterized in that, The cross-departmental collaborative scheduling and management module is also used to: conduct two-way data interaction with the traffic police traffic control system and the operation support monitoring system through the cross-departmental data interface using an encrypted transmission protocol; summarize the construction progress feedback data according to a preset cycle to generate a maintenance execution report; the closed-loop feedback parameters include prediction model correction coefficients and optimization constraint correction values, the prediction model correction coefficients are used to adjust the output deviation of the multi-factor coupled prediction network, and the optimization constraint correction values ​​are used to update the time constraint parameters and resource constraint parameters of the multi-objective constrained genetic optimization algorithm.

Images