Road informatization intelligent management and control system

Abstract

The invention discloses a highway informatization intelligent management and control system, and relates to the technical field of intelligent construction, and the system comprises the steps: collecting multi-source data in real time, uploading the multi-source data to an edge calculation node for timestamp alignment and format standardization, and inputting the multi-source data to a space-time Transform fusion model, extracting spatial-temporal features of sensor data and semantic features of video pictures through a multi-head attention mechanism, and generating a dynamic risk score; and updating a virtual construction progress model of the digital twin, superposing the adjusted construction standard to a real scene through an AR terminal, guiding a worker to carry out deviation correction, and encrypting and storing a closed-loop processing record to a block chain. According to the method, a pulse code driven distributed learning framework is adopted to greatly improve the anomaly detection efficiency, and quick response and global optimization of equipment cluster behaviors are realized through event triggered feature compression and an adaptive weight updating mechanism.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent construction, and particularly to a highway informatization intelligent control system. Background Art

[0002] In the field of intelligent construction, construction safety monitoring is realized through multi-modal sensing technology, and the collaborative analysis of vibration, temperature and video data is the core evaluation method. The existing technology uses independent sensors and video systems to construct a risk assessment model, and realizes multi-source data synchronization through linear interpolation or fixed time windows. For example, the vibration spectrum and video object detection results are fused to generate a risk score. However, such methods have significant defects in complex construction scenarios (such as multi-device collaborative operations and dynamic environmental interference).

[0003] The core problem of traditional methods lies in the insufficient spatio-temporal alignment accuracy of multi-source heterogeneous data: there is a millisecond-level clock deviation between sensor data and video streams, and existing interpolation algorithms cause key feature distortion in non-uniform sampling scenarios; at the same time, multi-modal fusion relies on static weight allocation and cannot dynamically adapt to sudden changes in the construction environment, resulting in a significant increase in the false positive rate of anomaly detection. Summary of the Invention

[0004] In view of the above existing problems, the present invention is proposed.

[0005] Therefore, the present invention provides a highway informatization intelligent control system to solve the problems of lagging construction safety risk assessment and high false positive rate caused by insufficient spatio-temporal alignment accuracy of multi-source heterogeneous data.

[0006] To solve the above technical problems, the present invention provides the following technical solutions: The present invention provides a highway informatization intelligent control system, which includes: collecting multi-source data in real time, uploading it to an edge computing node for timestamp alignment and format standardization, inputting it into a spatio-temporal Transformer fusion model, extracting spatio-temporal features of sensor data and semantic features of video images through a multi-head attention mechanism, and generating a dynamic risk score; triggering a hierarchical warning signal based on the dynamic risk score, and executing AR projection instructions, red alert push, rectification work order generation and measurement and payment locking; encrypting and uploading the abnormal mode features of the equipment cluster of the current project to the cloud through a federated learning platform, comparing and learning with the global feature library shared by historical projects, and performing cross-project optimized group anomaly detection; performing online inference on the real-time construction data stream, and automatically generating an instruction set including adjustment parameters when it is detected that the behavior of the equipment cluster deviates from the safety threshold optimized by federated learning; updating the virtual construction progress model of the digital twin, and superimposing the adjusted construction standards on the real scene through an AR terminal to guide workers to correct deviations, and at the same time encrypting and storing the closed-loop processing records in a blockchain.

[0007] As a preferred solution of the highway informatization intelligent control system of the present invention, wherein: the generation of the dynamic risk score is specifically carried out as follows, Adopt a hierarchical feature fusion architecture as the spatio-temporal Transformer fusion model; Capture the spatial relationship between devices through spatial attention flow to generate topological features; Apply a causal mask to extract the temporal dependence relationship of device operation parameters and establish a dynamic feature expression across time steps; Dynamically fuse the spatio-temporal features of sensor data and the semantic features of video images through a gating mechanism.

[0008] As a preferred solution of the highway informatization intelligent control system of the present invention, wherein: the hierarchical early warning signal refers to constructing a three-level dynamic threshold and dynamically adjusting the threshold range according to the real-time construction stage to trigger corresponding measures.

[0009] As a preferred solution of the highway informatization intelligent control system of the present invention, wherein: the abnormal mode feature of the device cluster refers to the spatio-temporal pulse diagram mode feature generated by the multi-modal feature vector after the gating fusion of the sensor topological feature extracted by the spatial attention flow and the temporal dynamic feature extracted by the temporal attention flow, combined with the video semantic feature.

[0010] As a preferred solution of the highway informatization intelligent control system of the present invention, wherein: the group anomaly detection for cross-project optimization is specifically carried out as follows, Collect the operation index data of multiple projects, preprocess the data and construct a multi-project spatio-temporal feature matrix; Analyze the spatio-temporal correlation of the indicators between projects and extract cross-project spatio-temporal correlation features; Build an anomaly detection model for cross-project optimization based on transfer learning, and perform cross-project feature transfer by dynamically adjusting the parameters of the anomaly detection model; Train the anomaly detection model by combining the group behavior pattern and historical anomaly data to generate an anomaly probability distribution; Dynamically adjust the anomaly threshold according to the real-time detection data and output the group anomaly detection result.

[0011] As a preferred solution of the highway informatization intelligent control system of the present invention, wherein: the virtual construction progress model of the digital twin is specifically carried out as follows, Collect multi-source heterogeneous data at the construction site, integrate the BIM model, sensor monitoring data and resource scheduling plan to construct a dynamically updated virtual construction scenario; Based on the principle of discrete event simulation, a construction progress logical relationship model is established. An initial progress plan is generated through time series analysis and resource constraint optimization. Combining real-time construction data with historical deviation rules, the task node parameters in the virtual construction scenario are dynamically corrected to generate progress prediction and risk warning results; Adjust the resource allocation and task priorities through a multi-objective optimization algorithm, and output the optimal construction progress plan.

[0012] As a preferred solution of the highway informatization intelligent control system described in the present invention, wherein: the timestamp alignment and format standardization include defining a reference time series, synchronizing sensor data and video streams using the reference signal interpolation alignment method, and integrating video semantic features and sensor data into a unified data tensor.

[0013] As a preferred solution of the highway informatization intelligent control system described in the present invention, wherein: the multi-source data includes the vibration spectrum of the roller, the temperature curve of the paver, the prestress tension value, and the construction images captured by the video monitoring equipment.

[0014] As a preferred solution of the highway informatization intelligent control system described in the present invention, wherein: the abnormal mode characteristics of the equipment cluster include the spatio-temporal distribution map of the deviation of the roller vibration spectrum from the mean value and the coordinates of the abnormal temperature gradient area of the paver.

[0015] As a preferred solution of the highway informatization intelligent control system described in the present invention, wherein: the safety threshold refers to the dynamic discrimination boundary optimized by federated learning; The instruction set includes the corrected value of the roller vibration frequency and the upper limit of the paver travel speed, and is sent to the corresponding equipment controller through the edge computing node to execute dynamic regulation.

[0016] The beneficial effects of the present invention are as follows: Through the high-precision spatio-temporal alignment algorithm and the adaptive multi-modal fusion architecture, the real-time performance and accuracy of the construction safety risk assessment are significantly optimized. Specifically, the spatio-temporal feature alignment mechanism based on physical constraints effectively suppresses the noise interference in the multi-source data synchronization process and ensures the integrity of the vibration conduction characteristics of the equipment cluster; the dynamic gating fusion module automatically adjusts the contribution weights of sensor and video semantic features through environmental perception, and realizes the collaborative optimization of multi-modal information under complex working conditions; the distributed learning framework driven by pulse coding greatly improves the anomaly detection efficiency, and through the event-triggered feature compression and adaptive weight update mechanism, realizes the rapid response and global optimization of the equipment cluster behavior. Thereby significantly reducing the probability of misjudgment of safety risks and at the same time improving the execution timeliness of early warning instructions. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the accompanying drawings required for the description of the embodiments. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other accompanying drawings can also be obtained based on these drawings.

[0018] Figure 1 It is a flowchart of the highway informatization intelligent control system.

[0019] Figure 2 It is a flowchart of multi-modal feature fusion of the highway informatization intelligent control system.

[0020] Figure 3 It is a schematic diagram of federated optimization interaction of the highway informatization intelligent control system.

[0021] Figure 4 It is a working principle diagram of virtual-real synchronization of the highway informatization intelligent control system. Specific Embodiments

[0022] To make the above objects, features, and advantages of the present invention more obvious and understandable, the following will make a detailed description of the specific embodiments of the present invention in conjunction with the accompanying drawings of the specification.

[0023] In the following description, many specific details are set forth to fully understand the present invention. However, the present invention can also be implemented in other ways different from those described herein. Those skilled in the art can make similar extensions without departing from the connotation of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed below.

[0024] Secondly, the so-called "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that can be included in at least one implementation manner of the present invention. The "in one embodiment" that appears in different places in this specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment that mutually excludes other embodiments.

[0025] Embodiment 1, referring to Figures 1 to 4 , which is the first embodiment of the present invention. This embodiment provides a highway informatization intelligent control system, including the following steps: S1. Real-time collect multi-source data, upload it to the edge computing node for timestamp alignment and format standardization, input it into the spatio-temporal Transformer fusion model, extract the spatio-temporal features of sensor data and the semantic features of video images through the multi-head attention mechanism, and generate a dynamic risk score.

[0026] Specifically, it includes the following steps: Deploy a sensor network and video stream devices at the hardware layer for multi-source data collection.

[0027] Among them, the multi-source data includes the vibration spectrum of the roller, the temperature curve of the paver, the prestress tension value, and the construction images captured by the video monitoring equipment.

[0028] Define the data transmission protocol for sensor data and video stream: Specifically, for sensor data: encapsulate the vibration spectrum, temperature curve, and tension value into the Cap’n Proto binary format and upload them to the edge node through the 5G NR Uu interface, with an end-to-end delay ≤ 30 ms. For the video stream: use H.265 encoding, slice it into 1-second HLS segments, and transmit it to the edge node through the RTSP protocol.

[0029] Perform data preprocessing on the sensor data and video stream.

[0030] Specifically, perform FFT transformation on the vibration spectrum of the roller to extract the energy value in the 0-100 Hz frequency band. Perform moving average filtering (window length 10 seconds) on the temperature curve of the paver. Use the YOLOv7-tiny model to detect key objects (tower cranes, workers, material piles) in the construction images in real time, and output the Bounding Box coordinates and class labels.

[0031] Based on the time characteristics of the sensor data and video stream after data preprocessing, determine the starting point and step size of the reference time series, and define the discrete time series : ; In the formula, (alignment accuracy), is the time index, is the length of the reference time series, indicating that there are time points in the reference time series (because starts from 0); Adopt the reference signal interpolation alignment method to calculate the interpolation of the sensor data at , expressed as: ; In the formula, represents the interpolation of the aligned sensor data at the reference time point , represents the measured value of the sensor data at the time point, represents the th time point in the reference time series, represents the time point of the sensor data, that is, the timestamp of the sensor data, Indicates the number of sensor data points for interpolation calculation, Indicates pi, Indicates the time index of sensor data; It should be noted that, Indicates the Sinc function, which is used for interpolation calculation.

[0032] Specifically, according to the sensor data and the reference time point calculate the interpolation weights based on the time difference between them; When the value of the Sinc function is 1.

[0033] When the value of the Sinc function gradually decays, indicating that data points farther from the target time point contribute less to the interpolation result.

[0034] Use the model to extract the Bounding Box coordinates of key objects (such as tower cranes, workers) in each frame, synchronize their timestamps with the sensor data, and construct a unified data tensor ; Among them, Indicates the set of real numbers, (number of sensor channels + number of video semantic features), data block length (corresponding to a 1-second window).

[0035] After normalizing and partitioning the unified data tensor input it into the spatio-temporal Transformer fusion model to extract features through spatial attention flow, temporal attention flow, and multimodal fusion.

[0036] It should be noted that normalizing and partitioning means calculating the mean and standard deviation for each column of the unified data tensor, that is, each feature dimension, and then normalizing the data by subtracting the mean from each feature value and dividing by the standard deviation to obtain the normalized data tensor; then perform partitioning, the purpose of which is to divide the data into blocks of fixed length for subsequent model processing. The specific operation is to set the block length according to requirements. For example, two hundred corresponds to a 1-second window, and the normalized data tensor is divided into multiple fixed-length data blocks along the time dimension. Each data block contains a certain number of time points and feature dimensions; finally, the divided data blocks are stored or passed to the subsequent model for feature extraction and fusion.

[0037] Furthermore, after the data preprocessing is completed, the sensor data and video data are respectively converted into feature matrices for subsequent fusion and analysis. Specifically, the sensor data is represented as , after the video data is processed by the ResNet-18 model, it is represented as ; Among them, represents the dimension of the sensor features (e.g., physical quantities measured by 10 sensors), represents the length of the feature vector output by ResNet-18 (e.g., 1024 dimensions).

[0038] Specifically, by using spatial attention flow to capture the spatial relationship between devices (such as device position coordinates), the topological features of sensor data are extracted and represented as: ; In the formula, represents the output of the attention mechanism, is the query matrix, representing the target features to be focused on, is the key matrix, representing the feature representation of the input data, is the value matrix, representing the actual feature values of the input data, represents the query matrix and the transpose matrix of the key matrix represents the symbol of matrix transpose, represents the square root of the feature dimension of the key matrix , represents the subscript of , used to represent the dimension of the key matrix represents the mask matrix based on the physical distance between devices (set to when the distance > 10m), represents the weighted sum of the value matrix according to the attention weights; By using temporal attention flow to capture the temporal dependence relationship of multi-sensor data, the temporal dynamic features of sensor data are extracted and represented as: ; In the formula, represents the causal mask (only allowing historical information to participate); The video semantic features (1024-dimensional vector extracted by ResNet-18) are fused with the sensor features through a gating mechanism and represented as: ; ; In the formula, represents the gating weight vector, used to dynamically adjust the fusion ratio of sensor features and video semantic features, represents the Sigmoid activation function, denotes the weight matrix, which is used to map the concatenated feature vectors to the gated weight space. denotes the concatenated vector of sensor features and video semantic features. denotes the bias vector, which is used to adjust the baseline value of the gated weights. denotes the fused multi-modal feature vector, which contains the dynamic weighted combination of sensor features and video semantic features. denotes the weighted result of sensor features, where controls the contribution ratio of sensor features. denotes the weighted result of video semantic features, where controls the contribution ratio of video features. denotes the element-wise multiplication (Hadamard product), that is, the elements at the corresponding positions of two vectors are multiplied. It should be noted that the sensor topology and temporal dynamic features extracted by the spatial attention flow and temporal attention flow are the main sources of sensor features in multi-modal fusion. Through the gating mechanism, the fusion ratio of sensor features and video semantic features is dynamically adjusted to ensure that the fused features can fully utilize the information of both modalities.

[0039] Based on the fused multi-modal feature vector , a dynamic risk score is generated , which is expressed as: ; In the formula, denotes the weight coefficient, which is used to adjust the contribution of the difference between the current feature and the safety feature. is the summation symbol, indicating the summation over all feature dimensions. denotes the weight coefficient, which is used to adjust the difference contribution of the th feature dimension. denotes the value of the th feature dimension at time . is the safety feature, representing the feature value in the ideal safe state. is the L2 norm, representing the Euclidean distance between the current feature and the safety feature . denotes the weight coefficient, which is used to adjust the contribution of the difference between the current feature distribution and the historical feature distribution. is the KL divergence, representing the difference between the current feature distribution and the historical feature distribution .

[0040] It should be noted that the weight coefficients and It can be determined by empirical values or optimization algorithms. For example, in a construction scenario, it can be set according to historical data or expert experience 、 indicates that the contribution of the current feature difference is greater, and it can also be optimized through machine learning: for example, using cross-validation or gradient descent method to optimize the values of and based on training data.

[0041] For the weight coefficient it can be determined by a time decay function or expert experience. Specifically, if an exponential decay function is used, where is the decay coefficient, indicating that the weight of a time point farther from the current time is smaller; for determination by expert experience: fixed weights can be assigned to different time points according to the importance of the construction stage.

[0042] For the safety feature it can be determined by historical data statistics. For example, based on past normal construction data, calculate the mean or median of the feature values as , and it can also be set by expert experience. For example, according to construction specifications or safety standards, define the specific value of .

[0043] For the KL divergence it can be determined by probability distribution estimation. For example, use kernel density estimation (KDE) or Gaussian mixture model (GMM) to fit the current and historical feature distributions, calculate the KL divergence, and it can also be determined by historical data statistics. For example, based on past normal construction data, construct the historical feature distribution .

[0044] S2. Trigger a graded warning signal based on the dynamic risk score, execute the AR projection instruction, push the red alert, generate a rectification work order, and lock the measurement and payment.

[0045] Specifically, it includes the following steps: Construct a graded threshold through a scoring formula: ; It should be noted that indicates the existence of potential risks, which need attention but have not reached the emergency level, indicates a relatively high risk and immediate measures need to be taken, indicates an extremely high risk and construction must be stopped immediately and emergency treatment measures must be taken; Among them, the classification threshold can be dynamically adjusted according to the real-time construction stage (such as pouring, hoisting) to adapt to the risk characteristics and tolerance of different stages. The specific implementation method can be to preset the threshold range of different construction stages and automatically switch the corresponding threshold according to the real-time construction stage.

[0046] Furthermore, when the dynamic risk score reaches a certain classification threshold, the corresponding classification warning signal will be triggered and the following operations will be executed: AR projection instruction: Through augmented reality (AR), project risk information or rectification instructions onto the construction site to guide the operation of workers.

[0047] Red alert push: Send a red alert to relevant personnel to remind them to take immediate action.

[0048] Rectification work order generation: Automatically generate a rectification work order to clarify the content to be rectified and the responsible person.

[0049] Measurement and payment lock: Before the risk is lifted, lock the measurement and payment of related projects to ensure the implementation of rectification measures.

[0050] It should be noted that the classification threshold is the basis for judging and triggering the classification warning signal. When reaching a certain threshold, the corresponding warning signal will be automatically triggered and the corresponding operations will be executed.

[0051] For example: If , then trigger , and execute the AR projection instruction and rectification work order generation.

[0052] If , then trigger , and execute the red alert push and measurement and payment lock.

[0053] If , then trigger , and execute all warning and rectification measures.

[0054] Preferably, through this mechanism, corresponding warning and rectification measures can be automatically executed according to the risk level to ensure construction safety.

[0055] S3. Encrypt and upload the abnormal mode characteristics of the equipment cluster of the current project to the cloud through the federated learning platform, and compare and learn with the global feature library shared by historical projects to generate a cross-project optimized group anomaly detection model.

[0056] Specifically, it includes the following steps: Extract the abnormal mode characteristics of the equipment cluster, including the deviation of the vibration spectrum of the roller from the mean value Spatial-temporal distribution map, coordinates of abnormal temperature gradient regions of the paver.

[0057] Combined with the dynamic risk scoring sequence and spatio-temporal Transformer features are input into the pulse coding layer, and the continuous features are converted into a pulse sequence through the time difference threshold algorithm. Only when the feature change rate exceeds the threshold is a pulse triggered; The pulse sequence is expressed as: ; In the formula, represents the pulse state of the th feature at time , represents the value of the th fused feature at time , is a time variable, representing a certain moment within the integration interval, represents the th fused feature at time 's rate of change (derivative), is the pulse trigger threshold. Only when the feature change rate exceeds the threshold is a pulse emitted, significantly reducing the data transmission volume (compression rate ≥ 90%), is the length of the time integration interval, indicating the integration of the feature change rate over the most recent time, represents the integration interval; The pulse sequence and the sensor topology relationship (device spatial coordinates) are dynamically constructed into a spatio-temporal pulse map : Among them, the pulse sequence refers to multiple pulse states within a period of time.

[0058] Spatio-temporal pulse map is expressed as: ; In the formula, is a node, representing a single device in the device cluster, carrying a pulse sequence, is an edge, representing the physical connection or signal coupling relationship between devices (such as the vibration conduction path), represents the weight; Furthermore, the weight will be dynamically updated based on the physical distance and pulse synchronization between devices, and is expressed as: ; In the formula, represents at time , device and device The weight value between represents the device and the device the physical distance between is the time length of the summation interval, represents at the time variable when the pulse state of the device ; represents at the time variable when the pulse state of the device ; represents the time variable at any moment within the summation interval, represents from the current time pushing forward the position of time; It should be noted that the greater the weight the stronger the behavioral correlation between the device and the device , and at the same time the device and the device represent two independent construction devices (such as a roller and a paver).

[0059] Encrypt and upload the spatio-temporal pulse diagram to the cloud federated learning platform, and adopt an event-driven mechanism (only trigger upload when a significant pulse pattern is detected), replacing the traditional periodic update.

[0060] Among them, the cloud federated platform adopts an asynchronous update mechanism, and each edge node only triggers model upload (event-driven) when a significant pulse pattern is detected, rather than a fixed period.

[0061] The cloud updates the global model through the spike-timing dependent plasticity (STDP) rule, and the weight adjustment rule is: ; In the formula, represents at time when the change amount of the weight between the device and the device , represents the learning rate (update step size), represents at time when the pulse state of the device , represents the current time pushing forward the position of time; Compare the pulse pattern of the current project with the global feature library of historical projects for cross-project optimized population anomaly detection.

[0062] S4. Perform online inference on the real-time construction data stream. When it is detected that the behavior of the equipment cluster deviates from the safety threshold optimized by federated learning, an instruction set containing adjustment parameters is automatically generated.

[0063] Specifically, it includes the following steps: Model the group anomaly detection as a quantum Markov decision process, including the state space: the pulse synchronization matrix of the equipment cluster ; Action space: the instruction set for adjustment parameters (such as vibration amplitude, temperature threshold); Reward function: designed based on the decreasing rate of the dynamic risk score of.

[0064] Furthermore, use the quantum approximate optimization algorithm (QAOA) to solve the optimal instruction strategy, breaking through the local optimum limitation of classical reinforcement learning.

[0065] When it is detected that the synchronization matrix deviates from the safety threshold optimized by federated learning, an instruction set is generated through a quantum annealing processor; Specifically, the construction and optimization of the safety threshold are first based on historical normal construction data to construct the joint probability distribution of equipment characteristic parameters through kernel density estimation as the basic threshold, and rely on the federated learning platform to aggregate the topological feature weight matrices of the spatio-temporal pulse diagrams of each project, and use the pulse timing-dependent plasticity rule to dynamically update the global model parameters, and calculate the reference value of the safety threshold in combination with the equipment correlation weight increment; on this basis, introduce the construction stage adaptation coefficient to dynamically adjust the threshold range, load the preset threshold coefficient according to the real-time construction stage and fuse the environmental compensation factor of the temperature and humidity sensor for threshold compensation optimization, and finally realize the dynamic optimization and deviation detection of the safety threshold through a multi-dimensional joint determination mechanism, specifically including the triple condition collaborative determination that the Euclidean distance between the spatio-temporal Transformer fusion feature vector and the safety reference exceeds the dynamic compensation threshold, the KL divergence between the current feature distribution and the global feature library exceeds the preset probability deviation, and the pulse synchronization index of the equipment cluster breaks through the weighted sum of the correlation weights, forming a dynamic safety threshold system optimized by federated learning to provide a quantitative basis for the deviation detection of the equipment cluster behavior.

[0066] Among them, the instruction set includes the corrected value of the vibratory frequency of the roller and the upper limit of the traveling speed of the paver, and is sent to the corresponding equipment controller through the edge computing node to perform dynamic regulation.

[0067] S5. Update the virtual construction progress model of the digital twin, and superimpose the adjusted construction standards on the real scene through the AR terminal to guide the workers to correct the deviation, and at the same time encrypt and store the closed-loop processing records in the blockchain.

[0068] Specifically, it includes the following steps: Update the virtual construction progress model of the digital twin according to the instruction set; Specifically, the virtual construction progress model of the digital twin is a dynamic system for simulating and predicting the construction progress, and its state is described by the following variables: : The construction state vector at time , including equipment state, construction progress, environmental parameters, etc.

[0069] : The control input vector at time , including equipment parameter adjustment instructions (such as vibration frequency, temperature threshold).

[0070] : The output vector at time , including construction progress, risk score, AR projection path, etc.

[0071] The state update equation of the virtual construction progress model is expressed as:

[0072] In the formula: is the state transition function, which describes how the construction state changes with time, control input and external disturbances, is the external disturbance vector, including environmental noise, equipment anomalies, etc.; It should be noted that the acquisition of the state transition function needs to be realized through a multi-source data modeling and hybrid drive framework. Specifically, differential equations or discrete event models are constructed based on the physical characteristics of construction equipment (such as vibration energy transfer, material mechanics parameters), combined with historical anomaly data of the federated learning platform (such as equipment failure modes, environmental disturbance records), and spatio-temporal Transformers are used to extract the correlation features of multi-equipment clusters, and data-driven state prediction logic is trained through time series models such as LSTM or Neural ODE; Furthermore, real disturbances (such as sudden changes in temperature and humidity, abnormal equipment shutdown) are injected into the digital twin environment, the actual construction progress data feedback by the AR terminal is used to close-loop calibrate the model parameters, and the global features of multiple projects are aggregated through the federated learning framework to optimize the weights, and finally deployed to the edge computing node to receive control instructions and external disturbances detected by sensors in real time, realizing dynamic iterative update to ensure that the state transition process satisfies both physical laws and data generalization.

[0073] The output equation of the virtual construction progress model is expressed as:

[0074] In the formula: Represents the output function, which describes how the construction status and control inputs are mapped to outputs (such as construction progress, risk score).

[0075] It should be noted that the acquisition of the output function depends on the multi-modal feature fusion and rule-adaptive hybrid architecture. First, define the output targets according to the construction safety specifications (such as risk score formula, AR path planning rules), combine the sensor data (vibration, temperature) and video semantic features, dynamically weight and generate the fusion features through the gated fusion mechanism, and train a lightweight regression model (such as random forest or lightweight neural network) to map to the specific outputs (such as risk score, projection path). At the same time, measure the distribution shift between the construction status and the safety benchmark based on KL divergence, and dynamically optimize the risk threshold and path avoidance strategy through reinforcement learning; Simulate extreme scenarios (such as equipment failure, sudden rainfall) in the digital twin model to verify the output robustness, and finally deploy it to the cloud-edge collaborative architecture to support real-time calculation of the risk score and trigger dynamic re-planning of the AR path, and automatically encrypt and store the warning records to the IOTA Tangle blockchain through smart contracts to ensure the compliance, auditability and real-time response ability of the output logic.

[0076] Overlay the adjusted construction standards (such as the vibration frequency of the roller) to the real scene through the AR terminal; Dynamically re-plan the AR projection path (avoiding high-risk areas); Encrypt and store the closed-loop processing records (such as adjusted construction parameters, worker operation records) to the blockchain to achieve fee-free and high-concurrency evidence storage (based on the IOTA Tangle architecture).

[0077] In summary, through the high-precision spatio-temporal alignment algorithm and the adaptive multi-modal fusion architecture, the present invention significantly optimizes the real-time performance and accuracy of construction safety risk assessment. Specifically, the spatio-temporal feature alignment mechanism based on physical constraints effectively suppresses the noise interference in the multi-source data synchronization process and ensures the integrity of the vibration conduction characteristics of the equipment cluster; the dynamic gated fusion module automatically adjusts the contribution weights of the sensor and video semantic features through environmental perception, and realizes the collaborative optimization of multi-modal information under complex working conditions; the pulse-coded driven distributed learning framework greatly improves the anomaly detection efficiency, and through the event-triggered feature compression and adaptive weight update mechanism, realizes the rapid response and global optimization of the behavior of the equipment cluster. Thereby significantly reducing the probability of misjudgment of safety risks and at the same time improving the execution timeliness of warning instructions.

[0078] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention, and they should all be covered by the scope of the claims of the present invention.

Claims

1. A highway information-based intelligent management and control system, characterized by: include, Collect multi-source data in real time, upload it to the edge computing node for timestamp alignment and format standardization, input it into the spatiotemporal Transformer fusion model, extract the spatiotemporal features of sensor data and the semantic features of video images through the multi-head attention mechanism, and generate dynamic risk scores; Trigger graded warning signals based on dynamic risk scores, execute AR projection instructions, push red alarms, generate rectification work orders, and lock metering payments; The abnormal pattern features of the device cluster of the current project are encrypted and uploaded to the cloud through the federated learning platform, and compared with the global feature library shared by the historical projects to perform cross-project optimized group anomaly detection. Perform online reasoning on real-time construction data streams, and automatically generate instruction sets containing adjustment parameters when it detects that the behavior of the equipment cluster deviates from the safety threshold of federated learning optimization; Update the virtual construction progress model of the digital twin, and superimpose the adjusted construction standards on the real scene through the AR terminal to guide workers to correct deviations. At the same time, the closed-loop processing records are encrypted and stored in the blockchain.

2. The highway information intelligent management and control system according to claim 1, characterized in that: The specific steps of generating a dynamic risk score are as follows: Adopting a hierarchical feature fusion architecture as the spatiotemporal Transformer fusion model; Capture the spatial relationship between devices through spatial attention flow and generate topological features; Apply causal masks to extract the temporal dependencies of device operating parameters and establish dynamic feature expressions across time steps; The spatiotemporal features of sensor data and the semantic features of video images are dynamically fused through a gating mechanism.

3. The highway information intelligent management and control system according to claim 1, characterized in that: The hierarchical warning signal refers to constructing a three-level dynamic threshold, and dynamically adjusting the threshold range according to the real-time construction stage to trigger corresponding measures.

4. The highway information intelligent management and control system according to claim 2, characterized in that: The device cluster abnormal pattern features refer to the sensor topology features extracted by the spatial attention stream and the temporal dynamic features extracted by the temporal attention stream, combined with the multimodal feature vector after the gating fusion of the video semantic features, and the spatiotemporal pulse graph pattern features generated by the pulse coding layer.

5. The highway information intelligent management and control system according to claim 4, characterized in that: The specific steps of the cross-project optimized group anomaly detection are as follows: Collect the operating indicator data of multiple projects, pre-process the data and build a multi-project spatiotemporal feature matrix; Analyze the spatiotemporal correlation of indicators between projects and extract the spatiotemporal correlation characteristics across projects; Build a cross-project optimized anomaly detection model based on transfer learning, and perform cross-project feature migration by dynamically adjusting the anomaly detection model parameters; Training the anomaly detection model by combining group behavior patterns and historical anomaly data to generate anomaly probability distribution; The anomaly threshold is dynamically adjusted according to real-time detection data, and the group anomaly detection results are output.

6. The highway information intelligent management and control system according to claim 5, characterized in that: The virtual construction progress model of the digital twin has the following specific steps: Collect multi-source heterogeneous data from the construction site, integrate BIM models, sensor monitoring data and resource scheduling plans, and build a dynamically updated virtual construction scene; Based on the discrete event simulation principle, a construction progress logical relationship model is established, and an initial progress plan is generated through time series analysis and resource constraint optimization. By combining real-time construction data with historical deviation rules, the task node parameters in the virtual construction scene are dynamically corrected to generate progress prediction and risk warning results. Adjust resource allocation and task priority through multi-objective optimization algorithm to output the optimal construction schedule.

7. The highway information intelligent management and control system according to claim 1, characterized in that: The multi-source data include roller vibration spectrum, paver temperature curve, prestressed tension value, and construction images captured by video surveillance equipment.

8. The highway information intelligent management and control system according to claim 1, characterized in that: The timestamp alignment and format standardization include defining a reference time sequence and using a reference signal interpolation alignment method to synchronize sensor data and video streams.

9. The highway information intelligent management and control system according to claim 1, characterized in that: The abnormal mode characteristics of the equipment cluster include the deviation of the vibration spectrum of the roller from the mean The spatiotemporal distribution map and the coordinates of the abnormal temperature gradient area of ​​the paver.

10. The highway information intelligent management and control system according to claim 1, characterized in that: The safety threshold refers to a dynamic discrimination boundary optimized by federated learning; The instruction set includes the vibration frequency correction value of the roller and the upper limit of the paver's travel speed, and is sent to the corresponding equipment controller through the edge computing node to perform dynamic control.

Images