Communication engineering construction dynamic optimization method based on multi-dimensional perception

Abstract

The invention discloses a communication engineering construction dynamic optimization method based on multi-dimensional perception, and the method comprises the steps: collecting the multi-dimensional perception data of a construction site in real time through a multi-source heterogeneous sensor network, including environment parameters, equipment operation states, construction progress and personnel behavior data; performing space-time alignment processing on the multi-dimensional sensing data, and constructing a space-time associated dynamic construction digital twinborn body; establishing a dynamic optimization model based on a reinforcement learning algorithm, taking construction efficiency maximization, resource loss minimization and safety risk minimization as target optimization functions, and embedding risk constraint conditions; inputting the dynamic construction digital twin into the dynamic optimization model, and outputting a multi-dimensional parameter optimization instruction set comprising equipment scheduling parameters, construction path parameters and resource configuration parameters; and performing real-time performance evaluation on the multi-dimensional parameter optimization instruction set through edge computing nodes, and dynamically adjusting weight parameters of the model to form closed-loop feedback control. The purpose of cooperatively improving the construction efficiency, the safety and the economical efficiency is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of communication engineering, and particularly relates to a dynamic optimization method for communication engineering construction based on multi-dimensional perception. Background Art

[0002] With the large-scale deployment of 5G network construction and communication infrastructure, communication engineering construction faces severe challenges in a complex dynamic environment. Traditional construction optimization methods mainly rely on manual experience and static planning models, and there are the following technical bottlenecks: Single data perception dimension: Existing technologies mostly use single-type sensors (such as position or temperature and humidity sensors), lacking the multi-dimensional real-time perception ability of equipment operation status, personnel behavior, and environmental parameters, resulting in serious data island phenomena; Insufficient dynamic response ability: Traditional static planning models cannot adapt to the real-time changes at the construction site (such as sudden equipment failures, weather changes), and it is difficult to achieve dynamic balance optimization of construction efficiency, resource consumption, and safety risks; Risk prediction lag: Existing safety warning systems mostly rely on threshold trigger mechanisms, lacking the ability to model the multi-dimensional risk conduction chain of equipment-person-environment, resulting in delayed risk response; Lack of collaborative control: The resource scheduling, path planning, and safety control modules operate independently, lacking a global collaborative mechanism, and the construction efficiency improvement rate is insufficient.

[0003] Although there are current studies attempting to introduce digital twin technology, there are generally problems such as low physical simulation accuracy and lag in dynamic model updates, which are difficult to meet the high-precision construction requirements of communication engineering. In addition, the optimization method based on the rule engine is prone to policy oscillation in complex scenarios, seriously affecting construction continuity. Summary of the Invention

[0004] In order to overcome the deficiencies of the prior art, the present invention provides a dynamic optimization method for communication engineering construction based on multi-dimensional perception, which is used to solve the problems of low efficiency, resource waste, and safety hazards caused by the complex and dynamic environment in communication engineering construction, so as to achieve the purpose of synergistically improving construction efficiency, safety, and economy.

[0005] To solve the above problems, the technical solutions adopted by the present invention are as follows: A dynamic optimization method for communication engineering construction based on multi-dimensional perception, comprising the following steps: Real-time collect multi-dimensional perception data of the construction site through a multi-source heterogeneous sensor network, including environmental parameters, equipment operation status, construction progress, and personnel behavior data; Perform spatio-temporal alignment processing on the multi-dimensional perception data to construct a spatio-temporally correlated dynamic construction digital twin; A dynamic optimization model is established based on the reinforcement learning algorithm, with the maximization of construction efficiency, the minimization of resource consumption and safety risks as the objective optimization function, and risk constraint conditions are embedded; The dynamic construction digital twin is input into the dynamic optimization model, and a multi-dimensional parameter optimization instruction set including equipment scheduling parameters, construction path parameters and resource allocation parameters is output; Through the edge computing node, the real-time performance evaluation of the multi-dimensional parameter optimization instruction set is carried out, and the weight parameters of the dynamic optimization model are dynamically adjusted to form a closed-loop feedback control.

[0006] As a preferred embodiment of the present invention, the multi-source heterogeneous sensor network includes: Vibration sensors and GNSS positioning modules deployed on construction machinery; Temperature and humidity sensors, noise sensors and image acquisition devices set on the construction site; Wearable construction personnel positioning terminals and physiological parameter monitoring devices; RFID tracking system for material transport vehicles.

[0007] As a preferred embodiment of the present invention, constructing a spatio-temporal associated dynamic construction digital twin includes: Performing spatio-temporal calibration on the multi-dimensional perception data through an improved Kalman filtering algorithm to eliminate the clock deviation between devices and the difference in the spatial coordinate system; Establishing a multi-dimensional association rule base of equipment-person-environment based on the knowledge graph; Using the digital twin engine to generate a physically accurate three-dimensional visual construction scene; Among them, an adaptive noise covariance matrix including the variance of environmental noise, the variance of equipment noise, and the variance of human noise is introduced in the improved Kalman filtering algorithm.

[0008] As a preferred embodiment of the present invention, when establishing a multi-dimensional association rule base of equipment-person-environment, it includes: Defining the entity types and association attributes of the construction scene based on the dynamic ontology modeling technology; Using an improved Apriori algorithm to mine the frequent association patterns of multi-source heterogeneous data, and dynamically adjusting the confidence of the association rules through a time decay factor; Constructing a multi-dimensional association tensor of equipment-person-environment, extracting implicit association rules through a tensor decomposition algorithm, and dynamically adjusting the rule weights based on a fuzzy logic controller; Deploying a real-time rule conflict detection mechanism, using a three-step gradient descent method to eliminate the logical contradiction between the equipment scheduling instructions and the environmental constraints, and generating a multi-dimensional association rule instance library with spatio-temporal tags.

[0009] As a preferred embodiment of the present invention, when generating a three-dimensional visual construction scenario, it includes: Construct a hybrid physical engine based on the coupling of the discrete element method and the finite element method to perform multi-field joint simulation of the movement trajectory of construction machinery, the stress distribution of materials, and geological deformation; Among them, the discrete element particle diameter is determined by the maximum movement speed of the equipment, the simulation step size, and the number of particle contacts per step.

[0010] As a preferred embodiment of the present invention, when establishing a dynamic optimization model, it includes: Construct a hierarchical reinforcement learning architecture including an upper decision-making layer, a lower execution layer, and a risk prediction layer; Among them, the upper decision-making layer uses the deep deterministic policy gradient algorithm to generate a global optimization strategy; The lower execution layer realizes the optimal allocation of local resources based on the Q-learning algorithm; The risk prediction layer integrates a Bayesian network to real-time evaluate the construction safety risk value; Among them, the deep deterministic policy gradient algorithm adopts a double-delayed deep deterministic policy gradient architecture, and the target network smoothing coefficient τ = 0.005.

[0011] As a preferred embodiment of the present invention, when realizing the optimal allocation of local resources, it includes: Construct a dynamic multi-dimensional state space based on a variational autoencoder, and map the vibration spectrum of construction machinery, the topology of the material transportation path, and the fault warning parameters to the state vector orthogonal basis; Design a multi-objective reward function that combines resource utilization rate, task delay, and safety risk; Deploy a double memory pool experience replay mechanism: store the vibration spectrum characteristics of the equipment in the short term, store the environmental noise data in the long term, and use priority sampling training; Introduce a dynamic ε-greedy strategy to adaptively adjust the exploration probability according to the construction progress; Real-time update the Q-value matrix through the edge computing node, and trigger the Shapley value game equilibrium to generate a spatio-temporal constraint instruction set when there is a resource allocation conflict among multiple devices.

[0012] As a preferred embodiment of the present invention, when real-time evaluating the construction safety risk value, it includes: Construct a dynamic topological Bayesian network, map the equipment vibration spectrum, personnel positioning offset, and meteorological parameters to multi-dimensional evidence variables, and establish a device-person-environment risk conduction chain through a causal inference engine; Based on the sliding time window data stream, real-time update the conditional probability table, and use the variational Bayesian inference algorithm to dynamically optimize the dependence strength weights between nodes; Quantify the deviation between the actual construction status and the safety baseline through KL divergence. When the joint probability of multiple risk nodes is detected to exceed the threshold, trigger the emergency plan matching based on the convolutional neural network; Generate a spatio-temporal correlation risk heat map. The edge computing node feeds back the risk gradient descent direction and real-time corrects the action selection probability distribution of the reinforcement learning policy network.

[0013] As a preferred embodiment of the present invention, when generating a multi-dimensional parameter optimization instruction set, it includes: Based on the spatio-temporal correlation relationship and risk gradient descent direction of the risk heat map, construct a three-dimensional collaborative optimization model of geology-equipment-personnel, and embed the safety risk entropy value into the reward function correction term of the Q-learning algorithm; Impose spatial safety distance constraints on the equipment scheduling parameters based on the risk conduction chain prediction result of the dynamic topology Bayesian network.

[0014] As a preferred embodiment of the present invention, when performing real-time performance evaluation, it includes: Build a dynamic performance simulation sandbox based on the digital twin engine, load the multi-dimensional parameter optimization instruction set and inject real-time perception data streams to generate three-dimensional performance evaluation indicators; Segmentally verify the performance of the multi-dimensional parameter optimization instruction set based on a sliding time window. Calculate the deviation between the actual construction trajectory and the digital twin prediction trajectory within each time window through the grey relational analysis method.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) High-precision dynamic perception and modeling Achieve high-precision spatial positioning and time series alignment through a multi-source heterogeneous sensor network (vibration, GNSS, RFID, etc.), with significantly improved accuracy compared to traditional single-source perception schemes; Improve the Kalman filter algorithm by introducing an adaptive noise covariance matrix (adjustment range 0.5 - 1.2). The abnormal data recognition rate reaches 94.3% in a heavy machinery interference environment, which is 7.2% higher than the fixed parameter scheme; The discrete element-finite element coupled physical engine supports sub-meter-level construction machinery trajectory simulation, and the dynamic LOD technology significantly improves the rendering efficiency.

[0016] (2) Multi-objective collaborative optimization ability The hierarchical reinforcement learning architecture (DDPG + Q-learning + Bayesian network) realizes three-dimensional collaborative optimization of construction efficiency, resource consumption, and safety risk: The construction efficiency improvement rate reaches 23.5% (experimental data), and the resource consumption is reduced by 15.8%; Risk conduction chain modeling shortens the safety accident early warning time to within 30 seconds, with the response speed increased by 6 times compared to traditional methods; The dynamic ε-greedy strategy adaptively adjusts the exploration probability according to the construction progress, reducing the number of policy oscillations to 1.1 times per hour to ensure construction continuity.

[0017] (3) Closed-loop self-correction mechanism The real-time performance evaluation of edge computing nodes combined with the lightweight LSTM model enables online iteration of weight parameters (cycle ≤ 5 seconds) to dynamically adapt to environmental changes; The Q-learning correction driven by the risk heat map significantly improves the response speed of resource evacuation in high-risk areas, greatly reducing the violation rate of safety distance constraints; The digital twin sandbox verification mechanism controls the construction trajectory deviation within the safety threshold, greatly improving the stability of the optimization effect compared to the non-closed-loop control scheme.

[0018] Experimental comparison data

[0019] Note: The experimental data is from a base station construction project and is the test result of continuous operation for 72 hours.

[0020] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments. Description of the drawings

[0021] Figure 1 It is a step diagram of the dynamic optimization method for communication engineering construction based on multi-dimensional perception; Figure 2 It is a flow chart for constructing a dynamic construction digital twin with spatio-temporal correlation; Figure 3 It is a logic block diagram of the dynamic optimization model. Specific embodiments

[0022] See Figure 1 , the dynamic optimization method for communication engineering construction provided by the present invention includes the following steps: Step S1: Real-time collect multi-dimensional perception data of the construction site through a multi-source heterogeneous sensor network. The multi-dimensional perception data includes environmental parameter data, equipment operation status data, construction progress data, and personnel behavior data; Step S2: Perform spatio-temporal alignment processing on the multi-dimensional perception data to construct a dynamic construction digital twin with spatio-temporal correlation; Step S3: Establish a dynamic optimization model based on the reinforcement learning algorithm. The dynamic optimization model takes the maximization of construction efficiency, the minimization of resource consumption and safety risks as the objective optimization function, and embeds risk constraint conditions; Step S4: Input the dynamic construction digital twin into the dynamic optimization model to generate a multi-dimensional parameter optimization instruction set, which includes a collaborative optimization plan for equipment scheduling parameters, construction path parameters, and resource allocation parameters; Step S5: Conduct real-time performance evaluation on the multi-dimensional parameter optimization instruction set through edge computing nodes, and dynamically adjust the weight parameters of the dynamic optimization model according to the evaluation results to form a closed-loop feedback control.

[0023] In the above Step S1, the multi-source heterogeneous sensor network includes: Vibration sensors and GNSS positioning modules deployed on construction machinery; Temperature and humidity sensors, noise sensors, and image acquisition devices set on the construction site; Wearable construction worker positioning terminals and physiological parameter monitoring devices; RFID tracking systems for material transport vehicles.

[0024] Specifically, environmental parameter data refers to the physical environmental indicators that affect construction safety and efficiency within the construction site, including: meteorological parameters, spatial state parameters.

[0025] Among them, meteorological parameters include: temperature, humidity data (collected by temperature and humidity sensors), noise data (collected by noise sensors), wind speed, rainfall (collected by meteorological sensors).

[0026] Spatial state parameters include: geological deformation data (such as soil humidity, geological hardness).

[0027] Equipment operation status data refers to the real-time performance indicators of construction machinery and transportation equipment, including: mechanical body parameters, transportation status parameters, and fault warning parameters.

[0028] Among them, mechanical body parameters include: vibration spectrum (collected by vibration sensors, including frequency and amplitude), positioning coordinates and movement trajectories (collected by GNSS positioning modules).

[0029] Transportation status parameters include: topology of material transportation paths (the RFID tracking system records the real-time positions and path planning of vehicles), maximum movement speed of equipment.

[0030] Fault warning parameters include: abnormal equipment temperature (collected by temperature sensors), hydraulic system pressure fluctuations (comparison of pressure sensor data with historical baselines).

[0031] Construction progress data refers to the quantitative indicators that reflect the actual progress of the project and the deviation from the plan, including: time dimension parameters, resource consumption parameters, and spatial progress parameters.

[0032] Among them, the time dimension parameters include: task completion timestamp (generated by the spatio-temporal alignment of the BIM model and GNSS positioning data), and process delay duration.

[0033] The resource consumption parameters include: material usage (tracked by the RFID system for the unloading frequency of transport vehicles and material consumption), and cumulative equipment operation duration (analyzed by correlating GNSS positioning data with equipment start-stop logs).

[0034] The spatial progress parameters include: completed volume of the structure (calculated by fitting discrete element simulation and image point cloud data), and geological excavation depth (compared by GNSS elevation data and the BIM model).

[0035] Personnel behavior data refers to the operation status and safety-related indicators of construction personnel, including: physiological parameters, behavior trajectory parameters, safety status parameters, and interaction behavior parameters.

[0036] Among them, the physiological parameters include: heart rate (collected in real time by wearable physiological monitoring devices).

[0037] The behavior trajectory parameters include: personnel positioning coordinates, personnel positioning offset (collected by wearable terminals), and personnel movement speed (obtained from positioning data).

[0038] The safety status parameters include: safety equipment wearing data (such as the proportion of the duration without wearing safety equipment), and fatigue index (calculated comprehensively by heart rate variability (HRV) and continuous operation duration).

[0039] The interaction behavior parameters include: equipment operation data (such as operation frequency, equipment operation events and their intervals (such as start, stop, adjustment, etc.)).

[0040] See Figure 2 , in the above step S2, constructing a spatio-temporal correlated dynamic construction digital twin includes: Performing spatio-temporal calibration on multi-dimensional perception data through an improved Kalman filter algorithm to eliminate clock deviation between devices and differences in spatial coordinate systems; Constructing a construction element association model based on knowledge graph technology to establish a multi-dimensional association rule library for equipment-personnel-environment; Using a digital twin engine to generate a three-dimensional visual construction scene with physical accuracy; Among them, an adaptive noise covariance matrix is introduced in the improved Kalman filter algorithm, as shown in formula (1): (1); In the formula, is the covariance matrix, is the environmental dynamic coefficient, which is adjusted in real time according to the noise level at the construction site, and the adjustment range is 0.5 - 1.2; is a diagonal matrix, is the variance of environmental noise, is the variance of device noise, is the variance of human-made noise.

[0041] Specifically, the variance of environmental noise is calculated from data such as temperature, humidity, and noise sensors, = is the noise weight, is the noise variance, is the temperature weight, is the temperature variance, is the humidity weight, is the humidity variance.

[0042] Specifically, the variance of device noise is calculated from the vibration sensor (frequency, amplitude) and GNSS positioning error. Based on the amplitude sequence and frequency sequence collected by the vibration sensor, the dynamic weighted vibration variance is calculated: ; In the formula, is the dynamic weighted vibration variance, is the amplitude sampling value, is the mean amplitude within the sliding time window, is the vibration frequency at the corresponding moment, is the reference frequency (e.g., the rated vibration frequency of the device), is the total number of amplitude data points collected by the vibration sensor within the sliding time window; According to the positioning coordinate sequence of the GNSS module, the three-dimensional positioning error variance is calculated: ; In the formula, , , are the positioning errors in each direction; , , , is the actual coordinate, which can be calibrated through the BIM model.

[0043] Combining the vibration and GNSS positioning error variances, weighted fusion is performed through the dynamic weight coefficient : ; In the formula, , which is dynamically adjusted according to the device type.

[0044] Specifically, the variance of human-made noise Calculated from data such as personnel positioning offset and operation frequency, including: obtaining the variance of positioning offset, the fluctuation of operation frequency, and the abnormality degree of safety status.

[0045] Among them, based on the deviation value sequence of the personnel positioning coordinates from the planned operation path , calculate the variance of positioning offset: ; In the formula, is the variance of positioning offset, is the total number of personnel positioning data points collected by the wearable positioning terminal within the sliding time window; , is the personnel positioning coordinate, is the planned operation path coordinate.

[0046] Based on the equipment operation interval time sequence , calculate the variance of operation frequency fluctuation: ; In the formula, is the variance of operation frequency fluctuation, is the average operation interval time, is the number of interval times of equipment operation events (such as start, stop, adjustment, etc.) within the sliding time window.

[0047] Based on the proportion of the duration of not wearing safety equipment and the deviation of the fatigue index from the baseline value, calculate the variance of safety status abnormality: ; In the formula, is the variance of safety status abnormality, is the proportion of the duration of not wearing safety equipment, is the deviation of the fatigue index from the baseline value, , are weight coefficients (such as , ).

[0048] Fuse the variance of positioning offset, the variance of operation frequency fluctuation, and the variance of safety status abnormality through dynamic weight coefficients to generate the variance of artificial noise : ; In the formula, is the variance of artificial noise, is the dynamic weight coefficient, is the index, is the weight coefficient corresponding to the variance of positioning offset ( ), is the weight coefficient corresponding to the variance of the operation frequency fluctuation ( ), is the weight coefficient corresponding to the variance of the safety status abnormality degree ( ), corresponding to the variance of the positioning offset, corresponding to the variance of the operation frequency fluctuation, corresponding to the variance of the safety status abnormality degree.

[0049] Constraint conditions: .

[0050] Dynamic weight coefficient is optimized according to the real-time construction scenario. For example: When it is detected that personnel frequently deviate from the path ( increase), is increased; When the operation frequency of the equipment is abnormal ( increase), is increased.

[0051] Specifically, the comparison of data fusion accuracy under different environmental dynamic coefficients:

[0052] Test environment: 200 groups of sample data are collected at the construction site with heavy machinery interference. Using the adaptive range adjustment compared with the fixed parameter scheme, the spatial positioning accuracy is improved by 34%, and the time series alignment efficiency is increased by 33%. Among them, When the range is 0.5 - 1.2, the positioning error and time series alignment error are the smallest, and the abnormal data recognition rate is the highest.

[0053] Furthermore, when establishing the multi-dimensional association rule base of equipment - personnel - environment, it includes: Defining the entity types and associated attributes of the construction scenario based on the dynamic ontology modeling technology. The entity types include vibration-sensitive equipment, high-risk operation personnel, and weather-sensitive areas. The associated attributes include spatial dependence, operation coupling coefficient, and risk conduction factor; Using the improved Apriori algorithm to mine the frequent association patterns of multi-source heterogeneous data, and dynamically adjusting the confidence of the association rules through the time decay factor. The calculation of its dynamic confidence threshold is shown in formula (2): (2); In the formula, is the dynamic confidence threshold, is the initial confidence, is the time decay coefficient, is the data aging interval, is the environmental sensitivity coefficient, and are the actual environmental parameter value and the predicted environmental parameter value respectively, is the standard deviation of environmental parameters, is the natural constant; Construct a multi-dimensional correlation tensor of equipment-person-environment, extract implicit correlation rules through a tensor decomposition algorithm, and dynamically adjust the rule weights based on a fuzzy logic controller; Deploy a real-time rule conflict detection mechanism, use the three-step gradient descent method to eliminate the logical contradiction between the equipment scheduling instruction and the environmental constraint, and generate a multi-dimensional correlation rule instance library with spatio-temporal labels.

[0054] Specifically, when the difference between the actual environmental parameters (such as temperature, noise) and the predicted values is significant, the absolute value of the numerator term increases, directly affecting the adjustment of the confidence threshold: Positive difference ( ): The actual environmental value is higher than the prediction, the confidence threshold increases, and rules with higher confidence are required to be accepted; Negative difference ( ): The actual environmental value is lower than the prediction, the confidence threshold decreases, and the rule screening conditions are relaxed.

[0055] The environmental sensitivity coefficient The larger the value, the more significant the impact of environmental differences on the confidence. For example, in meteorologically sensitive areas (such as rainstorm warnings, strong wind warnings), automatically increases to quickly respond to environmental mutations.

[0056] Standard deviation normalization: The difference value is standardized by the standard deviation of environmental parameters to avoid deviations caused by different parameter dimensions.

[0057] Over time, the confidence threshold of historical data gradually decays ( term), ensuring that the model relies more on recent data and adapts to dynamic construction scenarios.

[0058] The present invention uses an improved Apriori algorithm to mine the correlation rules of multi-source heterogeneous data, and the dynamic confidence threshold is used to screen frequent item sets: Threshold self-adaptation: When environmental mutations (such as sudden strong noise) cause to increase, only high-confidence correlation rules (such as "abnormal equipment vibration → safety risk") are retained; Noise filtering: When the environmental difference is small, decreases, allowing more potential rules to participate in the analysis to mine implicit correlations (such as "personnel stay duration + temperature and humidity → fatigue risk").

[0059] Example: Assume the actual temperature value at a certain moment , the predicted value , the standard deviation , the parameter , then: ; Assume , minutes, substitute into formula 2: ; At this time, the confidence threshold is significantly improved, and the system only accepts high-confidence rules to avoid misjudgment in high-temperature environments.

[0060] Furthermore, when generating a three-dimensional visual construction scene, it includes: Construct a hybrid physical engine based on the coupling of the discrete element method and the finite element method, and conduct multi-physical field joint simulations on the movement trajectories of construction machinery, the stress distribution of materials, and geological deformations. The discrete element particle diameter , as shown in formula 3: (3); In the formula, is the maximum movement speed of the equipment, is the simulation step size, is the number of particle contacts per step; Specifically, the simulation step size is dynamically adjusted according to the construction progress (such as the task delay duration ). If the progress lags behind, shorten to increase the simulation frequency (such as adjusting from 1 second to 0.5 seconds) and accelerate the update speed of the digital twin.

[0061] Geological deformation data (such as soil moisture and geological hardness) affects the number of particle contacts per step . For example, in high-humidity soil, the viscosity of particles increases, value increases, and the higher the geological hardness, the greater the value, reflecting more particle contacts.

[0062] Adopt ray tracing and point cloud fusion technology to construct a multi-level rendering pipeline, perform sub-meter spatial registration on GNSS positioning data and BIM models, and realize dynamic deformation visualization by injecting equipment vibration spectrum parameters through vertex shaders; Deploy dynamic LOD (Level of Detail) technology, and dynamically adjust the model detail level based on the relative movement speed obtained from the mechanical body parameters and personnel behavior data, as shown in formula 4: (4); In the formula, is the level of detail, is the screen pixel density, is the frequency threshold of the human eye's visual persistence, is the floor function symbol; Integrate a real-time collision detection algorithm. When the spatial overlap between the working radius of the device and environmental obstacles exceeds the safety threshold , trigger warning coloring based on the stress criterion and generate a risk evolution animation with time sequence marks.

[0063] Specifically, calculate the relative speed through the moving speed of personnel and the mechanical movement trajectory. For example, when personnel approach high-speed machinery increases, trigger an increase in the level of model detail ( rises).

[0064] See Figure 3 , in step S3 above, when establishing a dynamic optimization model based on the reinforcement learning algorithm, it includes: Construct a hierarchical reinforcement learning architecture including a decision-making upper layer, an execution lower layer, and a risk prediction layer; Among them, the decision-making upper layer uses the deep deterministic policy gradient algorithm to generate a global optimization policy; The execution lower layer realizes the optimal allocation of local resources based on the Q-learning algorithm; The risk prediction layer integrates a Bayesian network to evaluate the construction safety risk value in real time; Among them, the deep deterministic policy gradient algorithm adopts a double-delayed deep deterministic policy gradient architecture, sets the policy network update frequency to 1 / 3 of the value network update frequency, and the target network smoothing coefficient τ = 0.005.

[0065] Specifically, verify the influence of the selection of the τ value on the construction optimization effect through a comparative experiment:

[0066] Experimental conditions: Run in series for 72 hours in the base station construction scenario, compare the optimization effects of construction machinery scheduling schemes under different τ values. The results show that when τ = 0.005, the best balance is achieved in terms of construction efficiency improvement, loss reduction, and strategy stability.

[0067] Furthermore, when realizing the optimal allocation of local resources, it includes: Construct a dynamic multi-dimensional state space, map the vibration spectrum of construction machinery, the topology of the material transportation path, and the fault warning parameters to the orthogonal basis of the state vector, and perform feature dimensionality reduction on the high-dimensional state space through a variational autoencoder; Design a multi-objective reward function, integrating the resource utilization rate improvement coefficient and the task delay penalty factor and the safety risk coefficient , and its reward value is calculated as shown in Formula 5: (5); In the formula, is the reward value at time , is the actual resource utilization rate, is the baseline utilization rate, is the task delay time, is the type of safety risk index, is the type of safety risk index's partial derivative with respect to time , is the number of categories of safety risk indices; Deploy a dual experience replay mechanism, store the device vibration spectrum characteristics and environmental noise data in the short-term memory pool and long-term memory pool respectively, and adopt a prioritized sampling strategy to extract training samples; Specifically, the vibration spectrum, the topology of the material transportation path, and the fault warning parameters are used as the orthogonal basis of the state vector, which affects the actual resource utilization rate : ; In the formula, is the th orthogonal basis component of the state vector (such as the corresponding component of the vibration spectrum, the corresponding component of the topology of the material transportation path, and the corresponding component of the fault warning parameter); is the th weight coefficient, which is dynamically optimized through reinforcement learning; is the number of orthogonal basis components, is the bias term, is the Sigmoid activation function, which normalizes the output to [0,1].

[0068] Fault warning parameters such as abnormal device temperature and pressure fluctuations in the hydraulic system participate in the reward calculation through the safety risk coefficient ( ): Quantify the deviation between the actual pressure of the hydraulic system and the safety baseline as the pressure fluctuation safety risk index: ; In the formula, is the pressure fluctuation safety risk index, is the maximum allowable pressure deviation; Apply the pressure fluctuation safety risk index to the reward value calculation as follows: ; Introduce a dynamic ε-greedy strategy and adaptively adjust the exploration probability based on the construction progress as shown in Equation (6): (6); wherein is the exploration probability value, is the initial exploration probability value, is the minimum exploration probability threshold, is the exploration decay exponent, is the maximum value of the construction progress; The Q-value matrix is updated in real time through the edge computing node. When detecting multi-device resource allocation conflicts, trigger a game equilibrium based on the Shapley value to generate a resource reallocation instruction set with spatio-temporal constraints.

[0069] Specifically, the construction progress includes: time dimension parameters (task completion timestamp, process delay duration), spatial progress parameters (completed volume of the structure, geological excavation depth).

[0070] Progress-driven attenuation: The higher the construction progress , the smaller the term, and the exploration probability value

[0071] decreases, and the system is more inclined to utilize the known optimal strategy. Dynamic response: If the construction progress stalls due to equipment failure or environmental mutation ( does not increase as expected), the exploration probability value

[0072] is maintained at a high level to encourage exploration of new strategies to break through the bottleneck. Furthermore, when real-time evaluating the construction safety risk value, it includes: Construct a dynamic topological Bayesian network, map the device vibration spectrum, personnel positioning offset and meteorological parameters into multi-dimensional evidence variables of risk nodes, and establish a three-dimensional risk conduction chain of equipment failure, personnel behavior mistakes and environmental mutation through a causal inference engine; (7); wherein is the posterior probability of hypothesis under the evidence , is the prior probability of the node corresponding to the th evidence variable under the hypothesis , is the The gradient change rate of an evidence variable, is the time decay factor, is other possible hypotheses except the hypothesis ; is the corresponding node prior probability of the th evidence variable under other possible hypotheses, is the number of evidence variables; Deploy risk entropy value fusion, quantify the deviation between the actual construction state and the safety baseline through KL divergence, and trigger the emergency plan matching based on the convolutional neural network when the joint probability of multiple risk nodes exceeds the threshold ; Generate a risk heat map with spatio-temporal correlation relationships, feedback the risk gradient descent direction to the construction machinery control system through edge computing nodes, and real-time correct the action selection probability distribution of the reinforcement learning policy network.

[0073] Specifically, take meteorological parameters as evidence variables , and its gradient change rate ( , such as the change of temperature, humidity, noise, wind speed, rainfall over time) affects the calculation of the posterior probability of the Bayesian network and is used for risk conduction chain modeling.

[0074] Take the equipment vibration spectrum as an evidence variable , and its gradient change rate ( , such as the change of frequency and amplitude over time) affects the calculation of the posterior probability of the Bayesian network and is used for risk conduction chain modeling.

[0075] Take the personnel positioning offset as an evidence variable , and its gradient change rate ( , such as the change of positioning offset over time) affects the calculation of the posterior probability of the Bayesian network, and the risk conduction chain is modeled through the Bayesian network.

[0076] In step S4 above, when generating the multi-dimensional parameter optimization instruction set, it includes: Based on the spatio-temporal correlation relationship and risk gradient descent direction of the risk heat map, construct a three-dimensional collaborative optimization model of geology-equipment-personnel, and embed the safety risk entropy value into the reward function correction term of the Q-learning algorithm. Its correction function is shown in formula 8: (8); In the formula, is the corrected reward value at time , is the risk constraint intensity coefficient, is the KL divergence between the real-time risk distribution and the safety baseline, The time-varying gradient of the risk indicator; Based on the prediction results of the risk conduction chain by the dynamic topological Bayesian network, spatial safety distance constraints are imposed on the device scheduling parameters.

[0077] Specifically, the geological deformation data changes the real-time risk distribution through the risk conduction chain and the safety baseline The difference is quantified by the KL divergence: ; In the formula, is the risk hypothesis (such as too high soil humidity, too low geological hardness); is the current risk probability; is the preset probability of the safety baseline; The KL divergence quantifies the deviation between the actual risk distribution and the safety baseline, and corrects the reward value of Q-learning.

[0078] The device vibration spectrum predicts the risk conduction chain through the Bayesian network, generates safety distance constraints (such as the minimum interval between devices), and embeds the reward correction of Q-learning.

[0079] The deviation of the fatigue index from the baseline value and the proportion of the duration without wearing safety equipment affect the KL divergence calculation through the safety risk entropy value ( ), and dynamically correct the reward function: Quantify the deviation of the fatigue index from the baseline value as the fatigue risk entropy value: ; In the formula, is the fatigue risk entropy value, is the maximum allowable deviation value of the fatigue index; Map the proportion of the duration without wearing safety equipment directly to the wearing risk entropy value: ; In the formula, is the wearing risk entropy value; Fuse the fatigue risk entropy value and the wearing risk entropy value into the comprehensive risk entropy value through the dynamic weight : ; Weight allocation: ; The weight is dynamically adjusted according to the change rate of the risk entropy value to quickly respond to sudden risks.

[0080] Integrate risk entropy value embedding with KL divergence calculation to correct the deviation measure between real-time risk distribution and safety baseline: ; Correction mechanism: When (high-risk personnel status), the KL divergence value is amplified to strengthen risk punishment; When (low-risk personnel status), the KL divergence maintains the original calculated value.

[0081] In step S5 above, when performing real-time performance evaluation and dynamically adjusting the dynamic optimization model, it includes: Build a dynamic performance simulation sandbox based on the digital twin engine, load the multi-dimensional parameter optimization instruction set and inject real-time perception data stream to generate three-dimensional performance evaluation indicators including resource consumption rate, construction progress deviation amount, and risk entropy value; Perform segmented performance verification on the multi-dimensional parameter optimization instruction set based on the sliding time window mechanism. Calculate the deviation degree between the actual construction trajectory and the digital twin prediction trajectory within each time window through the grey relational analysis method, as shown in formula 9: (9); In the formula, is the deviation degree, is the actual value of the -th dimension parameter, is the digital twin prediction value of the -th dimension, is the environmental noise tolerance threshold, is a very small constant to prevent division by zero, is the number of dimensions, is the -th dimension weight; When the deviation degree exceeds the preset safety boundary, trigger parameter sensitivity analysis based on the generative adversarial network to identify the key parameter dimensions that cause performance decay; Predict the time-series impact of weight parameter adjustment through the lightweight LSTM model deployed on the edge side, generate a feedback signal including the gradient correction direction and adjustment step size, and drive the dynamic optimization model to perform online parameter iteration.

[0082] Specifically, the deviation degree calculation parameters include: resource consumption rate, construction progress deviation amount, and risk entropy value parameters. Among them, the resource consumption rate parameters include: material usage rate (obtained from the material usage amount), equipment cumulative operation rate (obtained from the equipment cumulative operation duration), the construction progress deviation amount parameters include: task completion time deviation (obtained from the task completion timestamp), and the risk entropy value parameters include: fatigue risk entropy value, wearing risk entropy value.

[0083] The above embodiments are only preferred embodiments of the present invention, and the scope of protection of the present invention cannot be limited thereby. Any non-substantive changes and substitutions made by those skilled in the art based on the present invention fall within the scope of protection required by the present invention.

Claims

1. A dynamic optimization method for communication engineering construction based on multi-dimensional perception, characterized in that It includes the following steps: Collect multi-dimensional perception data of the construction site in real time through a multi-source heterogeneous sensor network, including environmental parameters, equipment operation status, construction progress, and personnel behavior data; Perform spatio-temporal alignment processing on the multi-dimensional perception data to construct a spatio-temporally correlated dynamic construction digital twin; Establish a dynamic optimization model based on the reinforcement learning algorithm, with the maximization of construction efficiency, the minimization of resource consumption and safety risks as the objective optimization function, and embed risk constraint conditions; Input the dynamic construction digital twin into the dynamic optimization model, and output a multi-dimensional parameter optimization instruction set including equipment scheduling parameters, construction path parameters, and resource allocation parameters; Perform real-time effectiveness evaluation on the multi-dimensional parameter optimization instruction set through edge computing nodes, dynamically adjust the weight parameters of the dynamic optimization model, and form a closed-loop feedback control.

2. The dynamic optimization method for communication engineering construction based on multi-dimensional perception according to claim 1, characterized in that The multi-source heterogeneous sensor network includes: Vibration sensors and GNSS positioning modules deployed on construction machinery; Temperature and humidity sensors, noise sensors, and image acquisition devices set on the construction site; Wearable construction personnel positioning terminals and physiological parameter monitoring devices; RFID tracking systems for material transport vehicles.

3. The dynamic optimization method for communication engineering construction based on multi-dimensional perception according to claim 1, wherein Constructing a spatio-temporally correlated dynamic construction digital twin includes: Perform spatio-temporal calibration on the multi-dimensional perception data through an improved Kalman filter algorithm to eliminate clock deviations between devices and differences in spatial coordinate systems; Establish a multi-dimensional association rule library for equipment-person-environment based on a knowledge graph; Generate a physically accurate three-dimensional visual construction scene using a digital twin engine; Among them, the improved Kalman filter algorithm introduces an adaptive noise covariance matrix including the variance of environmental noise, the variance of equipment noise, and the variance of human noise.

4. The dynamic optimization method for communication engineering construction based on multi-dimensional perception according to claim 3, characterized in that When establishing a multi-dimensional association rule library for equipment-person-environment, it includes: Define the entity types and association attributes of the construction scene based on dynamic ontology modeling technology; Adopt an improved Apriori algorithm to mine frequent association patterns of multi-source heterogeneous data, and dynamically adjust the confidence of association rules through a time decay factor; Construct a multi-dimensional association tensor for equipment-person-environment, extract implicit association rules through a tensor decomposition algorithm, and dynamically adjust the rule weights based on a fuzzy logic controller; Deploy a real-time rule conflict detection mechanism, and use the three-step gradient descent method to eliminate the logical contradiction between equipment scheduling instructions and environmental constraints, and generate a multi-dimensional association rule instance library with spatio-temporal tags.

5. The dynamic optimization method for communication engineering construction based on multi-dimensional perception according to claim 3, wherein When generating a three-dimensional visual construction scene, it includes: Construct a hybrid physical engine based on the coupling of the discrete element method and the finite element method to perform multi-field joint simulations of construction machinery movement trajectories, material stress distributions, and geological deformations; Among them, the discrete element particle diameter is determined by the maximum movement speed of the equipment, the simulation step size, and the number of particle contacts per step.

6. The dynamic optimization method for communication engineering construction based on multi-dimensional perception according to claim 1, characterized in that, When establishing a dynamic optimization model, it includes: Construct a hierarchical reinforcement learning architecture including a decision-making upper layer, an execution lower layer, and a risk prediction layer; Among them, the decision-making upper layer uses the deep deterministic policy gradient algorithm to generate a global optimization strategy; The execution lower layer realizes the optimal allocation of local resources based on the Q-learning algorithm; The risk prediction layer integrates a Bayesian network to real-time evaluate the construction safety risk value; Among them, the deep deterministic policy gradient algorithm adopts a double-delayed deep deterministic policy gradient architecture, and the target network smoothing coefficient τ = 0.

005.

7. The dynamic optimization method for communication engineering construction based on multi-dimensional perception according to claim 6, characterized in that, When implementing the optimal allocation of local resources, it includes: Constructing a dynamic multi-dimensional state space based on a variational autoencoder, and mapping the vibration spectrum of construction machinery, the topology of the material transportation path, and the fault warning parameters to the state vector orthogonal basis; Designing a multi-objective reward function that integrates resource utilization rate, task delay, and safety risk; Deploying a double-memory pool experience replay mechanism: storing the vibration spectrum characteristics of the device in the short term and the environmental noise data in the long term, and adopting prioritized sampling training; Introducing a dynamic ε-greedy strategy to adaptively adjust the exploration probability according to the construction progress; Updating the Q-value matrix in real time through edge computing nodes, and triggering the Shapley value game equilibrium to generate a spatio-temporal constraint instruction set when there are resource allocation conflicts among multiple devices.

8. The dynamic optimization method for communication engineering construction based on multi-dimensional perception according to claim 7, characterized in that When evaluating the construction safety risk value in real time, it includes: Constructing a dynamic topological Bayesian network, mapping the device vibration spectrum, the personnel positioning offset, and the meteorological parameters to multi-dimensional evidence variables, and establishing a device-person-environment risk conduction chain through a causal inference engine; Updating the conditional probability table in real time based on the sliding time window data stream, and dynamically optimizing the dependence strength weights between nodes using the variational Bayesian inference algorithm; Quantifying the deviation degree between the actual construction state and the safety baseline through KL divergence, and triggering the emergency plan matching based on a convolutional neural network when the joint probability of multiple risk nodes exceeds the threshold; Generating a spatio-temporal correlation risk heat map, and the edge computing node feedbacks the risk gradient descent direction and corrects the action selection probability distribution of the reinforcement learning policy network in real time.

9. The dynamic optimization method for communication engineering construction based on multi-dimensional perception according to claim 8, wherein When generating a multi-dimensional parameter optimization instruction set, it includes: Based on the spatio-temporal correlation relationship and the risk gradient descent direction of the risk heat map, constructing a three-dimensional collaborative optimization model of geology-device-personnel, and embedding the safety risk entropy value into the reward function correction term of the Q-learning algorithm; Applying a spatial safety distance constraint to the device scheduling parameters based on the risk conduction chain prediction result of the dynamic topological Bayesian network.

10. The dynamic optimization method for communication engineering construction based on multi-dimensional perception according to claim 1, characterized in that When conducting real-time performance evaluation, it includes: Constructing a dynamic performance simulation sandbox based on a digital twin engine, loading the multi-dimensional parameter optimization instruction set and injecting the real-time perception data stream to generate three-dimensional performance evaluation indicators; Verifying the performance of the multi-dimensional parameter optimization instruction set in segments based on the sliding time window, and calculating the deviation degree between the actual construction trajectory and the digital twin prediction trajectory through the grey relational analysis method within each time window.

Images