Beam field production state real-time monitoring and analyzing system and method based on big data

Abstract

The invention relates to the field, and discloses a beam field production state real-time monitoring and analyzing system and method based on big data. Comprising a data acquisition unit, a data stream processing unit, a digital twin modeling unit, a reinforcement learning decision unit, an edge-cloud collaboration unit and a visual feedback unit. Multi-source heterogeneous data of beam body positioning, hoisting equipment, energy consumption monitoring and a scheduling system are collected, and beam field operation data flow aligned through a time sequence is generated through ETL process standardization processing and flow type aggregation. Constructing a digital twinborn scene based on a BIM model, and dynamically mapping operation data to form a beam field state matrix; and a safe and efficient beam moving and production scheduling scheme is generated by combining conditions such as equipment load, path avoidance, process time window and energy consumption constraint. A low-time-delay decision is realized by deploying a lightweight model at an edge node, and a moving track and early warning information are displayed in a BIM scene in a three-dimensional manner, so that real-time monitoring and dynamic optimization of beam field production are realized.

Description

Technical Field

[0001] This invention relates to the field of digital engineering construction, specifically to a real-time monitoring and analysis system and method for beam yard production status based on big data. Background Technology

[0002] In recent years, precast beam yards have been a crucial link in the construction of large-scale infrastructure projects such as railways, highways, and urban rail transit, directly impacting the overall project schedule and cost control due to their production scheduling efficiency and safety status. With the expansion of beam yard scale, the increase in equipment types, and the frequent overlap of operations, traditional management methods relying on manual inspections, experience-based scheduling, and offline statistics are no longer sufficient to meet the demands for real-time and refined management. Therefore, it is necessary to leverage information and intelligent technologies to achieve dynamic monitoring and scheduling optimization of the entire beam yard production process.

[0003] Existing beam yard management systems mostly employ independent sensor networks or simple PLC control architectures, capable of only basic monitoring of equipment operation status. They lack the ability to comprehensively analyze multi-dimensional data such as beam storage location, hoisting path, energy consumption level, and operation sequence. In the scheduling process, tasks are often assigned using static rules or manually set priorities, failing to effectively consider the dynamic relationships between multiple factors such as equipment load, path conflicts, and energy consumption constraints. Furthermore, issues such as inconsistent data acquisition port standards, lagging model updates, and low visualization levels lead to information silos, slow response times, and resource waste in the beam yard production process, limiting overall construction efficiency and safety. To address these issues, we propose a real-time monitoring and analysis system and method for beam yard production status based on big data. This system aims to solve the problems of data silos, limited analysis dimensions, lack of dynamic optimization in scheduling, and slow response times in existing technologies, achieving refined, real-time management and intelligent scheduling of the beam yard production process. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a real-time monitoring and analysis system and method for beam yard production status based on big data, in order to solve the technical problems existing in the prior art.

[0005] The above-mentioned technical objective of the present invention is achieved through the following technical solution: The real-time monitoring and analysis system for beam yard production status based on big data includes: a data acquisition unit, which collects multi-source heterogeneous data generated by beam positioning sensors, hoisting equipment monitoring terminals, energy consumption monitors and scheduling command systems in the beam yard, and performs standardized extraction and transformation through an ETL data processing flow designed for the beam yard scenario. The standardization process includes: unifying the beam coding format, specifying the sensor sampling frequency, using IoT communication protocols to connect data, and mapping data fields from different sources to a unified structured format to achieve standardized access to multi-source data; The data stream processing unit is used to construct a data processing channel based on a stream processing framework to perform real-time aggregation and time synchronization of the multi-source data. The time synchronization is based on an event-time mechanism. By setting the sliding window length, window step size, and maximum allowable time deviation, when the difference between the recorded time and the window reference time does not exceed the maximum allowable time deviation, the record participates in the aggregation process to ensure the time alignment accuracy of data collected by different devices, thereby generating a time-aligned beam yard operation data stream. The digital twin modeling unit is used to build a three-dimensional digital twin scene of the beam yard based on the BIM model and map the running data stream into a state matrix that is dynamically updated over time. The state matrix is ​​a three-dimensional data structure, divided into grids along the spatial dimension. The first and second dimensions correspond to the horizontal and vertical partitions of the beam yard space, respectively. The third dimension is the channel dimension, which is used to record various feature data such as beam number, spatial location, operation status and energy consumption information, to reflect the real-time production status of the beam yard. A reinforcement learning decision-making unit is used to generate beam transport paths and production scheduling schemes based on the state matrix. The reinforcement learning decision unit includes: a) Feature extraction module, used to extract spatial and topological features from the state matrix using a convolutional residual network to form a feature map for policy training; b) The policy training module employs an improved deep Q-learning algorithm for scheduling optimization. The improvements include: Improve the learning performance of rare but high-value samples by setting sample weights; Simulation training is conducted using a multi-process parallel approach to improve policy convergence speed; Based on constraints such as equipment load limit, path safety distance, process priority, process time window, and energy consumption limit, action instructions that do not meet the requirements are filtered out, and the optimal solution that meets the on-site safety and efficiency requirements is generated. The reinforcement learning decision unit defines an action set A in each state to describe the optional beam transport instruction set, and the action set serves as the action input for policy training and Q-value update; The Q-value update satisfies the temporal difference objective; Among them, the system determines the conflict flag under the path safety distance constraint, which is used for reward function calculation; Edge-cloud collaborative units are used to deploy lightweight inference models on edge computing nodes to achieve low-latency policy execution; The lightweight model is optimized through network pruning and parameter quantization to ensure that the decision response time meets the preset response time threshold. Only key abnormal samples and state data with large deviations are uploaded to the cloud for retraining and parameter synchronization, thus forming a closed-loop optimization architecture of cloud training, edge execution and data feedback. The visualization feedback unit is used to display the optimal beam moving path, equipment operating status and abnormal warning information in a three-dimensional manner in the BIM digital twin scenario, so as to realize the visualization feedback and real-time monitoring of scheduling results.

[0006] Preferably, the standardized extraction and transformation of the data acquisition unit includes the following steps: a) Unified data format: A unified data standard model is established for data from different sources. The beam number adopts a three-segment coding rule of "project number - beam type number - component serial number". Equipment data is encapsulated in the form of field - value. All collected data uses the timestamp field as the primary key index. b) Heterogeneous data adaptation: For four types of data sources, namely beam positioning, hoisting equipment, energy consumption metering and production scheduling, corresponding data mapping tables are set up to convert the original fields into a unified set of fields, which includes: beam number, equipment number, three-dimensional coordinates, equipment status, energy consumption parameters and scheduling instructions. c) Unified sampling frequency: The sampling frequency for positioning and device data is uniformly set to once per second; d) Time synchronization: The clocks of each acquisition device are calibrated using an IoT time synchronization protocol to ensure that the time deviation does not exceed 1 second; e) Communication access: The device operating status and scheduling instructions are sent using the message queue transmission protocol, and the energy consumption and register data are sent using the fieldbus communication protocol. The data is then aggregated into a standardized data stream in a unified format through the data gateway. f) Integrity check: When the time interval between adjacent records is higher than the preset deviation threshold, the record is determined to be invalid and transferred to the exception cache; when the interval is lower than or equal to the preset deviation threshold, it is considered a valid record and processing continues. The data stream processing unit performs time synchronization processing based on the sliding window algorithm; Let the window length be Window step size is The maximum time drift threshold is Data record timestamp is The window center time is

[0007] when At that time, this data does not participate in window aggregation; when At that time, the data participates in the aggregation calculation, and the aggregation result is based on the window center time. Output by index.

[0008] Preferably, the digital twin modeling unit discretizes the beam field space into a regular grid, with grid resolutions of respectively... and The spatial coordinates of the beam are , The mapping is to the matrix index is , The calculation method is as follows:

[0009]

[0010] in, Indicates rounding down; The state matrix for dimensional tensor, number of channels It corresponds to four channels: beam number, spatial coordinates, operational status, and energy consumption information, with an update cycle of [missing information]. Second; When data for a certain channel is missing, linear interpolation is used for calculation:

[0011] in , This is the interpolation result; The energy consumption information in the state matrix includes energy consumption parameters calculated based on the device power data collected by the energy consumption monitor and the running time, which are used for energy consumption constraint determination in reinforcement learning algorithms.

[0012] Preferably, the channel data of the feature extraction module is standardized, and the standardization formula is as follows:

[0013] in, The original channel value; This is the channel moving average; For the channel sliding standard deviation; when When =0, let =0.

[0014] Preferably, the action set A is defined as a triple of beam-moving commands:

[0015] in, For equipment number, Indicates the starting position index; Indicates the target location index; when = At this time, the action instruction is not generated; Each action in the action set A It serves as the action input in reinforcement learning policy training and Q-value update, and participates in temporal difference calculation; The equipment load constraint is defined as follows:

[0016] When the equipment load Above the upper limit At that time, the action is marked as illegal; When the equipment load Lower than or equal to the upper limit At that time, the action is legal and retained; The path safety distance constraint is defined as follows:

[0017] in, For two path nodes in a digital twin or BIM scenario, this is an object used to calculate the safe distance between them; , For nodes Spatial coordinates on a two-dimensional plane, the coordinate system is consistent with BIM or digital twin, and the origin and axis are specified in the specification; , For nodes Spatial coordinates in the same coordinate system; For nodes and The Euclidean distance; The safety distance threshold is the minimum allowable distance preset by the system, used to avoid interference and collision between equipment, beams, and paths; when < At that time, the actions related to that path are deemed illegal; when ≥ At that time, the action is deemed legal; The reinforcement learning decision unit establishes energy consumption constraints based on the energy consumption parameters in the state matrix, and the constraints are defined as follows:

[0018] in, For action The corresponding energy consumption value is obtained by multiplying the equipment power by the execution time of the action, that is:

[0019] in, The energy consumption threshold set for the system is used to limit high-energy-consuming actions; when > At that time, the action was deemed illegal; when ≤ The action is retained at that time; The process window constraint in the reinforcement learning decision unit is defined as follows:

[0020] in, The earliest allowed start time for the process; This is the latest allowed completion time for the process; The actual execution time of the system plan; When execution time Execution earlier than Start or later At that time, the action was deemed illegal; When the execution time is within the interval [ , If the action is within the specified range, it is deemed legal and retained.

[0021] Preferably, the reward function for policy training is defined as:

[0022] in, This represents the path length. Waiting time; Energy consumption value; This is a conflict symbol; This refers to the equipment idle rate. , , , , These are non-negative weight parameters; when > Additional penalties will be added at that time.

[0023] when This item will not be added at this time; The conflict flag in the reinforcement learning decision unit is defined as:

[0024] in, This is an indicator function that takes the value 1 when the condition inside the parentheses is true, and 0 otherwise. Path node and The Euclidean distance between them; This is the safe distance threshold; When distance Below the safe distance threshold When a conflict is detected, =1; When distance Higher than or equal to the safe distance threshold At that time, it was determined that there was no conflict. =0; The objective of the time-series difference is:

[0025] The loss function is:

[0026] The system adopts a parameter update rule based on gradient descent:

[0027] in, For instant rewards; Discount factor; These are the current network parameters; For the target network parameters, The learning rate is a preset hyperparameter used to control the step size of each parameter update. Indicates the parameter The gradient of the loss function; when using mini-batch training, the loss function is calculated based on the mean squared error of the batch samples; The target network parameters are adjusted using a soft update method:

[0028] in, This is the soft update coefficient, with a value ranging from 0 to 1; when When the value exceeds the preset upper limit, the upper limit value is used. If the value is below the preset lower limit, the lower limit value will be used.

[0029] Preferably, the reinforcement learning decision unit adopts a sample-priority sampling mechanism based on weighted experience replay, and its sample priority is defined as:

[0030] in, For the first Temporal difference error of each sample; Adjustment factor for priority; Prioritize samples; The target Q value for the sample; This is the current network output; The sample sampling probability is defined as:

[0031] When error When, sample priority is taken ,in This is the lower limit threshold for the error. The reinforcement learning decision unit generates training samples through multi-process parallel simulation, with a simulation quantity of [number missing]. , No. The number of sampling steps for each simulation process is The total number of sampling steps per day is defined as follows:

[0032] and satisfy ≥ Each simulation process executes the state transition function in parallel. The generated sample data is uniformly stored in the experience replay buffer for use by the priority sampling mechanism.

[0033] Preferably, the edge-cloud collaborative unit is used to implement low-latency policy execution and abnormal data filtering and uploading, and its execution process includes: Edge inference execution involves deploying a lightweight inference model optimized through network pruning and parameter quantization on edge computing nodes, when the actual response time... Higher than the preset response threshold When this happens, the system automatically calls the strategy from the previous version; when Less than or equal to When the current policy is applied, execute the current policy. Data deviation filtering and uploading: Edge nodes filter key samples based on temporal difference error and state deviation, and upload them to the cloud for retraining and parameter synchronization. The judgment criteria are as follows:

[0034] in, This refers to timing difference error; This is the error threshold; This is the current state vector; For the predicted state vector; This is the deviation threshold; when both are below the threshold, the data sample is not uploaded. Anomaly detection and judgment: Abnormal samples are identified through standardized score calculation.

[0035] when When, it is marked as an abnormal sample, when When this occurs, it is not marked as an exception; among which Standardized threshold; Cloud-based parameter updates and data uploads enable model retraining and parameter synchronization in the cloud, forming a closed-loop optimization structure of cloud training, edge execution, and data feedback to maintain the model's long-term adaptability and real-time performance.

[0036] Preferably, the visualization feedback unit realizes the aggregated display of beam yard status and the visualization output of paths based on the BIM digital twin model, and the method includes: Window aggregation calculations aggregate data over time windows, and the average value is defined as follows:

[0037] in, These are the sample values ​​within the window. The number of samples; when When this happens, discard the aggregated results of that window to ensure data stability; Path and energy consumption visualization using path sequences:

[0038] As the input for rendering, For discrete path node indexes The number of walks is the distance from the path. When the distance between any adjacent nodes Below the safe distance threshold When this occurs, use a conflict style identifier; When the corresponding action consumes energy Above the energy consumption threshold When energy consumption is used, it should be indicated in an energy consumption style. The rendering output includes the beam's movement trajectory, equipment operation heatmap, and abnormal alarm information; The 3D dynamic display and visualization feedback unit maps the aggregated results to the BIM 3D model, realizing the dynamic display of path trajectory, energy consumption distribution and equipment utilization. Combined with the reinforcement learning optimization results, it provides real-time scheduling decision support for beam yard managers.

[0039] A method for real-time monitoring and scheduling optimization of beam yard production status based on big data includes the following steps: S1: Collect multi-source heterogeneous data generated by beam positioning sensors, hoisting equipment monitoring terminals, energy consumption monitors and scheduling command systems in the beam yard; extract, transform and standardize the multi-source data through an ETL data processing flow designed for the beam yard scenario to achieve unified structured management of multi-source heterogeneous data; S2: Construct a streaming data processing channel to perform time alignment and window aggregation on monitoring data from different sources, generating a time-synchronized beam yard operation data stream to provide dynamic input for subsequent digital twin modeling; S3: Construct a digital twin scene of the beam yard based on the BIM 3D model, and map the processed operation data into the model in real time to form a dynamic scene model that reflects the beam number, location, operation status and energy consumption parameters. S4: Utilize an artificial intelligence decision engine to perform state analysis on the digital twin model, and generate a beam transport path and production scheduling scheme that meets safety and efficiency requirements based on conditions such as equipment load limits, path avoidance distances, process time windows, and energy consumption constraints. S5: Deploy lightweight models on edge computing nodes for rapid decision-making and execution. When data anomalies or policy deviations are detected, key status data are uploaded to the cloud for model updates and parameter synchronization, forming a collaborative closed loop between cloud training and edge execution. S6: Map the scheduling results and operating status to the BIM digital twin 3D model to display the beam movement trajectory, equipment operating status and abnormal early warning information in a 3D manner, so as to realize real-time monitoring and dynamic optimization of the beam yard production status.

[0040] In summary, the present invention has the following main beneficial effects: By setting up a data acquisition and processing unit, the system achieves unified extraction, transformation, and standardized access to multi-source heterogeneous data generated by beam positioning sensors, hoisting equipment monitoring terminals, and energy consumption monitors. This forms a time-aligned beam yard operation data stream, effectively solving the problems of inconsistent data acquisition standards, insufficient real-time performance, and information silos in existing technologies. This design ensures the accuracy and consistency of data during transmission and fusion, providing a high-quality data foundation for subsequent digital twin modeling and intelligent scheduling.

[0041] By setting up a digital twin modeling and decision-making unit, a dynamically updated digital twin scenario of the beam yard is constructed based on the BIM 3D model. Combined with an intelligent decision-making module, it comprehensively assesses multiple constraints such as equipment load, path avoidance, process time windows, and energy consumption limits, thereby improving the intelligence level and safe collaborative capabilities of beam yard scheduling. This unit can perceive the status of beams and equipment in real time, generating optimal beam movement and production scheduling plans, enabling the overall beam yard operation to run efficiently within a safe range.

[0042] By setting up edge-cloud collaboration and visualization feedback units, rapid decision-making and execution are achieved at edge computing nodes, and abnormal data samples are uploaded to the cloud for retraining and parameter synchronization, forming a closed-loop optimization architecture. Simultaneously, the beam trajectory, energy consumption distribution, and early warning information are displayed in a 3D manner within a digital twin scenario, improving system response speed, decision-making adaptability, and monitoring intuitiveness. This design enables collaborative management of the entire beam yard production process, from data perception and intelligent decision-making to dynamic display, significantly improving production efficiency and operational safety. Attached Figure Description

[0043] Figure 1 This is a system block diagram of the present invention; Figure 2 This is a flowchart of the data standardization process of the present invention; Figure 3 This is a schematic diagram of the digital twin state matrix mapping of the present invention. Detailed Implementation

[0044] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0045] Example 1 refer to Figures 1-3 A real-time monitoring and analysis system for beam yard production status based on big data includes: The data acquisition unit is used to collect multi-source heterogeneous data generated by beam positioning sensors, hoisting equipment monitoring terminals, energy consumption monitors and scheduling command systems in the beam yard, and to perform standardized extraction and transformation through an ETL data processing flow designed for the beam yard scenario. The standardized process includes: unifying the beam body coding format, specifying the sensor sampling frequency, using the Internet of Things communication protocol to connect data, and mapping data fields from different sources to a unified structured format to achieve standardized access to multi-source data; The data stream processing unit is used to construct a data processing channel based on a stream processing framework to perform real-time aggregation and time synchronization of the multi-source data. The time synchronization is based on an event-time mechanism. By setting the sliding window length, window step size, and maximum allowable time deviation, when the difference between the recorded time and the window reference time does not exceed the maximum allowable time deviation, the record participates in the aggregation process to ensure the time alignment accuracy of data collected by different devices, thereby generating a time-aligned beam yard operation data stream. The digital twin modeling unit is used to build a three-dimensional digital twin scene of the beam yard based on the BIM model and map the running data stream into a state matrix that is dynamically updated over time. The state matrix is ​​a three-dimensional data structure, divided into grids along the spatial dimension. The first and second dimensions correspond to the horizontal and vertical partitions of the beam yard space, respectively. The third dimension is the channel dimension, which is used to record various feature data such as beam number, spatial location, operation status and energy consumption information, to reflect the real-time production status of the beam yard. A reinforcement learning decision-making unit is used to generate beam transport paths and production scheduling schemes based on the state matrix. The reinforcement learning decision unit includes: a) Feature extraction module, used to extract spatial and topological features from the state matrix using a convolutional residual network to form a feature map for policy training; b) The policy training module employs an improved deep Q-learning algorithm for scheduling optimization, the improvements including: Improve the learning performance of rare but high-value samples by setting sample weights; Simulation training is conducted using a multi-process parallel approach to improve policy convergence speed; Based on constraints such as equipment load limit, path safety distance, process priority, process time window, and energy consumption limit, action instructions that do not meet the requirements are filtered out, and the optimal solution that meets the on-site safety and efficiency requirements is generated. The reinforcement learning decision unit defines an action set A in each state to describe the optional beam transport instruction set, and the action set serves as the action input for policy training and Q-value update; The Q-value update satisfies the temporal difference objective; Among them, the system determines the conflict flag under the path safety distance constraint, which is used for reward function calculation; Edge-cloud collaborative units are used to deploy lightweight inference models on edge computing nodes to achieve low-latency policy execution; The lightweight model is optimized through network pruning and parameter quantization to ensure that the decision response time meets the preset response time threshold. Only key abnormal samples and state data with large deviations are uploaded to the cloud for retraining and parameter synchronization, thus forming a closed-loop optimization architecture of cloud training, edge execution and data feedback. The visualization feedback unit is used to display the optimal beam moving path, equipment operating status and abnormal warning information in a three-dimensional manner in the BIM digital twin scenario, so as to realize the visualization feedback and real-time monitoring of scheduling results.

[0046] The standardized extraction and transformation of the data acquisition unit includes the following steps: a) Unified data format: A unified data standard model is established for data from different sources. The beam number adopts a three-segment coding rule of "project number-beam type number-component serial number". Equipment data is encapsulated in the form of field-value. All collected data uses the timestamp field as the primary key index. b) Heterogeneous data adaptation: For four types of data sources, namely beam positioning, hoisting equipment, energy consumption metering and production scheduling, corresponding data mapping tables are set up to convert the original fields into a unified set of fields, which includes: beam number, equipment number, three-dimensional coordinates, equipment status, energy consumption parameters and scheduling instructions. c) Unified sampling frequency: The sampling frequency for positioning and device data is uniformly set to once per second; d) Time synchronization: The clocks of each acquisition device are calibrated using an IoT time synchronization protocol to ensure that the time deviation does not exceed 1 second; e) Communication access: The device operating status and scheduling instructions are sent using the message queue transmission protocol, and the energy consumption and register data are sent using the fieldbus communication protocol. The data is then aggregated into a standardized data stream in a unified format through the data gateway. f) Integrity check: When the time interval between adjacent records is higher than the preset deviation threshold, the record is determined to be invalid and transferred to the exception cache; when the interval is lower than or equal to the preset deviation threshold, it is considered a valid record and processing continues. The data stream processing unit performs time synchronization processing based on the sliding window algorithm; Let the window length be Window step size is The maximum time drift threshold is Data record timestamp is The window center time is

[0047] when At that time, this data does not participate in window aggregation; when At that time, the data participates in the aggregation calculation, and the aggregation result is based on the window center time. Output by index; In the stream processing layer, a time-based window aggregation mechanism is used to achieve time alignment, with a maximum permissible deviation set. When the data timestamp difference between different devices is less than Each event is considered as a single event window, thereby generating a time-synchronized beam yard operation data stream.

[0048] This mechanism ensures that asynchronously acquired data from different sensors are synchronized within milliseconds, providing high-precision dynamic input for subsequent digital twin models. By setting up data acquisition and stream processing units, the effect of unifying multi-source data formats and improving data synchronization accuracy is achieved, thus realizing the standardization and real-time acquisition of beam yard data. The digital twin modeling unit discretizes the beam field space into a regular grid with grid resolutions of [resolutions to be filled in]. and The spatial coordinates of the beam are , The mapping is to the matrix index is , The calculation method is as follows:

[0049]

[0050] in, Indicates rounding down; The state matrix for dimensional tensor, number of channels It corresponds to four channels: beam number, spatial coordinates, operational status, and energy consumption information, with an update cycle of [missing information]. Second; When data for a certain channel is missing, linear interpolation is used for calculation:

[0051] in , This is the interpolation result; The energy consumption information in the state matrix includes energy consumption parameters calculated based on the device power data and running time collected by the energy consumption monitor, which are used for energy consumption constraint determination in reinforcement learning algorithms. The digital twin model achieves real-time mapping between the beam object and the sensor data through continuous communication with the data acquisition layer; When the beam displacement sensor reports a change in position, the corresponding three-dimensional coordinates of the beam in the model are updated synchronously. When the energy consumption data of the equipment fluctuates, the color change of the energy consumption layer reflects the operating load status; The dynamic refresh cycle of the model is determined by the output frequency of the stream processing unit, which is usually 1 to 3 seconds; By setting up this unit, the physical entity and virtual model of the beam yard can be updated synchronously, achieving the effect of visual and dynamic monitoring of the beam yard's production status, and providing basic data support for subsequent intelligent decision-making. The channel data of the feature extraction module is standardized, and the standardization formula is as follows:

[0052] in, The original channel value; This is the channel moving average; For the channel sliding standard deviation; when When =0, let =0; Each action in the action set A It serves as the action input in reinforcement learning policy training and Q-value update, and participates in temporal difference calculation; When the equipment load Lower than or equal to the upper limit At that time, the action is legal and retained; The path safety distance constraint is defined as follows:

[0053] in, For two path nodes in a digital twin or BIM scenario, this is an object used to calculate the safe distance between them; , For nodes Spatial coordinates on a two-dimensional plane, the coordinate system is consistent with BIM or digital twin, and the origin and axis are specified in the specification; , For nodes Spatial coordinates in the same coordinate system; For nodes and The Euclidean distance; The safety distance threshold is the minimum allowable distance preset by the system, used to avoid interference and collision between equipment, beams, and paths; when < At that time, the actions related to that path are deemed illegal; when ≥ At that time, the action is deemed legal; The reinforcement learning decision unit establishes energy consumption constraints based on the energy consumption parameters in the state matrix, and the constraints are defined as follows:

[0054] in, For action The corresponding energy consumption value is obtained by multiplying the equipment power by the execution time of the action, that is:

[0055] in, The energy consumption threshold set for the system is used to limit high-energy-consuming actions; when > At that time, the action was deemed illegal; when ≤ The action is retained at that time; The process window constraint in the reinforcement learning decision unit is defined as follows:

[0056] in, The earliest allowed start time for the process; This is the latest allowed completion time for the process; The actual execution time of the system plan; When execution time Execution earlier than Start or later At that time, the action was deemed illegal; When the execution time is within the interval [ , If the action is within the specified time frame, it is deemed legal and retained. The reward function for training the policy is defined as follows:

[0057] in, This represents the path length. Waiting time; Energy consumption value; This is a conflict symbol; This refers to the equipment idle rate. , , , , These are non-negative weight parameters; when > Additional penalties will be added at that time.

[0058] when This item will not be added at this time; The conflict flag in the reinforcement learning decision unit is defined as:

[0059] in, This is an indicator function that takes the value 1 when the condition inside the parentheses is true, and 0 otherwise. Path node and The Euclidean distance between them; This is the safe distance threshold; When distance Below the safe distance threshold When a conflict is detected, =1; When distance Higher than or equal to the safe distance threshold At that time, it was determined that there was no conflict. =0; The objective of the time-series difference is:

[0060] The loss function is:

[0061] The system adopts a parameter update rule based on gradient descent:

[0062] in, For instant rewards; Discount factor; These are the current network parameters; For the target network parameters, The learning rate is a preset hyperparameter used to control the step size of each parameter update. Indicates the parameter The gradient of the loss function; when using mini-batch training, the loss function is calculated based on the mean squared error of the batch samples; The target network parameters are adjusted using a soft update method:

[0063] in, This is the soft update coefficient, with a value between 0 and 1; when When the value exceeds the preset upper limit, the upper limit value is used. When the value is below the preset lower limit, the lower limit value is used. In each decision-making loop, the system reads the state matrix and generates a set of candidate actions; The Q-value network is used to predict the reward of each action and select the optimal action to execute. Once the execution result is returned, the immediate reward is calculated, which includes a weighted result of various constraint indicators.

[0064] By setting up reinforcement learning decision-making units, adaptive scheduling optimization under multiple constraints was achieved, resulting in reduced total beam-moving costs and improved production efficiency. Experiments show that this system, tested at the Xiongxin High-Speed ​​Railway beam yard, reduced total beam-moving energy consumption by approximately 11.1% and increased average operating efficiency by about 15%, as shown in the table below:

[0065] The system of this invention significantly reduces overall energy consumption and conflict rate under the same operating conditions by setting up reinforcement learning decision-making units and edge-cloud collaborative mechanisms; The total energy consumption for beam movement was reduced by approximately 11.1%, and the average operating efficiency was increased by approximately 15%. The system response latency has been reduced to approximately 300 ms; This result proves that the system can stably achieve adaptive scheduling optimization in complex construction scenarios and has good prospects for industrial applications.

[0066] The reinforcement learning decision unit adopts a sample-priority sampling mechanism based on weighted experience replay, and its sample priority is defined as:

[0067] in, For the first Temporal difference error of each sample; Adjustment factor for priority; Prioritize samples; The target Q value for the sample; This is the current network output; The sample sampling probability is defined as:

[0068] When error When, sample priority is taken ,in This is the lower limit threshold for the error. The reinforcement learning decision unit generates training samples through multi-process parallel simulation, with a simulation quantity of [number missing]. , No. The number of sampling steps for each simulation process is The total number of sampling steps per day is defined as follows:

[0069] and satisfy ≥ Each simulation process executes the state transition function in parallel. The generated sample data is uniformly stored in the experience replay buffer for use by the priority sampling mechanism; The edge-cloud collaboration unit is used to achieve low-latency execution of policies and uploading of abnormal data. Its execution process includes: Edge inference execution involves deploying a lightweight inference model optimized through network pruning and parameter quantization on edge computing nodes, when the actual response time... Higher than the preset response threshold When this happens, the system automatically calls the strategy from the previous version; when Less than or equal to When the current policy is applied, execute the current policy. Data deviation filtering and uploading: Edge nodes filter key samples based on temporal difference error and state deviation, and upload them to the cloud for retraining and parameter synchronization. The judgment criteria are as follows:

[0070] in, This refers to timing difference error; This is the error threshold; This is the current state vector; For the predicted state vector; This is the deviation threshold; when both are below the threshold, the data sample is not uploaded. Anomaly detection and judgment: Abnormal samples are identified through standardized score calculation.

[0071] when When marked as an abnormal sample, When not marked as an exception; Standardized threshold; Cloud-based parameter updates and data uploads enable model retraining and parameter synchronization in the cloud, forming a closed-loop optimization structure of cloud training, edge execution, and data feedback to maintain the model's long-term adaptability and real-time performance. Edge nodes are responsible for fast local inference, ensuring a decision response time of less than 500 ms; The cloud periodically aggregates data uploaded from the edge, retrains the model, and distributes the latest parameters. This mechanism ensures the system's long-term self-learning capability and global consistency; By setting up this unit, the system response speed can be improved, network latency can be reduced, and the model can be continuously optimized, enabling the collaborative operation of real-time edge decision-making and global optimization in the cloud.

[0072] The visualization feedback unit, based on the BIM digital twin model, realizes the aggregated display of beam yard status and the visualization output of paths. The method includes: Window aggregation calculations aggregate data over time windows, and the average value is defined as follows:

[0073] in, These are the sample values ​​within the window. For the number of samples; when When this happens, discard the aggregated results of that window to ensure data stability; Path and energy consumption visualization using path sequences:

[0074] As the input for rendering, For discrete path node indexes The number of walks is the distance from the path. When the distance between any adjacent nodes Below the safe distance threshold When this occurs, use a conflict style identifier; When the corresponding action consumes energy Above the energy consumption threshold When energy consumption is used, it should be indicated in an energy consumption style. The rendering output includes the beam's movement trajectory, equipment operation heatmap, and abnormal alarm information; The three-dimensional dynamic display and visualization feedback unit map the aggregation results to the BIM three-dimensional model to realize the dynamic display of path trajectory, energy consumption distribution and equipment utilization. Combined with the reinforcement learning optimization results, it provides real-time scheduling decision support for beam yard managers. This unit is visualized based on a BIM digital twin model; the system displays the beam trajectory sequence. Projected into three-dimensional space, the equipment's operating status, path risks, and energy consumption distribution are reflected through visual means such as color and lighting effects. When the distance between any two beams Below the safety threshold The system will automatically issue a collision warning at that time. When energy consumption exceeds the threshold, the path is highlighted in red; Users can view scheduling plans, execution records, and energy consumption statistics through a visual interface; By setting up a visual feedback unit, a three-dimensional dynamic display of the beam yard's operating status was achieved, which improved information transparency and operational intuitiveness, providing managers with real-time decision-making support.

[0075] This system achieves full-process digitalization and intelligent scheduling of beam yard production status through a closed-loop structure of data acquisition, modeling, decision-making, execution, and feedback. The collaborative mechanism between units ensures that the system has real-time data, adaptive decision-making, and intuitive monitoring. The system operation results show that the method of the present invention can dynamically generate the optimal beam moving path under complex working conditions, effectively reducing energy consumption, reducing the risk of conflict and improving resource utilization.

[0076] Example 2 This invention relates to a method for real-time monitoring and scheduling optimization of beam yard production status based on big data, comprising the following steps: S1 data acquisition and standardized access include: Data acquisition configuration: (1) Beam positioning: Positioning base stations are set up in the beam fabrication area, beam storage area and key nodes of the beam transportation route, and positioning tags are installed on the beam or beam transportation equipment, with a sampling frequency of 1~5 Hz; (2) Lifting equipment: Collect data on lifting weight, boom length, slewing angle, travel speed, and operating status (no load / lifting / traveling / beam lowering, etc.) from the equipment PLC. (3) Energy consumption monitoring: Electric drive equipment is connected to a three-phase energy meter, and diesel equipment is connected to a fuel consumption pulse / flow meter, with a sampling frequency of 0.5~1 Hz; (4) Scheduling instructions: Obtain task ID, process type, job window, priority and target location from the MES / planning system.

[0077] ETL Standardized Access: The process employs a device source, IoT gateway, message queue, data bus, and standard tables. Fields are standardized as follows: timestamp, source device ID, beam ID, 3D position (X / Y / Z), device status, lifting weight / arm length, travel speed, instantaneous power / fuel consumption, task ID, target location, and anomaly code. Timestamps are standardized to ISO 8601 accuracy to milliseconds. The coordinate system adopts the BIM model coordinate system. Device and beam coding uses a three-part format: "Project-Beam Type-Serial Number." Outlier handling includes: missing data filling (previous value preservation / linear interpolation), outlier removal (based on engineering thresholds and device self-diagnostic codes), and recording data quality markers (normal / missing / abnormal).

[0078] By unifying sampling frequency, timestamps, coordinate systems, and field specifications, it has achieved consistent access to cross-vendor devices and multi-source data, significantly reducing the difficulty of subsequent fusion and avoiding data conflicts such as "multiple versions of the same event," thus providing an accurate and continuous data foundation for real-time processing and twin modeling.

[0079] S2: Streaming, time alignment, and window aggregation include: Real-time channel setup: Deploy message queues (such as Kafka) and stream processing engines (such as Flink / similar) between edge nodes and the central server. Create topics for each type of data; set up consumer groups to ensure ordered consumption and backpressure processing, ensuring no packet loss under high traffic. Time alignment strategy: The system adopts an event-time driven approach, using the original timestamp of each record for calculation; a maximum allowable time deviation threshold is set (100~200 ms is commonly used in engineering), and late data within the threshold can still be included in the same window; data exceeding the threshold enters the compensation channel, is marked separately, and is prompted to be compensated on the visualization end; Window aggregation aperture: The commonly used sliding window length is 2 seconds and the sliding step size is 1 second (which can be adjusted according to the site). Within the window, the following are completed: positioning trajectory smoothing (such as moving average), energy consumption statistics (window cumulative / average), equipment status stability check (anti-jitter). The output is a time-synchronized beam yard operation data stream, which includes: time index, beam / equipment multi-dimensional status, energy consumption summary, and task context. By aligning events with time and aggregating windows, the problem of timing misalignment caused by asynchronous sampling of different devices and network jitter is solved; high-frequency noise is transformed into a stable structured data stream, providing low-latency, comparable state snapshots at the same time for subsequent twin mapping and decision analysis.

[0080] S3: Real-time mapping of BIM digital twin includes: model preparation and coordinate consistency. Import the beam yard BIM model; create a mapping table to bind the output fields to the BIM attributes; ensure that the site coordinates correspond one-to-one with the BIM coordinates; Dynamic scene updates: After the data stream enters the twin engine, the object attributes are updated according to a refresh cycle (1~3 seconds): beam position and attitude, equipment status, energy consumption heat map, and task execution progress bar. Historical playback and multi-view switching are supported. By binding data, object attributes, and 3D scenes layer by layer, the on-site status is mapped in real time to an interactive 3D space, achieving a WYSIWYG monitoring experience. Managers can view location, status, energy consumption, and progress without switching between multiple systems, significantly reducing communication and interpretation costs.

[0081] S4: Constrained scheduling optimization and solution generation includes: constraint system and triggering logic, setting four types of engineering constraints: (1) Equipment load limit: Rated load of equipment, working range table and safety margin. If the limit is violated, the plan will be rejected. (2) Path avoidance distance: Minimum safe distance between oncoming / intersecting vehicles (e.g., 3~5 m, determined according to the width of the equipment and the passageway). If the standard is not met, it is determined that there is a risk of collision. (3) Process time window: the earliest start and latest finish time allowed for a process. Any unplanned advance or delay must be rescheduled. (4) Energy consumption constraints: For the same task, prioritize the option with the shortest path, less idle load, and fewer start-stops; if necessary, limit peak energy consumption to avoid peak electricity consumption. Solution generation and verification: The method engine automatically generates candidate beam-moving paths and operation sequences based on real-time status and the aforementioned constraints. Each candidate solution undergoes rule verification: solutions that do not meet any hard constraints (safety / load / time window) are directly eliminated; among solutions that meet the hard constraints, those with fewer equipment changes, shorter aisle occupancy, lower historical conflict rate, and lower energy consumption estimates are prioritized as output. After generation, the paths are pushed to edge nodes for execution, and the solution trajectory and key checkpoints (turns, meeting points, hoisting points) are displayed on the visualization platform. The above process is essentially a task allocation and path orchestration under multiple constraints: first ensuring safety and time windows, and then making trade-offs between efficiency and energy consumption within the feasible domain. This approach can reliably reduce the time loss caused by back-and-forth reversing / waiting / yielding on-site, and reduce unnecessary peak energy consumption, reflecting an engineering strategy that prioritizes safety, balances efficiency, and controls costs.

[0082] S5: Edge-cloud collaborative closed loop includes rapid edge execution: After receiving the solution, the edge node combines local sensing and human-machine interface (HMI) commands to perform local transmission and monitoring; a preset response time threshold is set (e.g., ≤500 ms, quantized according to the on-site link). When the timeout occurs or the link is abnormal, it will degrade to the previous stable solution or prompt manual takeover; the edge side supports breakpoint resume and local caching to ensure that critical actions can still be completed under temporary network jitter. Exceptions and data reflow: Set upload trigger conditions to control edge-cloud traffic: When significant status deviations occur (positioning drift out of bounds, abnormal energy consumption jumps, device status anomaly codes) or multiple manual interventions occur, edge nodes package key segments (including status, instructions, anomaly flags, and timestamps) and upload them to the cloud, incorporating them into incremental updates of the model and rule base. The cloud performs daily / weekly reviews of all samples, generating parameter adjustment suggestions and blacklists / whitelists (high-risk routes, congested periods, sensitive procedures), which are then synchronized to the edge side. By adopting the structure of "fast decision-making at the edge + slow learning in the cloud", low-latency execution is guaranteed, while the system can continuously absorb on-site experience to iterate rules and parameters, ensuring that the system maintains stable performance under long-term disturbances such as seasonal changes, construction phase switching, and equipment depreciation.

[0083] S6: 3D visualization feedback and operation and maintenance linkage overlays scheduling results and operating status onto the BIM scene: drawing trajectory lines and direction arrows for beams in transit; displaying operating loads of equipment with status lights / thermal layers; highlighting conflict risk nodes with prominent identification; and highlighting / pop-up notifications for energy consumption exceeding limits or time window violations. The page provides a task list, 3D positioning, and operation suggestion linkage capability, allowing for one-click location of problem points; The system automatically generates daily / weekly work reports in the background, with key indicators including: number of beam moving tasks, average work time, waiting time, path conflict events, total energy consumption / single energy consumption, equipment utilization rate, and timeliness of anomaly handling. It supports quick retrieval and playback by task ID / beam ID / equipment ID, enabling closed-loop traceability and accountability.

[0084] 3D visualization makes the process of planning, execution, and results transparent and visible. Combined with KPI monitoring in the form of reports, managers can quickly locate bottlenecks (such as a section of road that is prone to congestion, or a piece of equipment that is frequently waiting / consuming high energy), and make fine-grained adjustments to construction plans and on-site right-of-way based on this.

[0085] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A real-time monitoring and analysis system for beam yard production status based on big data, characterized in that: include: The data acquisition unit is used to collect multi-source heterogeneous data generated by beam positioning sensors, hoisting equipment monitoring terminals, energy consumption monitors and scheduling command systems in the beam yard. It performs standardized extraction and transformation through an ETL data processing flow designed for the beam yard scenario. The standardization process includes: unifying the beam coding format, specifying the sensor sampling frequency, using IoT communication protocols to connect data, and mapping data fields from different sources to a unified structured format to achieve standardized access to multi-source data. The data stream processing unit is used to construct a data processing channel based on a stream processing framework to perform real-time aggregation and time synchronization of the multi-source data. The time synchronization is based on an event-time mechanism. By setting the sliding window length, window step size, and maximum allowable time deviation, when the difference between the recorded time and the window reference time does not exceed the maximum allowable time deviation, the record participates in the aggregation process to ensure the time alignment accuracy of data collected by different devices, thereby generating a time-aligned beam yard operation data stream. The digital twin modeling unit is used to build a three-dimensional digital twin scene of the beam yard based on the BIM model and map the running data stream into a state matrix that is dynamically updated over time. The state matrix is ​​a three-dimensional data structure, divided into grids along the spatial dimension. The first and second dimensions correspond to the horizontal and vertical partitions of the beam yard space, respectively. The third dimension is the channel dimension, which is used to record various feature data such as beam number, spatial location, operation status and energy consumption information, to reflect the real-time production status of the beam yard. A reinforcement learning decision-making unit is used to generate beam transport paths and production scheduling schemes based on the state matrix. The reinforcement learning decision unit includes: a) Feature extraction module, used to extract spatial and topological features from the state matrix using a convolutional residual network to form a feature map for policy training; b) The policy training module employs an improved deep Q-learning algorithm for scheduling optimization. The improvements include: Improve the learning performance of rare but high-value samples by setting sample weights; Simulation training is conducted using a multi-process parallel approach to improve policy convergence speed; Based on constraints such as equipment load limit, path safety distance, process priority, process time window, and energy consumption limit, action instructions that do not meet the requirements are filtered out, and the optimal solution that meets the on-site safety and efficiency requirements is generated. The reinforcement learning decision unit defines an action set A in each state to describe the optional beam transport instruction set, and the action set serves as the action input for policy training and Q-value update; The Q-value update satisfies the temporal difference objective; Among them, the system determines the conflict flag under the path safety distance constraint, which is used for reward function calculation; Edge-cloud collaborative units are used to deploy lightweight inference models on edge computing nodes to achieve low-latency policy execution; The lightweight model is optimized through network pruning and parameter quantization to ensure that the decision response time meets the preset response time threshold. Only key abnormal samples and state data with large deviations are uploaded to the cloud for retraining and parameter synchronization, thus forming a closed-loop optimization architecture of cloud training, edge execution and data feedback. The visualization feedback unit is used to display the optimal beam moving path, equipment operating status and abnormal warning information in a three-dimensional manner in the BIM digital twin scenario, so as to realize the visualization feedback and real-time monitoring of scheduling results.

2. The real-time monitoring and analysis system for beam yard production status based on big data as described in claim 1, characterized in that, The standardized extraction and transformation of the data acquisition unit includes the following steps: a) Unified data format: A unified data standard model is established for data from different sources. The beam number adopts a three-segment coding rule of "project number - beam type number - component serial number". Equipment data is encapsulated in the form of field - value. All collected data uses the timestamp field as the primary key index. b) Heterogeneous data adaptation: For four types of data sources, namely beam positioning, hoisting equipment, energy consumption metering and production scheduling, corresponding data mapping tables are set up to convert the original fields into a unified set of fields, which includes: beam number, equipment number, three-dimensional coordinates, equipment status, energy consumption parameters and scheduling instructions. c) Unified sampling frequency: The sampling frequency for positioning and device data is uniformly set to once per second; d) Time synchronization: The clocks of each acquisition device are calibrated using an IoT time synchronization protocol to ensure that the time deviation does not exceed 1 second; e) Communication access: The device operating status and scheduling instructions are sent using the message queue transmission protocol, and the energy consumption and register data are sent using the fieldbus communication protocol. The data is then aggregated into a standardized data stream in a unified format through the data gateway. f) Integrity check: When the time interval between adjacent records is higher than the preset deviation threshold, the record is determined to be invalid and transferred to the exception cache; when the interval is lower than or equal to the preset deviation threshold, it is considered a valid record and processing continues. The data stream processing unit performs time synchronization processing based on the sliding window algorithm; Let the window length be Window step size is The maximum time drift threshold is Data record timestamp is The window center time is when When this data is in use, it does not participate in window aggregation; when... At that time, the data participates in the aggregation calculation, and the aggregation result is based on the window center time. Output by index.

3. The real-time monitoring and analysis system for beam yard production status based on big data as described in claim 2, characterized in that, The digital twin modeling unit discretizes the beam field space into a regular grid with grid resolutions of [resolutions to be filled in]. and The spatial coordinates of the beam are , The mapping is to the matrix index is , The calculation method is as follows: ; ;in, Indicates rounding down; The state matrix for dimensional tensor, number of channels It corresponds to four channels: beam number, spatial coordinates, operational status, and energy consumption information, with an update cycle of [missing information]. Second; When data for a certain channel is missing, linear interpolation is used for calculation: in , This is the interpolation result; The energy consumption information in the state matrix includes energy consumption parameters calculated based on the device power data collected by the energy consumption monitor and the running time, which are used for energy consumption constraint determination in reinforcement learning algorithms.

4. The real-time monitoring and analysis system for beam yard production status based on big data as described in claim 3, characterized in that, The channel data of the feature extraction module is standardized, and the standardization formula is as follows: in, The original channel value; This is the channel moving average; For the channel sliding standard deviation; when When =0, let =0.

5. The real-time monitoring and analysis system for beam yard production status based on big data as described in claim 4, characterized in that, The action set A is defined as a triple of beam-moving commands: in, For equipment number, Indicates the starting position index; Indicates the target location index; when = At this time, the action instruction is not generated; Each action in the action set A It serves as the action input in reinforcement learning policy training and Q-value update, and participates in temporal difference calculation; The equipment load constraint is defined as follows: When the equipment load Above the upper limit At that time, the action is marked as illegal; When the equipment load Lower than or equal to the upper limit At that time, the action is legal and retained; The path safety distance constraint is defined as follows: in, For two path nodes in a digital twin or BIM scenario, this is an object used to calculate the safe distance between them; , For nodes Spatial coordinates on a two-dimensional plane, the coordinate system is consistent with BIM or digital twin, and the origin and axis are specified in the specification; , For nodes Spatial coordinates in the same coordinate system; For nodes and The Euclidean distance; The safety distance threshold is the minimum allowable distance preset by the system, used to avoid interference and collision between equipment, beams, and paths; when < At that time, the actions related to that path are deemed illegal; when ≥ At that time, the action is deemed legal; The reinforcement learning decision unit establishes energy consumption constraints based on the energy consumption parameters in the state matrix, and the constraints are defined as follows: in, For action The corresponding energy consumption value is obtained by multiplying the equipment power by the execution time of the action, that is: in, The energy consumption threshold set for the system is used to limit high-energy-consuming actions; when > At that time, the action was deemed illegal; when ≤ The action is retained at that time; The process window constraint in the reinforcement learning decision unit is defined as follows: in, The earliest allowed start time for the process; This is the latest allowed completion time for the process; The actual execution time of the system plan; When execution time Execution earlier than Start or later At that time, the action was deemed illegal; When the execution time is within the interval [ , If the action is within the specified range, it is deemed legal and retained.

6. The real-time monitoring and analysis system for beam yard production status based on big data as described in claim 5, characterized in that, The reward function for training the policy is defined as follows: in, This represents the path length. Waiting time; Energy consumption value; This is a conflict symbol; This refers to the equipment idle rate. , , , , These are non-negative weight parameters; when > Additional penalties will be added at that time. when This item will not be added at this time; The conflict flag in the reinforcement learning decision unit is defined as: in, This is an indicator function that takes the value 1 when the condition inside the parentheses is true, and 0 otherwise. Path node and The Euclidean distance between them; This is the safe distance threshold; When distance Below the safe distance threshold When a conflict is detected, =1; When distance Higher than or equal to the safe distance threshold At that time, it was determined that there was no conflict. =0; The objective of the time-series difference is: The loss function is: The system adopts a parameter update rule based on gradient descent: in, For instant rewards; Discount factor; These are the current network parameters; For the target network parameters, The learning rate is a preset hyperparameter used to control the step size of each parameter update. Indicates the parameter The gradient of the loss function; when using mini-batch training, the loss function is calculated based on the mean squared error of the batch samples; The target network parameters are adjusted using a soft update method: in, This is the soft update coefficient, with a value ranging from 0 to 1; when When the value exceeds the preset upper limit, the upper limit value is used. If the value is below the preset lower limit, the lower limit value will be used.

7. The real-time monitoring and analysis system for beam yard production status based on big data as described in claim 6, characterized in that, The reinforcement learning decision unit adopts a sample-priority sampling mechanism based on weighted experience replay, and its sample priority is defined as: in, For the first Temporal difference error of each sample; Adjustment factor for priority; Prioritize samples; The target Q value for the sample; This is the current network output; The sample sampling probability is defined as: When error When, sample priority is taken ,in This is the lower limit threshold for the error. The reinforcement learning decision unit generates training samples through multi-process parallel simulation, with a simulation quantity of [number missing]. , No. The number of sampling steps for each simulation process is The total number of sampling steps per day is defined as follows: and satisfy ≥ Each simulation process executes the state transition function in parallel. The generated sample data is uniformly stored in the experience replay buffer for use by the priority sampling mechanism.

8. The real-time monitoring and analysis system for beam yard production status based on big data according to claim 7, characterized in that, The edge-cloud collaboration unit is used to achieve low-latency execution of policies and uploading of abnormal data. Its execution process includes: Edge inference execution involves deploying a lightweight inference model optimized through network pruning and parameter quantization on edge computing nodes, when the actual response time... Higher than the preset response threshold When this happens, the system automatically calls the strategy from the previous version; when Less than or equal to When the current policy is applied, execute the current policy. Data deviation filtering and uploading: Edge nodes filter key samples based on temporal difference error and state deviation, and upload them to the cloud for retraining and parameter synchronization. The judgment criteria are as follows: in, This refers to timing difference error; This is the error threshold; This is the current state vector; For the predicted state vector; This is the deviation threshold; when both are below the threshold, the data sample is not uploaded. Anomaly detection and judgment: Abnormal samples are identified through standardized score calculation. when When marked as an abnormal sample, When not marked as an exception; Standardized threshold; Cloud-based parameter updates and data uploads enable model retraining and parameter synchronization in the cloud, forming a closed-loop optimization structure of cloud training, edge execution, and data feedback to maintain the model's long-term adaptability and real-time performance.

9. The real-time monitoring and analysis system for beam yard production status based on big data as described in claim 8, characterized in that, The visualization feedback unit, based on the BIM digital twin model, realizes the aggregated display of beam yard status and the visualization output of paths. The method includes: Window aggregation calculations aggregate data over time windows, and the average value is defined as follows: in, These are the sample values ​​within the window. For the number of samples; when When this happens, discard the aggregated results of that window to ensure data stability; Path and energy consumption visualization using path sequences: As the input for rendering, For discrete path node indexes The number of walks is the distance from the path. When the distance between any adjacent nodes Below the safe distance threshold When this occurs, use a conflict style identifier; When the corresponding action consumes energy Above the energy consumption threshold When energy consumption is used, it should be indicated in an energy consumption style. The rendering output includes the beam's movement trajectory, equipment operation heatmap, and abnormal alarm information; The 3D dynamic display and visualization feedback unit maps the aggregated results to the BIM 3D model, realizing the dynamic display of path trajectory, energy consumption distribution and equipment utilization. Combined with the reinforcement learning optimization results, it provides real-time scheduling decision support for beam yard managers.

10. A method for real-time monitoring and scheduling optimization of beam yard production status based on big data, characterized in that, Includes the following steps: S1: Collect multi-source heterogeneous data generated by beam positioning sensors, hoisting equipment monitoring terminals, energy consumption monitors, and dispatch command systems in the beam yard; By using an ETL data processing flow designed for beam yard scenarios, the multi-source data is extracted, transformed, and standardized for access, thereby achieving unified structured management of multi-source heterogeneous data. S2: Construct a streaming data processing channel to perform time alignment and window aggregation on monitoring data from different sources, generating a time-synchronized beam yard operation data stream to provide dynamic input for subsequent digital twin modeling; S3: Construct a digital twin scene of the beam yard based on the BIM 3D model, and map the processed operation data into the model in real time to form a dynamic scene model that reflects the beam number, location, operation status and energy consumption parameters. S4: Utilize an artificial intelligence decision engine to perform state analysis on the digital twin model, and generate a beam transport path and production scheduling scheme that meets safety and efficiency requirements based on conditions such as equipment load limits, path avoidance distances, process time windows, and energy consumption constraints. S5: Deploy lightweight models on edge computing nodes for rapid decision-making and execution. When data anomalies or policy deviations are detected, key status data are uploaded to the cloud for model updates and parameter synchronization, forming a collaborative closed loop between cloud training and edge execution. S6: Map the scheduling results and operating status to the BIM digital twin 3D model to display the beam movement trajectory, equipment operating status and abnormal early warning information in a 3D manner, so as to realize real-time monitoring and dynamic optimization of the beam yard production status.

Images