A real-time state assembled component production progress prediction deviation correction scheduling method

Abstract

This invention discloses a real-time state-based prefabricated component production progress prediction and correction scheduling method, belonging to the field of intelligent manufacturing and production scheduling technology for prefabricated building components. It mainly addresses the problems of insufficient coupling between progress prediction and correction scheduling in existing technologies, the open-loop structure of the scheduling system, and the inability to adapt to various real-time disturbances during production. This method models the production scheduling process as a Markov decision process, collecting real-time data from the production line in each scheduling cycle and encoding it into a state vector. A long short-term memory network is used to continuously predict the progress and delay risk of subsequent processes. When the delay risk exceeds the limit, a deep Q-learning decision network outputs correction actions and generates a rescheduling plan for execution. Simultaneously, the execution results are fed back to the prediction model and decision network for iterative optimization. This invention constructs a fully closed-loop adaptive scheduling system, effectively improving the adaptability of prefabricated component production scheduling to uncertain environments.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing and production scheduling technology for prefabricated building components, specifically a real-time status-based prefabricated component production progress prediction and correction scheduling method. Background Technology

[0002] The production of prefabricated components involves multiple processes, including mold cleaning, rebar tying, concrete pouring, curing, and demolding. These processes are subject to strict time dependencies and resource constraints. Due to the rigid lower limit of concrete curing time and the initial setting time window, production scheduling is highly sensitive to time deviations. In actual production, frequent disturbances such as equipment failures, fluctuations in personnel efficiency, and material delays cause actual production progress to deviate from the initial production plan. Therefore, achieving dynamic prediction and scheduling correction of production progress under uncertain conditions is one of the key technical challenges in this field.

[0003] Currently, some research has applied intelligent optimization algorithms to the field of prefabricated component production scheduling. For example, existing technology, CN115330179B, discloses a prefabricated building production scheduling optimization method considering resource constraints. Its technical solution first obtains several initial production plans based on a heuristic algorithm, assuming balanced load across production lines. Then, it analyzes the production process diagram under the premise of limited production resources and component production constraints. Finally, it iteratively filters and modifies the initial production plans based on a genetic algorithm model in an evolutionary environment to obtain the optimal production scheduling plan. This solution can solve problems such as resource conflicts and untimely order delivery. However, its scheduling plan is a static plan generated offline in one go. Once a disturbance occurs during production, the original plan deviates from the actual production state and cannot be dynamically adjusted according to real-time progress. Another example is an existing technology, CN115204497B, which discloses a prefabricated component production scheduling optimization method and system based on reinforcement learning. This solution acquires real-time and historical production data to establish a scheduling model, transforms the solution of the optimization objective into a solution based on a deep reinforcement learning model, iteratively updates the model through an experience replay pool, and outputs the optimal scheduling strategy. This scheme introduces reinforcement learning, which can dynamically adapt to external disturbances such as design changes and emergency order insertions. However, once the model is trained offline, it is used in a fixed manner. Subsequent scheduling decisions are not used to further update the model's predictive capabilities, and prediction and scheduling remain two separate processes.

[0004] Comprehensive analysis reveals the following shortcomings in existing technologies: During the scheduling process, there is a lack of effective coupling mechanism between progress prediction and corrective scheduling—prediction results cannot directly drive the automatic generation of scheduling decisions, and the execution results of scheduling decisions cannot be fed back to the prediction model to correct its prediction accuracy. Existing solutions are mostly open-loop or semi-open-loop structures of "data acquisition → display → manual decision-making" or "offline training → fixed strategy," lacking a closed-loop scheduling framework that integrates real-time perception, dynamic prediction, intelligent decision-making, and feedback learning. Therefore, how to construct a closed-loop scheduling system that couples and co-evolves the progress prediction model and the scheduling decision network has become a pressing technical problem to be solved in this field. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a real-time state-based prefabricated component production schedule prediction and correction scheduling method. This method models the dynamic scheduling problem of prefabricated component production lines as a Markov decision process, constructs a closed-loop method consisting of perception, prediction, decision, execution, and feedback, and achieves deep coupling between prediction and decision by directly encoding the output of the rolling prediction model as the state input of the decision network. At the same time, the actual execution results are fed back to the prediction model and decision network for incremental training and gradient updates, enabling the system to have the ability to continuously adapt and evolve. Compared with existing technologies, this invention realizes a paradigm shift from open-loop static scheduling to closed-loop adaptive scheduling, which can effectively improve the adaptability of prefabricated component production scheduling systems to uncertain environments.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a real-time status-based prefabricated component production progress prediction and correction scheduling method, the method comprising the following steps:

[0007] Step 1: In each scheduling cycle, collect the real-time start time of the process, the real-time end time of the process, the current resource occupancy status, and multimodal sensor data of each workstation in the prefabricated component production line.

[0008] Step 2: Encode the collected real-time start time of the process, real-time end time of the process, current resource occupancy status, and process state features extracted based on the multimodal sensor data into a state vector of a Markov decision process;

[0009] Step 3: Input the state vector into the rolling prediction model based on the long short-term memory network. The rolling prediction model outputs the predicted start time, predicted end time of each subsequent process and the predicted completion time of the entire prefabricated component production line, and calculates the delay probability of each process.

[0010] Step 4: When the delay probability of any process exceeds the preset delay risk threshold, the state vector is input into a decision network based on deep Q-learning. The decision network outputs a correction action, which is selected from the action space. The action space includes process speed adjustment action, resource reallocation action, process sequence adjustment action, maintenance time compression action, emergency resource call action, and no adjustment action.

[0011] Step 5: Convert the correction action into an executable rescheduling scheme, and send the rescheduling scheme to the manufacturing execution system for execution;

[0012] Step 6: Collect the actual process completion time after the rescheduling scheme is executed, calculate the prediction error between the actual process completion time and the predicted end time, store the prediction error in the error cache pool, and incrementally train the rolling prediction model with a preset prediction model update cycle.

[0013] Step 7: Based on the reward value obtained after performing the correction action and the new state vector after the state transition, generate an experience sample, store the experience sample in the experience replay pool, and perform gradient updates on the decision network at a preset decision network update cycle.

[0014] Step 8: Repeat steps 1 through 7 until all components are produced.

[0015] By collecting real-time process time, resource occupancy status, and multimodal sensor data in each scheduling cycle, and constructing a state vector based on a Markov decision process, the delay probability is calculated using a rolling prediction model. When the delay risk exceeds the limit, the decision network automatically outputs corrective actions and converts them into a rescheduling plan for execution. At the same time, the prediction error after execution is fed back for incremental training, and the experience sample is fed back for gradient updates. This forms a complete closed loop of perception, prediction, decision-making, execution, and feedback, enabling the scheduling method to have continuous adaptive evolution capabilities and avoiding the disconnect between the scheduling plan and the actual production status caused by the open-loop structure.

[0016] Furthermore, in step three, the rolling prediction model based on the long short-term memory network adopts a three-layer parallel long short-term memory encoder architecture, which includes:

[0017] A microscopic long short-term memory encoder with a time step of minutes inputs sensor timing data for each process, including concrete pouring rate data and vibration frequency data, and outputs a state feature vector within the process.

[0018] A meso-level long short-term memory encoder with process as the step size takes process-level event data as input, including process start time, process end time and resource occupancy status data, and outputs inter-process flow feature vector.

[0019] A macroscopic long short-term memory encoder with order batch as the step size takes historical order data as input, including order type data, batch size data and delivery cycle data, and outputs an order-level periodic feature vector.

[0020] The rolling prediction model also includes an attention fusion module, which weights and fuses the intra-process state feature vector, the inter-process flow feature vector, and the order-level periodic feature vector. The attention weight of the attention fusion module is adaptively determined by the sensitivity of the current production stage.

[0021] By employing a three-layer parallel long short-term memory encoder to process minute-level time-series data, process-level flow data, and order batch-level periodic data respectively, and utilizing an attention fusion module to adaptively adjust the fusion weights of the three-layer features according to the current production stage, the progress patterns of multiple time scales in the production of prefabricated components can be fully explored, thereby improving the prediction accuracy of the rolling prediction model for complex process scenarios.

[0022] Furthermore, in step four, the reward function of the decision network based on deep Q-learning is calculated according to the following formula:

[0023]

[0024] Where R represents the total output value of the reward function, which is dimensionless. This represents the weighting coefficient for items completed on schedule; it is dimensionless. This represents the reward value for on-time completion. A positive value is taken when the actual completion time is earlier than or equal to the planned completion time, and a negative value is taken when the actual completion time is later than the planned completion time. The negative value is directly proportional to the square of the delay time. This represents the weighting coefficient for the resource utilization efficiency term; it is dimensionless. This represents the resource utilization efficiency bonus value, calculated based on the deviation between the actual working hours utilization rate and the target utilization rate for each workstation. The weighting coefficients for the scheduling stability term are dimensionless. This represents the scheduling stability reward value, which is a negative correlation function of the degree of difference between the current scheduling scheme and the previous scheduling scheme. This represents the weighting coefficient for the process compliance item; it is dimensionless. This represents the penalty value for process compliance. It takes a large negative value when any of the following constraints are violated: minimum curing time constraint, initial concrete setting time window constraint, or process sequence dependency constraint.

[0025] By combining on-time completion rewards, resource utilization efficiency rewards, scheduling stability rewards, and process compliance penalties into a reward function with weighted coefficients, and setting the delay penalty to be directly proportional to the square of the delay time, and taking a large negative value when violating the curing lower limit, initial setting window, or process dependency constraints, the decision network can be guided to strictly follow the process specifications of prefabricated component production while optimizing the construction period and resource utilization, thus avoiding the generation of scheduling schemes that do not meet the process requirements.

[0026] Furthermore, in step two, the state vector includes a real-time workstation progress vector, a resource occupancy matrix, and a rolling prediction output vector;

[0027] The real-time progress vector of each workstation includes the process start time, process end time, current execution status, elapsed time and remaining estimated time. The current execution status includes not started, in progress, completed and interrupted states.

[0028] The rows of the resource occupancy matrix correspond to resource types, and the columns correspond to time slots. The resource types include mold table resources, crane equipment resources, personnel resources, and curing kiln resources. Each matrix element of the resource occupancy matrix represents the occupancy status of the corresponding resource in the corresponding time slot and the expected release time.

[0029] The rolling prediction output vector includes the predicted start time and end time of each subsequent process, the predicted completion time of the entire prefabricated component production line, and the delay probability of each process.

[0030] By encoding the real-time progress vector of each workstation, the resource occupancy matrix, and the rolling prediction output vector into a state vector, where the resource occupancy matrix records the occupancy status and expected release time with resource type as the row and time slot as the column, the decision network can fully perceive the progress of each workstation, the future availability of each resource, and prediction information at each scheduling moment, providing a complete state basis for corrective decisions.

[0031] Furthermore, in step four, the decision network adopts a priority experience replay mechanism: each experience sample in the experience replay pool is assigned a sampling priority according to the severity of the delay. The longer the delay time or the more delayed processes, the higher the sampling priority of the experience sample. The decision network samples experience samples from the experience replay pool according to the sampling priority during each update. The quantitative index of the severity of the delay is calculated based on the deviation between the actual completion time and the planned completion time, as well as the number of subsequent processes affected by the delay.

[0032] By employing a priority experience replay mechanism in the decision network's experience replay, sampling priorities are allocated according to the severity of the delay. Experience samples with longer delay times or more subsequent processes affected by the delay receive higher sampling probabilities. This enables the decision network to learn from high-risk scenarios first, accelerates the convergence speed for severe delay problems, and improves the response quality of the scheduling system under emergency disturbances.

[0033] Furthermore, in step one, the multimodal sensing data includes process execution image data collected by the workstation camera, vibration spectrum data of the vibration process collected by the accelerometer, material arrival status data collected by the radio frequency identification reader, and curing temperature data collected by the temperature sensor.

[0034] In step two, the process state features extracted based on the multimodal sensing data include: rebar binding density features and concrete pouring flatness features extracted from the process execution image data using the YOLOv8 target detection model; time-frequency domain features extracted from the vibration spectrum data of the vibration process using a one-dimensional convolutional neural network; and material state features and temperature features extracted from the material arrival status data and the curing temperature data.

[0035] By collecting workstation image data, vibration spectrum data, radio frequency identification material arrival data, and curing temperature data as multimodal sensing data, and using the YOLOv8 model to extract the characteristics of rebar binding density and pouring flatness, and a one-dimensional convolutional neural network to extract vibration time-frequency domain features, it is possible to achieve refined perception of the process execution status and reduce misjudgment of status caused by relying solely on time data.

[0036] Furthermore, the method also includes a bottleneck resource dynamic matching step: constructing a resource-process bipartite graph, wherein the resource-process bipartite graph includes a resource node set and a process node set, wherein each resource node in the resource node set carries the current occupancy status and the expected release time, and each process node in the process node set carries the resource demand type and the demand duration, and the resource nodes and the process nodes are connected by an allocatable edge;

[0037] In each scheduling cycle, the occupancy status and expected release time of each resource node in the resource node set are updated based on real-time data. The matching degree score between each process node and the candidate resource node is calculated. The matching degree score takes into account spatial distance factors, current load factors and historical matching success rate factors. The Hungarian algorithm is used to solve the maximum weight perfect matching of the resource-process bipartite graph to obtain the optimal resource allocation scheme.

[0038] When the current load of any resource node exceeds the preset load threshold, an early warning message is output, and an alternative resource node is automatically searched.

[0039] By constructing a resource-process bipartite graph, with resource nodes carrying their occupancy status and expected release time, and process nodes carrying their resource demand type and duration, the Hungarian algorithm is used to solve for the maximum weight perfect matching to obtain the optimal resource allocation. When the resource load exceeds the limit, alternative resources are automatically searched, which can realize proactive early warning and dynamic reallocation of bottleneck resources, avoiding production interruptions caused by passive response to resource conflicts.

[0040] Furthermore, in step four, the curing time compression action in the correction action is limited by the minimum curing time limit for concrete specified in the process specification. When the decision network selects the curing time compression action, the compressed curing time shall not be lower than the minimum curing time limit.

[0041] The output duration of the maintenance time compression action is determined in the following way:

[0042] Obtain the concrete type parameters and ambient temperature parameters of the current component, calculate the safe compression coefficient based on the concrete type parameters and ambient temperature parameters, and use the product of the standard curing time and the safe compression coefficient as the curing time after compression.

[0043] By setting a rigid constraint that the curing time compression action must not be lower than the minimum curing time of concrete, and calculating the safe compression coefficient based on the concrete type and ambient temperature to determine the curing time after compression, it is possible to ensure that the strength of the components meets the demolding requirements when compressing the curing time to catch up with the construction schedule, and to prevent quality defects caused by excessive compression.

[0044] Furthermore, the method also includes a production line resilience quantification assessment step: after executing the rescheduling scheme, the resilience index is defined as the probability that the prefabricated component production line can still be completed on schedule under the current state and remaining resource conditions;

[0045] The resilience index is calculated by Monte Carlo simulation, specifically by performing N random disturbance simulations on the remaining production stage. The types of random disturbances include equipment failure disturbances, material delay disturbances, and personnel absence disturbances. The disturbance intensity of each random disturbance simulation is sampled according to historical statistical distribution. The number of times the project can be completed on schedule in the N random disturbance simulations is counted, and the ratio of the number of times the project can be completed on schedule to N is used as the estimated value of the resilience index.

[0046] When the resilience index is lower than the preset resilience threshold, emergency resources are automatically activated. These emergency resources include backup molds, overtime personnel, and outsourced production capacity.

[0047] By calculating the resilience index after the rescheduling scheme is executed, that is, by applying random disturbances such as equipment failure, material delay, and personnel absence to the remaining production stage using Monte Carlo simulation, and calculating the ratio of the number of times the work is completed on schedule to the total number of simulations, and automatically calling up backup modules, overtime personnel, or outsourced capacity when the resilience index is lower than the threshold, the ability of the scheduling scheme to resist disturbances can be quantitatively evaluated, and the scheduling upgrade from passive response to proactive self-healing can be achieved.

[0048] Furthermore, the method also includes offline simulation verification and online migration steps for scheduling strategies: in a digital twin environment, a variety of simulation scenarios are generated using domain randomization technology, and the decision network is trained in the simulation scenarios to obtain an initial strategy model. The domain randomization technology includes randomizing process time parameters, resource availability parameters, and order priority parameters.

[0049] The initial strategy model is deployed to the mirror sandbox of the digital twin system, and the mirror sandbox is driven by real-time data from the prefabricated component production line. The decision-making effect of the initial strategy model is verified in the virtual environment without interfering with real production.

[0050] When the cumulative reward value of the initial strategy model in the mirror sandbox is higher than the manual scheduling benchmark for M consecutive scheduling cycles, the gradual deployment process is initiated. The gradual deployment process includes: first applying the output of the initial strategy model to non-critical processes, and then gradually expanding the application scope.

[0051] If an anomaly is detected at any stage, the system will automatically revert to the previous stage's scheduling strategy or manual scheduling mode.

[0052] By using domain randomization technology in a digital twin environment to generate diverse simulation scenarios to train the decision network, the initial strategy is deployed in a mirror sandbox for real-time data verification. After successful verification, it is gradually deployed online—first applied to non-critical processes and then gradually expanded in scope, with automatic rollback in case of anomalies. This can solve the problem of uncertainty in the effect of migrating simulation strategies to the real production line, and ensure the safety and reliability of the intelligent scheduling system in actual deployment.

[0053] Compared with existing technologies, this real-time state-based prefabricated component production progress prediction and correction scheduling method has the following advantages:

[0054] I. This invention constructs a closed-loop scheduling framework covering the entire process of real-time perception, dynamic prediction, intelligent decision-making, execution, and feedback. It models the dynamic scheduling problem of prefabricated component production as a Markov decision process, directly encoding the output of the rolling prediction model as the state input of the decision network, thus achieving deep coupling between progress prediction and corrective scheduling. At the same time, the execution results of scheduling decisions are fed back to the prediction model and the decision network respectively, enabling incremental training and gradient updates. This effectively overcomes the open-loop structural defects of the separation between prediction and scheduling, adapts to various real-time disturbances in the production process, avoids the disconnect between the scheduling scheme and the actual production state, and realizes a paradigm upgrade from static offline scheduling to dynamic adaptive scheduling.

[0055] Second, this invention achieves accurate prediction of production progress through a rolling prediction model with parallel coding at multiple time scales. Combined with a reward function that optimizes multi-dimensional goals, it guides the decision network to output compliant corrective actions. At the same time, it is equipped with dynamic matching of bottleneck resources, quantitative assessment of production line resilience, offline simulation verification, and a gradual online mechanism. This can ensure the project schedule while strictly adhering to the process specifications for prefabricated component production, avoiding component quality defects caused by rushing to meet deadlines, optimizing the efficiency of production line resource utilization, and improving the anti-disturbance capability of the scheduling scheme and the safety and stability of on-site deployment.

[0056] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description

[0057] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0058] Figure 1 This is a schematic diagram of the overall process of the real-time status prefabricated component production progress prediction, correction and scheduling method of the present invention.

[0059] Figure 2 This is a schematic diagram of the architecture of the three-layer parallel long short-term memory encoder of the present invention;

[0060] Figure 3 This is a schematic diagram of the matching process of the resource-process bipartite graph of the present invention. Detailed Implementation

[0061] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0062] Example

[0063] This embodiment provides a complete implementation description of a real-time prefabricated component production progress prediction and correction scheduling method. In related technologies, prefabricated component production scheduling often adopts a static open-loop scheduling mode, which cannot adapt to real-time disturbances such as equipment failures, material delays, and personnel fluctuations during the production process. This results in insufficient progress prediction accuracy, a lack of closed-loop feedback mechanism for correction decisions, and a tendency for the scheduling plan to become disconnected from the actual production status. The following embodiment provides a detailed description of the entire scheduling method.

[0064] This embodiment is applied to an automated production line for precast concrete assembled components. The production line includes eight core stations: rebar processing, formwork assembly, rebar tying, concrete pouring, vibration, curing, demolding, and finished product inspection. In this embodiment, the scheduling cycle is set to 15 minutes, the preset delay risk threshold is set to 0.7, the preset load threshold is set to 0.85, the preset toughness threshold is set to 0.6, the prediction model update cycle is set to 24 hours, the decision network update cycle is set to 12 hours, the number of random disturbance simulations N in the Monte Carlo simulation is set to 1,000, and the continuous achievement scheduling cycle M for mirror sandbox verification is set to 20.

[0065] like Figure 1 As shown, the real-time status prefabricated component production progress prediction and correction scheduling method of this embodiment is executed cyclically according to the following steps until all components are produced.

[0066] Step 1, Real-time Data Acquisition:

[0067] At the start of each scheduling cycle, a full-dimensional data collection operation is performed on the production line. The collected data is divided into four categories: process time data, resource occupancy status data, multimodal sensor data, and basic order data.

[0068] Specifically, the process time data includes the real-time start time and real-time end time of the process currently being performed at each workstation in the production line. For processes in progress, the execution time is recorded; for processes that have not yet started, the planned start time and planned end time are recorded.

[0069] Resource occupancy status data includes the real-time occupancy status of mold table resources, crane equipment resources, personnel resources, and curing kiln resources, recording the current occupied workstation, estimated release time, and available capacity margin for each resource.

[0070] Multimodal sensing data includes process execution image data collected by workstation cameras, vibration spectrum data of the compaction process collected by accelerometers, material arrival status data collected by RFID readers, and curing temperature data collected by temperature sensors. Specifically, workstation cameras are deployed directly above each core workstation, with a frame rate of ten frames per second, and the collected image data covers the entire work area of ​​the workstation; accelerometers are installed at the working end of the compaction equipment, with a sampling frequency of 1 kHz, to collect vibration acceleration data throughout the compaction process and convert it into vibration spectrum data; RFID readers are deployed at the material inlet of each workstation to identify the arrival status and batch information of tagged steel bars, formwork, and concrete materials; and temperature sensors are deployed in different areas inside the curing kiln, with a sampling interval of one minute, to collect real-time ambient temperature and internal component temperature data inside the curing kiln.

[0071] The basic order data includes the type of the current production order, batch size, delivery cycle, and priority information.

[0072] Step 2, State Vector Encoding:

[0073] All data collected in step one is encoded into a state vector of a Markov decision process. The state vector consists of three parts: the real-time progress vector of the workstation, the resource occupancy matrix, and the rolling prediction output vector, while also incorporating process state features extracted from multimodal sensor data.

[0074] Specifically, the dimensions of the real-time progress vector for each workstation correspond one-to-one with the number of workstations on the production line. In this embodiment, it is eight-dimensional, with each dimension corresponding to the progress information of one workstation. Specifically, it includes the start time of the process, the end time of the process, the current execution status, the elapsed time, and the remaining estimated time for the corresponding workstation. The current execution status is divided into four categories: not started, in progress, completed, and interrupted. Each category is encoded with a fixed value: not started is encoded as zero, in progress as one, completed as two, and interrupted as three.

[0075] The rows of the resource occupancy matrix correspond to resource types. In this embodiment, resource types are divided into four categories: mold platform resources, crane equipment resources, personnel resources, and curing kiln resources. Therefore, the matrix has four rows. The columns of the matrix correspond to time slots. The duration of a time slot is consistent with the scheduling cycle and is set to 15 minutes. The number of time slots covers the remaining production cycle of the current order. In this embodiment, there are 96 time slots, corresponding to a 24-hour production cycle. Each element of the resource occupancy matrix represents the occupancy status and expected release time of the corresponding resource in the corresponding time slot. The occupancy status is divided into idle status and occupied status, encoded as zero and one, respectively. When the occupied status is one, the element also records the expected release time of the resource.

[0076] For the collected multimodal sensor data, feature extraction was performed to obtain process status features. Specifically, for the process execution image data, the YOLOv8 object detection model was used to extract rebar binding density features and concrete pouring smoothness features. The YOLOv8 model was pre-trained and fine-tuned on the prefabricated component production process image dataset. The input image size was set to 640x640, the confidence threshold was set to 0.5, and the output rebar binding density feature was the number of rebar nodes per unit area, and the concrete pouring smoothness feature was the elevation variance value of the pouring surface. For the vibration spectrum data of the vibration process, a one-dimensional convolutional neural network was used to extract time-frequency domain features. The one-dimensional convolutional neural network contains three convolutional layers, two pooling layers, and one fully connected layer. The input is a 1000-dimensional vibration spectrum sequence, and the output is a 64-dimensional time-frequency domain feature vector, which is used to characterize the compaction and compliance of the vibration process. Based on the material arrival status data and curing temperature data, the material state characteristics and temperature characteristics are directly extracted. The material state characteristics are coded as the material arrival status, with zero for not in place, 0.5 for partially in place, and one for fully in place. The temperature characteristics are the difference between the real-time temperature inside the curing kiln and the standard curing temperature.

[0077] Finally, the real-time progress vector of the workstation, the flattened vector of the resource occupancy matrix, and the extracted process state feature vector are concatenated to complete the encoding of the Markov decision process state vector. In this embodiment, the final encoded state vector has a dimension of 1024.

[0078] Dimensions and coding rules of each component:

[0079] Workstation real-time progress vector: 40 dimensions (8 workstations × 5 fields), with fields being process start time, end time, execution status, elapsed time, and remaining time; continuous time features are normalized to zero to one using min-max, and discrete execution status is encoded using integers 0 / 1 / 2 / 3.

[0080] Resource occupancy matrix: Flattened into 384 dimensions (4 resource types × 96 time slots × 2 fields), with fields representing occupancy status (0 / 1 encoding) and estimated release time (normalized to 0~1).

[0081] Multimodal feature vector: 600 dimensions, where:

[0082] YOLOv8n model: Input size 640×640, confidence threshold 0.5, output two-dimensional visual features (rebar binding density, concrete pouring flatness), normalized to zero to one;

[0083] One-dimensional convolutional neural network: The structure is Conv1d(1,16,3)-MaxPool1d(2)-Conv1d(16,32,3)-MaxPool1d(2)-Conv1d(32,64,3)-Flatten-Linear(7808,64), the input is a 1,000-dimensional vibration spectrum, and the output is a 64-dimensional time-frequency feature;

[0084] The material state characteristics are one-dimensional, the temperature characteristics are one-dimensional, and the remaining auxiliary characteristics are 532-dimensional.

[0085] Step 3, Progress Rolling Forecast and Delay Probability Calculation:

[0086] The state vector obtained from step two is input into the rolling prediction model based on a long short-term memory network. The model outputs the predicted start time and predicted end time of each subsequent process, the predicted completion time of the entire production line, and calculates the delay probability of each process.

[0087] like Figure 2 As shown, the rolling prediction model employs a three-layer parallel long short-term memory encoder architecture, coupled with an attention fusion module and a prediction output head. The three-layer parallel long short-term memory encoder consists of a micro-long short-term memory encoder, a meso-long short-term memory encoder, and a macro-long short-term memory encoder.

[0088] Specifically, the microscopic long short-term memory encoder uses a minute-based time step to input sensor timing data for each process, including concrete pouring rate data and vibration frequency data. The time step is set to one minute, and the input sequence length is set to sixty, corresponding to the sensor timing data of the past sixty minutes. The encoder contains two long short-term memory network layers, with a hidden layer dimension of 128, and outputs a 64-dimensional in-process state feature vector to characterize the real-time operation status and execution efficiency within the process.

[0089] The meso-level Long Short-Term Memory (LSTM) encoder uses process steps as the step size, inputting process-level event data, including process start time, process end time, and resource occupancy status data. The input sequence length is set to eight, corresponding to the eight core processes of the production line. The encoder contains two LSM network layers, with a hidden layer dimension of 128, and outputs a 64-dimensional inter-process flow feature vector to characterize the connection efficiency and flow bottlenecks between processes.

[0090] The macroscopic long short-term memory encoder uses order batches as the step size, inputting historical order data, including order type data, batch size data, and delivery cycle data. The input sequence length is set to twelve, corresponding to the data of the past twelve completed order batches. The encoder contains two long short-term memory network layers, with a hidden layer dimension of 128, and outputs a 64-dimensional order-level periodic feature vector to characterize the periodic pattern and overall progress trend of order batch production.

[0091] The attention fusion module performs weighted fusion of the feature vectors output by the three-layer encoder. The attention weights of the attention fusion module are adaptively determined by the sensitivity of the current production stage. Specifically, the calculation process of the attention weights is as follows: First, the progress completion rate of the current production stage is calculated, which is the ratio of the number of completed processes to the total number of processes. Then, the initial weights of the three-layer features are adaptively adjusted according to the progress completion rate. When the progress completion rate in the early stage of production is lower than 0.3, the initial weights of macro features are set to 0.5, meso features to 0.3, and micro features to 0.2. When the progress completion rate in the middle stage of production is between 0.3 and 0.7, the initial weights of meso features are set to 0.5, macro features to 0.25, and micro features to 0.25. When the progress completion rate in the later stage of production is higher than 0.7, the initial weights of micro features are set to 0.5, meso features to 0.3, and macro features to 0.2. Then, the initial weights are normalized using the softmax function to obtain the final attention weights. The three feature vectors are multiplied by their corresponding attention weights and then summed to obtain the fused total feature vector, which has a dimension of 64.

[0092] The three-layer encoder uniformly adopts a two-layer LSTM structure, with a hidden layer dimension of 128 and an activation function of tanh. The attention weights are adaptively based on the following criteria: in the early stage of production, the focus is on the overall trend of the order; in the middle stage, the focus is on the efficiency of the process flow; and in the later stage, the focus is on the execution accuracy of the single process. Therefore, the initial weights are set in segments according to the progress completion rate, and the final weights are obtained after softmax normalization.

[0093] The fused total feature vector is input to the prediction output head, which contains two fully connected layers with a hidden layer dimension of 256. The output includes the predicted start and end times of each subsequent process, the predicted completion time of the entire production line, and the delay probability of each process. The delay probability of each process is calculated using a Gaussian cumulative function based on the deviation between the predicted and planned completion times, combined with the probability distribution of historical process hours. The delay risk threshold is determined based on historical delay loss statistics over the past twelve months, ranging from 0.5 to 0.9; in this embodiment, it is set to 0.7. When the predicted completion time exceeds the planned completion time by one standard deviation, the delay probability is 0.84. When the predicted completion time is later than the planned completion time, the delay probability monotonically increases with the increase of the time deviation.

[0094] Error caching pool mechanism: capacity of 10,000 records, adopting a first-in-first-out replacement strategy; incremental training strategy: executed at 24:00 every day, the training set is all data in the cache pool, the optimization objective is to minimize the mean squared error, the Adam optimizer learning rate is 0.0001, the batch size is 32, the iteration is 10 rounds, and only the weights of the prediction output head and the attention fusion module are updated.

[0095] Step 4, Corrective Action Decision:

[0096] The delay probability of each process output in step three is monitored in real time. When the delay probability of any process exceeds the preset delay risk threshold of 0.7, the state vector encoded in step two is input into the decision network based on deep Q-learning, and the decision network outputs the optimal correction action.

[0097] Specifically, the decision network adopts a dual-depth Q-network architecture, which includes an online Q-network and a target Q-network. The two networks have completely identical structures, each containing three fully connected layers. The input layer dimension is the same as the state vector dimension, which is 1024 dimensions. The hidden layer dimension is set to 512 dimensions, and the output layer dimension is the same as the number of actions in the action space.

[0098] Decision network structure: three fully connected layers, ReLU activation function; experience replay pool capacity: 100,000 data points; gradient update rule: Adam optimizer learning rate: 0.0002, batch size: 64; target network update frequency: soft update performed every five online network updates, soft update coefficient: 0.001; priority calculation formula for experience replay: The sampling probability is directly proportional to the P-value.

[0099] The action space includes six types of optional actions: process speed adjustment actions, resource reallocation actions, process sequence adjustment actions, maintenance time compression actions, emergency resource mobilization actions, and no adjustment actions.

[0100] Specifically, the process speed adjustment actions include two sub-actions: increasing the number of workers at each workstation and increasing equipment operating power. The adjustment range is divided into three levels: low, medium, and high. The resource reallocation actions include three sub-actions: cross-workstation allocation of mold table resources, crane equipment resources, and personnel resources, prioritizing the allocation of idle resources to processes with delay risks. The process sequence adjustment actions include adjusting the execution order of processes without strict dependencies, changing parallel processes from sequential to parallel execution. The curing time compression action is used to compress the curing time of concrete components. The compressed curing time must not be lower than the minimum curing time limit for concrete specified in the process specification. In this embodiment, the minimum curing time limit is set to eight hours. The emergency resource mobilization actions include three sub-actions: mobilizing spare mold tables, activating overtime personnel, and coordinating with external production capacity. The no-adjustment action is used in scenarios where the probability of delay does not exceed the threshold, or where other actions cannot bring positive benefits.

[0101] The training and decision-making process of the decision network is optimized based on the reward function, which is calculated using the following formula:

[0102]

[0103] Here, R represents the total output value of the reward function, which is dimensionless and is the core objective of decision network optimization. The higher the total output value, the better the overall effect of the corresponding corrective action. This represents the weighting coefficient for items completed on schedule. It is dimensionless and set to 0.4 in this embodiment. It is used to adjust the proportion of the target completed on schedule in the overall reward. This represents the reward value for timely completion. It is dimensionless and takes a positive value when the actual completion time is earlier than or equal to the planned completion time. The value is directly proportional to the length of time the work is completed ahead of schedule. It takes a negative value when the actual completion time is later than the planned completion time. The absolute value of the negative value is directly proportional to the square of the delay time. It is used to impose a strong penalty on delays. The longer the delay, the stronger the penalty becomes exponentially. This represents the weighting coefficient of the resource utilization efficiency item. It is dimensionless and is set to 0.25 in this embodiment. It is used to adjust the proportion of the resource utilization efficiency target in the overall reward. This represents the reward value for resource utilization efficiency. It is dimensionless and is calculated based on the deviation between the actual working hours utilization rate and the target utilization rate of each workstation. The smaller the deviation between the actual working hours utilization rate and the target utilization rate, the higher the reward value. When the actual working hours utilization rate exceeds the upper limit of the target utilization rate or falls below the lower limit threshold, the reward value is negative. This represents the weight coefficient of the scheduling stability item. It is dimensionless and is set to 0.2 in this embodiment. It is used to adjust the proportion of the scheduling stability target in the overall reward. This represents the scheduling stability reward value, which is dimensionless and is a negative correlation function of the degree of difference between the current scheduling scheme and the previous scheduling scheme. The smaller the difference between the two scheduling schemes, the higher the reward value, which is used to avoid production order chaos caused by frequent and large-scale adjustments to the scheduling scheme. This represents the weighting coefficient for the process compliance item. It is dimensionless and set to 0.15 in this embodiment. It is used to adjust the proportion of the process compliance target in the overall reward. This represents the process compliance penalty value, which is dimensionless. When the corrective action violates any of the following constraints: the lower limit of curing time, the initial setting time window of concrete, or the process sequence dependency constraint, it takes a large negative value. In this embodiment, it is set to -100. When all constraints are satisfied, the value is zero. It is used to force the decision network to follow the production process specifications and avoid outputting corrective actions that do not meet quality requirements.

[0104] The weighting coefficients are determined based on the analytic hierarchy process (AHP) and the constraints. value range 0.3~0.5 0.2~0.3 0.15~0.25 0.1~0.2; Sub-reward calculation and normalization:

[0105] : in advance During the postponement Normalized to [-100, 100];

[0106] Normalized to [0, 100];

[0107] D represents the degree of difference between the two schemes. To maximize the possible difference, normalize to [-100, 0];

[0108] When violating the rules, the value is -100; when complying with the rules, the value is 0.

[0109] The decision network employs a priority experience replay mechanism for training and optimization. Specifically, each experience sample in the experience replay pool is assigned a sampling priority according to the severity of the delay; experience samples with longer delay times or more delayed processes have higher sampling priorities. The quantitative index of delay severity is calculated based on the deviation between the actual completion time and the planned completion time, as well as the number of subsequent processes affected by the delay. The calculation formula is as follows:

[0110]

[0111] Where P represents the priority value of the empirical sample; the higher the value, the higher the sampling priority. α represents the weighting coefficient of the time deviation term, which is dimensionless and set to 0.6 in this embodiment. This indicates the actual completion time of the process, in minutes. The planned completion time of the process is expressed in minutes. β represents the weighting coefficient of the affected process item, which is dimensionless and set to 0.4 in this embodiment. This represents the number of subsequent processes affected by the current process delay, and is dimensionless. During each update, the decision network samples experience samples from the experience replay pool according to sampling priority. Samples with higher priority values ​​have a greater probability of being sampled, ensuring the network learns first from high-risk delay scenarios and improves its response to emergency disturbances.

[0112] The output duration of the curing time compression action is determined using a fixed method. Specifically, the concrete type parameters and ambient temperature parameters of the current component are first obtained. The concrete type parameters include the concrete strength grade, initial setting time, and final setting time. The ambient temperature parameter is the real-time average temperature inside the curing kiln. Based on the concrete type parameters and ambient temperature parameters, the safe compression coefficient is calculated. The formula for calculating the safe compression coefficient is as follows:

[0113]

[0114] Where k represents the safety compression coefficient, which is dimensionless and ranges from 0.6 to 1.0, ensuring that the curing time after compression is not less than the minimum curing time limit. The base compression coefficient is dimensionless and determined based on the concrete strength grade. In this embodiment, the base compression coefficient of C30 concrete is set to 0.8, the base compression coefficient of C35 concrete is set to 0.85, and the base compression coefficient of C40 and above strength grade concrete is set to 0.9. This represents the temperature correction factor, which is dimensionless and determined based on the ambient temperature of the curing environment. When the curing temperature is within the standard curing temperature range of 20 degrees Celsius ± 2 degrees Celsius, the temperature correction factor is set to 1.0. When the curing temperature is higher than 22 degrees Celsius, the temperature correction factor decreases linearly with increasing temperature, with a minimum of 0.9. When the curing temperature is lower than 18 degrees Celsius, the temperature correction factor increases linearly with decreasing temperature, with a maximum of 1.1. This represents the strength correction factor, dimensionless, determined based on the ratio of the concrete's current strength to its design strength. When the current strength exceeds 70% of the design strength, the strength correction factor is set to 0.95; when the current strength is between 50% and 70% of the design strength, the strength correction factor is set to 1.0; and when the current strength is less than 50% of the design strength, the strength correction factor is set to 1.05. The product of the standard curing time and the safe compression coefficient is used as the curing time after compression. Simultaneously, it is verified that the curing time after compression must not be less than the minimum curing time limit of eight hours. If the calculated result is lower than the lower limit, the lower limit value is directly taken as the final curing time after compression.

[0115] The minimum curing time is determined based on GB50666-2011 "Code for Construction of Concrete Structures"; the temperature correction factor is calculated as follows: when the temperature is >22℃. =1.0-0.01×(T-22), minimum 0.9; <18℃ =1.0+0.01×(18-T), up to 1.1.

[0116] Step 5: Rescheduling scheme conversion and execution:

[0117] The corrective actions output by the decision network are converted into executable rescheduling schemes. The conversion process involves first breaking down the corrective actions into specific work instructions for each workstation, including instructions for adjusting process execution time, resource allocation, work parameter adjustments, and maintenance duration adjustments. Then, all work instructions are verified to meet process and resource constraints, eliminating those that do not meet the constraints, and supplementing the preconditions and time nodes for instruction execution. The final rescheduling scheme includes updated process plans, resource allocation plans, work parameter requirements, and quality control points for each workstation. The rescheduling scheme is then distributed to the manufacturing execution system (MES) on the production line via industrial Ethernet. The MES breaks down the scheme into work tasks for each workstation and pushes them to the operating terminals and automated equipment at each workstation, driving the production line to execute production operations according to the rescheduling scheme.

[0118] Step 6, Incremental Training of the Rolling Prediction Model:

[0119] After the rescheduling scheme is executed, the actual completion time of each process is collected, and the prediction error between the actual completion time and the predicted end time output in step three is calculated. The formula for calculating the prediction error is:

[0120]

[0121] Where e represents the prediction error for a single process, in minutes. This indicates the actual completion time of the process, in minutes. This represents the predicted completion time of the process output by the rolling prediction model, in minutes. The calculated prediction error is stored in an error cache pool with a fixed storage capacity (10,000 records in this embodiment). When the storage capacity is full, the earliest stored error data is replaced according to the first-in, first-out (FIFO) principle. The rolling prediction model is incrementally trained with a preset prediction model update cycle of 24 hours. The training set for incremental training consists of all data in the error cache pool, along with the state vector data and actual completion data within the corresponding cycle. The optimization objective of the training is to minimize the mean squared error between the predicted completion time and the actual completion time. The optimizer used is the Adam optimizer, with a learning rate of 0.0001, a training batch size of 32, and 10 training iterations. After training, the weight parameters of the rolling prediction model are updated.

[0122] Step 7, Decision Network Gradient Update:

[0123] After the corrective action is executed, the new state vector of the production line after the corrective action is obtained. Combined with the reward value obtained during the execution process and the state transition after the action, experience samples are generated. The format of the experience sample is: current state vector, executed corrective action, obtained reward value, new state vector after transition, and termination flag indicating whether production is complete. The generated experience samples are stored in an experience replay pool with a fixed storage capacity (100,000 samples in this embodiment). When the storage capacity is full, the earliest stored experience sample is replaced according to the first-in, first-out principle. The decision network is updated using a gradient at a preset decision network update cycle of twelve hours. During the gradient update, a priority experience replay mechanism is used to sample experience samples from the experience replay pool. The batch size is set to 64. The time difference error of the online Q network is calculated based on the sampled samples. The weight parameters of the online Q network are updated using gradient descent. The weight parameters of the target Q network are soft-updated every five updates from the online Q network, with a soft-update coefficient set to 0.001 to ensure the stability of the training process.

[0124] Step 8, cyclical execution and production termination:

[0125] Repeat all operations from step one to step seven, executing the complete process once per scheduling cycle, until all prefabricated components for the current order are produced and all finished products are inspected and put into storage, then terminate the scheduling process for the current order.

[0126] Bottleneck resource dynamic matching steps:

[0127] This embodiment also includes a bottleneck resource dynamic matching step, which is executed synchronously with the data acquisition in step one during each scheduling cycle. For example... Figure 3As shown, a resource-process bipartite graph is first constructed, comprising a set of resource nodes and a set of process nodes. Each resource node in the resource node set carries its current occupancy status and estimated release time. The type of resource node is consistent with the core resource type of the production line, including mold table resource nodes, crane equipment resource nodes, personnel resource nodes, and curing kiln resource nodes. Each process node in the process node set carries its resource requirement type and requirement duration. Each process node corresponds to all processes currently to be executed on the production line. Resource nodes and process nodes are connected by distributable edges; a distributable edge exists between two nodes only if the resource type of a resource node satisfies the resource requirement of a process node.

[0128] In each scheduling cycle, based on the real-time data collected in step one, the occupancy status and estimated release time of each resource node in the resource node set are updated, and the resource demand information of each process node in the process node set is also updated. Then, the matching score between each process node and the candidate resource node is calculated. The matching score comprehensively considers spatial distance, current load, and historical matching success rate factors. The calculation formula is as follows:

[0129]

[0130] Where S represents the matching score between the process node and the candidate resource node. The higher the score, the higher the matching priority. The weighting coefficient for the spatial distance term is dimensionless and is set to 0.3 in this embodiment. This represents the spatial distance matching score, which is dimensionless and calculated based on the spatial distance between the workstation where the resource is located and the workstation where the process is located. The closer the distance, the higher the score, and the value ranges from zero to one. This represents the weight coefficient of the current load item. It is dimensionless and is set to 0.4 in this embodiment. This represents the load matching score, which is dimensionless and calculated based on the current load rate of the resource node. The lower the load rate, the higher the score, and the value ranges from zero to one. This represents the weight coefficient of the historical matching success rate item. It is dimensionless and is set to 0.3 in this embodiment. This represents the historical matching score, which is dimensionless and calculated based on the ratio of the number of successful historical matches between the resource and this type of process to the total number of matches. The higher the ratio, the higher the score, and the value ranges from zero to one.

[0131] The Hungarian algorithm is used to solve the maximum weight perfect matching problem in the resource-process bipartite graph, obtaining the optimal resource allocation scheme. This optimal resource allocation scheme is synchronized to the rescheduling scheme and then sent to the manufacturing execution system for execution. Simultaneously, the current load of each resource node is monitored in real time. When the current load of any resource node exceeds a preset load threshold of 0.85%, a resource overload warning is output, and a search is automatically conducted for alternative resource nodes of the same type. Resource allocation suggestions are generated and incorporated into the rescheduling scheme.

[0132] Resource-process bipartite graph construction rules: resource nodes include ID, type, location, current load, and estimated release time; process nodes include ID, workstation, required resource type, required duration, and earliest start time; allocable edges exist only when the resource type matches and the estimated release time is less than or equal to the earliest start time of the process.

[0133] Matching degree sub-item calculation: (d is the distance between workstations) (Maximum distance of the production line). (L represents the current resource load rate); Historical successful matches / Total matches.

[0134] Production line resilience quantitative assessment steps:

[0135] This embodiment also includes a production line resilience quantification assessment step, which is executed synchronously with the data acquisition in step six after each rescheduling scheme is completed. Specifically, the resilience index is defined as the probability that the prefabricated component production line can still be completed on schedule under the current state and remaining resource conditions. The resilience index is calculated through Monte Carlo simulation. The specific implementation process is as follows: First, all processes in the remaining production stage are determined, along with the corresponding process time distribution, resource availability distribution, and historical statistical distribution of disturbance types and intensity. The types of random disturbances include equipment failure disturbances, material delay disturbances, and personnel absence disturbances. The disturbance intensity of each disturbance is fitted based on the historical statistical data of the production line over the past twelve months to obtain the corresponding probability distribution. One thousand random disturbance simulations are performed on the remaining production stage. During each simulation, the intensity of the corresponding disturbance is sampled based on the probability distribution, the time and resource availability of the remaining processes are updated, the completion time after simulation is calculated, and it is determined whether the project can be completed on schedule. The resilience index is estimated by calculating the number of times the project can be completed on schedule out of 1,000 random disturbance simulations. The ratio of the number of times the project can be completed on schedule to the total number of simulations is used as the estimated value, ranging from zero to one. When the calculated resilience index is lower than the preset resilience threshold of 0.6, the emergency resource allocation process is automatically triggered, an emergency resource allocation plan is generated, and it is included in the rescheduling plan. Emergency resources include backup modules, overtime personnel, and outsourced production capacity.

[0136] Monte Carlo simulation parameters: the disturbance distribution is equipment failure (exponential distribution λ=0.001 times / minute), material delay (normal distribution μ=30 minutes, σ=15 minutes), and personnel absence (binomial distribution p=0.05); each simulation is independently sampled, and the resilience threshold is determined based on historical production data statistics, with a value range of 0.5 to 0.8, and 0.6 is used in this embodiment.

[0137] Offline simulation verification and online migration steps of scheduling strategy:

[0138] This embodiment also includes offline simulation verification and online migration steps for the scheduling strategy. These steps are performed before the method is formally deployed to the real production line, and periodic strategy optimization is conducted during production line downtime for maintenance. Specifically, a digital twin environment for the production line is first constructed. This digital twin environment fully maps the workstation layout, equipment parameters, process specifications, resource configuration, and historical production data of the real production line. Within the digital twin environment, domain randomization technology is used to generate diverse simulation scenarios. This technology includes randomizing process time parameters, resource availability parameters, and order priority parameters. The range of parameter randomization is determined based on the historical data fluctuation range of the real production line, ensuring that the simulation scenarios cover most possible operating conditions in real production. A decision network is trained in the generated diverse simulation scenarios to obtain an initial strategy model. The training process employs a reward function and a priority experience replay mechanism consistent with online updates. The training iteration rounds are set to one thousand rounds. After training, the weight parameters of the initial strategy model are saved. The initial strategy model is deployed to a mirror sandbox within the digital twin system. The mirror sandbox synchronizes data in real-time with the manufacturing execution system (MES) of the real production line, using real-time production line data to drive the mirror sandbox's operation. This virtual environment validates the decision-making effectiveness of the initial strategy model without interfering with normal production on the real production line. The performance of the initial strategy model in the mirror sandbox is continuously monitored. When the model's cumulative reward value exceeds the manual scheduling baseline for twenty consecutive scheduling cycles, a gradual rollout process is initiated. This process consists of three phases: Phase 1 applies the initial strategy model's output only to non-critical processes, while critical processes remain under manual scheduling. After 72 hours of continuous operation without anomalies, the process proceeds to Phase 2. Phase 2 expands the model's application to all non-critical processes and some low-risk critical processes. After 168 hours of continuous operation without anomalies, the process proceeds to Phase 3. Phase 3 achieves full application of the model to all processes across the entire production line. At any stage of the gradual rollout, if an abnormal model output is detected, or if a significant disturbance occurs on the production line, the system automatically reverts to the previous stage's scheduling strategy. In extreme cases, it directly switches to manual scheduling mode to ensure production safety.

[0139] The core database and execution logic of this embodiment are as follows:

[0140] Database structure: production real-time database (fields: ID, timestamp, workstation ID, value, unit), process constraint database (source: GB50666-2011, enterprise process manual, fields: constraint ID, type, threshold, judgment logic), resource status database (fields: resource ID, type, status, expected release time); real-time data is entered into the database every 15 minutes and outliers are removed after 3σ verification.

[0141] Rescheduling and MES interface: adopts OPCUA protocol, interface fields: instruction ID, workstation ID, job type, start time, end time, parameters; rescheduling generation process: action parsing to constraint verification to resource conflict detection to generate job instruction.

[0142] Closed-loop update process: Pre-update is triggered when the error cache pool is full of 1,000 entries; full incremental training of the prediction model is performed at 24:00 every day; and gradient update is performed on the decision network every 12 hours.

[0143] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A real-time status-based prefabricated component production progress prediction and correction scheduling method, characterized in that, The method includes the following steps: Step 1: In each scheduling cycle, collect the real-time start time of the process, the real-time end time of the process, the current resource occupancy status, and multimodal sensor data of each workstation in the prefabricated component production line. Step 2: Encode the collected real-time start time of the process, real-time end time of the process, current resource occupancy status, and process state features extracted based on the multimodal sensor data into a state vector of a Markov decision process; Step 3: Input the state vector into the rolling prediction model based on the long short-term memory network. The rolling prediction model outputs the predicted start time, predicted end time of each subsequent process and the predicted completion time of the entire prefabricated component production line, and calculates the delay probability of each process. Step 4: When the delay probability of any process exceeds the preset delay risk threshold, the state vector is input into a decision network based on deep Q-learning. The decision network outputs a correction action, which is selected from the action space. The action space includes process speed adjustment action, resource reallocation action, process sequence adjustment action, maintenance time compression action, emergency resource call action, and no adjustment action. Step 5: Convert the correction action into an executable rescheduling scheme, and send the rescheduling scheme to the manufacturing execution system for execution; Step 6: Collect the actual process completion time after the rescheduling scheme is executed, calculate the prediction error between the actual process completion time and the predicted end time, store the prediction error in the error cache pool, and incrementally train the rolling prediction model with a preset prediction model update cycle. Step 7: Based on the reward value obtained after performing the correction action and the new state vector after the state transition, generate an experience sample, store the experience sample in the experience replay pool, and perform gradient updates on the decision network at a preset decision network update cycle. Step 8: Repeat steps 1 through 7 until all components are produced.

2. The real-time status-based prefabricated component production progress prediction and correction scheduling method according to claim 1, characterized in that, In step three, the rolling prediction model based on the long short-term memory network adopts a three-layer parallel long short-term memory encoder architecture, which includes: A microscopic long short-term memory encoder with a time step of minutes inputs sensor timing data for each process, including concrete pouring rate data and vibration frequency data, and outputs a state feature vector within the process. A meso-level long short-term memory encoder with process as the step size takes process-level event data as input, including process start time, process end time and resource occupancy status data, and outputs inter-process flow feature vector. A macroscopic long short-term memory encoder with order batch as the step size takes historical order data as input, including order type data, batch size data and delivery cycle data, and outputs an order-level periodic feature vector. The rolling prediction model also includes an attention fusion module, which weights and fuses the intra-process state feature vector, the inter-process flow feature vector, and the order-level periodic feature vector. The attention weight of the attention fusion module is adaptively determined by the sensitivity of the current production stage.

3. The real-time status-based prefabricated component production progress prediction and correction scheduling method according to claim 1, characterized in that, In step four, the reward function of the decision network based on deep Q-learning is calculated according to the following formula: ； Where R represents the total output value of the reward function, which is dimensionless. This represents the weighting coefficient for items completed on schedule; it is dimensionless. This represents the reward value for on-time completion. A positive value is taken when the actual completion time is earlier than or equal to the planned completion time, and a negative value is taken when the actual completion time is later than the planned completion time. The negative value is directly proportional to the square of the delay time. This represents the weighting coefficient for the resource utilization efficiency term; it is dimensionless. This represents the resource utilization efficiency bonus value, calculated based on the deviation between the actual working hours utilization rate and the target utilization rate for each workstation. The weighting coefficients for the scheduling stability term are dimensionless. This represents the scheduling stability reward value, which is a negative correlation function of the degree of difference between the current scheduling scheme and the previous scheduling scheme. This represents the weighting coefficient for the process compliance item; it is dimensionless. This represents the penalty value for process compliance. It takes a large negative value when any of the following constraints are violated: minimum curing time constraint, initial concrete setting time window constraint, or process sequence dependency constraint.

4. The real-time status-based prefabricated component production progress prediction and correction scheduling method according to claim 1, characterized in that, In step two, the state vector includes a workstation real-time progress vector, a resource occupancy matrix, and a rolling prediction output vector. The real-time progress vector of each workstation includes the process start time, process end time, current execution status, elapsed time and remaining estimated time. The current execution status includes not started, in progress, completed and interrupted states. The rows of the resource occupancy matrix correspond to resource types, and the columns correspond to time slots. The resource types include mold table resources, crane equipment resources, personnel resources, and curing kiln resources. Each matrix element of the resource occupancy matrix represents the occupancy status of the corresponding resource in the corresponding time slot and the expected release time. The rolling prediction output vector includes the predicted start time and end time of each subsequent process, the predicted completion time of the entire prefabricated component production line, and the delay probability of each process.

5. The real-time status-based prefabricated component production progress prediction and correction scheduling method according to claim 1, characterized in that, In step four, the decision network adopts a priority experience replay mechanism: each experience sample in the experience replay pool is assigned a sampling priority according to the severity of the delay. The longer the delay time or the more delayed processes, the higher the sampling priority of the experience sample. The decision network samples experience samples from the experience replay pool according to the sampling priority during each update. The quantitative index of the severity of the delay is calculated based on the deviation between the actual completion time and the planned completion time, as well as the number of subsequent processes affected by the delay.

6. The real-time status-based prefabricated component production progress prediction and correction scheduling method according to claim 1, characterized in that, In step one, the multimodal sensing data includes process execution image data collected by the workstation camera, vibration spectrum data of the vibration process collected by the accelerometer, material arrival status data collected by the radio frequency identification reader, and curing temperature data collected by the temperature sensor. In step two, the process state features extracted based on the multimodal sensing data include: rebar binding density features and concrete pouring flatness features extracted from the process execution image data using the YOLOv8 target detection model; time-frequency domain features extracted from the vibration spectrum data of the vibration process using a one-dimensional convolutional neural network; and material state features and temperature features extracted from the material arrival status data and the curing temperature data.

7. The real-time status-based prefabricated component production progress prediction and correction scheduling method according to claim 1, characterized in that, The method further includes a bottleneck resource dynamic matching step: constructing a resource-process bipartite graph, wherein the resource-process bipartite graph contains a set of resource nodes and a set of process nodes, wherein each resource node in the set of resource nodes carries the current occupancy status and the expected release time, and each process node in the set of process nodes carries the resource demand type and the demand duration, and the resource nodes and the process nodes are connected by an allocatable edge; In each scheduling cycle, the occupancy status and expected release time of each resource node in the resource node set are updated based on real-time data. The matching degree score between each process node and the candidate resource node is calculated. The matching degree score takes into account spatial distance factors, current load factors and historical matching success rate factors. The Hungarian algorithm is used to solve the maximum weight perfect matching of the resource-process bipartite graph to obtain the optimal resource allocation scheme. When the current load of any resource node exceeds the preset load threshold, an early warning message is output, and an alternative resource node is automatically searched.

8. The real-time status-based prefabricated component production progress prediction and correction scheduling method according to claim 1, characterized in that, In step four, the curing time compression action in the correction action is limited by the minimum curing time limit for concrete specified in the process specification. When the decision network selects the curing time compression action, the compressed curing time shall not be lower than the minimum curing time limit. The output duration of the maintenance time compression action is determined in the following way: Obtain the concrete type parameters and ambient temperature parameters of the current component, calculate the safe compression coefficient based on the concrete type parameters and ambient temperature parameters, and use the product of the standard curing time and the safe compression coefficient as the curing time after compression.

9. The real-time status-based prefabricated component production progress prediction and correction scheduling method according to claim 1, characterized in that, The method also includes a production line resilience quantification assessment step: after executing the rescheduling scheme, the resilience index is defined as the probability that the prefabricated component production line can still be completed on schedule under the current state and remaining resource conditions; The resilience index is calculated by Monte Carlo simulation, specifically by performing N random disturbance simulations on the remaining production stage. The types of random disturbances include equipment failure disturbances, material delay disturbances, and personnel absence disturbances. The disturbance intensity of each random disturbance simulation is sampled according to historical statistical distribution. The number of times the project can be completed on schedule in the N random disturbance simulations is counted, and the ratio of the number of times the project can be completed on schedule to N is used as the estimated value of the resilience index. When the resilience index is lower than the preset resilience threshold, emergency resources are automatically activated. These emergency resources include backup molds, overtime personnel, and outsourced production capacity.

10. The real-time status-based prefabricated component production progress prediction and correction scheduling method according to claim 1, characterized in that, The method also includes offline simulation verification and online migration steps for scheduling strategies: in a digital twin environment, a variety of simulation scenarios are generated using domain randomization technology, and the decision network is trained in the simulation scenarios to obtain an initial strategy model. The domain randomization technology includes randomizing process time parameters, resource availability parameters, and order priority parameters. The initial strategy model is deployed to the mirror sandbox of the digital twin system, and the mirror sandbox is driven by real-time data from the prefabricated component production line. The decision-making effect of the initial strategy model is verified in the virtual environment without interfering with real production. When the cumulative reward value of the initial strategy model in the mirror sandbox is higher than the manual scheduling benchmark for M consecutive scheduling cycles, the gradual deployment process is initiated. The gradual deployment process includes: first applying the output of the initial strategy model to non-critical processes, and then gradually expanding the application scope. If an anomaly is detected at any stage, the system will automatically revert to the scheduling strategy of the previous stage or the manual scheduling mode.

Images