Intelligent scheduling method and system for factory production plan

Abstract

The invention relates to the technical field of intelligent scheduling, and discloses an intelligent scheduling method and system for a factory production plan, and the method comprises the steps: carrying out the three-dimensional scanning and sensor data collection of a factory environment, and obtaining a digital twin model and a real-time production data flow; according to the digital twinborn model and the real-time production data flow, performing prototype comparison learning processing on a double-tower encoder to obtain a pre-training scheduling model; performing knowledge constraint and knowledge distillation processing on the pre-trained scheduling model to obtain a knowledge enhanced scheduling model; according to the knowledge enhanced scheduling model, constructing a double-loop coordinated scheduling system comprising a production intelligent control loop and a resource dynamic allocation loop; and performing multi-objective optimization and predictive scheduling calculation based on the double-loop coordinated scheduling system to obtain a production plan scheduling scheme. Therefore, the problem of knowledge migration of a traditional scheduling system when production conditions change is effectively solved, and the scheduling accuracy in a dynamic production environment is improved.

Description

Technical Field

[0001] This application relates to the field of intelligent scheduling technology, and particularly to an intelligent scheduling method and system for factory production plans. Background Art

[0002] In the digital factory environment, traditional production scheduling methods usually adopt fixed models based on rules or mathematical optimization. Although they perform well in stable production environments, they show obvious limitations in the complex scenarios of multiple production lines and frequently changing production conditions. On the one hand, traditional scheduling systems lack the ability to adapt to dynamic changes. When the production environment changes, the scheduling system needs to be redesigned and adjusted, and knowledge transfer is difficult, resulting in a decline in system performance or even collapse. On the other hand, when switching between normal production and emergency scheduling, traditional methods often show serious production fluctuations and cannot make a smooth transition, seriously affecting the continuity and stability of production.

[0003] Existing production scheduling technologies usually adopt a passive response strategy when dealing with situations such as sudden orders, equipment failures, or material shortages, lacking foresight and prediction capabilities, resulting in unreasonable resource allocation and low production efficiency. Especially in a mixed assembly line environment, the coordinated scheduling of equipment and the logistics system has become a key technical problem. Existing methods often only focus on the optimization of a single goal, such as minimizing the completion time, while ignoring multi-dimensional goals such as energy consumption, production cost, and equipment utilization rate, and it is difficult to achieve global optimality. At the same time, traditional scheduling systems usually separate the physical production environment from scheduling decisions, lacking a real-time feedback mechanism, resulting in the disconnection between the scheduling plan and the actual production situation, reducing the feasibility and execution efficiency of the plan. Summary of the Invention

[0004] This application provides an intelligent scheduling method and system for factory production plans, thereby effectively solving the knowledge transfer problem of traditional scheduling systems when production conditions change and improving the scheduling accuracy in a dynamic production environment.

[0005] In the first aspect of this application, an intelligent scheduling method for factory production plans is provided. The intelligent scheduling method for factory production plans includes: Performing three-dimensional scanning and sensor data collection on the factory environment to obtain a digital twin model and real-time production data streams; Performing prototype comparison learning processing on a dual tower encoder according to the digital twin model and the real-time production data streams to obtain a pre-trained scheduling model; Performing knowledge constraint and knowledge distillation processing on the pre-trained scheduling model to obtain a knowledge-enhanced scheduling model; Constructing a dual-loop coordinated scheduling system including a production intelligent control loop and a resource dynamic allocation loop according to the knowledge-enhanced scheduling model; Based on the dual-loop coordinated scheduling system, multi-objective optimization and predictive scheduling calculations are performed to obtain a production plan scheduling scheme.

[0006] The second aspect of this application provides an intelligent scheduling system for factory production plans. The intelligent scheduling system for factory production plans includes: An acquisition module, configured to perform three-dimensional scanning and sensor data acquisition on the factory environment to obtain a digital twin model and real-time production data streams; A contrastive learning module, configured to perform prototype contrastive learning processing on a dual-tower encoder according to the digital twin model and the real-time production data streams to obtain a pre-trained scheduling model; A knowledge distillation module, configured to perform knowledge constraint and knowledge distillation processing on the pre-trained scheduling model to obtain a knowledge-enhanced scheduling model; A construction module, configured to construct a dual-loop coordinated scheduling system including a production intelligent control loop and a resource dynamic allocation loop according to the knowledge-enhanced scheduling model; A calculation module, configured to perform multi-objective optimization and predictive scheduling calculations based on the dual-loop coordinated scheduling system to obtain a production plan scheduling scheme.

[0007] Compared with the prior art, this application has the following beneficial effects: By establishing a virtual model of the production environment through digital twin technology and combining prototype contrastive learning pre-training and a knowledge constraint mechanism, the scheduling system can quickly adapt to new production scenarios, effectively solve the problem of knowledge transfer in traditional scheduling systems when production conditions change, and improve the scheduling accuracy in a dynamic production environment. The dual-loop coordinated scheduling system constructed based on the knowledge-enhanced scheduling model can, through an adaptive gain coordination mechanism, adjust scheduling parameters in real time during the conversion of different production modes such as normal production and emergency scheduling, eliminate production fluctuations in traditional methods during mode switching, and use a multi-objective intelligent optimization algorithm to simultaneously optimize four key indicators: makespan, energy consumption, production cost, and equipment utilization rate. Compared with traditional single-objective optimization methods, it improves the execution efficiency of production plans, reduces resource waste, reduces energy consumption, and improves equipment utilization rate. By simulating and verifying candidate scheduling schemes through a digital twin model, the risk of implementing scheduling schemes is reduced. At the same time, through a scheduling execution and feedback mechanism, execution deviations are monitored in real time and re-scheduling is triggered when necessary to ensure the feasibility and stability of the scheduling scheme. Through a predictive scheduling strategy and an intelligent resource limiter, the system can proactively adjust production plans. Especially when facing sudden orders, by precisely adjusting the priority of resource allocation, recording each scheduling decision and its actual execution effect through a scheduling knowledge accumulation module, and continuously optimizing the knowledge-enhanced scheduling model in a sample increment manner, the system can continuously learn and evolve during long-term operation, and its adaptability and decision-making quality continue to improve. Description of the Drawings

[0008] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0009] The structures, proportions, sizes, etc. shown in the drawings of this specification are only used to cooperate with the content disclosed in the specification for those who are familiar with this technology to understand and read, and are not used to limit the limiting conditions under which the present invention can be implemented. Therefore, they do not have technical essential significance. Any modification of the structure, change of the proportional relationship or adjustment of the size, without affecting the effects that the present invention can produce and the purposes that can be achieved, should still fall within the scope that can be covered by the technical content disclosed by the present invention.

[0010] Figure 1 is a schematic flowchart of the intelligent scheduling method for the factory production plan provided by the embodiment of the present invention; Figure 2 is a schematic block diagram of the structure of the intelligent scheduling system for the factory production plan provided by the embodiment of the present invention. Detailed implementation manners

[0011] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts fall within the scope of protection of the present invention.

[0012] The flowchart shown in the drawings is only an example illustration, and does not necessarily include all the content and operations / steps, nor does it necessarily execute in the described order. For example, some operations / steps can also be decomposed, combined or partially merged. Therefore, the actual execution order may be changed according to the actual situation.

[0013] It should also be understood that the terms used in the specification of this application are only for the purpose of describing specific embodiments and are not intended to limit this application. As used in the specification of this application and the appended claims, unless the context clearly indicates otherwise, the singular forms "a", "an" and "the" are intended to include the plural forms.

[0014] It should be further understood that the term "and / or" used in the specification of this application and the appended claims refers to any combination and all possible combinations of one or more of the related listed items, and includes these combinations. Please refer toFigure 1 , an embodiment of the intelligent scheduling method for the factory production plan in the embodiment of the present application includes: Step 100: Perform three-dimensional scanning and sensor data collection on the factory environment to obtain a digital twin model and real-time production data streams; It can be understood that the execution subject of the present application can be an intelligent scheduling system for the factory production plan, or a terminal or a server, and specific limitations are not made here. The embodiment of the present application takes the server as the execution subject as an example for illustration.

[0015] Specifically, the overall layout of the factory is measured by three-dimensional laser scanning to obtain the geometric model of the factory. High-precision laser scanning equipment is used to construct the factory spatial structure using point cloud data to ensure that the measurement accuracy is within the range of ±5mm, thereby accurately reflecting the layout of the production workshop, equipment distribution and spatial relationship. The IoT sensor group is deployed on the target production equipment in the factory geometric model. The sensor group collects the operating parameters of the equipment in real time, including multi-dimensional data such as the temperature, pressure, vibration frequency, energy consumption and output of the equipment. In order to ensure the efficiency and accuracy of data collection, the sampling frequency of all sensors is set to 100Hz. At the same time, based on radio frequency identification technology, identification tags are installed on production materials, and with the help of a reader network distributed at key locations in the factory, the flow trajectory of the materials is continuously monitored to form material location tracking data, and the positioning accuracy of the data reaches ±10cm. During the execution of the production process, a special process execution data acquisition module is used to accurately record each time node of the production process, including the start time and end time of the process, the information of the personnel involved in the operation, and the quality inspection results, etc., to form a process execution data set. The equipment operation parameter data set, material location tracking data, and process execution data set are transmitted to the edge computing server through the industrial communication network for data cleaning and time synchronization. All data are transmitted to the edge computing server using a transmission method that complies with the OPC UA standard protocol. In the edge computing server, data cleaning operations are performed to eliminate outliers and redundant data, and ensure that data integrity reaches more than 99.5%. Then time synchronization is performed to align data from different sources according to a unified time reference, thereby eliminating data deviations between different sensors. After data cleaning and time synchronization, a multi-source heterogeneous data set is formed to describe the status of production equipment within the factory, the flow of materials, and the execution details of the production process. The factory geometry model and multi-source heterogeneous data sets are input into the digital twin engine. Through the dynamic mapping mechanism of the digital twin engine, the real-time status of the physical factory is synchronously mapped into the virtual model. In this process, the digital twin engine can store the static geometry information of the factory, and integrate dynamic data such as the real-time operating status of the equipment, the location information of the materials, the process flow and production parameters to form a highly simulated virtual factory model. This model can be continuously updated during the production process, reflecting the operating status of the factory in real time and providing visual data support.

[0016] Step 200: Perform prototype comparative learning processing on the dual-tower encoder according to the digital twin model and the real-time production data stream to obtain a pre-trained scheduling model; Specifically, for anomaly detection of real-time production data streams, multiple anomaly detection algorithms are adopted for different types of sensor data, including using the Z-score method to detect numerical anomalies, identifying temporal pattern anomalies through the sliding window clustering method, and combining expert rules to verify logical anomalies, so as to eliminate error data, abnormal fluctuations, and invalid samples, and make the data integrity index reach more than 99.5%. After completing the anomaly detection, the missing data in the cleaned production data set is complemented. The multivariable tensor completion algorithm is used to reconstruct the missing data to ensure the integrity and temporal continuity of the data. Among them, the tensor completion algorithm infers the missing sensor values based on the information of the global production state and minimizes the completion error by optimizing the objective function, so as to ensure that the root mean square error of the completed data is better than 0.05. The complete production data set is standardized to eliminate the numerical scale differences of different physical quantities, thereby improving the comparability of the data and the stability of the calculation. In the standardization process, for production parameters with different dimensions, a method combining Min-Max standardization and Z-score standardization is adopted to convert all numerical values to a unified scale space to avoid affecting the convergence of subsequent learning due to different numerical ranges. Based on the production batch and the process flow node, the standardized data set is segmented into time series to construct a production cycle unit data set with clear semantics. In this process, using the identification information of the production batch and the key time nodes in the process flow, the continuous data stream is divided into multiple independent production cycle units according to the process, and it is ensured that each unit contains complete start, execution, and end information, forming a structured representation of the production process. After obtaining the production cycle unit data set, multi-dimensional feature extraction is carried out to obtain feature vectors describing the production state and process evolution. Statistical features are extracted from the time domain, including parameters such as mean, variance, kurtosis, and skewness, to reflect the overall trend and distribution characteristics of the production data. Spectrum features are extracted from the frequency domain. The power spectral density, main frequency component, and harmonic components of the signal are calculated using the Fourier transform to reveal the periodic patterns in the production process. Wavelet features are extracted from the time-frequency domain. The wavelet transform is used to decompose the time series to obtain the instantaneous dynamic change features at different time scales. To improve the model's understanding ability of complex production processes, the correlation features between devices are extracted. By calculating the cross-correlation coefficient between the sensor data of different devices, the coupling degree between devices is measured. At the same time, the production process transition features are extracted to analyze the transition patterns between different process stages to identify potential production rules. After obtaining all the features, matrix mapping is performed on the time domain statistical features, frequency domain spectrum features, time-frequency domain wavelet features, device correlation features, and production process transition features to construct a high-dimensional production feature vector matrix, where each row represents a time point or a production batch, and each column corresponds to a feature dimension, forming structured data that can be input into a machine learning model.Based on the production feature vector matrix, a two-tower encoder is constructed and trained for prototype contrast learning. The two-tower encoder consists of a device status encoder and a production plan encoder. Each encoder uses a four-layer fully connected neural network with hidden layer dimensions of 512, 256, 128, and 64 in sequence, and LeakyReLU is used as the activation function to prevent gradient vanishing. During training, the prototype vectors of production modes are defined. Each prototype vector is initialized by the weighted average of the embedded representations of samples of the same class and is dynamically updated during training. The contrast loss function is used to calculate the cosine similarity between the sample embedding vector and the prototype vector, to maximize the similarity between samples of the same class and the prototype vector, and to minimize the similarity between samples of different classes. At the same time, a temperature parameter is introduced into the loss function to control the learning difficulty. In the unsupervised pre-training stage, a random masking strategy is adopted to mask some features of the production feature vector matrix, and the model is trained to predict the masked feature values, so that the model can learn the internal relationship between features. In the supervised fine-tuning stage, using the existing historical scheduling decision samples, the gradient descent method is used to optimize the model parameters to learn the mapping relationship between the optimal scheduling decision and the production status. To improve the model's adaptability to complex scenarios, during training, the batch hard sample mining technique is used to identify the boundary samples with difficult classification in each training batch and increase their weights in the loss function to improve the model's ability to identify difficult-to-classify samples. At the same time, the generalization performance of the model is evaluated through cross-validation, and Bayesian optimization is used to automatically adjust the hyperparameters, including the learning rate, batch size, and number of network layers, etc., to ensure the optimal convergence of the training process. After 100,000 batches of iterative training, the pre-trained scheduling model is finally obtained.

[0017] Build a four-layer fully connected neural network based on device status data to obtain a device status encoder. The input of this device status encoder is the operating parameters of the device, including key status variables such as temperature, pressure, vibration frequency, energy consumption, and production output. In terms of the network structure, a four-layer fully connected architecture is adopted, with the number of neurons in each layer being 512, 256, 128, and 64 respectively. And LeakyReLU is used as the activation function in each layer to prevent gradient vanishing, while batch normalization technology is utilized to improve the training stability. At the same time, build another independent four-layer fully connected neural network based on production plan data to obtain a production plan encoder. The input of this production plan encoder is the scheduling information related to production tasks, including information such as production batches, process flows, resource allocations, and task priorities, and shares the same number of layers and neuron structure with the device status encoder to ensure the consistency of the information expression ability between the encoders. Cascade the device status encoder and the production plan encoder to form a complete two-tower encoder. In this process, the output vectors of the two encoders are concatenated in the feature space and dimension-matched through a linear transformation layer, so that the finally encoded production features can retain the key information of both the device status and the production plan at the same time, and achieve effective fusion between different data modalities, thereby enhancing the model's ability to distinguish complex production patterns. After building the two-tower encoder, perform weighted average calculation on the sample sets of different production patterns to obtain the prototype vectors of each type of production pattern. In this process, classify the production patterns, including normal production patterns, emergency scheduling patterns, equipment maintenance patterns, etc. The prototype vector of each type of production pattern is obtained by weighted averaging the sample embedding vectors belonging to this type, and the prototype vector is dynamically adjusted during the training process to make it more accurately represent the feature distribution of this type of samples. Calculate the cosine similarity between the sample embedding vector and each production pattern prototype vector, and build a prototype contrast loss function based on this. The goal of this loss function is to maximize the similarity between the same-class samples and the corresponding prototype vectors and minimize the similarity between different-class samples. At the same time, in order to control the learning difficulty, a temperature parameter is introduced into the loss function to adjust the gradient change range of the similarity distribution. After building the prototype contrast loss function, perform random masking processing on the production feature vector matrix to generate a pre-training data set. In this process, randomly select some features of the input data for masking, and the masking ratio is set to 15%, and let the model predict the masked feature values during training to achieve unsupervised pre-training. This strategy can enable the model to learn the internal correlation between different features, improve the model's robustness to missing information, and capture key production patterns without relying on labeled data. After completing the unsupervised pre-training, obtain the initialized scheduling model.Input the labeled historical scheduling decision samples into the initial scheduling model, and perform gradient descent optimization. Use the cross-entropy loss function to calculate the error between the predicted decision and the true scheduling decision, and adjust the weight parameters of the neural network through backpropagation to gradually improve the accuracy of the scheduling decision. At the same time, in each training batch, evaluate the difficulty of the classification samples and adjust their weights in the loss function according to the classification difficulty of the samples. Assign smaller weights to the easily classified samples and larger weights to the difficult-to-classify boundary samples to enhance the discriminative ability of the model in complex production environments. After 100,000 batches of iterative training, finally obtain the pre-trained scheduling model.

[0018] Step 300: Perform knowledge constraint and knowledge distillation processing on the pre-trained scheduling model to obtain a knowledge-enhanced scheduling model; It should be noted that a domain knowledge base containing production process constraint rules, equipment capacity parameters, and quality standards is constructed. These knowledge are transformed into the form of IF-THEN logical expressions, and each rule is converted from natural language or technical specifications into a structured representation that can be input into the model. These expressions represent the constraint conditions in the production process, the operating capabilities of equipment, and the requirements of quality control, and provide hard constraints for the model, enabling the scheduling system to follow certain norms and rules when making production decisions. The TransE algorithm is used to perform triple embedding processing on the rules in the domain knowledge base. Each rule is represented as a triple, that is, (subject, relation, object). For example, "Equipment A (subject) needs to be shut down when the temperature exceeds X (relation) (object)". Through the TransE algorithm, the triples of the rules are mapped into a low-dimensional space, and a vector representation is generated for each element in the triple. These rule vectors are used to control the direction of model decision-making in the subsequent learning process. The rule vector representation can effectively capture the knowledge in the production process and provide constraint information for the model, enabling it to follow this prior knowledge when generating scheduling decisions. Perform rule attention mechanism processing on the rule vector representation based on the current production state. According to the current production environment state, calculate the correlation between each rule and the current production situation, and then weight the rule vectors according to the correlation to form a dynamic knowledge constraint layer. Use the pre-trained scheduling model as the teacher model. This model has been trained with a large amount of data and can generate accurate scheduling decisions. Construct a student model, and make it more lightweight by reducing the number of parameters of the pre-trained scheduling model to adapt to resource-constrained environments. Reduce the number of parameters of the student model to a preset target value. By controlling the structural parameters such as the number of layers of the student model and the number of neurons in each layer, the student model can maintain high performance while reducing the computational overhead and storage requirements. During the knowledge distillation process, the training objective of the student model is to minimize the KL divergence between the output of the student model and the output of the teacher model, so that the output distribution of the student model is as close as possible to the output distribution of the teacher model. This process enables the student model to obtain better generalization ability and get a lightweight scheduling model by transferring the knowledge learned by the teacher model during pre-training. In this way, the student model can maintain high accuracy when making scheduling decisions and has a fast inference speed, meeting the real-time requirements in actual production. Quantify the contribution degree of the input features of the lightweight scheduling model. By using methods such as integrated gradients, evaluate the contribution degree of each input feature to the final scheduling decision to obtain the feature importance ranking. In this process, the model calculates the influence of each feature on the decision result, identifies the features that are important for the scheduling decision, and provides a basis for subsequent incremental learning.Feature importance ranking helps optimize the performance of the model and improve its interpretability, enabling the scheduling system to provide a more transparent and traceable decision-making process in the face of a complex production environment. Based on the results of feature importance ranking, incremental learning is performed on the lightweight scheduling model. Based on new production data and actual scheduling decision samples, the model parameters are gradually optimized to ensure that the model can continuously adapt to changes in the production environment and avoid the occurrence of catastrophic forgetting. During the incremental learning process, only the part of the model related to the new data is updated, thus maintaining the stability of other parts and ensuring the long-term effectiveness of the model. At the same time, to ensure the compliance and reliability of the model, formal verification is performed on the lightweight scheduling model. Through mathematical proofs or formal methods, it is verified whether the scheduling decisions of the model comply with the constraint rules in the production process, especially whether they follow the process constraints, equipment capabilities, and quality standards in the domain knowledge base, etc., to ensure that each scheduling decision does not violate the hard rules. A knowledge-enhanced scheduling model is obtained.

[0019] Step 400: Construct a dual-loop coordinated scheduling system including a production intelligent control loop and a resource dynamic allocation loop according to the knowledge-enhanced scheduling model; Specifically, based on the integration of the knowledge-enhanced scheduling model and the digital twin model, a coordination system including a production intelligent control loop and a resource dynamic allocation loop is constructed. In this process, the production intelligent control loop will become the core component of the system. This loop contains multiple sub-modules. The state perception module integrates the digital twin model with the real-time production data stream to monitor the operation status of equipment, material flow, process execution, etc. in the production process in real time, and collects key parameter information such as the temperature, pressure, and energy consumption of the equipment to obtain various data of the current production environment. Through the continuous update of these data, the state perception module reflects any abnormal situations occurring in the production process in real time and provides detailed state information for subsequent pattern recognition and decision-making. The state perception module will be combined with the pattern recognition module, which analyzes the features extracted from the production data. By comparing with historical data and dynamically updating the production environment model, it identifies which type of pattern the current production state belongs to, such as normal production, emergency scheduling, or equipment failure, etc. The output of the pattern recognition module will directly affect the decision-making of the control execution module. The control execution module decides the specific execution of the scheduling strategy based on the identified production pattern and the changes in the production environment, so as to flexibly adjust the production process in a dynamic environment to ensure the continuity and stability of production. At the same time, the resource dynamic allocation loop effectively monitors, predicts demand, and optimizes the allocation of human, equipment, material, and energy resources required in the production process. In this loop, the usage of production resources is monitored in real time, and the consumption of resources is recorded through sensors and data collection devices to ensure the transparency of all resource usage. Through the demand prediction module, according to historical data and the current production progress, the resource demand situation within a certain period of time in the future is predicted, and a preliminary judgment on resource allocation is made for each production unit in advance to avoid the situation of resource shortage or waste. Based on the prediction results, the allocation optimization module dynamically adjusts the resource allocation plan through an optimization algorithm to optimize the allocation of resources among various production links, ensure the maximization of production efficiency, and reduce resource waste. In order to achieve the efficient coordination of the two loops, the mutual relationship between the production intelligent control loop and the resource dynamic allocation loop is coordinated. By calculating the change amount of the production state, an adaptive gain coordination mechanism is obtained, which dynamically adjusts the coordination gain between the loops through a mathematical model. The calculation method of the coordination gain is based on the change amount of the production state and its impact on the scheduling efficiency, enabling the system to flexibly adjust the response speed of the loops according to the changes in the production environment to ensure the overall stability and scheduling efficiency of the system. The setting of the coordination gain enables the production intelligent control loop to quickly respond to changes in the production pattern, while the resource dynamic allocation loop can dynamically adjust the allocation ratio and priority of resources according to the coordination gain, optimize resource usage, and reduce scheduling fluctuations in the production process. On this basis, an intelligent resource limiter is introduced to dynamically calculate the priority in resource allocation through real-time calculation and analysis of the changes in the production state.The calculation of resource priority is based on multiple factors, including resource utilization rate, cost coefficient, and scarcity level. Different priorities are assigned to each resource in the form of weighted coefficients to ensure that critical resources, such as production equipment, materials, or human resources, can be preferentially allocated in case of emergencies, thus guaranteeing the smooth execution of production plans and the ability to respond to emergencies. Based on the design of an adaptive gain coordination mechanism and an intelligent resource limiter, data between the production intelligent control loop and the resource dynamic allocation loop is synchronously processed to form the loop synchronization mechanism of the system. In this mechanism, decision-making information between the two loops is synchronized through a shared data storage and transmission channel, ensuring that when the decision of one loop changes, the other loop can respond in a timely manner and adjust its resource allocation plan. The core purpose of the loop synchronization mechanism is to ensure data consistency between the two loops, enabling the scheduling system to continuously operate in a changing production environment and provide real-time feedback on status and resource changes, avoiding decision-making conflicts or inconsistencies. Priority arbitration is performed on the decision-making conflicts between the production intelligent control loop and the resource dynamic allocation loop to obtain a dual-loop coordinated scheduling system. The arbitration mechanism dynamically adjusts the priority of conflict decisions according to the priorities of each loop, the urgency of the production environment, and the specific requirements of production tasks, ensuring the most appropriate decisions are made according to the actual needs of production and avoiding disorderly resource scheduling and instability during the production process.

[0020] Step 500: Perform multi-objective optimization and predictive scheduling calculations based on the dual-loop coordinated scheduling system to obtain a production plan scheduling scheme.

[0021] Specifically, a multi-objective evaluation system is defined, which includes four key indicators: the makespan, the total energy consumption, the production cost, and the equipment utilization rate. The makespan represents the longest completion time of all production tasks and affects the overall production cycle; the total energy consumption reflects the energy use efficiency in the production process and is directly related to the sustainability of production; the production cost is an important indicator to measure production efficiency, and a production process with a low cost is more competitive; the equipment utilization rate measures the production efficiency of equipment, and a higher equipment utilization rate represents a higher resource use efficiency. To optimize these objectives simultaneously, a hybrid crossover operator and an adaptive mutation operator are used to initialize the multi-objective evolutionary algorithm. Through the crossover operator and the mutation operator, the algorithm can explore more solution spaces, thereby obtaining diverse solutions, and adjust the mutation degree through the adaptive mutation operator to avoid falling into local optima, ensuring the global search ability of the solutions and generating an initial population. Based on the dual-loop coordinated scheduling system, time series prediction processing is carried out. The production process is monitored in real time through the digital twin model, and time series prediction is performed using historical data and the current production status to predict the state changes of equipment, the material flow situation, and the resource demand trend in the future for a period of time. Through the predictive scheduling strategy, the production plan is adjusted in advance to ensure that the production process can proceed smoothly. Especially in the face of emergencies or changes in the production environment, predictive scheduling can respond in advance and reduce fluctuations in production. After obtaining the predictive scheduling strategy, the initial population is optimized according to this strategy, and the individuals in the population are screened and adjusted using the information provided by the time series prediction to generate the target population. The candidate solutions in the target population represent the production scheduling plan optimized based on the current production environment and prediction strategy and can better meet the requirements of multi-objective optimization. The candidate scheduling solutions in the target population are input into the digital twin model for simulation verification. The digital twin model simulates the execution effects of different scheduling solutions in the actual production environment through a real-time virtual factory model and real-time production data streams. The purpose of the simulation verification is to evaluate the feasibility of the candidate scheduling solutions and ensure that the selected solutions can be successfully executed in actual production without violating process constraints or resource conflicts. During the simulation process, the digital twin model adjusts the production plan according to real-time feedback, evaluates the execution effects of each solution, and provides support for subsequent selection of the optimal solution. According to the feasibility evaluation results obtained from the simulation, the optimal scheduling solution is selected. The optimal scheduling solution is decomposed into specific execution instructions and sent to each production unit, and at the same time, the execution process is monitored in real time through the scheduling execution and feedback mechanism. During the execution process, the progress of each production link is continuously tracked, and the deviations during the execution process are monitored. Once it is found that the execution deviation exceeds the predetermined threshold, the system triggers the rescheduling mechanism to ensure the continuity and stability of the production process. Through the real-time feedback mechanism, production scheduling can respond more flexibly to emergencies, avoiding production stagnation or resource waste.After each scheduling execution, record the scheduling decision and its actual execution effect, and continuously optimize the scheduling model through sample incremental processing. By recording and analyzing the execution of each scheduling, the system accumulates new scheduling samples and conducts incremental learning in combination with real-time data. This process can improve the accuracy of the scheduling model and make it gradually adapt to different production environments and complex production modes. Over time, the model will be continuously updated to form a knowledge update dataset based on real-time data and historical experience. The knowledge update dataset is used to continuously optimize the knowledge-enhanced scheduling model. By continuously learning new scheduling patterns and production changes, the knowledge-enhanced scheduling model proposes a more accurate production plan scheduling scheme according to the changing production conditions. After multiple iterative optimizations, the finally output production plan scheduling scheme has high execution efficiency, can effectively reduce costs, improve equipment utilization rate, and ensure the rationality of energy consumption, thus achieving comprehensive production optimization.

[0022] In the embodiments of the present application, a virtual model of the production environment is established through digital twin technology, combined with prototype comparison learning for pre-training and knowledge constraint mechanisms, enabling the scheduling system to quickly adapt to new production scenarios, effectively solving the knowledge transfer problem of traditional scheduling systems when production conditions change, and improving the scheduling accuracy in a dynamic production environment. The dual-loop coordinated scheduling system constructed based on the knowledge-enhanced scheduling model can, through the adaptive gain coordination mechanism, adjust scheduling parameters in real time during the conversion of different production modes such as normal production and emergency scheduling, eliminating production fluctuations during mode switching in traditional methods. The multi-objective intelligent optimization algorithm is used to simultaneously optimize four key indicators: makespan, energy consumption, production cost, and equipment utilization rate. Compared with traditional single-objective optimization methods, it improves the execution efficiency of the production plan, reduces resource waste, reduces energy consumption, and improves equipment utilization rate. The simulation verification of the candidate scheduling scheme through the digital twin model reduces the risk of implementing the scheduling scheme. At the same time, through the scheduling execution and feedback mechanism, the execution deviation is monitored in real time and rescheduling is triggered when necessary to ensure the feasibility and stability of the scheduling scheme. Through the predictive scheduling strategy and intelligent resource limiter, the system can prospectively adjust the production plan. Especially when facing sudden orders, by precisely adjusting the resource allocation priority, recording each scheduling decision and its actual execution effect through the scheduling knowledge accumulation module, and continuously optimizing the knowledge-enhanced scheduling model in a sample incremental manner, the system continuously learns and evolves during long-term operation, and the adaptability and decision-making quality are continuously improved.

[0023] In a specific embodiment, the process of executing step 100 may specifically include the following steps: Perform three-dimensional laser scanning measurement on the factory layout to obtain a factory geometric model; Deploy an Internet of Things sensor group for the target production equipment in the factory geometric model, and collect a dataset of equipment operation parameters through the Internet of Things sensor group; Install identification tags on production materials based on radio frequency identification technology to obtain material position tracking data, and record time points and perform quality inspections during the execution of production processes to obtain a process execution dataset; Transmit the equipment operation parameter dataset, material position tracking data, and process execution dataset to the edge computing server through the industrial communication network for data cleaning and time synchronization processing to obtain a multi-source heterogeneous dataset; Input the multi-source heterogeneous dataset and the factory geometric model into the digital twin engine for dynamic mapping to obtain a digital twin model and a real-time production data stream.

[0024] Specifically, use 3D laser scanning technology to comprehensively measure the entire factory environment. Adopt high-precision lidar equipment to construct a geometric model of the factory through point cloud data. During this process, the laser scanning equipment obtains the internal structure of the factory with millimeter-level accuracy (±5mm), including the spatial distribution of workshops, production lines, equipment, and channels, and uses a multi-view stitching algorithm to fuse scan data from different angles to generate a 3D factory model. To ensure the accuracy of the model, an error correction algorithm, such as the least squares fitting method, is used to optimize the data of multiple measurement points to minimize geometric errors and obtain a geometric model that truly restores the factory layout. After obtaining the factory geometric model, deploy an Internet of Things sensor group on key target production equipment to collect a dataset of equipment operation parameters in real time. These sensors include temperature sensors (measuring temperature changes inside the equipment), pressure sensors (monitoring the pressure status of hydraulic or pneumatic systems), acceleration sensors (used for vibration analysis to detect abnormal vibrations of the equipment), current sensors (monitoring the motor load), energy consumption sensors (recording the power consumption of the equipment), etc., and the sampling frequency of the data is set to 100Hz to ensure high-precision capture of key parameters during the production process. To improve the reliability of data collection, a multi-sensor data fusion algorithm is used, that is, the weighted average method is used: Among them, represents the fusion state at the current time of, is the data collected by sensor , is the weight of this sensor (calculating the optimal weight based on historical data), is the number of sensors. This method can effectively reduce the error of a single sensor and improve the system's ability to perceive the status of the equipment. At the same time, radio frequency identification technology is used to install identification tags on production materials, and an RFID reader network is set up inside the factory. When the material passes through a specific area, the reader can record the location data of the material in real time, with a positioning accuracy of ±10cm. The RFID tag of each production material stores key information, such as material number, production batch, storage location and current status. All data are recorded in the material location tracking data set, and during the process execution, the timestamp recording system is used to record the start time, end time and operator of each production step. At the same time, the product is quality inspected in combination with the industrial visual inspection system, defects are automatically identified, and the inspection results are stored in the process execution data set. All collected data, including equipment operation parameter data sets, material location tracking data and process execution data sets, are transmitted to the edge computing server through the industrial communication network for data processing. The industrial communication protocol uses OPC UA to ensure compatibility between different data sources. In the edge computing server, data cleaning is performed, and anomaly detection algorithms are used to eliminate invalid data, for example: in, For the cleaned data, is the original data set, is the mean, The purpose of cleaning is to remove outliers and noise to improve data accuracy. Time synchronization is performed to eliminate clock deviations between different data sources, using a dynamic time warping algorithm: in, and Represent the time series data points of the two data streams, Calculate their alignment errors, and align the time series of different sensors through iterative optimization to form a multi-source heterogeneous data set with high consistency and time series continuity. After completing data preprocessing, the multi-source heterogeneous data set and the factory geometry model are input into the digital twin engine for dynamic mapping to build a digital twin model and real-time production data flow. The core of the digital twin engine is the data mapping function: in, Representative time The digital twin status at all times, is the factory geometry model, is the multi-source heterogeneous data at the current time, is a mapping function that projects the operating state of a physical device into a virtual space through a machine learning model or a physical simulation model, enabling the digital twin system to reflect the state of the production process in real time.

[0025] In a specific embodiment, the process of executing step 200 may specifically include the following steps: Perform anomaly detection processing on the real-time production data stream to obtain a cleaned production data set; Perform multivariate tensor completion on the missing data in the cleaned production data set to obtain a complete production data set; Perform normalization processing on the complete production data set to obtain a normalized data set, and perform time series segmentation processing on the normalized data set based on production batches and process flow nodes to obtain production cycle unit data sets; Perform multi-dimensional feature extraction on the production cycle unit data sets to obtain time-domain statistical features, frequency-domain spectrum features, time-frequency domain wavelet features, device association features, and production process transformation features; Perform matrix mapping on the time-domain statistical features, frequency-domain spectrum features, time-frequency domain wavelet features, device association features, and production process transformation features to obtain a production feature vector matrix; Perform prototype contrast learning processing on the dual tower encoder based on the production feature vector matrix to obtain a pre-trained scheduling model.

[0026] Specifically, perform anomaly detection processing on the real-time production data stream to ensure the accuracy and stability of the data. In an industrial environment, the data collected by sensors has problems such as noise, mutations, or missing values. Multiple data cleaning methods are used to identify and correct outliers. To detect abnormal data, an adaptive anomaly detection method based on a sliding window is used, and an anomaly scoring function is defined: where represents the anomaly score of the data at time is the data value at the current time point, is the mean value within the sliding window, is the standard deviation, is a very small value to prevent the denominator from being zero. If exceeds the set threshold, then the data point is determined to be an outlier and is interpolated and corrected. After completing the anomaly detection, perform multivariate tensor completion on the cleaned production data set to fill in the missing data and maintain the integrity of the data. In industrial production data, there is a certain correlation between multiple sensors. The method of multivariate tensor decomposition is used to predict the missing values, and a completion objective function is defined: where where represents an observed value in the data matrix, and is the feature matrix after low-rank decomposition, is the weight matrix, indicating whether the value is missing, is the regularization parameter to prevent overfitting. By optimizing this objective function, the missing data is inferred using the known data, so that the reconstructed data matrix can retain the structural relationship of the original data as much as possible, and a complete production data set is obtained. After filling in the missing data, the complete production data set is standardized to eliminate the scale differences between different physical quantities and improve the computational stability of the model. Z-score standardization is adopted: where, represents the standardized data, and are the mean and standard deviation of the th variable respectively. After standardization, the data set is segmented into time series according to production batches and process flow nodes, and a production cycle division function is defined: where, represents the th production cycle unit, and are the start and end time points of this production cycle respectively. All data is divided into time series units with clear production stages. Multidimensional feature extraction is performed on the production cycle unit data set, including time-domain statistical features, frequency-domain spectrum features, time-frequency domain wavelet features, equipment association features, and production process conversion features. In terms of time-domain statistical features, descriptive statistics such as mean, standard deviation, skewness, and kurtosis are calculated for each production cycle. In frequency-domain analysis, the fast Fourier transform is used to calculate the main frequency components of the signal to obtain the frequency-domain signal . By analyzing the spectrum features, the periodic change patterns in the production process are identified. In order to capture the sudden change information in the production process, continuous wavelet transform is used to calculate the time-frequency domain features: where, represents the transformation result at scale and position , is the mother wavelet function, and the local features of the signal changing with time are captured through this transformation. The correlation between equipment is calculated using mutual information: where, measures the information correlation between variables and , is the joint probability distribution, and are the marginal probability distributions respectively. The device association features are used to judge the influence relationship between different devices and improve the scheduling optimization ability. After the above feature extraction, the time-domain statistical features, frequency-domain spectrum features, time-frequency domain wavelet features, device association features and production process conversion features are matrix-mapped to obtain the production feature vector matrix: wherein, is the high-dimensional feature matrix, represents the vector of the th feature dimension. This matrix constitutes the input data for training the dual tower encoder. Based on the production feature vector matrix, the dual tower encoder is trained for prototype contrast learning. The dual tower encoder includes a device status encoder and a production plan encoder, which respectively learn the feature embeddings of the device status and the production plan. The prototype contrast loss function is used to optimize the model: wherein, represents the similarity between the sample and its prototype vector of the category it belongs to, is the temperature parameter, which controls the sensitivity of the model to difficult-to-classify samples. After multiple rounds of iterative training, a pre-trained scheduling model is finally obtained, which can efficiently distinguish different production modes.

[0027] In a specific embodiment, the process of obtaining the pre-trained scheduling model by performing prototype contrast learning on the dual tower encoder based on the production feature vector matrix may specifically include the following steps: Construct a four-layer fully connected neural network based on the device status data to obtain a device status encoder, and construct a four-layer fully connected neural network based on the production plan data to obtain a production plan encoder; Perform a cascading process on the device status encoder and the production plan encoder to obtain a dual tower encoder; Perform a weighted average calculation on the sample sets of different production modes to obtain the prototype vector of each production mode, and calculate the cosine similarity between the sample embedding vector and the prototype vector of each production mode to obtain the prototype contrast loss function; Perform a random masking process on the production feature vector matrix based on the prototype contrast loss function to obtain a pre-trained data set, and perform unsupervised pre-training by predicting the masked feature values to obtain an initialized scheduling model; Input the labeled historical scheduling decision samples into the initialized scheduling model for gradient descent optimization processing to obtain the fine-tuned model parameters, and perform difficulty evaluation and weight adjustment on the classification samples in each training batch to obtain the pre-trained scheduling model.

[0028] Specifically, two four-layer fully connected neural networks are constructed, which are used to extract device status features and production plan features respectively. For the device status encoder, its input is the operating parameters of the device, such as temperature, pressure, vibration frequency, energy consumption, etc. Define the input vector as , where is the number of devices, is the feature dimension of each device. The device status encoder adopts a four-layer fully connected structure, and the output of its -th layer is expressed as: where, is the activation output of the -th layer, is the weight matrix, is the bias term, adopts Leaky ReLU as the activation function to enhance the non-linear expression ability. Similarly, for the production plan encoder, its input is the planned information of the production task, including batch number, process route, resource allocation, and scheduling priority, etc. The input vector is defined as , where is the number of production batches, is the feature dimension of the production plan. A four-layer fully connected neural network is also adopted. Define the output of the -th layer as: During the training process, the two encoders respectively learn the feature distributions of the device status and the production plan. Subsequently, a two-tower encoder is obtained through a concatenation operation. The output of the device status encoder is concatenated with the output of the production plan encoder and fused through a fully connected layer. Define the final feature vector representation of the two-tower encoder as: where, is the fusion weight matrix, is the bias term, and are the output of the final layers of the two encoders respectively. After obtaining the two-tower encoder, prototype contrast learning is carried out, and the weighted average calculation is performed on the sample sets of different production modes to obtain the prototype vector of each production mode. Suppose there are types of production modes, and the prototype vector of each mode is calculated by the weighted mean of the sample embedding vectors belonging to this category: where, represents all samples under category , is the sample The weight, is the corresponding embedding vector. To calculate the similarity between the sample embedding vector and the production mode prototype vector, cosine similarity is used for calculation: Based on the cosine similarity, a prototype comparison loss function is constructed. Its goal is to maximize the similarity between the sample embedding and the prototype of the correct category, while minimizing its similarity with the prototypes of other categories: Among them, is the temperature parameter, which is used to control the category discrimination degree, represents the sample 's true category. This loss function can improve the model's ability to distinguish different production modes and make it more generalizable. During the training process, to enhance the robustness of the model, random masking processing is adopted, that is, randomly masking some feature values in the production feature vector matrix, so that the model must still be able to perform effective inference in the case of missing some features. Define the masking matrix : Among them, is the binary mask matrix, represents element-wise multiplication, and the masking ratio is set to 15%. Through predicting the masked feature values, unsupervised pre-training is carried out, and the optimization objective is: Among them, represents the masked feature index, is the feature value predicted by the model. Unsupervised pre-training can improve the model's learning ability of the internal relationship between production features, and finally obtain an initialized scheduling model. After completing the unsupervised pre-training, the labeled historical scheduling decision samples are input into the initialized scheduling model for gradient descent optimization, so that it can learn the optimal mapping relationship of the scheduling decision. The optimization objective function is defined as: Among them, is the scheduling decision predicted by the model, is the true scheduling decision. To further improve the generalization ability of the model, during the training process, the difficulty of the classification samples in each training batch is evaluated, and a difficult sample weight adjustment mechanism is defined: Among them, is the adjustment coefficient, Used to assign different weights to samples in the loss function, so that difficult-to-classify samples are given higher attention during the optimization process. After multiple rounds of iterative training, the obtained pre-trained scheduling model can accurately identify different production modes and provide optimization support for intelligent scheduling.

[0029] In a specific embodiment, the process of executing step 300 may specifically include the following steps: Convert the production process constraint rules, equipment capacity parameters, and quality standards into IF-THEN logical expression forms to obtain a domain knowledge base; Use the TransE algorithm to perform triple embedding processing on the rules in the domain knowledge base to obtain rule vector representations, and perform rule attention mechanism processing on the rule vector representations based on the current production status to obtain a knowledge constraint layer; Use the pre-trained scheduling model as the teacher model, and use the model with the number of parameters of the pre-trained scheduling model reduced to the preset target value as the student model, and perform knowledge distillation processing on the student model to obtain a lightweight scheduling model; Perform contribution degree quantization processing on the input features of the lightweight scheduling model to obtain a feature importance ranking, and perform incremental learning and formal verification on the lightweight scheduling model based on the feature importance ranking to obtain a knowledge-enhanced scheduling model.

[0030] Specifically, construct a domain knowledge base, which contains the constraint conditions, equipment performance limitations, and quality requirements in the production process. These rules are represented in the form of IF-THEN logical expressions. These rules cover multiple aspects, such as process flow constraints, equipment usage limitations, product quality standards, etc. Through the rule extraction process, a structured domain knowledge base is formed. In order to enable the model to effectively utilize these rules, the TransE algorithm is used to perform triple embedding on the rules in the domain knowledge base, and each rule is represented in the form of "(subject, relation, object)". For all rules, use the TransE objective Among them, respectively represent the embedding vectors of the subject, relation, and object in the rule triple, is the optimization objective. By minimizing this objective function, the embedded rule vectors can maintain the semantic consistency of the rules. In order to optimize the use of rules, a rule attention mechanism is introduced during the model inference process, and the rule attention weight calculation formula is defined: Among them, is the weight of rule , represents the rule vector and the current production status The similarity between them is ensured, and the rules most relevant to the current state are given higher weights to construct the knowledge constraint layer. After completing the construction of the knowledge constraint layer, knowledge distillation is carried out. The pre-trained scheduling model is used as the teacher model, a student model with a reduced number of parameters is constructed, and the knowledge of the teacher model is transferred to the student model through the knowledge distillation method. Define the probability distribution output by the teacher model as... where is the output score of the teacher model, is the temperature parameter, which controls the smoothness of knowledge distillation. The output distribution of the student model is defined as: where is the output of the student model. The knowledge distillation loss function is defined as: The optimization goal is to make the prediction distribution of the student model as close as possible to the teacher model, so as to learn the feature representation ability of the teacher model, while reducing the computational complexity and obtaining a lightweight scheduling model. Quantify the contribution degree of the input features of the lightweight scheduling model to evaluate the impact of different features on the scheduling decision. Use the integrated gradient algorithm to calculate the feature importance: where represents the contribution degree of feature , is the baseline input, is the prediction function of the scheduling model. By calculating the integrated gradients of different features, the feature importance ranking is obtained, and the top 30% features with the most significant impact on the decision are selected to reduce the computational burden of the model. On this basis, based on the feature importance ranking, incremental learning is carried out on the lightweight scheduling model to make it adapt to the changing production environment. Use the online update strategy: where is the model parameter at time , is the learning rate, is the loss function, represents the newly added sample data. Through incremental learning, the model can continuously update its knowledge and adapt to new production scenarios. To ensure the reliability of the model, formal verification is carried out on the knowledge-enhanced scheduling model, that is, to verify whether it complies with the production rules and constraint conditions. Define the formal constraint detection function. Among them, represents the scheduling decision Whether all production constraints are met. If not, the scheduling strategy needs to be readjusted. On this basis, a knowledge-enhanced scheduling model is finally obtained, which has higher reliability and adaptability and can optimize scheduling decisions in a complex production environment.

[0031] In a specific embodiment, the process of executing step 400 may specifically include the following steps: Construct a production intelligent control loop including a state perception module, a pattern recognition module, and a control execution module based on the knowledge-enhanced scheduling model and the digital twin model; Monitor, predict demand, and optimize the allocation of human, equipment, material, and energy resources to construct a resource dynamic allocation loop; Calculate the coordination gain of the production intelligent control loop and the resource dynamic allocation loop based on the production state change amount to obtain an adaptive gain coordination mechanism, and calculate the resource priority of the production intelligent control loop and the resource dynamic allocation loop to obtain an intelligent resource limiter; Based on the adaptive gain coordination mechanism and the intelligent resource limiter, perform data synchronization processing on the production intelligent control loop and the resource dynamic allocation loop to obtain a loop synchronization mechanism; Perform priority arbitration on the decision conflicts of the production intelligent control loop and the resource dynamic allocation loop to obtain a dual-loop coordinated scheduling system.

[0032] Specifically, based on the knowledge-enhanced scheduling model and the digital twin model, a production intelligent control loop including a state perception module, a pattern recognition module, and a control execution module is established. Relying on digital twin technology, the state perception module uses Internet of Things sensors to collect key data such as equipment status, material flow, and energy consumption in real time, and establishes a dynamic feature representation in combination with historical production data. After state perception is completed, the pattern recognition module uses a prototype contrast learning model to calculate the similarity between the current production state and known patterns, and determines whether the current system is in normal production, equipment overload, emergency scheduling, maintenance and repair, etc. based on the similarity level. Once pattern recognition is completed, the system calls corresponding control strategies to enable the production plan to dynamically adapt to the current state. In the control execution module, based on the state recognition result, the scheduling system executes adjustment instructions, such as reducing the working intensity of high-load equipment, reallocating task loads, or optimizing the material flow path, to ensure the stable and efficient production process. At the same time, the resource dynamic allocation loop monitors, forecasts demand, and optimizes the allocation of human, equipment, materials, and energy. During the production process, the consumption rates, availabilities, and supply states of different resources fluctuate over time. Therefore, prediction algorithms are used to estimate future resource requirements in advance and dynamically adjust the resource allocation plan. Based on the long short-term memory network (LSTM), the resource requirements at future times are predicted to ensure that each production unit can reasonably obtain the required resources, thereby improving production stability and resource utilization efficiency. Based on the prediction results, the resource allocation strategy is automatically adjusted through an optimization algorithm to avoid resource waste or production bottlenecks. To ensure the coordinated operation of the production intelligent control loop and the resource dynamic allocation loop, an adaptive gain coordination mechanism is introduced. This mechanism dynamically adjusts the control gain based on the change amount of the production state, enabling the system to respond flexibly to emergencies. When the production state fluctuates significantly, the system needs to adjust the control parameters more quickly, while when the production state is relatively stable, the system should maintain a lower adjustment intensity to reduce unnecessary changes. Therefore, the gain function adopts an exponential decay form: Wherein, represents the coordination gain at the current moment, is the reference gain, is the adjustment factor, Represents the change in the current production status. This formula ensures that when the production status changes drastically, the system can respond quickly, and when the production status is stable, the adjustment strength naturally weakens to improve the stability and robustness of the system. To optimize the resource utilization efficiency, an intelligent resource limiter is designed. This limiter calculates the resource allocation priority based on the availability, shortage degree, and usage cost of various resources. The system allocates critical resources first in case of emergencies, such as the maintenance resources of high-priority equipment and the preferential supply of scarce materials, to ensure production continuity. For example, in the case of an emergency order, the system automatically adjusts the resource allocation plan, gives priority to ensuring the production tasks of critical orders, and reduces the resource occupancy of low-priority tasks, making the overall production plan more flexible. During the dual-loop operation, data synchronization is the key to ensuring the coordinated operation of the two. A loop synchronization mechanism is established to ensure that the production intelligent control loop and the resource dynamic allocation loop are consistent during data update and decision execution. The synchronization mechanism enables the scheduling decisions of the two loops to always be based on the latest production status information through shared data storage and real-time communication protocols, avoiding execution deviations caused by data lag. For example, when the production intelligent control loop adjusts the workload of a certain device, the resource dynamic allocation loop should update the available production capacity of this device in real time for corresponding resource reallocation. Since there may be scheduling conflicts between the production intelligent control loop and the resource dynamic allocation loop, for example, in some cases, the production plan hopes to prioritize the completion of high-value orders, while the resource allocation algorithm believes that the optimal strategy is to balance resource utilization. To prevent such conflicts from affecting the execution of the production plan, a priority arbitration mechanism is designed. This mechanism automatically adjusts the execution priority of different scheduling decisions by calculating the urgency, production benefits, and resource occupancy of scheduling tasks. Define the priority calculation formula: Wherein, Represents the priority of the scheduling task, Is the contribution degree of the task to the overall production efficiency, Is the urgency of the task, Is the current available resource quantity, Is the weight coefficient. Dynamically adjust the task execution order according to the calculated priority to ensure that critical tasks can be completed first while taking into account the efficient utilization of production resources.

[0033] In a specific embodiment, the process of executing step 500 may specifically include the following steps: Define a multi-objective evaluation system including the maximum completion time, total energy consumption, production cost, and equipment utilization rate, and initialize the multi-objective evolutionary algorithm using a hybrid crossover operator and an adaptive mutation operator to obtain an initial population; Based on the double-loop coordinated scheduling system, perform time series prediction processing to obtain a predictive scheduling strategy, and optimize the initial population according to the predictive scheduling strategy to obtain the target population; Input the candidate scheduling schemes in the target population into the digital twin model for simulation verification to obtain the feasibility evaluation results, and select the optimal scheduling scheme according to the feasibility evaluation results; Decompose the optimal scheduling scheme into execution instructions and monitor the deviations during the execution process to obtain the scheduling execution and feedback mechanism, and record and perform sample increment processing on each scheduling decision and its actual execution effect based on the scheduling execution and feedback mechanism to obtain the knowledge update data set; Use the knowledge update data set to continuously optimize the knowledge-enhanced scheduling model and output the production plan scheduling scheme.

[0034] Specifically, a multi-objective evaluation system is defined, which includes key indicators such as the makespan, total energy consumption, production cost, and equipment utilization rate, to ensure that the scheduling scheme can achieve the optimal state in multiple dimensions. The makespan is used to measure the longest time required to complete all production tasks and affects the efficiency of the overall production cycle; the total energy consumption reflects the energy utilization in the production process, and optimizing this indicator helps reduce energy consumption costs; the production cost includes raw material consumption, labor costs, and equipment maintenance expenses, which determine the overall economic benefits of production; the equipment utilization rate represents the usage efficiency of equipment resources. Too low a utilization rate means resource waste, while too high a utilization rate leads to equipment overload. In the initialization stage of the multi-objective optimization algorithm, a hybrid crossover operator and an adaptive mutation operator are adopted to enhance the global search ability and local optimization ability of the algorithm. Among them, the hybrid crossover operator combines gene segments among multiple candidate solutions to improve the diversity of solutions, while the adaptive mutation operator dynamically adjusts the mutation rate according to the convergence of the population to ensure that the algorithm maintains an appropriate exploration intensity at different stages. Each individual in the initialized population represents a possible scheduling scheme, and its gene encoding includes multiple dimensions such as task allocation, equipment scheduling, material flow, and personnel arrangement to form a complete set of initial solutions. After completing the population initialization, a dual-loop coordinated scheduling system is used for time series prediction processing to obtain the changing trend of the future production state. Based on historical production data, a long short-term memory network (LSTM) is used to predict the workload of equipment, material consumption trends, and order delivery requirements within a future time window to form a predictive scheduling strategy. The predictive scheduling strategy can identify potential future production bottlenecks and optimize resource allocation and task arrangements in advance to reduce sudden adjustments during the production process. On this basis, the predictive scheduling strategy is used to optimize the initialized population, that is, on the basis of the original candidate solutions, adjust the order of scheduling tasks, resource allocation ratios, and equipment switching times to improve the adaptability and executability of the overall scheduling scheme and form a target population. After obtaining the target population, the candidate scheduling scheme is input into the digital twin model for simulation verification to evaluate its feasibility. The digital twin system uses real-time production data and equipment operation models to simulate the actual execution effects of different scheduling schemes and calculate their impacts on the multi-objective evaluation system. For example, if the makespan of a certain scheme far exceeds the acceptable range, or the equipment utilization rate is too high, resulting in a potential risk of equipment damage, then this scheme will be judged as infeasible. During the simulation process, the following optimization objectives are used: Among them, is the target weight, and these weights are adjusted to balance the priorities among different targets. By sorting the simulation results of all candidate solutions, the optimal scheduling solution is selected to ensure that the finally executed solution is optimal in multiple dimensions. After determining the optimal scheduling solution, it is decomposed into specific execution instructions and sent to each production unit. Meanwhile, a scheduling execution and feedback mechanism is established to monitor the deviation in the execution process in real time. If the system detects that the actual production status deviates from the scheduling plan, for example, the equipment running load exceeds the expectation or a certain process is delayed due to material shortage, the system will automatically trigger an adjustment mechanism to minimize the impact of the deviation by reallocating tasks or adjusting the execution order of processes. The formula for calculating the deviation in the execution process is as follows: where represents the production status deviation at time , is the actual execution status, is the planned execution status. If the deviation exceeds the set threshold, the system will automatically adjust the production plan and send the adjusted new plan to each execution unit to ensure the continuity and stability of the production process. While the scheduling plan is being executed, each scheduling decision and its actual execution effect are recorded, and a knowledge update data set is constructed through the sample incremental processing method. This data set contains the execution effect of the scheduling plan, adjustment records, deviation information, etc. By continuously accumulating experience data, the system optimizes the future scheduling strategy. For example, if the system finds that a certain type of equipment is more likely to fail within a specific load range, in future scheduling, the system will automatically lower the load upper limit of this equipment to improve its stability. The knowledge update data set is used to continuously optimize the knowledge-enhanced scheduling model. The system inputs the latest scheduling data into the model through online learning and uses the incremental training method to continuously optimize the model parameters so that it can adapt to the changing production environment. Through this optimization process, the system dynamically adjusts the scheduling strategy, improves the overall production efficiency, and ensures the intelligence, self-adaptability, and high efficiency of the scheduling plan. For example, in a certain intelligent manufacturing factory, the system finds that the delivery time of a certain batch of orders is often delayed during the scheduling process. After knowledge update analysis, the system adjusts the scheduling strategy for similar orders and automatically increases equipment redundancy and resource buffering in future scheduling to make the production more flexible. Through this closed-loop optimization mechanism, the system can continuously evolve, improve the accuracy of the scheduling plan and production efficiency, and ensure the continuous optimization and stable operation of the intelligent scheduling system in a complex production environment.

[0035] The intelligent scheduling method for the factory production plan in the embodiment of the present application is described above. Next, the intelligent scheduling system 10 for the factory production plan in the embodiment of the present application will be described. Please refer to Figure 2, an embodiment of the intelligent scheduling system 10 for the factory production plan in the embodiments of the present application includes: An acquisition module 11, configured to perform three-dimensional scanning on the factory environment and acquire sensor data to obtain a digital twin model and real-time production data streams; A contrastive learning module 12, configured to perform prototype contrastive learning processing on a dual tower encoder according to the digital twin model and real-time production data streams to obtain a pre-trained scheduling model; A knowledge distillation module 13, configured to perform knowledge constraint and knowledge distillation processing on the pre-trained scheduling model to obtain a knowledge-enhanced scheduling model; A construction module 14, configured to construct a dual-loop coordinated scheduling system including a production intelligent control loop and a resource dynamic allocation loop according to the knowledge-enhanced scheduling model; A calculation module 15, configured to perform multi-objective optimization and predictive scheduling calculations based on the dual-loop coordinated scheduling system to obtain a production plan scheduling scheme.

[0036] Through the collaborative cooperation of the above-mentioned various components, a virtual model of the production environment is established through digital twin technology. Combining prototype contrastive learning pre-training and knowledge constraint mechanisms, the scheduling system can quickly adapt to new production scenarios, effectively solve the knowledge transfer problem of traditional scheduling systems when production conditions change, and improve the scheduling accuracy in a dynamic production environment. The dual-loop coordinated scheduling system constructed based on the knowledge-enhanced scheduling model can, through an adaptive gain coordination mechanism, adjust scheduling parameters in real time during the conversion of different production modes such as normal production and emergency scheduling, eliminate production fluctuations during mode switching of traditional methods, and use a multi-objective intelligent optimization algorithm to simultaneously optimize four key indicators: makespan, energy consumption, production cost, and equipment utilization rate. Compared with traditional single-objective optimization methods, it improves the execution efficiency of the production plan, reduces resource waste, reduces energy consumption, and improves equipment utilization rate. Through the digital twin model, the candidate scheduling scheme is simulated and verified, reducing the implementation risk of the scheduling scheme. At the same time, through the scheduling execution and feedback mechanism, the execution deviation is monitored in real time and rescheduling is triggered when necessary, ensuring the feasibility and stability of the scheduling scheme. Through the predictive scheduling strategy and intelligent resource limiter, the system can proactively adjust the production plan. Especially when facing sudden orders, by precisely adjusting the resource allocation priority, recording each scheduling decision and its actual execution effect through the scheduling knowledge accumulation module, and continuously optimizing the knowledge-enhanced scheduling model in a sample increment manner, the system continuously learns and evolves during long-term operation, and the adaptability and decision-making quality are continuously improved.

[0037] Those skilled in the art can clearly understand that for the convenience and conciseness of description, the specific working processes of the above-described system, system, and unit can refer to the corresponding processes in the foregoing method embodiments and will not be described herein again.

[0038] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing an electronic device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of this application. The foregoing storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical discs that can store program codes.

[0039] As described above, the above embodiments are only used to illustrate the technical solutions of this application, rather than to limit them; although this application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of various embodiments of this application.

Claims

1. An intelligent scheduling method for factory production plans, characterized in that Including: Performing three-dimensional scanning and sensor data acquisition on the factory environment to obtain a digital twin model and real-time production data streams; Performing prototype comparison learning on the twin towers encoder according to the digital twin model and the real-time production data streams to obtain a pre-trained scheduling model; Performing knowledge constraint and knowledge distillation on the pre-trained scheduling model to obtain a knowledge-enhanced scheduling model; Constructing a dual-loop coordinated scheduling system including a production intelligent control loop and a resource dynamic allocation loop according to the knowledge-enhanced scheduling model; Performing multi-objective optimization and predictive scheduling calculations based on the dual-loop coordinated scheduling system to obtain a production plan scheduling scheme.

2. The intelligent scheduling method for the factory production plan according to claim 1, characterized in that, The performing three-dimensional scanning and sensor data acquisition on the factory environment to obtain a digital twin model and real-time production data streams includes: Performing three-dimensional laser scanning measurement on the factory layout to obtain a factory geometric model; Deploying an Internet of Things sensor group on target production equipment in the factory geometric model and collecting a dataset of equipment operation parameters through the Internet of Things sensor group; Installing identification tags on production materials based on radio frequency identification technology to obtain material position tracking data, and recording time points and performing quality inspections on the production process execution to obtain a process execution dataset; Transmitting the equipment operation parameter dataset, the material position tracking data, and the process execution dataset to an edge computing server through an industrial communication network for data cleaning and time synchronization processing to obtain a multi-source heterogeneous dataset; Inputting the multi-source heterogeneous dataset and the factory geometric model into a digital twin engine for dynamic mapping to obtain a digital twin model and real-time production data streams.

3. The intelligent scheduling method for the factory production plan according to claim 1, characterized in that, The performing prototype comparison learning on the twin towers encoder according to the digital twin model and the real-time production data streams to obtain a pre-trained scheduling model includes: Performing anomaly detection on the real-time production data streams to obtain a cleaned production dataset; Performing multi-variable tensor completion on the missing data in the cleaned production dataset to obtain a complete production dataset; Performing standardization processing on the complete production dataset to obtain a standardized dataset, and performing time series segmentation on the standardized dataset based on production batches and process flow nodes to obtain production cycle unit datasets; Performing multi-dimensional feature extraction on the production cycle unit datasets to obtain time-domain statistical features, frequency-domain spectrum features, time-frequency domain wavelet features, equipment association features, and production process transition features; Performing matrix mapping on the time-domain statistical features, the frequency-domain spectrum features, the time-frequency domain wavelet features, the equipment association features, and the production process transition features to obtain a production feature vector matrix; Performing prototype comparison learning on the twin towers encoder based on the production feature vector matrix to obtain a pre-trained scheduling model.

4. The intelligent scheduling method for the factory production plan according to claim 3, characterized in that, The performing prototype comparison learning on the twin towers encoder based on the production feature vector matrix to obtain a pre-trained scheduling model includes: Constructing a four-layer fully connected neural network based on equipment status data to obtain an equipment status encoder, and constructing a four-layer fully connected neural network based on production plan data to obtain a production plan encoder; Cascade process the device status encoder and the production plan encoder to obtain a two-tower encoder; Perform weighted average calculation on the sample sets of different production modes to obtain the prototype vectors of each production mode, and calculate the cosine similarity between the sample embedding vectors and the prototype vectors of each production mode to obtain the prototype contrast loss function; Based on the prototype contrast loss function, perform random masking processing on the production feature vector matrix to obtain a pre-training data set, and perform unsupervised pre-training by predicting the masked feature values to obtain an initialized scheduling model; Input the labeled historical scheduling decision samples into the initialized scheduling model for gradient descent optimization processing to obtain the fine-tuned model parameters, and evaluate the difficulty and adjust the weights of the classification samples in each training batch to obtain a pre-trained scheduling model.

5. The intelligent scheduling method for the factory production plan according to claim 1, characterized in that, The knowledge constraint and knowledge distillation processing of the pre-trained scheduling model to obtain a knowledge-enhanced scheduling model includes: Convert the production process constraint rules, equipment capacity parameters, and quality standards into IF-THEN logical expression to obtain a domain knowledge base; Use the TransE algorithm to perform triple embedding processing on the rules in the domain knowledge base to obtain the rule vector representation, and perform rule attention mechanism processing on the rule vector representation based on the current production state to obtain a knowledge constraint layer; Use the pre-trained scheduling model as the teacher model, and use the model with the number of parameters of the pre-trained scheduling model reduced to the preset target value as the student model, and perform knowledge distillation processing on the student model to obtain a lightweight scheduling model; Quantify the contribution of the input features of the lightweight scheduling model to obtain the feature importance ranking, and perform incremental learning and formal verification on the lightweight scheduling model based on the feature importance ranking to obtain a knowledge-enhanced scheduling model.

6. The intelligent scheduling method for the factory production plan according to claim 1, wherein Construct a dual-loop coordinated scheduling system including a production intelligent control loop and a resource dynamic allocation loop according to the knowledge-enhanced scheduling model, including: Based on the knowledge-enhanced scheduling model and the digital twin model, construct a production intelligent control loop including a state perception module, a pattern recognition module, and a control execution module; Monitor, demand predict, and allocate and optimize human, equipment, material, and energy resources to construct a resource dynamic allocation loop; Based on the production state change amount, calculate the coordination gain of the production intelligent control loop and the resource dynamic allocation loop to obtain an adaptive gain coordination mechanism, and calculate the resource priority of the production intelligent control loop and the resource dynamic allocation loop to obtain an intelligent resource limiter; Based on the adaptive gain coordination mechanism and the intelligent resource limiter, perform data synchronization processing on the production intelligent control loop and the resource dynamic allocation loop to obtain a loop synchronization mechanism; Perform priority arbitration on the decision conflicts of the production intelligent control loop and the resource dynamic allocation loop to obtain a dual-loop coordinated scheduling system.

7. The intelligent scheduling method for the factory production plan according to claim 1, characterized in that Perform multi-objective optimization and predictive scheduling calculation based on the dual-loop coordinated scheduling system to obtain a production plan scheduling scheme, including: Define a multi-objective evaluation system that includes the maximum completion time, total energy consumption, production cost, and equipment utilization rate, and initialize the multi-objective evolutionary algorithm using a hybrid crossover operator and an adaptive mutation operator to obtain an initial population; Based on the dual-loop coordinated scheduling system, perform time series prediction processing to obtain a predictive scheduling strategy, and optimize the initial population according to the predictive scheduling strategy to obtain a target population; Input the candidate scheduling schemes in the target population into the digital twin model for simulation verification to obtain a feasibility evaluation result, and select the optimal scheduling scheme according to the feasibility evaluation result; Decompose the optimal scheduling scheme into execution instructions and monitor the deviation during the execution process to obtain a scheduling execution and feedback mechanism, and record and perform sample increment processing on each scheduling decision and its actual execution effect based on the scheduling execution and feedback mechanism to obtain a knowledge update data set; Use the knowledge update data set to continuously optimize the knowledge-enhanced scheduling model and output a production plan scheduling scheme.

8. An intelligent scheduling system for factory production plans, characterized in that, An intelligent scheduling method for executing the factory production plan according to any one of claims 1-7, wherein the intelligent scheduling system for the factory production plan includes: An acquisition module for performing three-dimensional scanning and sensor data acquisition on the factory environment to obtain a digital twin model and real-time production data streams; A contrast learning module for performing prototype contrast learning processing on the dual tower encoder according to the digital twin model and the real-time production data streams to obtain a pre-trained scheduling model; A knowledge distillation module for performing knowledge constraint and knowledge distillation processing on the pre-trained scheduling model to obtain a knowledge-enhanced scheduling model; A construction module for constructing a dual-loop coordinated scheduling system including a production intelligent control loop and a resource dynamic allocation loop according to the knowledge-enhanced scheduling model; A calculation module for performing multi-objective optimization and predictive scheduling calculations based on the dual-loop coordinated scheduling system to obtain a production plan scheduling scheme.

Images