Industrial intelligent collaborative control methods, systems, model training methods, media and electronic devices

By collaborating with a single industrial physical intelligent neural network model and a classical system, the problem of unclear responsibilities between process planning and execution control is solved, the connection between process adjustment and current physical conditions is realized, and the efficiency and stability of variable process manufacturing are improved.

CN123113577APending Publication Date: 2026-11-17SHANGHAI LINGJI INTELLIGENT TECH CO LTD
View PDF 6 Cites 0 Cited by

Patent Information

Application Number
CN202611456190.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-18
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

In existing technologies, when facing variable process manufacturing, the responsibilities between process planning and execution control are unclear, leading to substandard quality, redundant calculations and ineffective adjustments, and making it difficult to connect process adjustments with current physical conditions.

Method used

By employing a single industrial physical intelligent neural network model in collaboration with classical process algorithms, simulation, and control systems, macroscopic process planning and specific implementation schemes are generated through training on the correspondence between manufacturing tasks, process states, and physical results. During the feedback adjustment phase, the manufacturing state estimate is updated to ensure the coordinated implementation of process adjustment and execution control.

Benefits of technology

It achieves unified consideration of process decision-making and on-site feedback, reduces model calls and repeated verification, maintains the protection of hard constraints and executed content, improves the reliability of process adjustment and the stability of execution, and is suitable for various variable process manufacturing scenarios.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123113577A_ABST
    Figure CN123113577A_ABST
Patent Text Reader

Abstract

This invention relates to intelligent manufacturing, and discloses an industrial intelligent collaborative control method, system, model training method, medium, and electronic equipment for variable process manufacturing. The method involves a single industrial physical intelligent neural network model collaborating with classical process algorithms, an industrial world model, and classical control. This model uses the same shared network entity and the same set of neural network parameters to complete macro-level process planning, specific implementation generation, and manufacturing state cognition. Manufacturing state estimation is fed back to the same model to generate specific implementation schemes. Classical process algorithms generate action mask constraints for model output, the industrial world model verifies the physical consequences, and classical control executes. During feedback, actions and quality are first verified; if not met, the execution chain is diagnosed; if qualified, the cost is evaluated; if unreasonable, the specific implementation is adjusted by the same model; only when infeasibility is confirmed is the macro-level process replanned. During runtime, only the variable future is updated, while executed and submitted parts are retained, and the neural network parameters are fixed.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to intelligent manufacturing, process planning, industrial simulation, and automation control. Specifically, it relates to an industrial intelligent collaborative control method, system, model training method, medium, and electronic equipment for variable process manufacturing. In particular, it relates to an industrial intelligent collaborative control method, system, model training method, medium, and electronic equipment for variable process manufacturing, in which an industrial physical intelligent neural network model acts as the manufacturing decision-making agent, collaborating with classical process algorithms, classical simulation, and classical control systems to complete manufacturing tasks. Background Technology

[0002] CNC machining typically requires first determining the machining sequence, then selecting the tools, toolpath methods, and process parameters for each step, and finally converting the toolpath into machine tool motion. During machining, changes may occur in tool status, material load, and clamping. Simply executing the original program accurately may not guarantee that the modified process will still be suitable.

[0003] Existing process planning methods encompass various technical approaches. For example, patent document CN118709849A uses a genetic algorithm to solve for the process route; patent document CN114611379B employs a neural network to predict energy consumption and uses a genetic algorithm for process sequencing and optimization. The search algorithms or local prediction functions in these solutions have different responsibilities than unified manufacturing intelligence, which simultaneously undertakes overall process planning, current implementation, and manufacturing status awareness.

[0004] Regarding learning-based process decision-making, patent document US20260010142A1 discloses a process generation method combining policy networks, digital twins, and tree search; patent document US20260023331A1 discloses a method that calibrates the state based on monitoring data and calls upon a trained policy network to regenerate the remaining process. These solutions involve learning-based process generation, simulation evaluation, or monitoring feedback, but it is still necessary to clarify the specific tasks of the models used, the sharing relationships of neural network parameters, and their collaboration with the execution system.

[0005] Regarding action constraints, patent document US20220164636A1 discloses a method of inputting constraint vectors into a model and restricting the output using action masks; patent document CN121413980A employs action masks and knowledge guidance in reconfigurable workshop scheduling. Action masks and restricted choices in training and inference have already been applied. This invention focuses on the macro-planning, specific implementation, and state cognition of a single industrial physical intelligent neural network model, as well as the collaborative relationship between this model and classical process algorithms, classical simulations, and classical control; action masks are the means of integrating engineering constraints into this collaboration.

[0006] Furthermore, when manufacturing conditions change, process adjustments must align with the overall task, step-by-step coordination, and current physical conditions, while execution control must continue to stably achieve approved objectives. If feedback responsibilities are unclear, tracking errors, quality defects, and unreasonable costs may all be directly attributed to process design problems. In reality, quality defects can also stem from driver, control code, or execution chain failures. Repeatedly modifying the process without first verifying execution and quality will result in ineffective adjustments, redundant calculations, or conflicting objectives. On the other hand, content that has already been processed or submitted for execution cannot be arbitrarily modified, and process changes require reliable future effective boundaries. Therefore, it is necessary to clearly define the aforementioned responsibilities, data, and authority relationships between unified manufacturing intelligence and classical engineering capabilities. Summary of the Invention

[0007] In view of the shortcomings of the prior art, the purpose of this invention is to provide an industrial intelligent collaborative control method for variable process manufacturing.

[0008] According to the present invention, an industrial intelligent collaborative control method for variable process manufacturing is provided, in which an intelligent agent composed of a single industrial physical intelligent neural network model collaborates with a classical process algorithm operating according to deterministic rules, an industrial world model undertaking classical simulation, and a classical control system. The industrial physical intelligent neural network model is trained through the correspondence between manufacturing tasks, applicable process states, confirmed processes, and their physical results. Using the same shared network entity and the same set of neural network parameters, it completes macro-process planning, specific implementation generation, and manufacturing state recognition based on the manufacturing task and context. The manufacturing state estimate obtained from updating the industrial physical intelligent neural network model is used as input for the industrial physical intelligent neural network model to generate specific implementation schemes. The method includes the following steps: S1: Obtain manufacturing tasks, industrial resources, and known manufacturing conditions, and establish an initial manufacturing state estimate based on the known manufacturing conditions; S2: The classical process algorithm determines the optional processes, the allowable range of continuous process quantities, and the modifiable process fields based on the process dependencies of the manufacturing task, available resources, confirmed manufacturing conditions, and the current allowable modification range. In the output selection or decoding of the industrial physical intelligent neural network model, disallowed options are directly excluded, numerical ranges are limited, and the original values ​​of process fields that cannot be modified are maintained. Under the above constraints, the same industrial physical intelligent neural network model generates a macro-process scheme containing one or more macro-process steps and their sequence. Then, based on the manufacturing task, the complete macro-process scheme, the selected one or more macro-process steps to be designed, and their applicable manufacturing states, a corresponding specific implementation scheme is generated. S3: The classical process algorithm generates process execution parameters or process execution plans according to deterministic rules based on specific implementation schemes and industrial constraints. The industrial world model calculates the expected physical results and performs engineering verification on the specific implementation schemes of the selected single or multiple macroscopic process steps, or the complete process scheme containing the specific implementation schemes of each step. After passing the required verification, the design results corresponding to the verification scope are output, or the process execution parameters or process execution plans that have passed the required verification are handed over to the classical control system for execution. S4: After entering the feedback adjustment phase and obtaining the simulated physical observation results or the physical observation results of the entity, the specific implementation scheme, actual execution status, applicable expected physical results saved before observation, observed physical results and necessary manufacturing history corresponding to the process version and execution interval will be input into the same industrial physical intelligent neural network model to update the manufacturing state estimate. During the manufacturing state estimate update and subsequent process re-decision, the neural network parameters of the industrial physical intelligent neural network model will remain fixed. S5: The classical control system handles execution deviations within its authority. Based on actual execution and quality feedback, it verifies whether the target action has been achieved, and whether the execution quality and manufacturing quality of the corresponding stage meet the preset requirements. If the action is not achieved, the quality is unqualified, or the evidence is insufficient, it enters the driver, control code, and execution chain diagnosis. After confirming that the action has been achieved and the quality is qualified, if the costs such as time, energy, or resource consumption do not meet the preset evaluation conditions, the classical process algorithm updates the optional process, the allowable range of continuous process quantities, and the modifiable process fields based on the confirmed conditions and future modification permissions. Then, the same industrial physical intelligent neural network model estimates the current manufacturing status and regenerates the specific implementation scheme under the updated constraints. If there is evidence that the target of the corresponding macroscopic process step and its necessary connection requirements cannot be met within the allowable specific implementation scheme range, the same industrial physical intelligent neural network model replans the macroscopic process. The adjusted scheme is output after being generated and verified as required. During operation, only the future parts that are allowed to be modified are replaced, while the executed parts and the submitted parts that cannot be rewritten according to the execution protocol are retained.

[0009] Preferably, the classical process algorithm uses an action mask to connect the selectable process, the allowable range of continuous process quantities, and the modifiable process fields to the output selection or decoding of the same industrial physical intelligent neural network model; the action mask includes one or more of the following: allow or prohibit flags for discrete options, allowable range of continuous process quantities, or process fields that must maintain their original values, and updates subsequent restrictions as the selected process content is updated; priority is given to allowing adjustments to continuous process quantities that meet the implementation conditions, and when no feasible solution can be obtained within the current range, modification permissions for process methods, tools, or fixtures are gradually opened according to preset conditions; expanding the modification range does not remove the still valid hard constraints and the protection of executed and submitted content.

[0010] Preferably, in step S5, the preset evaluation conditions specify the corresponding step objectives, execution intervals, quality items, and cost reference ranges; the classic control system handles tracking errors and disturbances within the approved objectives, control limits, and safety boundaries; when an action is not achieved or the quality is unqualified, execution differences and quality evidence are recorded, and the drive, control code, process parameter mapping, and execution timing are checked, without directly identifying it as a process design problem; after confirming that the action has been achieved and the quality is qualified, the effective cost records for the same objective and corresponding interval are compared, and only when the cost does not conform to the cost reference range is the same industrial physical intelligent neural network model re-decision on the specific implementation scheme; when the execution chain is confirmed to be normal but the objective exceeds the available execution capacity after diagnosis, the confirmed capability constraints are fed back to request a feasible solution; safety interlocks and emergency handling do not need to wait for the inference of the industrial physical intelligent neural network model.

[0011] Preferably, the execution plan is divided into an executed portion, a committed portion, and a variable future portion. The committed portion is the part that has been submitted to the execution chain, has not yet been executed, and is not allowed to be rewritten in this replanning according to the current execution protocol. The variable future portion is the plan content that has not yet been executed and is currently allowed to be modified. The update scheme selects a future position (future semantic boundary) with a clear corresponding step or action and corresponding state and connection conditions without changing the executed portion and the committed portion. The variable future portion is generated and verified from this position. Before release, the execution plan version on which the candidate scheme is based, the actual position that has been executed (executed frontier), the endpoint of the part that has been accepted by the execution chain and is not rewritten according to the protocol (committed frontier), and the proposed switching position are checked to ensure that they are still applicable. If the conditions are met, the new plan version, associated results, and proposed switching position are registered as a complete update (atomic release). When the boundary is reached, the applicable state and connection conditions are checked again. Only if they match are switched. During the computation, the machine can continue to execute the approved content that is still safe and valid. Expired or failed verification candidates do not replace the formal plan.

[0012] Preferably, the manufacturing state estimation characterizes one or more of the following: material effective load, tooling state, dynamic stability, thermal state, geometric state, or tooling state, and can express the estimated value, range, probability, or unknown item of the effective state; the expected physical results and observed physical results correspond to each other according to the process version, applicable object, and execution interval, and the expected physical results used for corresponding comparison should be results that have been saved and are still applicable before the observed physical results are obtained, and the manufacturing history is limited to the information available at that time; when observations are lacking, a priori manufacturing state estimation with nominal assumptions or unknown labels is used, and the expected physical results are not directly identified as physical measurement results.

[0013] Preferably, the generation and verification of specific implementation schemes are carried out alternately, or multiple or all specific implementation schemes are generated first and then verified in a concentrated manner. The method adopted can be selected or switched according to the task or different stages of the same task. Specific implementation schemes can be generated in a certain order and then divided into verification batches as needed. The processes involved must still meet the actual manufacturing dependencies and state connection requirements. When generating subsequent specific implementation schemes in advance, the macro process scheme, known initial conditions, generated processes and intermediate manufacturing states clearly marked as assumptions can be read.

[0014] Preferably, the correspondence between the process generation results, process execution plans, and verification results and the tasks, resources, manufacturing status, and specific implementation schemes is recorded by using the recorded stability identifier, the content summary (input fingerprint) of the calculated input, and the version. When the specific implementation scheme changes, it is determined whether the existing results are still applicable based on the dependency relationship. If the results are still applicable, they are reused, and only the affected future results are regenerated or verified. Only the process execution plan that matches the input basis with the verification version and passes the required verification is delivered for execution.

[0015] Preferably, the process design before commencement, the scheme adjustment under simulation results, and the state cognition and process adjustment under physical feedback conditions all call the same neural network parameter version of the industrial physical intelligent neural network model, and independent manufacturing intelligent models are not established for the above purposes; the state update during operation does not modify the neural network parameters of the deployed industrial physical intelligent neural network model, and after long-term data accumulation, a new neural network parameter version is formed through offline training and verification and released separately.

[0016] Preferably, the industrial world model includes one or more of geometry, kinematics, dynamics, collision, material removal, force, heat, vibration, wear, and mass calculation; if a learning-type physics prediction auxiliary task is set, it is undertaken by the same industrial physics intelligent neural network model, and the formal engineering verification is still performed by the industrial world model according to the corresponding algorithm.

[0017] Preferably, the manufacturing task includes one or more of CNC machining, laser processing, welding, additive manufacturing, semiconductor manufacturing and packaging, electronic assembly, or precision assembly; in laser processing such as laser cutting and laser abrasion, the macroscopic process plan determines the processing contour or area, stages, and sequence, while the specific implementation plan determines the processing method, allowable laser power level, and scanning or feed speed; in welding, the macroscopic process plan determines the weld seam and stage sequence, while the specific implementation plan determines the method, tools, heat input, and speed; in additive manufacturing, the macroscopic process plan determines the layers, areas, and scanning sequence, while the specific implementation plan determines the energy, material flow rate, and scanning method; in semiconductor manufacturing and packaging, electronic assembly, and precision assembly, the macroscopic process plan determines the step objectives and sequence, while the specific implementation plan determines the tools, fixtures, formula, pressure, temperature, time, or action method.

[0018] The present invention provides a training method for an industrial physical intelligent neural network model for industrial intelligent collaborative control, used to train a single industrial physical intelligent neural network model, and organizing the generation of training samples through a data generator, comprising the following steps: M1: Obtain manufacturing tasks, industrial resources, confirmed excellent macro-process schemes, process applicability status, and confirmed excellent specific implementation schemes under that status; call the industrial world model to calculate the corresponding physical results; and construct macro-process planning supervision samples and specific implementation generation supervision samples. M2: The data generator performs constrained changes on the materials, tools, machine dynamics, thermal, geometric or tooling conditions of the simulation environment, calls the industrial world model to run the existing specific implementation scheme, retains the expected physical results under the nominal conditions before the change, and obtains the simulated physical results and the actual execution of the simulation after the change. M3: The data generator constructs effective manufacturing status labels based on the hidden conditions adopted, forming a manufacturing status supervision sample with the corresponding specific implementation scheme, actual execution status, expected physical results, observed physical results and necessary manufacturing history as inputs and the effective manufacturing status labels as supervision targets. The input of the manufacturing status supervision sample does not include the truth value of the hidden conditions and the effective manufacturing status labels. M4: Using deterministic constraint rules consistent with the inference phase, based on the tasks, resources, confirmed conditions, and allowable modification ranges available during the corresponding training decision, the allowable options, numerical ranges, and modifiable process fields of the process output are determined. Under the condition that the above constraints directly participate in the output selection or decoding of the industrial physical intelligent neural network model, the training loss between the predicted macroscopic process scheme and the specific implementation scheme and their respective applicable labels is calculated. M5: After task-based training of macro-process planning, specific implementation generation, and manufacturing state estimation, mixed sample joint training is performed. Each task updates the same shared network entity and the same set of neural network parameters. The manufacturing state estimation output and the manufacturing state input in the specific implementation generation task use the same manufacturing state definition. The same industrial physical intelligent neural network model is trained for inference. This model keeps the neural network parameters fixed during inference.

[0019] Preferably, the macro-level process planning and specific implementation generation capabilities are trained first, and then manufacturing state training is added, while the samples of each task are retained in the joint training. For manufacturing state supervision tasks, the hidden condition truth value is not used as the input of the industrial physical intelligent neural network model, but only as the basis for constructing the supervision label of the task. The specific implementation scheme label must be a reasonable process confirmed in the corresponding state. The original process that is no longer applicable after the change is not directly used as the excellent specific implementation scheme label of the state.

[0020] Preferably, the limitation of the process output is implemented by using an action mask to access the output selection or decoding of the industrial physical intelligent neural network model, and the corresponding process supervision loss is calculated based on the limited output; when the macroscopic process scheme label or the specific implementation scheme label conflicts with the corresponding action mask, the sample and constraint basis are checked first, and it is not used for the process supervision for the time being; the current answer to be predicted and future information are not used to generate the mask; the training and inference follow the corresponding constraint rules.

[0021] Preferably, one round of training uses fixed data with source and version information. Subsequently, selected entity records and simulation records are used for offline joint training to update the neural network parameters. The industrial world model is calibrated by using the difference between the applicable expected physical results and the credible entity observation physical results. The capability of the trained industrial physical intelligent neural network model is verified, and the prediction results of the calibrated industrial world model are verified. After the verification is passed, the corresponding parameters are saved as the new version. The runtime state update does not modify the neural network parameters of the deployed industrial physical intelligent neural network model.

[0022] Preferably, discrete outputs are classified or sequenced according to the encoding, and continuous process quantities and supervised state components are regressed. Components without reliable labels are not calculated for corresponding supervision errors. The macro-process planning training loss, specific implementation generation training loss, and manufacturing state estimation training loss are weighted and summed according to the set weights to obtain the total training objective.

[0023] The present invention provides an industrial intelligent collaborative control system for variable process manufacturing, comprising a task and resource management module, an industrial physical intelligence module consisting of an intelligent agent composed of a single industrial physical intelligence neural network model, a deterministic process generation and constraint module that runs classical process algorithms, an industrial world model module that undertakes classical simulation, a classical control module, and an observation module. The industrial physical intelligence neural network model in the industrial physical intelligence module is trained by the correspondence between manufacturing tasks, process applicability status, confirmed processes and their physical results. It generates macroscopic process schemes and specific implementation schemes with the same shared network subject and the same set of neural network parameters. The manufacturing status estimate is updated according to the specific implementation schemes corresponding to the process version and execution interval, the actual execution status, the expected physical results, the observed physical results and the necessary manufacturing history. The estimate is then used as the input for the industrial physical intelligence neural network model to generate specific implementation schemes. During the operation period, the neural network parameters of the industrial physical intelligence neural network model remain fixed. The deterministic process generation and constraint module determines and updates the optional process, the allowable range of continuous process quantity, and the modifiable process field based on process dependence, available resources, confirmed manufacturing conditions, and future modification permissions. The above constraints are directly used for the output selection or decoding when the same industrial physical intelligent neural network model generates macroscopic process schemes and specific implementation schemes, and the specific implementation schemes are expanded into process execution parameters or process execution plans. The industrial world model module calculates physical consequences and verifies candidate process schemes; the classical control module handles execution deviations within the approved target authority; and the observation module provides feedback with observation sources and execution ranges. When an action is not performed, the quality is unqualified, or there is insufficient evidence, the system enters the driver, control code, and execution chain diagnosis. If it is confirmed that the target action has been performed, the execution quality and the manufacturing quality of the corresponding stage meet the preset requirements, but the costs such as time, energy, or resource consumption do not meet the preset evaluation conditions, the same industrial physical intelligent neural network model adjusts the specific implementation scheme. If it is confirmed that the target of the corresponding macroscopic process step and its necessary connection requirements still cannot be met within the allowable specific implementation scheme, the same industrial physical intelligent neural network model re-plans the macroscopic process. The adjusted scheme is output after being generated and verified as required. During operation, only the future parts that can be modified are updated, while the executed parts and the submitted parts that cannot be rewritten according to the execution protocol are retained.

[0024] Preferably, the deterministic process generation and constraint module uses an action mask to select or decode the output of the same industrial physical intelligent neural network model that restricts access. The action mask includes one or more of the following: allow or prohibit flags for discrete options, allowable ranges for continuous process quantities, or process fields that must maintain their original values. The system updates subsequent constraints as the selected process content is updated. The system divides the operation plan into the executed part, the submitted part, and the variable future part. When updating, it starts from the future semantic boundary. Before publishing, it checks the version of the operation plan on which the candidate solution is based, the executed frontier, the submitted frontier, and the boundary to be effective. When the boundary is reached, the applicable manufacturing status and connection conditions are checked again. Switching is only performed if they match.

[0025] According to the present invention, a computer-readable storage medium storing a computer program is provided on the computer-readable storage medium, wherein the computer program, when executed by a processor, implements the steps of the industrial intelligent collaborative control method for variable process manufacturing.

[0026] An electronic device according to the present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the industrial intelligent collaborative control method for variable process manufacturing.

[0027] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention completes macro-process planning, specific implementation generation, and manufacturing status cognition through the same shared network entity and the same set of neural network parameters. The manufacturing status estimate is input into the same model to generate specific implementation schemes. This allows the same model to jointly consider the overall goal, the selected implementation, and on-site physical feedback when making process decisions, avoiding the fragmentation and inconsistency of multiple models. It determines whether a model needs to be called according to the aforementioned responsibilities and reuses the still applicable calculation results, reducing unnecessary model calls and repeated verifications, and supporting the connection between process adjustments and current physical conditions.

[0028] 2. This invention uses a classic process algorithm to dynamically generate action masks based on process dependencies, available resources, confirmed manufacturing conditions, and the current allowed modification range. The masks are then directly input into the model output for selection or decoding. Subsequent constraints are updated along with the selected process content, enabling the model to apply manufacturing experience within the process range allowed by the current deterministic rules. Candidates that are confirmed to be disallowed by the current deterministic rules are not allowed to enter subsequent verification, reducing the amount of engineering recalculation, while maintaining the protection of hard constraints and executed and submitted content.

[0029] 3. This invention expands high-level process intentions into process execution parameters or execution plans through classical process algorithms. The industrial world model calculates the expected physical results for the selected single step, multiple steps, or complete scheme and performs engineering verification. This combines the flexible experience of the learning model with deterministic engineering calculations to verify the feasibility and physical consequences of the candidates. The verification range corresponds to the output range, avoiding the mistaken assumption that the overall scheme is feasible when some verifications pass.

[0030] 4. This invention adopts a judgment process of action verification, quality confirmation, cost evaluation, and capability constraint feedback to determine whether the feedback should be handled by execution control or process design: First, it verifies whether the target action has been achieved and whether the execution quality and manufacturing quality are qualified. If not, it diagnoses the driver, control code, and execution chain. Only after the qualification is confirmed can the cost be evaluated. If the cost is unreasonable, the specific implementation is adjusted by the same model according to the cost evaluation branch. Only when it is confirmed that it is not feasible can the macro process be replanned. In this way, the execution diagnosis and process adjustment are diverted according to the existing evidence, reducing ineffective adjustments and duplicate calculations.

[0031] 5. This invention divides the operation plan into an executed part, a submitted part, and a variable future part. It generates and verifies the variable future starting from the future semantic boundary without changing the first two parts. Before release, it checks the version, executed position, and submitted position, and registers the new plan version, associated results, and the proposed effective boundary together. When the boundary is reached, the connection conditions are checked again. This ensures that process adjustments do not rewrite the content that has been completed or submitted for execution, maintains the consistency of the plan when online adjustments and on-site execution are carried out in parallel, and executes according to the connection conditions and safety rules.

[0032] 6. This invention records the stability identifier, input fingerprint, and version of the inputs used to generate process results, execution plans, and verification results. When specific implementation changes, applicable results can be reused based on dependencies. Only the affected future results are regenerated or verified, and the execution plan that matches the inputs with the verification version is delivered for execution. This reduces the scope of recalculation after adjustments, shortens the time required to regenerate the plan, and improves the reliability of delivery and execution.

[0033] 7. This invention constructs effective manufacturing state labels based on the actual conditions used in the simulated environment, hides the truth values ​​of these conditions in the state task input, calculates the error between the prediction result after applying the action mask and the supervision label using the action masking rule consistent with inference, and performs joint training with mixed samples after task-based training.

[0034] 8. This invention adopts a method of fixing neural network parameters during operation, accumulating data, and then training offline. The same industrial physical intelligent neural network model updates the manufacturing state estimation and necessary process schemes based on on-site feedback. After screening, long-term data is jointly trained and verified offline, and a new version is released separately. This avoids the risk of model drift and verification failure caused by modifying neural network parameters online, and supports the stability, verifiability, and maintainability of the manufacturing system in long-term operation.

[0035] 9. This invention adapts to various variable process manufacturing scenarios such as CNC machining, laser processing, welding, additive manufacturing, semiconductor manufacturing and packaging, electronic assembly and precision assembly through the same industrial physical intelligence architecture. It configures state definitions, physical models, measurement accuracy and safety thresholds according to the domain, and realizes unified collaboration of macro planning, specific implementation and state cognition. It has wide industrial applicability and scalability. Attached Figure Description

[0036] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 A schematic diagram of a collaborative system of a single industrial physical intelligent neural network model and classical algorithms, classical simulations, and classical control. Figure 2 A process design flowchart that allows selection of the generation range, verification range, and verification timing; Figure 3 A flowchart showing the responsibility branches for actions, quality, cost, and confirmed capability constraints; Figure 4 A diagram illustrating future plans for switching positions (semantic boundaries) and partial updates; Figure 5 A relational graph showing the use of both high-quality process monitoring samples and manufacturing status monitoring samples for training the same model; Figure 6 A diagram showing the relationship between simulation, entity data, and offline version updates; Figure 7 A flowchart illustrating the overall process of CNC pocket machining, from process generation to execution, step completion, and feedback adjustments. Figure 8 A schematic diagram illustrating the process of updating the state of the same model with fixed neural network parameters and adjusting the process as necessary. Figure 9 This is a schematic diagram illustrating the phased and multi-task joint training of the same model; Figure 10 This diagram illustrates the training and inference processes that use action masks to connect the same industrial physical agent with classical engineering capabilities. Detailed Implementation

[0037] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.

[0038] Example 1: In the following text, the industrial physical intelligence model, intelligent model, unified model, and manufacturing intelligence all refer to the same industrial physical intelligence neural network model; the intelligent agent refers to the role of this model in manufacturing decision-making. The industrial world model uses classical simulation algorithms to calculate the physical results of machining. Among them, engineering parameters such as materials, tools, and equipment are used for simulation calculations and can be corrected through engineering calibration; neural network parameters are used for model inference, updated through training, and remain fixed during deployment and operation. This embodiment uses the CAMX CNC intelligent machining system as a carrier, connecting target geometry, machine tool resources, deterministic machining algorithms, industrial world model, CNC controller, and measurement interface. The system uses action masks to specify the currently selectable processes, clamping, tools, methods, high-level process variable values, and allowed process content modifications, illustrating how classical engineering capabilities directly participate in the decision-making of the same model; the classification and access order of these masks are selected in this CAMX embodiment.

[0039] like Figure 1As shown, the system inputs the manufacturing task's objectives, requirements, and resource information into a single industrial physical intelligence model; the model proposes macroscopic solutions and specific implementations. The phrase "Learning from process and physical results" within the model frame indicates that these capabilities are formed during training; this model is still used after deployment, with on-site feedback used to update the status and necessary processes. The deterministic process algorithm expands candidate specific implementation schemes into process execution parameters, paths, or movements; the industrial world model calculates and verifies the physical consequences; the content that passes the check is handed over to classical control and machine learning, where the controller adjusts execution within approved goals and permissions. The system obtains actual actions and effective measurements through execution records and physical observations, and inputs this information, along with the corresponding processes and expected results, into the same model for status awareness and necessary process adjustments.

[0040] For example, machining a pocket measuring 60mm x 40mm, 8mm deep, and with an interior corner radius of 6mm from a 100mm x 80mm x 20mm sheet metal. The task clearly defines the pocket's position relative to the machining datum, dimensional tolerances, and surface requirements; the material is specified from the confirmed material record, and the tooling fixation method and fixture clearance space have been established. Available resources include face milling cutters, a 10mm diameter roughing end mill, and a 6mm diameter finishing end mill. Tool lengths, assembly, available spindle speeds, and machine tool limits are read from the resource record, not guessed from the model.

[0041] The project saves the target's precise geometry, machining features, initial blank, material parameter records used in this calculation (material parameter snapshots), machine tool kinematics, and tool and clamping records. The target geometry answers "what shape will it ultimately be made into?", the remaining material geometry answers "what shape is currently left?", and the material parameter snapshot answers "what material parameters are used?". Subsequent calculations read the corresponding records according to their purpose. The model reads the necessary information from these existing records as input for the current task, without directly rewriting the original records or creating a separate set of independently modifiable machine tool and workpiece records.

[0042] Next, the data generator prepares training samples. First, it selects a set of machining tasks with confirmed applicable processes, sharing the same structure as the aforementioned pocket tasks. Engineers, mature process libraries, or reliable production records confirm their respective macro-level solutions and specific implementations. As excellent process samples, they must have confirmation criteria for corresponding manufacturing quality requirements and selected evaluation conditions. The model's own judgment does not replace confirmation, nor is it required to prove global optimality. Each sample simultaneously saves its applicable materials, tooling status, clamping, and necessary conditions, denoted as Class A. New tasks generated by changing dimensions or tools require reconfirmation of the process and cannot directly inherit the "excellent" label.

[0043] The data generator, acting as a data preparation program, calls the deterministic process chain to generate process execution parameters, toolpaths, and motions under these conditions. It then calls the industrial world model to calculate time, remaining material geometry, geometry check results, and other physical indicators calculable by the current engineering configuration, forming a Class B physical results summary. Existing results are directly referenced when still applicable. The summary retains the nominal conditions and calculation version; complete material geometry data or waveforms are stored in separate records, and training samples store references to these records. The corresponding data version is fixed before each training round, and this version is directly read during the current training round, avoiding repeated solution of toolpaths and physical processes in each training batch.

[0044] To train state cognition, the data generator selects a known roughing scheme and saves its nominal expected result. Then, within the allowable engineering scope, it changes the defined and adjustable material or tooling conditions in the selected engineering model and still runs the original scheme. The generator knows the hidden conditions actually used, but the input to this manufacturing state query does not contain the truth values ​​of the hidden conditions; it only contains the original process, actual execution, nominal expectation, post-change observations, and existing history. The generator constructs valid state labels based on the hidden conditions, forming Class C samples. For example, two environments using the same roughing command may produce different load observation results when the environment is changed. The model learns the valid manufacturing states corresponding to these results without necessarily asserting that it is material hardening or tool wear. If the observations are insufficient to distinguish the cause, the labels and outputs use the comprehensive load state values ​​or ranges supported by the observations, without forcibly distinguishing individual causes that the observations cannot determine.

[0045] When preparing macro-level and specific implementation monitoring samples, the data generator reconstructs the action mask according to the resources, applicable conditions, and constraint versions at the decision point in time, and checks whether the excellent process label is still within the allowed range. When a label is excluded, the resource references, units, rule applicability, and sample confirmation basis are first checked, and it is temporarily not used for the monitoring of that process. Hard constraints are not relaxed in order to retain the label.

[0046] After machining, an S-type record is generated, saving the actual execution, the initial forecast, the effective measurements, and the final quality. Only confirmed process and state components are used for corresponding monitoring; model estimates are retained as estimates. Dimensional inspections after machining can serve as offline labels and evaluation criteria, but cannot be used as inputs already known during the roughing process.

[0047] Unified model training and deployment: This example uses task identifiers to distinguish invocations. Calling the same model to complete a specified task based on the task identifier and corresponding context is called a query. The unified network is denoted as Fθ, where θ represents all neural network parameters of the model. The task identifier indicates the type of problem to be solved, and the context provides available task, resource, process, or state information for this invocation. The macro-process query reads the task, target structure, resources, and necessary manufacturing state estimates, outputting a complete macro-process scheme. The specific implementation query also reads the complete macro-process scheme, selected steps, applicable manufacturing states, and allowable ranges, outputting candidate implementations for these steps. The manufacturing state query reads the corresponding specific implementation scheme, actual execution status, expected physical results, observed physical results, and necessary manufacturing history, outputting an updated manufacturing state estimate. Specific implementations can be called one step at a time or multiple steps can be organized for output. Different queries use the same neural network parameter version; the task identifier only determines which task is completed in this invocation.

[0048] The model is structured as "input organization, shared network entity, and output decoding." The input organization feeds the task identifier and current context into the network according to agreed-upon fields; the shared entity calculates internal results for use by each task; and the output decoding reads out the step sequence, specific implementation fields, or manufacturing state estimates for the current task. The same shared entity participates in the computation and joint training of all tasks, and input and output processing are integral parts of the same model.

[0049] The input-output relationship can be written as: M = Fθ(macro-process query identifier, manufacturing task, industrial resources, current manufacturing status estimate); R = Fθ(specific implementation query identifier, manufacturing task, M, selected macro-process step, applicable manufacturing status, allowable range); b' = Fθ(manufacturing status query identifier, R, actual execution status, expected physical result, observed physical result, necessary manufacturing history). Here, M represents the macro-process scheme; R represents the candidate specific implementation of the selected step in the specific implementation query, and the specific implementation version corresponding to the observation interval in the manufacturing status query; b' is the manufacturing status estimate updated based on the current valid information. Subsequently, the same Fθ generates a new specific implementation scheme based on b' when there is a basis for process adjustment.

[0050] Figure 9 This describes the training process of the same model from basic learning to unified deployment. The first box, "Basic Process Training," first uses a Class A superior process: the macro-level task provides the task and resources, requiring the model to output the confirmed sequence of steps; the specific implementation task then provides the macro-level solution, selected steps, and applicable conditions, requiring the model to output the confirmed tools, methods, and process variables. During training, the model's answer is compared with the sample answers, and the neural network parameters are adjusted using the training error. Here, "task division" means preparing different questions and answers to teach the same model separately, not training several models separately and then splicing them together.

[0051] The second box, "Adding State Training," continues using this model by adding C-class samples with B as the nominal baseline. For example, for the same pocket roughing scheme, given the predicted load results, effective observations after changing conditions, actual execution, and recent history, the model outputs a manufacturing state estimate, which is then compared with the corresponding labels provided by the generator. This error continues to update the same set of neural network parameters, retaining or interspersing basic process samples to ensure that state learning and existing process capabilities are maintained. The model cannot use hidden conditional truths or manufacturing state supervision labels as input for this manufacturing state query.

[0052] The third section, "Multi-Task Joint Training," allows three types of tasks—macroscopic, concrete implementation, and state—to continuously participate in the same training phase. The training procedure can mix samples of all three types in a single batch or alternate between different batches; each sample still uses its own task identifier, allowed context, and corresponding answer. The model obtains the macroscopic process planning training loss, the concrete implementation generation training loss, and the manufacturing state estimation training loss, respectively, and sums them according to set weights, for example: L = λmacro·Lmacro + λimplementation·Limplementation + λstate·Lstate. Here, L represents the total training loss; Lmacro, Limplementation, and Lstate represent the three training losses mentioned above, respectively, and λmacro, λimplementation, and λstate are the corresponding set weights. Backpropagation adjusts the same shared network entity and the same set of neural network parameters θ based on these errors, enabling the model to consider all three capabilities and support the coordination of manufacturing state estimation and process decision-making. Its effectiveness still needs to be validated through model capability verification.

[0053] In process training, an action mask is first calculated based on the task and the process content (process prefix) determined before the current decision. Then, the same model generates scores for the allowed options. For discrete fields, scores are normalized only within the allowed set, and the selection probability of excluded items is zero; cross-entropy loss is calculated with the probability of applying the label. The model output is converted into continuous process quantities within the allowed range, and regression error is calculated with the corresponding labels. For example, when the allowed velocity coefficient is from the lower limit to the upper limit, and the model output is a normalized value u, candidate velocity coefficients can be obtained by "lower limit + u × (upper limit - lower limit)," where u is between 0 and 1. If multiple variables influence each other, the allowed range of the remaining variables is determined based on the selected variables, or common constraints are set for the variable combinations; variables falling within their respective intervals do not automatically prove the feasibility of the combination.

[0054] When clamping, tooling, and method are generated sequentially, training can employ teacher-mandated methods: using the previously confirmed choices in the samples as the context for subsequent choices. For example, first predict clamping, then generate a tooling mask using the clamping labels of that sample; after predicting the tooling, generate a methoding mask using existing clamping and tooling labels. The answer to the currently predicted field cannot be used to determine its own admission criteria. When macro-training involves step labels, only previously given steps are used as prefix conditions. During deployment, these preceding labels are replaced with the actual content selected by the model; model validation also runs step-by-step according to the actual selections.

[0055] Model capability validation checks macroscopic, implementation, and state capabilities separately, and also verifies their compatibility on tasks not used for training: observations are passed to state queries, their actual estimated outputs are passed to implementation queries of the same model, and the process results are checked by the engineering chain. Hidden ground truth values ​​cannot be used to substitute for deployment-time state estimates to masquerade as effective compatibility. If a test fails, training is adjusted and validation is repeated. A unified deployment selects a validated version, and all three types of queries share the neural network parameters of that version. Subsequent offline updates involve retraining and validation after new records are selected, and then re-released; retraining is not performed for every new task or feedback.

[0056] Process generation and verification before commencement of work: Before commencing work, an initial state is established based on known material, tooling, and clamping information. Components such as wear and dynamic stability, which are not actually measured, are retained as nominal assumptions or unknown markers. The macroscopic process capabilities of the same model are invoked to obtain a sequence of steps, such as "machining the upper surface datum, roughing the pocket, finishing the pocket sidewalls and bottom, and dimensional inspection," clearly defining the target characteristics and completion conditions for each step. Here, macroscopic steps are divided according to task objectives; specific software can use one or more operation records to complete the corresponding steps.

[0057] For pocket roughing, the model reads the complete macroscopic scheme, the roughing steps of this design, the applicable remaining material geometry and resources, selects the roughing tool, available pocket toolpath methods, and allowable process strength and speed levels. The output expresses the process intent, such as retaining finishing allowance and adjusting the cutting load within a specified range; it does not directly output a series of unconstrained machine joint positions as an approval command.

[0058] This example uses a deterministic constraint and masking engine to aggregate rules from classic algorithms and still applicable engineering validation criteria, reusing project features, clamping, tools, machine capabilities, and input material records. The engine converts these rules into selectable process items, numerical ranges, and fields that must retain their original values ​​for selection by the same model. For example, if it is confirmed that the task in this pocket requires roughing before finishing, the mask for selecting the next macro step prevents finishing from being scheduled before necessary roughing; once roughing has been scheduled in the planning prefix, finishing can be designed without waiting for the entity's roughing to complete. Whether the entity can proceed to finishing is checked separately based on the actual progress and actual boundaries.

[0059] When generating this pocket roughing implementation, the known reachable directions, workspace, and explicit fixture occlusions of candidate clamping are first checked; the model selects from clampings that have never been excluded. Then, the tool selection is performed based on the clamping: tools that do not exist, are confirmed unavailable, or have explicitly incompatible geometry are not selected; roughing tools with a diameter of 10 mm are only retained if the tool length and assembly conditions are suitable. Subsequently, the engine reads the features, stages, tool types, and necessary geometric conditions supported by each machining method. Based on the selected clamping and tool formation method mask, the engine does not write a duplicate toolpath algorithm. After the model is selected level by level, the allowable range of process strength and speed is calculated according to the selected method, material, tool, and machine tool conditions. A quick check of unexcluded clamping and tools is performed, and the entire motion, collision, and machining verification is still required.

[0060] Figure 10 The top layer assigns the task and current available conditions to the classical algorithm and the same industrial physical agent, respectively. The left box reuses engineering data to form a mask, and the right box provides process selection scores. The two paths converge at "Select process by action mask in this example." For each selected field, the selection is simultaneously added to the model context and mask basis, subsequent constraints are updated, and the same model continues the selection process. The left training branch compares the prediction results after applying the action mask with the applicable supervision labels and updates the shared neural network parameters; in the diagram, "parameters" refers to the candidate high-level process variables, "precise parameters" refers to the expanded process execution parameters, and "model parameters" refers to the neural network parameters. The right inference branch sequentially expands the execution plan using the classical algorithm, verifies the consequences using classical simulation, and then executes according to the approved goals using classical control, keeping the neural network parameters of the deployed model fixed. At the bottom, execution and physical feedback are obtained. The same model updates the manufacturing status, and the system determines whether to adjust based on responsibility conditions; if there is a basis for adjustment, the classical algorithm updates the allowed range, and the model proposes candidates again.

[0061] Classic process algorithms read the confirmed tools, materials, and process templates, expanding high-level variables such as process strength and speed into process execution parameters such as depth of cut, stepover, feed rate, spindle speed, and allowance. These are the planned values; the actual execution values ​​are fed back by the machine. The impact of each variable is explicitly defined by the template. For example, in a local adjustment mode that only allows changes to the feed rate, the speed coefficient can be multiplied by the approved baseline feed rate, and the allowable range can be checked; this mode does not simultaneously change the target tool, depth of cut, stepover, and spindle speed.

[0062] The toolpath algorithm generates a theoretical path based on the pocket boundary, roughing allowance, tool, and the geometry of the input remaining material. The machine tool motion module converts this into the motion of each joint according to the actual kinematic chain and clamping, checking reachability, limits, and collisions. The material removal module calculates the remaining material along the effective motion, geometrically verifies and compares the remaining shape with the target, and checks for overcutting, allowance, and deviations. In engineering calculations, the material input for finishing must be consistent with the applicable output for roughing; it cannot be recalculated from the initial blank assumption.

[0063] The path, motion, material removal, and verification results are all associated with the input records used through stability identifiers, and input fingerprints and versions are saved so that it can be determined which results can still be reused after process adjustments. Stability identifiers identify the same record, versions distinguish updates, and input fingerprints reflect the input content on which the generated results depend; these bases and dependencies must be verified simultaneously during reuse. Only after all necessary checks pass will the system approve the corresponding scope for execution according to established rules; the model's own textual judgments cannot replace verification.

[0064] This pocket task can also be organized in different sequences. For example, a roughing implementation can be generated and verified first, followed by a finishing implementation; or all implementations of planarization, roughing, finishing, and inspection can be provided first, and then the entire solution can be handed over to the machining chain for unified verification by the world model; or roughing and finishing can be grouped together, and all implementations of that group can be generated and verified together. The verification of the complete design must cover the corresponding steps and their connections; passing a single roughing step does not replace passing the entire solution.

[0065] Figure 2These organizational methods are visualized as a reversible design flow. After the task forms a macro-level solution, the "Select Generation Scope, Verification Scope, and Timing" box determines what to generate, what to verify, and when to verify in this round. The verification scope can include multiple previously generated implementations and does not need to be the same as the newly added generation scope in this round. Existing applicable candidates can be reused directly, and the same model is only called once or multiple times when generation is required. At the "Start Engineering Verification" decision point, the "Do Not Verify for Now" arrow on the right allows other implementations to be generated first, so the next candidate does not need to wait for the previous step to be verified separately. When starting verification is selected, the deterministic process chain generates process execution parameters or process execution plans for the specific implementation scheme to be verified, and the world model checks the physical consequences and state connections of single-step, grouped, or complete schemes.

[0066] Determination of responsibility for on-site execution, quality, and costs: Figure 3 The feedback is divided into different responsibility paths. The top input field shows the actions, quality, and costs corresponding to the process version and execution range. The first judgment box checks the action implementation and quality; if these are not met or the evidence is insufficient, it proceeds to the left execution chain for diagnosis, and then reviews the results. The down arrow only enters the cost judgment when these conditions have been confirmed: if the cost is reasonable, it follows the maintain branch; if the evidence is insufficient, it reports; only if the cost is unreasonable does it proceed to the specific implementation adjustment within the same model. The lower left capacity constraint entry requires that the execution chain has been diagnosed and confirmed to be normal, and the target exceeds the available capacity; a single failure cannot replace this confirmation.

[0067] When normal position or speed tracking deviations occur during roughing, the controller automatically corrects them within the approved target and limits, maintaining the current process; the intelligent model is not required to replan due to deviations at each sampling point. Subsequently, the actual position, speed, and other execution results, as well as the processing quality to be met at this stage, are verified using the same process version and execution range.

[0068] If an action is not performed or fails to meet quality standards, such as when the motion trajectory deviates from the approved toolpath, or when the geometry or surface quality checks for this stage fail, first check the drive status, control code, process parameter mapping, and execution timing. Even if one position tracking indicator is qualified, it does not rule out other execution or control problems. Problems discovered should be handled by the responsible department; if evidence is insufficient, the process should remain in a pending diagnosis state and cannot be directly used to determine if the upper-level process design is flawed. Final inspection data after machining can only be used for judgment after acquisition and cannot be used as quality evidence known during rough machining.

[0069] The cost is evaluated only after confirming that the target action has been achieved and the execution quality and processing quality of this stage are acceptable. For example, if the roughing allowance and surface requirements for the pocket have been met, but the time, energy, or tool consumption to complete the same processing target exceeds the preset cost range, then adjustments are made at the process design level. Effective observations are used to update the manufacturing status estimate according to the corresponding process version, without considering cost anomalies. When the above-mentioned process adjustment basis is available, a more suitable process intensity or speed is proposed by querying the specific implementation of the same model, combined with the updated manufacturing status estimate and allowable range, and the quality, cost, and execution constraints are re-verified. If the cost is reasonable, the plan is maintained; the neural network parameter θ of the deployed model remains unchanged during this feedback adjustment process.

[0070] Another scenario requires upper-level cooperation: the control and execution chain is confirmed to be normal after diagnosis, but the approved target does exceed the current available capacity. The system delegates the confirmed capability constraints to the same model request for feasible implementation, instead of replacing capability diagnosis with execution failures of unknown cause.

[0071] Macro-level replanning is only considered when there is evidence confirming that the current macro-level steps cannot be completed within the permissible scope of concrete implementation. For example, if it is confirmed that the current step's objective is unachievable under permissible clamping and resource conditions, the system assigns the infeasibility basis, completed content, remaining material geometry, and available resources to the same model, rearranges the objectives and order of the remaining steps, and then generates concrete implementations, including clamping, according to the new macro-level scheme. The clamping scheme itself still belongs to the concrete implementation layer and cannot be considered a macro-level replanning simply because the clamping is adjusted. The new macro-level scheme must not overwrite executed or unmodifiable committed content according to the execution agreement and must still pass engineering checks.

[0072] After the system reaches the end boundary of the current step and confirms the completion conditions, it records the actual completed content and proceeds to the next step according to the macro plan. For example, after rough machining is completed, the finishing process reads the geometric state of the current effective remaining material, the state estimate of the same model, and the finishing target, and then generates or verifies the finishing achievement. After all steps are completed, the corresponding final inspection and record summary are performed, and the task ends; if it is not yet completed, the execution of the current step and the feedback loop continue.

[0073] When the corresponding step has been completed, it is still evaluated according to its original process version, and only the manufacturing status estimate of the applicable object is updated; when adjustments are needed, the variable futures that can be modified in the affected subsequent steps are the adjustment objects, and the completed processing content is not regenerated. When all tasks are completed, the feedback is used for result recording, subsequent tasks, and offline training. The new solution should not be regarded as the actual cost of changing the past.

[0074] Process intensity, speed, method, tools, and fixtures are not allowed to be modified at every moment. Deterministic constraints determine the current scope based on boundary conditions and resources, prioritizing adjustments with less impact and those that are feasible. If no feasible candidates are found in the current scope, the allowed modification scope is expanded according to preset conditions and confirmed evidence, or the reason for not finding feasible candidates is reported. The system limits the running time, number of attempts, or number of candidate iterations for process adjustment cycles, and reports the reason when the limit is reached. The model cannot be called indefinitely.

[0075] Once a process adjustment is initiated with justification, the modification range mask directly specifies which fields the model can currently modify. For example, in this case, within a machining block where the reliable remaining material geometry is not yet saved, only the velocity field, which does not change the geometry, is enabled, while clamping, tooling, method, and strength fields that affect the geometry are locked. Only at subsequent plan switching locations with reliable remaining material geometry and connection conditions are other variables considered for enabling according to the original rules. Locked fields retain their original values, but if the original value violates a valid hard constraint, it cannot be forcibly returned to the optional set. When the mask disallows the selection of any process item, allows empty intervals, or locks in conflict with hard constraints, it reports the specific reasons and maintains the plan only if the current plan is still safe and valid; otherwise, it waits or pauses according to established rules. The modification permissions can be relaxed for this instance, but it cannot exceed still valid hard constraints, nor can it replan the macro-process based solely on a single instance of an empty optional or a failed candidate.

[0076] Figure 8 This section specifically explains the runtime loop under fixed neural network parameters. The top section shows the uniformly deployed model; tasks, resources, and applicable states are then listed. Figure 2 The process generation and verification flow is as follows: the approved plan is handed over to classical control for execution, and the version, execution range, and pre-observation predictions are associated. New feedback is entered into the same model to update the manufacturing status, and then... Figure 3 Determine if adjustments are needed. The left branch indicates whether the current state can be maintained or requires further diagnosis; continue according to the execution rules or wait. The right branch generates a constrained implementation based on the new state, and returns to the execution box after generating an execution plan, verifying it, and publishing the new plan to be executed. Both return arrows obtain new feedback only after subsequent execution; the model does not fabricate observations itself. Throughout the loop, the neural network parameters θ remain fixed; state changes do not necessarily require process modifications, and process changes do not necessarily require retraining.

[0077] Future plans for relocation and partial updates: Figure 4The top section divides the rough processing plan into three segments: the executed portion is complete; the submitted portion has been accepted by the execution chain but not yet executed, and according to the current protocol, rewriting this plan is not allowed; only variable futures can be modified when conditions are met. The system records where the execution has progressed and where it has been submitted, forming the executed frontier and the submitted frontier. Pre-calculating the shape of the remaining material or generating a playback cache only indicates that the simulation data is ready, not that the actual machine has completed the corresponding processing, nor that the execution chain has accepted the corresponding plan.

[0078] To generate new solutions, the system selects a point where execution can continue after the submitted frontier as the plan switch point (future semantic boundary), verifies the materials, tools, clamping, and motion connection conditions at that point, and then generates and verifies subsequent candidate solutions without modifying the current plan. This point must correspond to a clearly defined step or action and its boundary state. Candidates can be selected by... Figure 2 The process can be organized step-by-step, in groups, or as a whole; it is only calculated in the background and is not yet part of the submitted portion.

[0079] When a new scheme only changes the allowed feed rate without altering the geometry, the still-applicable path can be reused, but the timing, execution constraints, and physical results affected by the speed must be re-examined. If changes to the depth of cut, stepover, or toolpath structure are required, the effective boundary must have a reliable and applicable input residual material geometry and corresponding generation capability; in this example configuration, speed adjustments that do not change the geometry are allowed at most when there is no confirmed intermediate residual material geometry record within the machining block.

[0080] For example, the system pre-sets a fixed identifier (semantic anchor) for a subsequent plan switching position, even though the current continuous toolpath has not yet reached that position. The system confirms that it falls within the submitted range and has adjustment capabilities, then, based on the material state and motion connection conditions at that location, it creates a candidate solution without altering the current plan, generates the remaining toolpath, and re-completes the necessary motion, collision, material removal, and quality checks. Executed and submitted portions are retained, and background calculations do not modify the currently running official records.

[0081] Figure 4The middle re-verification box checks the input version, material source, executed frontier, and submitted frontier before release. If the machine has exceeded the candidate's effective position during computation, the proposed replacement interval overlaps with an unmodifiable submitted portion, or resource changes render the original basis inapplicable, the candidate expires and requires a new boundary selection or re-verification. Only when the conditions are met does the atomic release box proceed, registering the new plan version, associated results, and proposed effective boundary as a complete update to avoid replacing only a portion. The bottom of the diagram illustrates approved content that can continue to be executed safely and effectively during computation; if verification fails, the reason is reported. The security system independently decides whether to slow down, pause, or take other predetermined actions. Simultaneously, the task, resources, rule version, and modification permissions on which the action mask used by the candidate is based are verified; when relevant basis changes, only the affected mask and candidate checks are updated. A candidate solution passing the old mask check does not guarantee its applicability during release or actual switching; the above conditions still need to be verified.

[0082] Figure 4 The final checkpoint re-verifies the connection conditions when the machine actually reaches the boundary. Only if they match will the system switch, making the new plan effective at that point. The verification checks the applicable materials, tools, clamping, and motion connection conditions at the intended effective location, not just the instantaneous state at the time of calculation initiation. If the actual conditions do not match, the system does not switch, and existing diagnostic and safety rules apply. Registering the new plan and its associated results together prevents the mixing of different versions; only after reaching the switching point and verifying the site conditions can the feasibility of the new plan be confirmed.

[0083] When changing clamps, a safety retreat is first performed in the original clamping state. Only after reaching the safe zone are the new workpiece placement, fixture status, and tool initialization applied. The clamping change is not plotted as a continuous cutting motion; subsequent machining is still re-verified based on the actual new clamping and the applicable remaining material geometry.

[0084] Record reuse, offline training, and engineering integration: Model invocation, training data extraction, and file saving are handled as background tasks by the application; the interface displays progress and diagnostics. The mechanical and kinematic solver continues to perform calculations but does not handle training set export, interface refresh, or repetitive format conversions for data saving. Manufacturing status estimation is a source-based estimate and does not directly overwrite project material parameters, tool definitions, or remaining material geometry records that need to be retained by version. The application invokes the constraint and mask engine to check the applicability of resources and machining methods; existing masks are reused when rules and input criteria remain unchanged, and are not recalculated in each servo cycle.

[0085] Figure 6The two channels at the top are for simulated and physical manufacturing records. Importing them into "Filter, Align, and Version" means first verifying the process version, execution range, data source, validity, and supervision label basis for each record, ensuring that the existing processes, expected results, observed results, and labels in the sample correspond, and then saving it as a fixed version of training data. Then, press... Figure 9 Jointly train, validate, and release the same model. After entering the run box below, keep the neural network parameters fixed and press... Figure 8 Update the status and necessary solutions; adding a new entity record does not mean immediate modification of the current model. Alternatively, during the offline training phase after basic capabilities are established, an industrial world model can be used as a constrained search or reinforcement learning environment to optimize time, quality, stability, and resource costs within feasible processes.

[0086] Figure 6 The other arrow on the right leads to industrial world model calibration: when reliable entity observations show systematic deviations from previously saved and conditionally matched predictions, this can be used to calibrate engineering parameters such as materials, tools, and equipment in the industrial world model. Calibration adjusts the engineering parameters used in the simulation, while offline training adjusts the neural network parameters. New neural network parameters are used for deployment after model capability verification; new engineering parameters are used for subsequent simulations after verification by comparing simulation results with reliable entity observations. The bottom of the figure requires retaining the original predictions and generation conditions for post-hoc verification of the predictions and actual results; post-calibration results should not overwrite the original predictions for comparison.

[0087] Example 2: This embodiment relates to welding manufacturing. In welding, the macroscopic plan determines the weld seam and stage sequence, while the specific implementation selects methods, tools, heat input, and speed. A single industrial physical intelligence model, using the same shared network entity and the same set of neural network parameters, completes macroscopic planning, specific implementation, and state awareness. Classical process algorithms generate action masks based on welding process dependencies, resources, confirmed physical boundaries, and modification permissions. The industrial world model calculates heat, force, deformation, and weld quality. The classical control system executes according to the approved objectives. During the feedback phase, actions and quality are first verified; if not met, the execution chain is diagnosed. If both are met but the cost is unreasonable, the specific implementation is adjusted using the same model. If there is evidence confirming infeasibility, the macroscopic process is replanned. During runtime, only the future parts that are allowed to be modified are updated.

[0088] Example 3: This embodiment relates to additive manufacturing. In additive manufacturing, the macroscopic plan determines layers, regions, and scanning sequence, while the specific implementation selects energy, material flow rate, and scanning method. The same model completes macroscopic planning, specific implementation, and state awareness; classical process algorithms generate action masks to constrain the scanning path, energy, and material flow rate; an industrial world model verifies the temperature field, molten pool, residual stress, and geometry; and classical control executes according to approved objectives. Feedback adjustment is similar to that in Embodiment 1, updating only the planned portions that have not yet been executed and are allowed to be modified.

[0089] Example 4: This embodiment relates to laser processing, semiconductor manufacturing and packaging, electronic assembly, and precision assembly. In semiconductor manufacturing and packaging, electronic assembly, and precision assembly, the macro-level scheme determines the step objectives and sequence, and the specific implementation selects the tools, fixtures, formulas, pressure, temperature, time, or action methods. For laser processing such as laser cutting and laser abrasion, the macro-level scheme determines the processing contour or area, stages, and sequence, and the specific implementation selects the processing method, allowable laser power level, and scanning or feed speed. Classical process algorithms allow options based on material, equipment capabilities, and confirmed conditions, and expand the selected high-level process variables into process execution parameters; the industrial world model verifies physical results such as material removal and thermal effects according to the configured physical model; and the classical control system executes according to the approved objectives. All of the above applications are planned, implemented, and state-aware by the same manufacturing intelligence, using deterministic process algorithms, engineering verification, and execution control in their respective fields. Stepwise, complete, or grouped generation and verification are organizational methods that can be selected and mixed for specific software and algorithms, and are not determined by industry names. Specific state definitions, physical models, measurement accuracy, and safety thresholds are configured according to corresponding engineering conditions.

[0090] The working principle of this invention is as follows: This invention forms a feedback loop according to the following process. A single industrial physical intelligence model, using the same shared network entity and the same set of neural network parameters, unifies macro-process planning, specific implementation generation, and manufacturing state cognition within the same model. Classical process algorithms generate action masks based on real-time tasks, resources, confirmed conditions, and modification permissions, directly integrating engineering hard constraints into the model output selection or decoding, allowing the model to utilize manufacturing experience within permissible limits. The industrial world model undertakes classical simulation, calculating the physical consequences of candidate solutions and performing engineering verification. The classical control system executes within approved goals and permissions, handling routine tracking deviations. The feedback phase first verifies whether the action has been implemented and whether the execution quality and manufacturing quality are up to standard; if not met or insufficient evidence is found, the execution chain is diagnosed, and it is not directly attributed to the process design; after confirming that the action has been implemented and the quality is up to standard, the time, energy, or resource costs are evaluated; only when the cost is unreasonable is the specific implementation adjusted by the same model according to the cost evaluation branch; if the execution chain is confirmed to be normal after diagnosis but the goal exceeds the available capacity, a feasible specific implementation scheme is requested according to the confirmed capacity constraints; only when there is evidence confirming that the macro-steps are infeasible within permissible limits is the macro-process replanned by the same model. During operation, the neural network parameters are fixed, and on-site feedback updates the manufacturing status and necessary process plans. Long-term data is filtered, trained offline, validated, and then new versions are released. This enables a trainable, verifiable, and safely adjustable manufacturing collaboration process.

[0091] Those skilled in the art will understand that, besides implementing the system and its various devices, modules, and units provided by this invention in the form of purely computer-readable program code, the same functions can be achieved entirely through logical programming of the method steps, making the system and its various devices, modules, and units of this invention function in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers. Therefore, the system and its various devices, modules, and units provided by this invention can be considered as a hardware component, and the devices, modules, and units included therein for implementing various functions can also be considered as structures within the hardware component; alternatively, the devices, modules, and units for implementing various functions can be considered as both software modules implementing the method and structures within the hardware component.

[0092] In the description of this application, it should be understood that the terms "upper", "lower", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this application.

[0093] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. An industrial intelligent collaborative control method for variable process manufacturing, characterized in that, An intelligent agent composed of a single industrial physical intelligent neural network model collaborates with classical process algorithms operating according to deterministic rules, an industrial world model undertaking classical simulation, and a classical control system. The industrial physical intelligent neural network model is trained through the correspondence between manufacturing tasks, applicable process states, confirmed processes, and their physical results. Using the same shared network entity and the same set of neural network parameters, it completes macro-level process planning, specific implementation generation, and manufacturing state cognition based on the manufacturing task and context. The manufacturing state estimate obtained from updating the industrial physical intelligent neural network model is used as input for generating specific implementation schemes. The method includes the following steps: S1: Obtain manufacturing tasks, industrial resources, and known manufacturing conditions, and establish an initial manufacturing state estimate based on the known manufacturing conditions; S2: The classical process algorithm determines the optional processes, the allowable range of continuous process quantities, and the modifiable process fields based on the process dependencies of the manufacturing task, available resources, confirmed manufacturing conditions, and the current allowable modification range. In the output selection or decoding of the industrial physical intelligent neural network model, disallowed options are directly excluded, numerical ranges are limited, and the original values ​​of process fields that cannot be modified are maintained. Under the above constraints, the same industrial physical intelligent neural network model generates a macro-process scheme containing one or more macro-process steps and their sequence. Then, based on the manufacturing task, the complete macro-process scheme, the selected one or more macro-process steps to be designed, and their applicable manufacturing states, a corresponding specific implementation scheme is generated. S3: The classical process algorithm generates process execution parameters or process execution plans according to deterministic rules based on specific implementation schemes and industrial constraints. The industrial world model calculates the expected physical results and performs engineering verification on the specific implementation schemes of the selected single or multiple macroscopic process steps, or the complete process scheme containing the specific implementation schemes of each step. After passing the required verification, the design results corresponding to the verification scope are output, or the process execution parameters or process execution plans that have passed the required verification are handed over to the classical control system for execution. S4: After entering the feedback adjustment phase and obtaining the simulated physical observation results or the physical observation results of the entity, the specific implementation scheme, actual execution status, applicable expected physical results saved before observation, observed physical results and necessary manufacturing history corresponding to the process version and execution interval will be input into the same industrial physical intelligent neural network model to update the manufacturing state estimate. During the manufacturing state estimate update and subsequent process re-decision, the neural network parameters of the industrial physical intelligent neural network model will remain fixed. S5: The classical control system handles execution deviations within its authority. Based on actual execution and quality feedback, it verifies whether the target action has been achieved, and whether the execution quality and manufacturing quality of the corresponding stage meet the preset requirements. If the action is not achieved, the quality is unqualified, or the evidence is insufficient, it enters the driver, control code, and execution chain diagnosis. After confirming that the action has been achieved and the quality is qualified, if the costs such as time, energy, or resource consumption do not meet the preset evaluation conditions, the classical process algorithm updates the optional process, the allowable range of continuous process quantities, and the modifiable process fields based on the confirmed conditions and future modification permissions. Then, the same industrial physical intelligent neural network model estimates the current manufacturing status and regenerates the specific implementation scheme under the updated constraints. If there is evidence that the target of the corresponding macroscopic process step and its necessary connection requirements cannot be met within the allowable specific implementation scheme range, the same industrial physical intelligent neural network model replans the macroscopic process. The adjusted scheme is output after being generated and verified as required. During operation, only the future parts that are allowed to be modified are replaced, while the executed parts and the submitted parts that cannot be rewritten according to the execution protocol are retained.

2. The industrial intelligent collaborative control method for variable process manufacturing according to claim 1, characterized in that, The classic process algorithm uses an action mask to connect the selectable process, the allowable range of continuous process quantities, and the modifiable process fields to the output selection or decoding of the same industrial physical intelligent neural network model. The action mask includes one or more of the following: allow or prohibit flags for discrete options, allowable range of continuous process quantities, or process fields that must maintain their original values. Subsequent restrictions are updated as the selected process content is updated. Adjustments to continuous process quantities that meet the implementation conditions are given priority. When no feasible solution can be obtained for the current range, the modification permissions for process methods, tools, or fixtures are gradually opened according to preset conditions. Expanding the scope of modifications does not remove the still-effective hard constraints and the protection of already executed and submitted content.

3. The industrial intelligent collaborative control method for variable process manufacturing according to claim 1, characterized in that, In S5, the preset evaluation conditions specify the corresponding step objectives, execution intervals, quality items, and cost reference ranges. The classic control system handles tracking errors and disturbances within the approved objectives, control limits, and safety boundaries. When an action is not achieved or the quality is unqualified, execution differences and quality evidence are recorded, and the drive, control code, process parameter mapping, and execution timing are checked, without directly identifying it as a process design problem. After confirming that the action has been achieved and the quality is qualified, the effective cost records for the same objective and corresponding interval are compared. Only when the cost does not meet the cost reference range is the same industrial physical intelligent neural network model used for further decision-making on the specific implementation scheme. When the execution chain is confirmed to be normal but the objective exceeds the available execution capacity, the confirmed capability constraints are fed back to request a feasible solution. Safety interlocks and emergency handling do not need to wait for the inference of the industrial physical intelligent neural network model.

4. The industrial intelligent collaborative control method for variable process manufacturing according to claim 1, characterized in that, It also has at least one of the following configurations: The execution plan is divided into three parts: the executed part, the submitted part, and the variable future part. The submitted part is the part that has been submitted to the execution chain, has not yet been executed, and is not allowed to be rewritten in this replanning according to the current execution protocol. The variable future part is the plan content that has not yet been executed and is currently allowed to be modified. The update plan, without changing the executed and committed parts, selects a future position (future semantic boundary) with a clearly defined corresponding step or action and corresponding state and connection conditions. From this position, it generates and verifies the variable future part. Before release, it checks whether the version of the runtime plan on which the candidate plan is based, the actual position already executed (executed frontier), the endpoint of the part of the execution chain that has been accepted and cannot be rewritten according to the protocol (committed frontier), and the proposed switching position are still applicable. If the conditions are met, the new plan version, associated results, and proposed switching position are registered as a complete update (atomic release). When the boundary is reached, the applicable state and connection conditions are checked again, and the switch is only performed if they match. During the computation, the machine can continue to execute the approved content that is still safe and valid. Expired or failed verification candidates do not replace the formal plan. The manufacturing state estimation characterizes one or more of the following: material effective load, tooling state, dynamic stability, thermal state, geometric state, or tooling state, and can express the estimated value, range, probability, or unknowns of the effective state; the expected physical results and observed physical results correspond according to the process version, applicable object, and execution interval. The expected physical results used for comparison should be results that have been saved and are still applicable before the observed physical results are obtained, and the manufacturing history is limited to the information available at that time; when observations are lacking, a priori manufacturing state estimation with nominal assumptions or unknown labels is used, and the expected physical results are not directly recognized as physical measurement results; The generation and verification of specific implementation schemes can be carried out alternately, or multiple or all specific implementation schemes can be generated first and then verified in a concentrated manner. The method adopted can be selected or switched according to the task or different stages of the same task. Specific implementation schemes can be generated in one order and then divided into verification batches as needed. The processes involved must still meet the actual manufacturing dependencies and state connection requirements. When generating subsequent specific implementation schemes in advance, the macro process scheme, known initial conditions, generated processes and intermediate manufacturing states clearly marked as assumptions can be read. By recording stability identifiers, calculating the content summary (input fingerprint) of inputs, and version, the correspondence between process generation results, process execution plans, and verification results and the tasks, resources, manufacturing status, and specific implementation schemes used is recorded. When the specific implementation plan changes, determine whether the existing results are still applicable based on the dependencies, reuse the still applicable results, and only regenerate or verify the affected future results; only deliver the process execution plan that matches the input basis with the verification version and passes the required verification for execution. Before commencement of work, the process design, scheme adjustment under simulation results, and state cognition and process adjustment under physical feedback conditions all call the same neural network parameter version of the industrial physical intelligent neural network model. No separate manufacturing intelligent models are established for the above purposes. During operation, the state update does not modify the neural network parameters of the deployed industrial physical intelligent neural network model. After long-term data accumulation, a new neural network parameter version is formed through offline training and verification and released separately. The industrial world model includes one or more of the following: geometry, kinematics, dynamics, collision, material removal, force, heat, vibration, wear, and mass calculation. If a learning-type physics prediction auxiliary task is set, it will be undertaken by the same industrial physics intelligent neural network model, while the formal engineering verification will still be performed by the industrial world model according to the corresponding algorithm. The manufacturing tasks include one or more of CNC machining, laser processing, welding, additive manufacturing, semiconductor manufacturing and packaging, electronic assembly, or precision assembly. In laser processing such as laser cutting and laser abrasion, the macroscopic process plan determines the processing contour or area, stages, and sequence, while the specific implementation plan determines the processing method, allowable laser power level, and scanning or feed speed. In welding, the macroscopic process plan determines the weld seam and stage sequence, while the specific implementation plan determines the method, tools, heat input, and speed. In additive manufacturing, the macroscopic process plan determines the layers, areas, and scanning sequence, while the specific implementation plan determines the energy, material flow rate, and scanning method. In semiconductor manufacturing and packaging, electronic assembly, and precision assembly, the macroscopic process plan determines the step objectives and sequence, while the specific implementation plan determines the tools, fixtures, formulas, pressure, temperature, time, or action methods.

5. A method for training an industrial physical intelligent neural network model for industrial intelligent collaborative control, characterized in that, To train a single industrial physical intelligence neural network model, the generation of training samples is organized through a data generator, including the following steps: M1: Obtain manufacturing tasks, industrial resources, confirmed excellent macro-process schemes, process applicability status, and confirmed excellent specific implementation schemes under that status; call the industrial world model to calculate the corresponding physical results; and construct macro-process planning supervision samples and specific implementation generation supervision samples. M2: The data generator performs constrained changes on the materials, tools, machine dynamics, thermal, geometric or tooling conditions of the simulation environment, calls the industrial world model to run the existing specific implementation scheme, retains the expected physical results under the nominal conditions before the change, and obtains the simulated physical results and the actual execution of the simulation after the change. M3: The data generator constructs a valid manufacturing status label based on the hidden conditions adopted, forming a manufacturing status supervision sample with the corresponding specific implementation scheme, actual execution status, expected physical results, observed physical results and necessary manufacturing history as inputs and the valid manufacturing status label as the supervision target. The input of the manufacturing status supervision sample does not include the hidden condition truth value and the valid manufacturing status label. M4: Using deterministic constraint rules consistent with the inference phase, based on the tasks, resources, confirmed conditions, and allowable modification ranges available during the corresponding training decision, the allowable options, numerical ranges, and modifiable process fields of the process output are determined. Under the condition that the above constraints directly participate in the output selection or decoding of the industrial physical intelligent neural network model, the training loss between the predicted macroscopic process scheme and the specific implementation scheme and their respective applicable labels is calculated. M5: After task-based training of macro-process planning, specific implementation generation, and manufacturing state estimation, mixed sample joint training is performed. Each task updates the same shared network entity and the same set of neural network parameters. The manufacturing state estimation output and the manufacturing state input in the specific implementation generation task use the same manufacturing state definition. The same industrial physical intelligent neural network model is trained for inference. This model keeps the neural network parameters fixed during inference.

6. The method for training an industrial physical intelligent neural network model for industrial intelligent collaborative control according to claim 5, characterized in that, It also has at least one of the following configurations: First, train the macro-process planning and specific implementation generation capabilities, then add manufacturing state training, and retain the task samples in the joint training; for manufacturing state supervision tasks, the hidden condition truth value is not used as the input of the industrial physical intelligent neural network model, but only as the basis for constructing the supervision label of the task. The specific implementation scheme label must be a reasonable process confirmed in the corresponding state. The original process that is no longer applicable after the change is not directly used as the excellent specific implementation scheme label of the state. The process output restrictions are implemented by using action masks to select or decode the output of the industrial physical intelligent neural network model, and the corresponding process supervision loss is calculated based on the restricted output. When the macroscopic process scheme label or the specific implementation scheme label conflicts with the corresponding action mask, the sample and constraint basis are checked first, and it is not used for the process supervision. The current answer to be predicted and future information are not used to generate the mask. Training and inference follow the corresponding constraint rules; The first round of training uses fixed data with source and version information. Subsequent rounds use selected entity records and simulation records for offline joint training to update the neural network parameters. The industrial world model is calibrated by using the difference between the applicable predicted physical results and the credible entity observation physical results. The trained industrial physical intelligent neural network model is then validated for capability verification. The calibrated industrial world model is then validated for prediction results. Once the validation is successful, the corresponding parameters are saved as the new version. During runtime, the state update does not modify the neural network parameters of the deployed industrial physical intelligent neural network model. Discrete outputs are classified or sequenced based on their encoding, while continuous process quantities and supervised state components are used to calculate regression losses. Components without reliable labels are not used to calculate corresponding supervision errors. The training losses from macroscopic process planning, concrete implementation generation, and manufacturing state estimation are weighted and summed according to set weights to obtain the overall training objective.

7. An industrial intelligent collaborative control system for variable process manufacturing, characterized in that, It includes a task and resource management module, an industrial physical intelligence module consisting of an intelligent agent composed of a single industrial physical intelligence neural network model, a deterministic process generation and constraint module that runs classical process algorithms, an industrial world model module that undertakes classical simulation, a classical control module, and an observation module. The industrial physical intelligence neural network model in the industrial physical intelligence module is trained by the correspondence between manufacturing tasks, process applicability status, confirmed processes and their physical results. It generates macroscopic process schemes and specific implementation schemes with the same shared network subject and the same set of neural network parameters. The manufacturing status estimate is updated according to the specific implementation schemes corresponding to the process version and execution interval, the actual execution status, the expected physical results, the observed physical results and the necessary manufacturing history. The estimate is then used as the input for the industrial physical intelligence neural network model to generate specific implementation schemes. During the operation period, the neural network parameters of the industrial physical intelligence neural network model remain fixed. The deterministic process generation and constraint module determines and updates the optional process, the allowable range of continuous process quantity, and the modifiable process field based on process dependence, available resources, confirmed manufacturing conditions, and future modification permissions. The above constraints are directly used for the output selection or decoding when the same industrial physical intelligent neural network model generates macroscopic process schemes and specific implementation schemes, and the specific implementation schemes are expanded into process execution parameters or process execution plans. The industrial world model module calculates physical consequences and verifies candidate process schemes; the classical control module handles execution deviations within the approved target authority; and the observation module provides feedback with observation sources and execution ranges. When an action is not performed, the quality is unqualified, or there is insufficient evidence, the system enters the driver, control code, and execution chain diagnosis. If it is confirmed that the target action has been performed, the execution quality and the manufacturing quality of the corresponding stage meet the preset requirements, but the costs such as time, energy, or resource consumption do not meet the preset evaluation conditions, the same industrial physical intelligent neural network model adjusts the specific implementation scheme. If it is confirmed that the target of the corresponding macroscopic process step and its necessary connection requirements still cannot be met within the allowable specific implementation scheme, the same industrial physical intelligent neural network model re-plans the macroscopic process. The adjusted scheme is output after being generated and verified as required. During operation, only the future parts that can be modified are updated, while the executed parts and the submitted parts that cannot be rewritten according to the execution protocol are retained.

8. The industrial intelligent collaborative control system for variable process manufacturing according to claim 7, characterized in that, It also includes the following configurations: The deterministic process generation and constraint module uses action masks to select or decode the output of the same industrial physical intelligent neural network model that restricts access. The action mask includes one or more of the following: allow or prohibit flags for discrete options, allowable ranges for continuous process quantities, or process fields that must maintain their original values. The system updates subsequent constraints as the selected process content is updated. The system divides the operation plan into the executed part, the submitted part, and the variable future part. When updating, it starts from the future semantic boundary. Before publishing, it checks the version of the operation plan on which the candidate solution is based, the executed frontier, the submitted frontier, and the boundary to be effective. When the boundary is reached, the applicable manufacturing status and connection conditions are checked again. Switching is only performed if they match.

9. A computer-readable storage medium storing a computer program, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the industrial intelligent collaborative control method for variable process manufacturing as described in any one of claims 1 to 4.

10. An electronic device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the industrial intelligent collaborative control method for variable process manufacturing as described in any one of claims 1 to 4.

Citation Information

Patent Citations

  • A data-driven energy-saving planning method for machining processes

    CN114611379B

  • Complex part process route planning method based on re-optimization genetic algorithm

    CN118709849A

  • Knowledge-guided deep reinforcement learning-based reconfigurable workshop dynamic scheduling method

    CN121413980A

  • Constraint based inference and machine learning system

    US20220164636A1

  • Autonomous Process Recipe Generation for Semiconductor Process Systems through Reinforcement Learning

    US20260010142A1