Simulation decision-making method, device, medium and product based on industrial world model

Abstract

This application relates to the field of industrial artificial intelligence technology, and provides a simulation decision-making method, device, medium, and product based on an industrial world model. The method includes: converting process-related text into template action codes based on an industrial ontology, and then submitting the template action codes to an industrial world model; using standardized multimodal data as input parameters to an industrial state representation model to obtain latent state features output by the industrial state representation model; using the latent state features and intervention action codes as input parameters to an industrial dynamic simulation model to obtain a multi-round simulation time series output by the industrial dynamic simulation model, and quantifying the multi-round simulation time series to obtain multi-dimensional quantitative simulation indicators; and using the multi-round simulation time series and multi-dimensional quantitative simulation indicators as input parameters to an industrial prediction and decision-making model to obtain the target decision instructions output by the industrial prediction and decision-making model under the constraints of process compliance and optimization target weights.

Description

Technical Field

[0001] This application relates to the field of industrial artificial intelligence technology, and in particular to a simulation decision-making method, device, medium and product based on an industrial world model. Background Technology

[0002] In the evolution of modern intelligent manufacturing and the Industrial Internet, with the deep integration of digital twins and artificial intelligence technologies, various discrete manufacturing and process manufacturing enterprises (such as automobile assembly, semiconductor wafer processing, precision machinery equipment, and chemical processes) are facing the dual challenges of highly flexible production lines and massive bursts of multi-source heterogeneous data. During production line operation, PLC (Programmable Logic Controller), industrial cameras, 3D laser scanners, various sensors (such as vibration sensors, temperature sensors, and pressure sensors), MES (Manufacturing Execution System) or ERP (Enterprise Resource Planning) and other enterprise information systems are continuously generating high-dimensional heterogeneous data such as one-dimensional sensor time series, two-dimensional images, three-dimensional point clouds, and unstructured chemical single text. To cope with production disturbances such as order insertions, sudden equipment failures, material batch fluctuations, and energy consumption control, enterprises urgently need industrial AI (Artificial Intelligence) systems with long-term time-series operating condition extrapolation, multi-objective decision optimization, and implementation scheduling control capabilities. The AI ​​system must achieve functions such as improving production line yield, predictive maintenance, and flexible production scheduling while strictly ensuring equipment load safety and process compliance.

[0003] In related technologies, the AI ​​system's equipment acquisition layer interfaces with PLCs and MES to collect time-series data; the data processing layer uses general data cleaning algorithms to complete simple missing value completion; the digital twin simulation layer provides static visualization based on a 3D geometric model, supporting only simple time-series trend prediction. Meanwhile, a single large-scale industrial model uses natural language interaction to analyze production needs and directly outputs text-based solutions for production scheduling, maintenance, or process optimization.

[0004] However, traditional digital twin simulations lack process semantics and rely solely on time-series data for digital extrapolation. The simulation results only output quantitative data such as energy consumption and yield. Furthermore, general industrial large models can only process text work orders and process documents and do not have the ability to simulate the physical dynamics of production lines. The output process optimization solutions contain a large number of "industrial illusions" that violate equipment load, thermodynamics, and capacity constraints, and cannot be directly applied to production lines. Summary of the Invention

[0005] This application provides simulation decision-making methods, devices, media, and products based on an industrial world model, which integrate process semantics, state representation, dynamic inference, and decision output into the same processing link. This enables each module to work collaboratively around a unified semantic input and a unified state expression. The simulation no longer relies solely on time-series data for digital inference, but can form executable decision instructions that avoid "industrial illusion" through constrained simulation and decision output.

[0006] In a first aspect, embodiments of this application provide a simulation decision-making method based on an industrial world model, including:

[0007] Based on the industrial ontology, process-related text is converted into template action codes, and the template action codes are then handed over to the industrial world model; the industrial world model includes an industrial state representation model, an industrial dynamic simulation model, and an industrial prediction and decision-making model.

[0008] Standardized multimodal data is used as input parameters for an industrial state characterization model to obtain the latent state features output by the model.

[0009] By using latent state features and intervention action codes as input parameters of the industrial dynamic simulation model, a multi-round simulation time series output by the industrial dynamic simulation model is obtained, and the multi-round simulation time series is quantized to obtain multi-dimensional quantitative simulation indicators.

[0010] By using multi-round simulation time series and multi-dimensional quantitative simulation indicators as input parameters for the industrial prediction and decision-making model, the target decision instructions output by the industrial prediction and decision-making model under the constraints of process compliance and optimization target weights are obtained.

[0011] In one possible implementation, the steps for acquiring standardized multimodal data include:

[0012] Obtain multimodal raw data from production line equipment;

[0013] Text data is obtained from an industrial information system, and text vectors are obtained by embedding the text data.

[0014] Standardized multimodal raw data and text vectors are cleaned and standardized to obtain standardized multimodal data.

[0015] In one possible implementation, the process-related text is converted into template action encoding based on an industrial ontology, including:

[0016] Semantic extraction and entity recognition are performed on process-related texts to obtain key business elements;

[0017] Threshold constraints, standardization scaling, and numerical fitting are applied to the numerical values ​​of key business elements to transform them into normalized values.

[0018] Based on the industrial ontology, constraint mapping and format unification are performed on the key business elements after numerical transformation to obtain standardized data;

[0019] Based on the pre-defined coding rules, template action codes are generated using standardized data.

[0020] In one possible implementation, standardized multimodal data is used as input parameters to an industrial state characterization model to obtain the latent state features output by the model, including:

[0021] Separate time-series data, visual data, and text vectors from standardized multimodal data;

[0022] Steady-state baseline features and abrupt failure features are extracted from time series data, and then the steady-state baseline features and abrupt failure features are concatenated into noise-free time series features.

[0023] Visual data is encoded and processed to filter image noise and obtain noise-free visual features;

[0024] Noise-free temporal features, noise-free visual features, and text vectors are used as input parameters for a cross-modal projection encoder that has undergone contrastive learning, thereby obtaining multiple latent features output by the cross-modal projection encoder.

[0025] Adaptive weighted aggregation is performed on multiple latent features to obtain latent state features.

[0026] In one possible implementation, latent state characteristics include equipment-level latent state characteristics and production line-level latent state characteristics; intervention action coding includes equipment-level intervention action coding and production line-level intervention action coding.

[0027] Industrial dynamic simulation models include stochastic state-space models, temporal Transformers, and physical information neural networks;

[0028] By using latent state features and intervention action codes as input parameters to the industrial dynamic simulation model, a multi-round simulation time sequence output by the industrial dynamic simulation model is obtained, including:

[0029] By using device-level latent state features and device-level intervention action encoding as input parameters of a stochastic state-space model, long-term time-series prediction features of multiple devices are obtained.

[0030] By using production line-level potential state features and production line-level intervention action encodings as input parameters of the temporal Transformer, global temporal features of the entire production line can be obtained.

[0031] The long-term prediction features and global time-series features of multiple devices are input into the physical information neural network to obtain the multi-round simulation time-series sequence output by the physical information neural network.

[0032] In one possible implementation, the method further includes:

[0033] Numerical analysis is performed on multi-dimensional quantitative simulation indicators to obtain structured numerical features;

[0034] By utilizing industrial knowledge graphs, causal reasoning and natural language semantic generation are performed on multi-dimensional quantitative simulation indicators carrying structured numerical features to obtain preliminary attribution conclusions for operating conditions.

[0035] Based on the process entity, the preliminary attribution conclusions of the operating conditions are revised to obtain a process semantic analysis report.

[0036] In one possible implementation, the multi-round simulation time series includes corrected long-time prediction features of multiple devices and corrected global time series features after multiple rounds of simulation.

[0037] Using multi-round simulation time series and multi-dimensional quantitative simulation indicators as input parameters for the industrial prediction and decision-making model, the target decision instructions output by the industrial prediction and decision-making model under the constraints of process compliance and optimization target weights are obtained, including:

[0038] By using the corrected long-term prediction features and multi-dimensional quantitative simulation indicators of multiple devices as input parameters for industrial MPC, the candidate control parameters output by industrial MPC for each device under process compliance constraints and optimization target weight limitations are obtained.

[0039] The global time series features and multi-dimensional quantitative simulation indicators after multiple rounds of correction are used as input parameters for the deep reinforcement learning algorithm to obtain the medium- and long-term candidate scheduling schemes output by the deep reinforcement learning algorithm under the constraints of process compliance constraints, optimization target weights and candidate control parameters output by the industrial MPC for each device.

[0040] Based on the candidate control parameters of each device and the medium- and long-term candidate scheduling schemes corresponding to each set of candidate control parameters, multiple candidate decision schemes are generated.

[0041] Filter multiple candidate decision-making options based on compliance filtering strategies;

[0042] Generate the target decision instruction based on the remaining candidate decision schemes after filtering.

[0043] In one possible implementation, multiple candidate decision options are filtered based on a compliance filtering strategy, including:

[0044] Filter multiple candidate decision schemes based on fundamental physical constraints;

[0045] The remaining candidate decision schemes after the initial filtering are then filtered a second time based on process standards, safety production regulations, and industry-specific rules.

[0046] In one possible implementation, generating a target decision instruction based on the remaining candidate decision schemes after filtering includes:

[0047] The remaining candidate decision schemes after filtering are broken down according to the execution level to obtain the decision instructions for each execution level;

[0048] Construct target decision instructions based on decision instructions from multiple execution levels.

[0049] In one possible implementation, the method further includes:

[0050] The target decision instructions are handed over to the intelligent agent, which then distributes the decision instructions at each execution level of the target decision instructions to the execution object.

[0051] In one possible implementation, the method further includes:

[0052] Obtain real production line operating data within a specified time period after executing the target decision instruction;

[0053] An online residual adaptive mechanism is triggered based on real production line operating data to adjust the model parameters of the industrial dynamic simulation model.

[0054] In one possible implementation, standardized multimodal data is used as input parameters to an industrial state characterization model to obtain the latent state features output by the model, including:

[0055] Standardized multimodal data is used as input parameters for an industrial state representation model to obtain the latent state features output by the industrial state representation model under the guidance of semantic attention-guided features; wherein, the semantic attention-guided features are generated by a large industrial model.

[0056] In one possible implementation, the latent state features include standard latent state features and lightweight latent state features;

[0057] The method also includes:

[0058] After the industrial state representation model outputs latent state features, the lightweight latent state features are handed over to the large industrial model to complete the cross-modal semantic alignment of standardized multimodal data.

[0059] In one possible implementation, the intervention action code is obtained as follows:

[0060] Before encoding potential state characteristics and intervention actions as input parameters for the industrial dynamic simulation model, it is determined whether the industrial prediction and decision-making model outputs the target decision instruction.

[0061] If not, generate an intervention action code that is empty;

[0062] If so, the latest target decision instructions output by the industrial forecasting and decision-making model are converted into intervention action codes based on the industrial ontology and template action codes.

[0063] In one possible implementation, the intervention action code is obtained as follows:

[0064] In response to production demand adjustment instructions, candidate adjustment schemes are generated based on knowledge graphs, and the candidate adjustment schemes are converted into intervention action codes according to industrial ontology and template action codes.

[0065] In one possible implementation, the method further includes:

[0066] Based on the noise-free visual features in the multi-round simulation time sequence and latent state features, a 3D simulation image of an industrial scene is rendered and displayed.

[0067] Secondly, embodiments of this application provide an electronic device, including: a memory and a processor;

[0068] The memory stores instructions that the computer executes;

[0069] The processor executes computer execution instructions stored in memory, causing the processor to perform the methods described above.

[0070] Thirdly, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the methods provided above.

[0071] Fourthly, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the method provided above.

[0072] The simulation decision-making method, equipment, medium, and product based on the industrial world model provided in this application first converts process-related text into template action codes through an industrial ontology. Standardized multimodal data is uniformly mapped to a shared latent space by an industrial state representation model to form latent state features. The latent state features and intervention action codes jointly drive the industrial dynamic simulation model to generate multi-round simulation time-series sequences, and further obtain multi-dimensional quantitative simulation indicators. Subsequently, the industrial prediction and decision-making model outputs target decision instructions under process compliance constraints and optimization target weight limitations, thereby incorporating process semantics, state representation, dynamic inference, and decision output into the same processing link. This processing method addresses the problems in existing industrial systems, such as the difficulty in directly entering the unified inference link for process text, the lack of consistent state expression for multimodal data, and the disconnect between simulation results and decision control. It enables each module to work collaboratively around a unified semantic input and a unified state expression, and forms executable decision instructions that avoid "industrial illusion" through constrained simulation and decision output. In addition, simulation and decision-making can be repeatedly iterated, continuously outputting decision instructions that can be implemented in production lines under adapted operating conditions. Attached Figure Description

[0073] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0074] Figure 1 A flowchart illustrating a simulation decision-making method based on an industrial world model provided in an embodiment of this application;

[0075] Figure 2 A flowchart illustrating a method for acquiring standardized multimodal data according to an embodiment of this application;

[0076] Figure 3 Provided for an embodiment of this application Figure 1 Detailed flowchart of step S101;

[0077] Figure 4 Provided for an embodiment of this application Figure 1 Detailed flowchart of step S102;

[0078] Figure 5 Provided for an embodiment of this application Figure 1 Detailed flowchart of step S103;

[0079] Figure 6 A flowchart illustrating a method for generating a process semantic analysis report according to an embodiment of this application;

[0080] Figure 7 Provided for an embodiment of this application Figure 1 Detailed flowchart of step S104;

[0081] Figure 8 This application provides an architectural diagram of the industrial AI system.

[0082] Figure 9 A schematic diagram of the structure of the simulation decision-making device based on the industrial world model provided in this application;

[0083] Figure 10 A schematic diagram of the structure of the electronic device provided in this application.

[0084] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0085] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application.

[0086] In the field of intelligent manufacturing, production line monitoring, process optimization, and scheduling control typically rely on the collaborative operation of digital twins, industrial intelligent analysis, and execution systems. Related applications cover electronics manufacturing, equipment processing, semiconductors, and process industries. These systems generally connect field devices, business management platforms, analytics platforms, and operator terminals to process multi-source data, including sensor timing data, visual images, and process text.

[0087] Existing industrial systems often employ a separate deployment approach for simulation, analysis, and scheduling. The bottom layer collects equipment status and production data, while the upper-layer modules respectively complete operating condition prediction, process analysis, and control command generation. Operators then adjust parameters and issue tasks based on simulation results or analysis conclusions.

[0088] In actual operation, simulation results of this type of technology often remain at the level of numerical output. Because the process text and control intent lack a unified semantic bridge with the simulation flow, the process text and control intent cannot be directly integrated into the simulation extrapolation process. Furthermore, multimodal data lacks a consistent state representation before entering the analysis module, causing information connection between different systems to rely on manual processing and conversion. With the continuous accumulation of on-site noise, equipment aging, and operating condition fluctuations, long-term extrapolation errors easily accumulate, leading to a gradual deviation between predicted results and the actual production line.

[0089] Meanwhile, the decision-making module fails to adequately utilize simulation timing, quantitative indicators, and process constraints. The generated scheduling or process instructions lack sufficient verification before issuance, potentially leading to "industrial illusions" that violate equipment load, thermodynamics, and capacity constraints. This can result in mismatches with optimization objectives, inconsistencies with process specifications, or poor adaptability. Such deficiencies reduce production line optimization efficiency, increase trial-and-error costs, and impact continuous and stable operation in complex industrial scenarios.

[0090] To address the aforementioned technical challenges, an industrial ontology is introduced as a unified semantic carrier between process text and model actions. First, process-related text is organized into parsable template action codes, which are then passed to the industrial world model for continuous simulation. Field equipment, industrial internet, core algorithm servers, and visualization terminals form the foundation for collaborative operation. Standardized multimodal data is input into the industrial state representation model to form latent state features. These latent state features, along with intervention action codes, jointly drive the industrial dynamic simulation model to output multi-round simulation time-series sequences, which are then quantified to obtain multi-dimensional quantitative simulation indicators.

[0091] To address the disconnect between simulation results and decision control, multi-round simulation time sequences and multi-dimensional quantitative simulation indicators are uniformly fed into the industrial prediction and decision model, generating target decision instructions under process compliance constraints and optimization target weight limitations. This integrates process semantics, state representation, dynamic deduction, and decision output into a single processing loop, improving the accuracy, reliability, and adaptability of simulation decisions in complex industrial scenarios.

[0092] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0093] Figure 1 A flowchart illustrating the simulation decision-making method based on the industrial world model provided in this application is shown below. Figure 1 As shown, the method includes the following steps S101 to S104.

[0094] Step S101: Based on the industrial ontology, process-related text is converted into template action codes, and the template action codes are handed over to the industrial world model; wherein, the industrial world model includes an industrial state representation model, an industrial dynamic simulation model, and an industrial prediction and decision-making model.

[0095] Here, process-related text refers to text information generated during production line operation that can affect equipment actions, process parameters, or scheduling behavior. For example, process-related text may include, but is not limited to, process optimization text schemes and semantic intervention instructions. Process optimization text schemes can contain multiple sets of composite action codes such as parameter adjustments, process changes, maintenance scheduling, and production line modifications, representing medium- to long-term systematic optimization actions. Process optimization text schemes can be generated based on industrial knowledge graphs and industrial ontology reasoning, and are industry-specific texts. Semantic intervention instructions are concise, scenario-specific natural language operation instructions output by industrial knowledge graphs and industrial ontology reasoning.

[0096] An industrial ontology is a proprietary business semantic specification for an enterprise. It defines unified semantic standards for industrial entities, attributes, and relationships, serving as a unified semantic bridge for interactions between industrial world models, large-scale industrial models, and intelligent agents. Industrial ontologies can include domain ontologies and business ontologies. Domain ontologies can include product ontologies, process ontologies, quality ontologies, equipment ontologies, etc.; business ontologies can include ontologies for planning management, production processes, procurement management, quality inspection, etc.

[0097] Template action encoding is a standardized intermediate representation constrained by the industrial ontology. It is used as template information for action input within the industrial world model and between the industrial world model and external models (such as large industrial models, intelligent agents, etc.). In other words, subsequent instructions involving action control can be parsed based on template action encoding, so that the instructions can be executed correctly.

[0098] The industrial state representation model is responsible for transforming multimodal data from industrial sites and industrial IT systems into a unified shared latent space to form a latent state representation.

[0099] The industrial dynamic simulation model is responsible for performing multiple rounds of time-series simulations under the combined effect of potential state representation and intervention actions.

[0100] Industrial forecasting and decision-making models are responsible for outputting executable instructions under constraints and target limitations.

[0101] In this application, the execution entity can be deployed on the core algorithm server or distributed in an architecture where edge computing nodes and the central server collaborate.

[0102] First, a large amount of pre-generated process-related text can be obtained. Then, relying on the industrial ontology, the process-related text is converted into template action codes. If there are multiple action statements in the text, an action chain can be formed that can be read step by step by the industrial world model. After obtaining the template action codes, the template action codes are handed over to the industrial world model.

[0103] Here, the introduction of industrial ontology means that process-related text no longer enters the analysis chain as raw natural language. Instead, it is transformed into structured, machine-processable encoding through unified semantic mapping, allowing control intentions, process constraints, and execution sequences in the natural language text to directly participate in subsequent simulations and decisions. In one possible embodiment, template action encoding can uniformly define fields such as: equipment identifier, controlled object, adjustment type, target value, execution cycle, constraint threshold, priority label, etc. For example: Equipment: Welding machine 03; Parameter: Preheating temperature; Action: Increase; Action: Increase; Target value: 180℃; Execution duration: 8 consecutive hours.

[0104] In this case, by relying on the industrial ontology to convert process-related text into template action codes, a consistent mapping relationship is established between the process-related text and the model input. The industrial state representation model, the industrial dynamic simulation model, and the industrial prediction and decision-making model can work around the same action semantics, reducing the link break caused by cross-system manual interpretation and format conversion.

[0105] In one possible embodiment, step S101 can be performed using a pre-trained industrial big model, which is trained on a general big model using industrial basic knowledge and domain knowledge of sub-sectors (such as home appliances, automobiles, machinery and equipment, and semiconductors). Multiple big models can be trained and combined into an industrial big model.

[0106] The industrial big data model can leverage industrial knowledge graphs to perform process semantic reasoning, simulation result interpretation, and decision compliance verification. In step S101, process-related text is handed over to the industrial big data model, which then converts the text into template action codes based on the industrial ontology. After obtaining the template action codes, they can be passed to the industrial world model.

[0107] Step S102: Use standardized multimodal data as input parameters for the industrial state characterization model to obtain the latent state features output by the industrial state characterization model.

[0108] Standardized multimodal data refers to multimodal data that has undergone standardized cleaning. The standardized cleaning method can be pre-set according to business needs. Multimodal data can include vibration, temperature, images, 3D point cloud data acquired from production line equipment, as well as text data acquired from industrial information systems (such as ERP systems, MES systems, etc.).

[0109] The industrial state characterization model is used to map data from different modalities into latent state features in a shared latent space. These latent state features are a comprehensive expression of equipment, processes, production line load, quality trends, and contextual constraints, and can be used as state inputs for subsequent industrial dynamic simulation models.

[0110] In practice, after standardized multimodal data is input into the industrial state characterization model, the industrial state characterization model outputs the potential state characteristics.

[0111] The latent state features can take the form of fixed-length vectors, state tensors, etc., as long as they can comprehensively represent the current production line state. This step ensures that data from different systems and devices have a unified state representation before entering the simulation. Subsequent models do not need to interpret the original formats of each modality separately, but instead directly perform continuous deduction based on the latent state features.

[0112] Step S103: Use the latent state features and intervention action codes as input parameters of the industrial dynamic simulation model to obtain the multi-round simulation time sequence output by the industrial dynamic simulation model, and quantize the multi-round simulation time sequence to obtain multi-dimensional quantitative simulation indicators.

[0113] Intervention action codes are control parameters that indicate actions such as fine-tuning, maintenance, or shutdown of equipment parameters in the production line. They can also be control parameters that indicate actions such as batch discharge adjustments, multi-AGV (Automated Guided Vehicle) scheduling, and overall line process modifications. The coding method for intervention action codes is the same as that for template action codes. Therefore, industrial dynamic simulation models can understand intervention action codes based on template action codes.

[0114] Intervention action codes can be generated automatically in the simulation and decision-making process, or they can be generated according to business needs when human intervention is involved. See the relevant description below for details.

[0115] The industrial dynamic simulation model is the core deduction module in the industrial world model. Its function is to simulate the impact of a given action on the evolution of equipment-level, process-level, and production line-level states over time, and output a multi-round simulation time sequence at multiple future moments.

[0116] A multi-round simulation time series is a set of predicted states arranged by time steps, which may include state variables such as key process parameters, capacity, energy consumption level, delivery cycle, quality risk value, equipment load value, and constraint violation flags.

[0117] Multidimensional quantitative simulation indicators are obtained by further quantification of the simulation time series from multiple rounds. They are used to form comparable and sortable decision evaluation inputs, and can also be used to generate analysis reports on the simulation process.

[0118] If multiple intervention action codes exist at the same time, multiple sets of simulations can be executed in parallel, and each intervention action code can correspond to multiple rounds of simulation time sequence.

[0119] After obtaining the time series sequence of multiple simulations, quantitative processing is performed on it. The quantitative processing can statistically analyze, map, and summarize key state quantities according to a preset evaluation window, such as calculating equipment wear, yield decay, failure probability, maximum temperature deviation, cumulative energy consumption, equipment utilization rate, work-in-process inventory, order completion rate, and constraint violation count.

[0120] After executing step S103, the impact of process actions on future states in the intervention action coding can be explicitly expanded into a continuously analyzable simulation time series, and further structured quantitative indicators can be formed, thereby forming input parameters that can be directly used by industrial forecasting and decision-making models.

[0121] Step S104: Use the multi-round simulation time sequence and multi-dimensional quantitative simulation indicators as input parameters of the industrial prediction and decision-making model to obtain the target decision instructions output by the industrial prediction and decision-making model under the constraints of process compliance and optimization target weights.

[0122] Among them, process compliance constraints are used to limit the decision instructions output by the model to meet process specifications, equipment capabilities, quality requirements, safety boundaries, and resource constraints.

[0123] The optimization target weights are used to characterize the relative priority of each optimization target in the current business phase. For example, if the optimization targets include delivery time, yield, and energy consumption, with yield having the highest priority, delivery time second, and energy consumption the lowest priority, then yield can be assigned the highest weight, delivery time a smaller weight, and energy consumption the lowest weight.

[0124] The target decision instruction can be an equipment control parameter adjustment instruction, a process path switching instruction, a batch sorting instruction, a task issuance instruction, a resource reallocation instruction, or one or more of these instructions.

[0125] Industrial forecasting and decision-making models can simultaneously read multi-round simulation time series and multi-dimensional quantitative simulation indicators. Within a virtual sandbox, they repeatedly test and iterate, solving for candidate decision schemes that satisfy multiple optimization objectives under process compliance constraints. Finally, they determine the optimal decision scheme from among these candidates and construct the target decision instruction. The optimization objective weights can be dynamically configured according to the current production stage. For example, when delivery deadlines are tight, the weight of order completion rate is increased; when quality fluctuates, the weight of quality stability is increased; and when energy consumption is limited, the weight of energy consumption indicators is increased.

[0126] In summary, process-related text is first converted into template action codes via an industrial ontology. Standardized multimodal data is then uniformly mapped to a shared latent space by an industrial state representation model to form latent state features. These latent state features, along with intervention action codes, jointly drive the industrial dynamic simulation model to generate multi-round simulation time-series sequences, further yielding multi-dimensional quantitative simulation indicators. Subsequently, the industrial prediction and decision-making model outputs target decision instructions under process compliance constraints and optimization target weight limitations. This integrates process semantics, state representation, dynamic inference, and decision output into a single processing link. This approach addresses the problems in existing industrial systems, such as the difficulty of directly integrating process text into a unified inference link, the lack of consistent state representation in multimodal data, and the disconnect between simulation results and decision control. It enables each module to work collaboratively around a unified semantic input and a unified state representation, and through constrained simulation and decision output, forms executable decision instructions that avoid "industrial illusion." Furthermore, simulation and decision-making can be iteratively repeated, continuously outputting actionable decision instructions tailored to specific operating conditions.

[0127] In one possible embodiment, see Figure 2 This is a flowchart illustrating a standardized multimodal data acquisition method provided in an embodiment of this application, as shown below. Figure 2 As shown, the method may include the following steps S201 to S203.

[0128] Step S201: Obtain multimodal raw data from production line equipment.

[0129] The production line equipment may include, but is not limited to, PLCs, CNC equipment, AGVs, industrial 3D cameras, vibration sensors, temperature sensors, pressure sensors, etc.

[0130] Multimodal raw data can include one-dimensional sensor time-series data, two-dimensional image data, and three-dimensional point cloud data.

[0131] Multimodal raw data from the production line equipment can be accessed through an industrial internet interface to a data acquisition gateway. The data acquisition gateway performs timestamp alignment and format encapsulation on different modal data before outputting the raw data stream.

[0132] Step S202: Obtain text data from the industrial information system and perform text embedding on the text data to obtain text vectors.

[0133] Industrial information systems may include, but are not limited to, MES, ERP, PLM (Product Lifecycle Management), CRM (Customer Relationship Management), SCADA (Supervisory Control and Data Acquisition), and other information systems related to enterprise production.

[0134] Text data can include, but is not limited to, structured or unstructured text data such as work orders, process documents, equipment ledgers, quality inspection records, and production schedules.

[0135] Text data is imported from industrial information systems via message interfaces, database interfaces, or file interfaces, and then processed into text vectors through text segmentation, term normalization, and semantic encoding.

[0136] Step S203: Standardize and clean the original multimodal data and text vectors to obtain standardized multimodal data.

[0137] After obtaining the original multimodal data and text vectors, standardized cleaning can be performed on them. Cleaning may include missing value imputation, outlier removal, dimension normalization, noise suppression, and vector dimension alignment. The cleaning method can be configured according to business needs and is not limited here.

[0138] This measure enables the unified organization of multimodal raw data and text data into standardized multimodal data, reducing the expression bias between different modalities and minimizing the interference of noise, missing data, and scale differences on subsequent model inputs, thereby improving the stability of industrial state representation and the reliability of inputs for subsequent simulation decisions.

[0139] In one possible embodiment, see Figure 3 This is provided as an embodiment of the present application. Figure 1 A detailed flowchart of step S101 is shown below. Figure 3 As shown, the method may include the following steps S301 to S304.

[0140] Step S301: Perform semantic extraction and entity recognition on process-related text to obtain key business elements.

[0141] Natural language processing technology is used to identify entities in process-related texts. Based on entity recognition, semantic extraction is used to further extract the relationships between entities and their intentions, thereby obtaining key business elements such as equipment, process parameters, adjustment range, and execution time.

[0142] Step S302: Apply threshold constraints, standardization scaling, and numerical fitting to the values ​​in the key business elements to transform them into normalized values.

[0143] After obtaining the key business elements, numerical fields and their semantic boundaries can be extracted from these elements based on industrial knowledge graphs and industrial entities. For values ​​such as temperature, pressure, rotational speed, flow rate, time, and proportion, abnormal data are eliminated according to the process allowable thresholds, and then uniformly converted according to the corresponding dimensions. Subsequently, linear fitting, piecewise fitting, or interpolation fitting methods are used to complete missing values ​​or smooth abnormal fluctuations, transforming the values ​​into normalized values ​​and thus eliminating differences in parameter dimensions.

[0144] Step S303: Based on the industrial ontology, perform constraint mapping and format unification on the key business elements after numerical transformation to obtain standardized data.

[0145] Based on the unified standards defined in the product, process, and quality domains of the industrial ontology, constraint mapping is performed on the text elements in key business elements to unify the various names for the same object into a standard name; the normalized values ​​in key business elements are formatted uniformly, including precision, units, and encoding formats, and after completing constraint mapping and format unification, standardized data is obtained.

[0146] Step S304: Generate template action codes based on the standard data according to the preset coding rules.

[0147] According to the preset coding rules, various information such as the action subject, action type, parameter value and constraint label in the standardized data are mapped to coding fields in sequence to generate template action codes for subsequent model calls.

[0148] Through the above measures, the control intent and parameter constraints in process-related texts can be expressed in a structured way, numerical fields can maintain scale consistency and controllable boundaries, the semantic correspondence between normative data and industrial ontology is more stable, and the parsingability and consistency of template action coding are improved, thereby providing a unified input for the state representation, dynamic simulation and decision output of subsequent industrial world models.

[0149] In one possible embodiment, see Figure 4 This is provided as an embodiment of the present application. Figure 1 A detailed flowchart of step S102 is shown below. Figure 4 As shown, the method may include the following steps S401 to S405.

[0150] Step S401: Separate time-series data, visual data, and text vectors from the standardized multimodal data.

[0151] In this application, the industrial state representation model can be composed of a time-series branch, a visual branch, and a cross-modal fusion branch. Therefore, for standardized multimodal data, it is necessary to separate time-series data, visual data, and text vectors so that the separated data can be sent to the corresponding branches for processing.

[0152] Step S402: Extract steady-state baseline features and abrupt failure features from the time series data, and concatenate the steady-state baseline features and abrupt failure features into noise-free time series features.

[0153] Among them, steady-state baseline characteristics refer to the characteristic patterns of various indicators when equipment and production lines are operating normally.

[0154] Sudden failure characteristics refer to the abnormal change patterns of various indicators when equipment or production lines suddenly deviate drastically from their normal operating state.

[0155] By extracting long-term steady-state baseline characteristics of equipment or production lines using encoders, the distribution of normal operating conditions is learned, and stable signals under normal conditions are distinguished. Furthermore, the normal timing curve is reconstructed using a decoder, and the reconstruction error is calculated. If the error exceeds a threshold, it can be identified as noise and filtered out. For example, noise such as sensor zero-point drift and power frequency electromagnetic interference can be identified and filtered out.

[0156] In one possible embodiment, the task of extracting steady-state baseline features can be achieved by any of the following algorithms: industrial temporal VAE (Variational Autoencoder), MAE (Masked Autoencoders), and an industrial version of SimCLR (A Simple Framework for Contrastive Learning of Visual Representations).

[0157] By modeling the temporal hidden state cycle based on the encoder using time-series data, short-term sudden pulses and instantaneous fluctuations are captured, thereby extracting the features of fault mutations and abnormal offsets as mutation fault features.

[0158] In one possible embodiment, the task of extracting abrupt failure features can be performed using a RSSM (Recurrent State Space Models) time-series state encoder.

[0159] After obtaining the steady-state baseline features and abrupt fault features, the two can be spliced ​​together to obtain noise-free time series features with high discriminative power and the removal of time series noise.

[0160] In this application, in order to enable the algorithm for extracting steady-state baseline features to better handle time-series data in industrial scenarios, the algorithm can be adaptively adjusted and constrained.

[0161] The adjustments may include: First, for intermittent equipment start-up and shutdown conditions, a segmented sliding window is added to model the three time intervals of shutdown, no-load, and full-load separately, avoiding interference from start-up fluctuations with the steady-state baseline. Second, for long-period slow drift scenarios (e.g., bearing aging, sensor temperature drift), the algorithm dynamically updates the steady-state baseline features, allowing for monthly incremental adjustments rather than relying on a single generated steady-state baseline feature. Third, for workshops with strong electromagnetic interference (e.g., welding, stamping, variable frequency motors), a high-frequency noise penalty term is added to the algorithm's loss function to automatically suppress 50 / 60Hz power frequency interference.

[0162] The constraints may include: First, sampling frequency constraints: support for multiple rates from 1Hz to 10kHz with automatic resampling alignment, and a unified window length across device timing sequences. Second, real-time constraints: the encoding time for a single device timing sequence should not exceed a preset duration threshold (e.g., 5 milliseconds) to meet the real-time monitoring requirements of the production line. Third, boundary constraints: extreme operating conditions (material shortage, emergency shutdown) are not directly filtered but are separately marked and retained as a mutation feature to avoid missed faults.

[0163] Step S403: Encode the visual data, filter image noise, and obtain noise-free visual features.

[0164] Visual data can include image data and 3D point cloud data, such as graphic data captured by industrial cameras in industrial scenarios, point cloud data of workpieces captured by 3D laser scanners, and images of equipment appearance.

[0165] By employing discrete vector quantization encoding, the image pixel space of the visual data is compressed and reconstructed. During the reconstruction process, irrelevant pixel noise such as dust, fog, lighting fluctuations, and equipment oil occlusion is filtered out, while preserving the contour features of the target subject in the visual data. Furthermore, based on an occlusion prediction mechanism, secondary feature extraction is performed on the reconstructed visual data to extract key effective features such as workpiece size contours, surface defects, and equipment geometry, while weakening background clutter interference. Through these two stages of processing, image noise can be filtered out, resulting in noise-free visual features.

[0166] In one possible embodiment, visual data can be discretely vector quantized and encoded using an industrial VQ-VAE (Vector Quantized Variational Autoencoder), and then the reconstructed visual features can be extracted using a JEPA (Joint-Embedding Predictive Architecture) self-supervised visual encoder to obtain noise-free visual features.

[0167] In this application, in order to enable the algorithm for generating noise-free visual features to better process visual data in industrial scenarios, the algorithm can be adaptively adjusted and constrained.

[0168] The adjustments may include: First, to adapt to dynamic lighting in the workshop, illumination-enhanced pre-training samples can be introduced into the JEPA self-supervised visual encoder to construct multi-brightness and multi-angle features, thereby enhancing the robustness of the JEPA self-supervised visual encoder. Second, to address issues such as metal reflection and oil occlusion, a local occlusion code reconstruction branch can be added to the industrial VQ-VAE, ensuring that the core features of the target object are not lost in occluded areas. Third, for multi-scale workpieces, a multi-scale feature pyramid can be used to adapt millisecond-level micro-defects and synchronously identify large positive-size equipment.

[0169] The constraints can include: First, resolution constraints: compatibility with visual images of multiple resolutions (e.g., compatible with images captured by 2K industrial cameras and 3D sparse point clouds), and uniform output of noise-free visual features of fixed dimensions. Second, robustness constraints: when the occlusion rate of the visual data is no greater than a preset threshold (e.g., 40%), defect features must still be fully extracted without losing the detection target. Third, the algorithm scale can be adaptively adjusted according to its deployment location. For example, algorithms deployed on low-computing-power edge devices can be pruned, while algorithms deployed in the cloud can retain their complete architecture and parameters.

[0170] Step S404: Use the noise-free temporal features, noise-free visual features, and text vectors as input parameters to the cross-modal projection encoder that has undergone contrastive learning, and obtain multiple latent features output by the cross-modal projection encoder.

[0171] The cross-modal projection encoder can be trained within a contrastive learning framework by constructing positive and negative sample pairs (positive samples are data from the same device under the same operating conditions, and negative samples are data from different devices or under abnormal operating conditions). The cross-modal projection encoder can minimize the distance between multimodal features within the same scene and increase the distance between features from different scenes.

[0172] By using a cross-modal projection encoder to uniformly map noiseless temporal features, noiseless visual features, and text vectors, different modalities are aligned in the same latent space, thereby outputting multiple latent features.

[0173] Step S405: Perform adaptive weighted aggregation on multiple latent features to obtain latent state features.

[0174] When performing adaptive weighted aggregation on multiple latent features, weight coefficients can be generated based on the consistency, confidence, or modal completeness of each latent feature with the current industrial state. These weight coefficients are then used to perform a weighted summation or weighted concatenation of the latent features before compression and mapping to obtain the latent state features. After weighted aggregation, the fusion weights corresponding to each modality can be recorded. For example, when the time-series data is of high quality, the weight corresponding to the time-series data automatically increases; when the visual data is clear and can clearly reflect equipment defects, the weight corresponding to the visual data automatically increases; and when it is necessary to refer to the process manual for judgment, the weight corresponding to the text vector automatically increases.

[0175] In this application, adaptive adjustments and constraints can be set to achieve better adaptive weighted aggregation results.

[0176] The adjustments may include: First, modal missing data compatibility: when a modality's data is missing (e.g., a camera blackout causing visual feature loss, or a sensor disconnection causing time-series data loss), the weight of the missing modality can be adaptively reduced. Second, industry-specific adaptation: for example, in automotive equipment scenarios, the weight of vibration time-series data and component visual data can be strengthened; in chemical processes, the weight of time-series data and process text can be strengthened.

[0177] The constraints may include: First, consistency constraints: within the same production batch, the deviation in the feature space distribution of the fused time-series data and visual data is controlled within a preset deviation threshold. Second, interpretability constraints: during the fusion process, the weight records of each single-modal data are retained to facilitate root cause tracing of faults.

[0178] Through the above measures, time-series data, visual data, and textual semantics are uniformly compressed into latent state features, enabling the industrial state representation model to output a more stable state representation. These latent state features are more robust to field noise and can provide continuous, aligned input representations for subsequent industrial dynamic simulation models, thereby improving the consistency of state inference and the reliability of the decision-making chain.

[0179] In one possible embodiment, the latent state characteristics include device-level latent state characteristics and production line-level latent state characteristics. Intervention action coding includes device-level intervention action coding and production line-level intervention action coding.

[0180] Among them, device-level latent state features are the latent state features of a single device, which can be constructed from the time-series data, visual data, and text vectors corresponding to a single device.

[0181] Production line-level latent state characteristics are the global latent state characteristics of the production line. They can be a combination of the device-level latent state characteristics of all devices globally, or they can be obtained by fusing the device-level latent state characteristics of all devices globally.

[0182] Equipment-level intervention action coding is intervention action coding at the equipment level, used to control operations such as parameter fine-tuning, maintenance, and shutdown of a single device.

[0183] Production line-level intervention action codes are global intervention action codes used to control operations such as batch scheduling adjustments, multi-AGV scheduling, and overall line process modifications.

[0184] See Figure 5 This is provided as an embodiment of the present application. Figure 1 A detailed flowchart of step S103 is shown below. Figure 5 As shown, when executing step S103, steps S501 to S503 can be executed.

[0185] Step S501: Use the device-level potential state features and device-level intervention action encoding as input parameters of the stochastic state-space model to obtain long-term time-series prediction features of multiple devices.

[0186] In this application, the industrial dynamic simulation model may include a stochastic state-space model, a temporal Transformer, and a physical information neural network; the stochastic state-space model, the temporal Transformer, and the physical information neural network are three branches of the industrial dynamic simulation model.

[0187] Among them, the stochastic state-space model is used to simulate and predict the operating state of a single device.

[0188] The Timing Transformer is used to simulate and predict the overall operational status of a production line. It supports long-term simulations of the entire production line's workflow, ranging from several days to several months, addressing the problem of error accumulation due to long-distance timing dependencies.

[0189] Physical information neural networks (PINs) are used to constrain the outputs of the first two branches using physical information. PSNs embed industrial physical equations into the model loss function, preventing simulations from generating false operating conditions that contradict common sense about production.

[0190] After obtaining the latent state features and intervention action codes, the device-level latent state features from the latent state features and the device-level intervention action codes from the intervention action codes can be input into the stochastic state space model, so that the stochastic state space model can perform simulation prediction for the operating state of each device, output the long-term prediction features of each device, and obtain the long-term prediction features of multiple devices.

[0191] In this process, a recursive structure can be used to separate steady-state baseline features and abrupt failure features from the equipment-level potential state characteristics, separating the long-term deterministic wear degradation trend of the equipment from the instantaneous random fluctuations caused by load disturbances. Furthermore, Gaussian random variables can be introduced to characterize uncontrollable random disturbances in the workshop, such as raw material tolerances, voltage fluctuations, and environmental temperature changes. Then, based on the historical degradation trend of the equipment and current random disturbances, the continuous change curves of the equipment in terms of vibration, temperature, and damage can be predicted, thereby outputting long-term time-series predictive features.

[0192] In one possible embodiment, RSSM (Recurrent State Space Model) can be used as an algorithm to generate long-term prediction features of devices in a stochastic state space model.

[0193] In this application, in order to enable the algorithm of the stochastic state-space model to better handle the potential state characteristics and intervention action encoding of industrial scenarios, the algorithm can be adaptively adjusted and constrained.

[0194] The adjustments may include: First, long-cycle degradation adaptation: For scenarios involving slow damage to bearings, motors, and cutting tools, add segmented sliding windows (the window duration can be configured as needed, such as days, weeks, or months), and model separately for three operating conditions: no-load, full-load, and intermittent shutdown. Second, compatibility with abrupt events: Automatically reset the long-term wear baseline and recalculate the initial degradation value after equipment overhaul or parts replacement. Third, when multiple devices are running in parallel on the production line, allocate an independent RSSM hidden state channel for each device to avoid interference between the degradation trends of different devices.

[0195] The constraints can include: First, numerical constraints: the wear trend must be monotonically non-negative; in other words, a result that violates industrial common sense—a decrease in wear value over time—is not allowed. Second, real-time inference constraints: the inference time for short-term inferences of a single device must be limited to a preset time threshold, which can be configured as needed; for example, the time threshold could be 10 milliseconds. Third, training adjustments: during training, the algorithm for the stochastic state-space model uses piecewise weighted loss, increasing the loss weight of late-stage degradation samples to prevent the algorithm from biasing towards normal stable operating conditions and underestimating aging faults; time-series data augmentation (e.g., time-series stretching, small-amplitude noise superposition) is introduced into small sample device fault data to supplement the training dataset.

[0196] Step S502: Use the production line-level potential state features and production line-level intervention action encoding as input parameters of the temporal Transformer to obtain the global temporal features of the entire production line.

[0197] The potential state features and intervention actions at the production line level are encoded and input into the temporal Transformer, which then performs a long-term extrapolation of the entire production line and outputs the global temporal features of the entire production line.

[0198] This process involves multi-scale temporal encoding, adding corresponding positional codes to the data at each time level in the potential state features of the production line. These positional codes can be minute-level, day-level, or month-level. By adding positional codes, the sequence of processes and the temporal relationship of material flow can be marked, assisting the temporal Transformer in completing simulation predictions. The temporal Transformer can implement multi-head self-attention computation to capture long-distance dependencies across processes and equipment. For example, fluctuations in upstream material input affect downstream assembly, and night shift capacity constraints affect the next day's production scheduling. During long-cycle temporal decoding, a global production line flow simulation sequence spanning several days to several months can be output, mitigating the accumulation of long-term errors in traditional cyclic models. After feature compression and extraction of the global production line flow simulation sequence, global temporal features of the entire production line can be output.

[0199] In this application, to enable the temporal Transformer to better handle the potential state features and intervention action encoding in industrial scenarios, the algorithm can be adaptively adjusted and constrained.

[0200] The adjustments may include: First, discrete process adaptation: for discrete manufacturing scenarios (such as automobile manufacturing, machinery equipment, etc.), adding process masks to shield attention associations for irrelevant processes that do not involve material flow. Second, continuous process chemical scenario optimization: lengthening the attention window to adapt to 24 / 7 uninterrupted long-cycle simulations of continuous chemical plants. Third, dynamic batch adaptation: dynamically adjusting the length of the input sequence when the production scheduling vertices increase or decrease, supporting simulations for variable order sizes.

[0201] The constraints may include: First, temporal causality constraints: only historical data is allowed to participate in prediction, and leakage of future time-series information is prohibited (strict causal masking). Second, periodic accuracy constraints: the cumulative error of monthly projected production capacity is controlled within a specified deviation range, which can be set as needed. For example, the specified deviation range can be within ±3%. Third, training adjustment: a phased pre-training process is adopted, where existing short-term time-series samples converge to the basic attention weights, and then monthly long-series data is loaded to fine-tune the attention weights. Temporal denoising pre-training is introduced to filter out abnormal outlier time-series samples such as holiday shutdowns and temporary equipment maintenance.

[0202] Step S503: Input the long-term prediction features and global time-series features of multiple devices into the physical information neural network to obtain the multi-round simulation time-series sequence output by the physical information neural network.

[0203] After obtaining long-term time-series prediction features and global time-series features from multiple devices, these features can be input into a physical information neural network (PEN). This PSN can load relevant industry-specific physical priors (e.g., mechanical stress equations, thermodynamic temperature rise equations, fluid pressure balance, chemical reaction kinetics equations, etc.) to correct spurious simulation values ​​in the long-term and global time-series features that do not conform to mechanical, thermal, and fluid dynamics principles. This results in a corrected global production line flow simulation sequence that lasts from several days to several months. A copy of this sequence can be stored for virtual sandbox visualization and backtracking verification of simulation results. Temporal pooling aggregation of the corrected global production line flow simulation sequence yields a multi-round simulation time-series sequence that has been validated and corrected according to physical rules.

[0204] In this application, step S503 can be performed using PINN (Physics-Informed Neural Networks).

[0205] In this application, in order to enable the physical information neural network to better handle long-term time-series prediction features and global time-series features in industrial scenarios, the algorithm can be adaptively adjusted and constrained.

[0206] The adjustments may include: First, setting up physical prior libraries for different industries, which can be switched as needed. For example, loading stress and wear mechanics equations for mechanical equipment scenarios; loading fluid and reaction thermodynamics equations for chemical engineering scenarios; and loading simplified physical models for home appliance assembly scenarios to reduce computational power consumption. Second, multi-physics coupling adaptation. For example, scenarios such as heat treatment and die casting can simultaneously embed two sets of equations for heat conduction and mechanical deformation. Third, dynamic updating of boundary conditions. For example, automatically updating the boundary parameters of physical equations when workshop ambient temperature, cooling water pressure, and raw material composition change.

[0207] The constraints may include: First, physical residual constraints. The residuals from the simulation results substituted into the physical equations must be lower than a preset residual threshold; if the threshold is exceeded, the simulation must be recalculated. Second, parameter boundary constraints. Physical quantities such as temperature, pressure, rotational speed, and stress must not exceed the rated upper and lower limits of the equipment.

[0208] Furthermore, if the industrial dynamic simulation model uses RSSM, Temporal Transformer, and PINN as three branches, a two-stage training strategy can be adopted. The first stage trains RSSM and Temporal Transformer using only real industrial data; the second stage freezes the model parameters of these two models and optimizes the physical calculation weights of PINN. In industrial scenarios with limited sample sizes, the proportion of physical loss can be increased, relying on physical priors to compensate for the lack of real-world operating data.

[0209] Through the above measures, equipment-level information and production line-level information can be entered into the corresponding modeling branches respectively, reducing the representation bias caused by the mixing of information at different scales. After the long-term time series prediction features and global time series features are fused by physical constraints, the output results have a higher degree of matching with the actual industrial operation mechanism, thereby improving the continuity, stability and reliability of the multi-round simulation time series sequence for subsequent decision-making.

[0210] In one possible embodiment, see Figure 6 This is a flowchart illustrating a method for generating a process semantic analysis report according to an embodiment of this application. Figure 6 As shown, the method may include the following steps S601 to S604.

[0211] Step S601: Perform numerical analysis on the multi-dimensional quantitative simulation indicators to obtain structured numerical features.

[0212] After obtaining multi-dimensional quantitative simulation indicators, these indicators can be segmented and grouped. Through clustering, outlier location, feature filtering, and correlation quantification, numerical analysis can be performed to obtain structured numerical features.

[0213] Step S602: Call the industrial knowledge graph to perform causal reasoning and natural language semantic generation on multi-dimensional quantitative simulation indicators carrying structured numerical features, and obtain preliminary attribution conclusions of the working conditions.

[0214] After obtaining structured numerical features, industrial knowledge graphs can be used to match the relationships between key business elements such as equipment, processes, and faults in multi-dimensional quantitative simulation indicators. This allows for causal reasoning and natural language semantic generation, leading to preliminary attribution conclusions about operating conditions. For example, explanations of the correlation mechanisms for abnormal temperature rise, cycle time fluctuations, or mass deviations can be provided.

[0215] Step S603: Based on the process entity, revise the preliminary attribution conclusions of the operating conditions to obtain the process semantic analysis report.

[0216] Among them, the process ontology is a part of the domain ontology in the industrial ontology. It is an explicit and formalized specification of the core probability and related knowledge of "process" in a specific industrial domain. Through structured modeling, it clearly defines the core elements such as process route, process definition, process parameters and quality standards and their internal relationships.

[0217] After obtaining the preliminary attribution conclusions of the operating conditions, the terminology, hierarchy, and causal chain in the preliminary attribution conclusions can be corrected through the industrial ontology, eliminating expressions that are inconsistent with process rules and generation laws, and supplementing constraints that conform to the process definition, thereby generating a process semantic analysis report.

[0218] In one possible embodiment, steps S601 to S603 can be completed by the industrial big model with the help of industrial knowledge graphs and process ontology.

[0219] By employing the above measures, semantic analysis and root cause reasoning of multi-dimensional quantitative simulation indicators can be performed using industrial knowledge graphs and process ontology. The final results are then output as report texts that are easy for process engineers to understand and verify. This allows multi-round simulation results to move beyond numerical displays and form analytical conclusions with process semantic constraints, thereby improving the interpretability of the simulation results.

[0220] In one possible embodiment, the multi-round simulation time series includes corrected long-term time series prediction features for multiple devices and multi-round corrected global time series features. The corrected long-term time series prediction features for the devices may include simulation trajectories of the device's temperature, vibration, pressure, and load.

[0221] See Figure 7 This is provided as an embodiment of the present application. Figure 1 A detailed flowchart of step S104 is shown below. Figure 7 As shown, when executing step S104, steps S701 to S705 can be executed.

[0222] Step S701: Use the corrected long-term prediction features and multi-dimensional quantitative simulation indicators of multiple devices as input parameters of the industrial MPC to obtain the candidate control parameters output by the industrial MPC for each device under process compliance constraints and optimization target weight limitations.

[0223] After obtaining the simulation time series from multiple rounds, the corrected long-term time-series prediction features and multi-dimensional quantitative simulation indicators of multiple devices can be input into the industrial MPC (Model Predictive Control). Under the constraints of process compliance and optimization target weights, the industrial MPC uses general real-time optimization algorithms for industrial processes and discrete equipment to solve for parameters that can be directly executed by the PLC, such as temperature, pressure, speed, and robot position, in the rolling time domain, thereby obtaining the candidate control parameters for each device.

[0224] To enable industrial MPC to better calculate candidate control parameters for each device, corresponding adjustments and constraints can be set.

[0225] The adjustments may include: First, for discrete manufacturing industries (such as machining and home appliance assembly): narrowing the rolling time domain to adapt to frequent start-stop and part-changing conditions; adding point-to-point smoothing constraints to avoid robot vibration impacting equipment. Second, disturbance adaptation: for workshop voltage fluctuations and raw material batch differences, introducing random disturbance terms from the sand table simulation into the MPC prediction model to compensate for disturbances in advance.

[0226] The constraints can include: First, hard boundary constraints: Temperature, pressure, speed, etc., are strictly limited within the rated range specified on the equipment nameplate. Inequality constraints are directly added during the solution process to ensure that the final output control parameters are within a safe range. Second, rate of change constraints: Limits are set on the single adjustment range of temperature, pressure, etc., to prevent sudden increases or decreases from damaging the equipment. Third, real-time delay constraints: The time taken for a single round of MPC solution is limited to not exceeding a preset solution duration threshold to meet the real-time closed-loop requirements of the operating conditions. Here, the solution duration threshold can be configured according to business needs. For example, the solution duration threshold is 200 milliseconds.

[0227] Industrial MPC has no offline training process; it uses online rolling solutions. Industrial MPC relies on a virtual sandbox to simulate various disturbance conditions in batches for parameter pre-tuning, avoiding repeated debugging after deployment. Because the control framework of MPC is universal, new production lines only need to import equipment physical parameter templates (such as volume, heat transfer coefficient, rated boundaries, etc.), and the framework automatically generates mechanism models and constraints, allowing for rapid reuse.

[0228] Step S702: Use the global time series features and multi-dimensional quantized simulation indicators after multiple rounds of correction as input parameters for the deep reinforcement learning algorithm to obtain the medium- and long-term candidate scheduling schemes output by the deep reinforcement learning algorithm under the constraints of process compliance constraints, optimization target weights, and candidate control parameters output by the industrial MPC for each device.

[0229] The global temporal features and multi-dimensional quantified simulation indicators, refined through multiple rounds of adjustments, are input into a deep reinforcement learning algorithm. Using multi-objective solution kernels such as weighted summation multi-objective, Pareto NSGA-II, and linear weighted optimization, medium- to long-term candidate scheduling schemes are calculated. During the calculation process, candidate control parameters output by the industrial MPC for each device can be used as the basic environmental action space. Medium- to long-term candidate scheduling schemes can include discrete action codes such as work order scheduling, AGV delivery, monthly maintenance, and process modification adjustments.

[0230] In one possible implementation, the deep reinforcement learning algorithm can employ any of the following: Dreamer (based on a world model), TD-MPC (Temporal Difference Model Predictive Control) (based on a world model), Proximal Policy Optimization (PPO) (an industrially adapted version), or Soft Actor-Critic (SAC) (an industrially adapted version). If an industrially adapted version of PPO or SAC is used, the multi-objective weighted reward function and hierarchical decision output mechanism must be retained.

[0231] To enable deep reinforcement learning algorithms to better compute medium- and long-term candidate scheduling schemes, appropriate adjustments and constraints can be set.

[0232] The adjustments may include: First, for equipment manufacturing scenarios with multiple product types and small batches: expanding the scope of operations and increasing cost-return items for work order switching and mold change. Second, for continuous production scenarios in process chemicals: reducing the weight of production scheduling switching and prioritizing optimization of long-term equipment maintenance and energy consumption balance. Third, for scenarios with a fault preference and multiple objectives: increasing the penalty weight for failure losses on older production lines, while focusing on capacity and manufacturing cost optimization for new production lines.

[0233] Constraints can include: First, resource capacity constraints. For example, the number of AGVs, the parallel capacity of workstations, and the upper limit of warehouse inventory cannot be exceeded. Second, temporal and causal constraints. For example, the order of work processes, the lower limit of maintenance intervals must be locked, and logically reversed solutions should not be generated. Third, safety constraints. For example, high-risk equipment must have a minimum maintenance cycle, and the strategy does not allow operation beyond the maintenance period.

[0234] Deep reinforcement learning algorithms can be pre-trained using virtual simulation. This stage of training is entirely completed within the virtual simulation, without occupying the real production line. The simulation generates massive amounts of fault, production change, and energy consumption sample data, solving the long-term problem of sample scarcity in real production lines. Training can be conducted in stages, first converging the basic scheduling strategy in the short term, and then expanding to monthly long-term fine-tuning.

[0235] Step S703: Based on the candidate control parameters of each device and the medium- and long-term candidate scheduling schemes corresponding to each set of candidate control parameters, generate multiple candidate decision schemes.

[0236] After obtaining the candidate control parameters for each piece of equipment and the medium- and long-term candidate scheduling schemes corresponding to each set of candidate control parameters, the candidate control parameters determined by each piece of equipment in the production line in a single decision, and the medium- and long-term candidate scheduling schemes corresponding to the candidate control parameters of all pieces of equipment in the production line in this current decision, can be combined into a single candidate decision scheme, thereby obtaining multiple candidate decision schemes.

[0237] Step S704: Filter multiple candidate decision schemes based on compliance filtering strategies.

[0238] Among them, the compliance filtering strategy is a strategy to verify the equipment threshold constraints (such as temperature, pressure, speed, load, etc.), basic physical constraints (such as mechanical and thermodynamic limits, etc.), and business rules (such as process sequence, material completeness, safe operation procedures, etc.) in the candidate decision scheme.

[0239] Compliance filtering strategies can filter out candidate decision-making schemes that do not meet the requirements in terms of device threshold constraints, physical constraints, or business rules.

[0240] Step S705: Generate the target decision instruction based on the remaining candidate decision schemes after filtering.

[0241] After obtaining the filtered candidate decision schemes, a target decision instruction can be generated based on the remaining candidate decision schemes. When at least two candidate decision schemes remain, a target decision instruction can be generated for each candidate decision scheme separately. Alternatively, the candidate decision scheme that performs better on the optimization objective with the highest optimization objective weight can be selected to generate the target decision instruction.

[0242] Through the above measures, the outputs of industrial MPC and deep reinforcement learning can be mutually verified and jointly constrained, thereby ensuring that the final target decision instruction simultaneously meets the requirements of local execution feasibility, global task weight, and process compliance. By adding candidate decision scheme combination and compliance filtering steps before output, the probability of inconsistent instructions being directly issued is reduced, and the stability and executability of decision results in complex industrial scenarios are improved.

[0243] In one possible embodiment, when performing step S704, multiple candidate decision schemes can be filtered based on basic physical constraints. Here, the basic physical constraints can be pre-configured based on experience. For example, the basic physical constraints can filter out schemes with excessive temperature, overload, or high safety risks.

[0244] Furthermore, the remaining candidate decision schemes after the initial filtering can be filtered a second time based on process standards, safety production specifications, and industry ontology rules. This allows for matching of the action semantics, parameter combinations, and execution order in the schemes. If there are problems such as incorrect connection with the process section, concurrent dangerous actions, missing protection conditions, or conflicts in ontology relationships, the scheme is deemed invalid.

[0245] In one possible embodiment, the remaining candidate decision schemes after initial filtering can be handed over to the industrial big data model, which will then perform secondary filtering based on process standards, safety production specifications, and industry ontology rules.

[0246] After the above two layers of filtering, the remaining candidate decision schemes all meet the requirements of physical executability, process consistency, and safety compliance.

[0247] This compliance filtering mechanism first excludes physically unfeasible solutions, and then eliminates non-compliant solutions at the process and safety levels, so that the final output decision can be executed normally and avoid safety risks, thereby improving the reliability and on-site adaptability of the target decision instructions.

[0248] In one possible embodiment, when generating the target decision instruction based on the remaining candidate decision schemes after filtering, the remaining candidate decision schemes can be split according to the execution level to obtain decision instructions for each execution level. Here, the execution level can be divided according to the business needs of the industrial scenario.

[0249] For example, the execution hierarchy can be divided into bottom, middle, and top layers. The decision instructions extracted from the candidate decision schemes can be bottom-level real-time control instructions, middle-level scheduling instructions, and top-level medium- and long-term planning instructions. Among them, bottom-level real-time control instructions can include hardware-directly executed parameters such as equipment speed, heating temperature, stamping pressure, and robot positioning; middle-level scheduling instructions can include instructions such as work order switching, AGV material delivery, and assistance with shift adjustments; and top-level medium- and long-term planning instructions can include monthly maintenance plans, production line process optimization schemes, and equipment replacement strategies.

[0250] Furthermore, target decision instructions can be constructed based on decision instructions from multiple execution levels. Decision instructions from multiple execution levels can be encoded in a unified format, including constraints, priority labels, and other information, thereby constructing the target decision instructions. For example, for lower-level real-time control instructions, constraints may include limits on parameter variation to prevent drastic adjustments; and alignment of instruction update cycles with industrial control scanning cycles. For mid-level scheduling instructions, constraints may include upper limits on workstation capacity and shift personnel configuration constraints. For upper-level medium- and long-term planning, constraints may include compliance with dual energy consumption control, safety production regulations, and annual order delivery targets.

[0251] Through the above measures, the filtered candidate decision schemes can be decomposed into instructions adapted to different execution objects, and then uniformly constructed into target decision instructions that can be directly issued. This enables equipment control, production line scheduling, and task execution to be implemented collaboratively within the same decision framework. Because the instructions at each level have been hierarchically decomposed and consistently organized before construction, the final output target decision instructions maintain a correspondence with the original candidate decision schemes and are easier for the control end to parse and execute.

[0252] In one possible embodiment, after obtaining the target decision instruction, the target decision instruction can be handed over to the intelligent agent, which then distributes the decision instructions of each execution level in the target decision instruction to the execution object.

[0253] After the target decision instructions are handed over to the agent, the agent extracts the decision instructions for each execution level and combines them with the address identifier of the execution object, device status and current task queue to generate a set of instructions to be issued, thereby completing the instruction issuance.

[0254] In this way, the target decision instructions can be parsed and distributed hierarchically in a unified intelligent agent, so that the simulation decision results can be smoothly mapped from the upper-level analysis output to the field execution objects, forming a closed-loop control basis, thereby enabling decision instructions at multiple execution levels to be executed consistently in the industrial field.

[0255] In one possible embodiment, after the intelligent agent issues a target decision instruction, it can obtain real production line operating condition data within a specified time period after executing the target decision instruction. The specified time period can be configured based on experience. For example, the specified time period can be several days, several weeks, or even several months, and it can be related to the content of the target decision instruction.

[0256] Furthermore, an online residual adaptive mechanism can be triggered based on real production line operating data to adjust the model parameters of the industrial dynamic simulation model.

[0257] In practical implementation, after the target decision instruction is executed, data such as production line equipment status, process results, and environmental disturbance data can be continuously collected according to a specified time period matched with the process cycle. The collected data is then time-aligned, outlier-removed, and normalized to form real production line operating condition data. Subsequently, this real production line operating condition data is compared item by item with the simulation prediction values ​​output by the industrial dynamic simulation model within the same time window to obtain the residual vector. Based on the residual amplitude, trend of change, and process sensitivity, the parameter correction amount is determined, and then the model parameters are iteratively adjusted in small steps using an online update algorithm.

[0258] Through the above measures, after the target decision-making instructions are implemented, real production line operating condition data is introduced back into the system, and residual feedback drives the online correction of model parameters. This enables the industrial dynamic simulation model to be continuously updated as equipment ages, operating conditions fluctuate, and process disturbances occur, thereby continuously reducing the virtual-real deviation in the simulation process and dynamically adapting to problems such as equipment wear and raw material batch distribution drift.

[0259] In one possible embodiment, when performing step S102, standardized multimodal data can be used as input parameters for the industrial state representation model to obtain the latent state features output by the industrial state representation model under the guidance of semantic attention-guided features.

[0260] The semantic attention-guided features are generated by the industrial big data model. After obtaining standardized multimodal data in step S101, the standardized multimodal data can be input into the industrial big data model. The model's attention mechanism is used to identify key information, and the identification result is used as the semantic attention-guided features. The semantic attention-guided features can take the form of a one-dimensional vector, a two-dimensional weight graph, or a multi-head attention weight set, etc.

[0261] This semantic attention-guided feature can be used as an attention gating parameter input into the industrial state representation model. It weights the channel responses of different modalities, giving higher response weights to features related to process constraints, anomalies, and state transitions, while suppressing features less relevant to the scenario. Guided and constrained by the semantic attention-guided feature, the industrial state representation model focuses on key information, thereby generating latent state features.

[0262] After implementing the above measures, the industrial state characterization model can be constrained by the semantic attention-guided features generated by the large industrial model when extracting potential state features. This makes the output features more consistent with the process semantics and the field state, reduces the interference of irrelevant noise on the characterization results, and provides more stable state input for subsequent simulation and deduction.

[0263] In one possible embodiment, the latent state features may include standard latent state features and lightweight latent state features. The standard latent state features are uncompressed latent state features. For example, standard latent state features can be generated through the aforementioned steps S401 to S405. After obtaining the standard latent state features, they can be compressed using sparse quantization VAEs, compressing the high-dimensional standard latent state features into low-dimensional lightweight latent state features. In this process, sparse encoding, numerical quantization, and removal of redundant and invalid dimensions can be performed on the standard latent state features, thereby compressing the thousands of dimensions of standard latent state features into lightweight latent state features of less than 100 dimensions. This lightweight latent state feature is suitable for business scenarios such as large-scale cloud simulation and real-time edge inference.

[0264] In this application, after the industrial state representation model outputs latent state features, the lightweight latent state features can be handed over to the large industrial model to complete the cross-modal semantic alignment of standardized multimodal data.

[0265] Through the above measures, after each time the industrial big data model provides semantic attention-guided features with standardized multimodal data, it can obtain lightweight latent state features corresponding to the standardized multimodal data. Through these lightweight latent state features, cross-modal semantic alignment is completed, thereby correcting its own attention mechanism so as to generate semantic attention-guided features that are more conducive to generating latent state features in the future.

[0266] In one possible embodiment, before executing step S103, it is necessary to obtain the intervention action code, which can determine whether the industrial forecasting and decision-making model outputs the target decision instruction.

[0267] On the one hand, if not, it means that there is currently no intervention action for the simulation system, and an intervention action code with no intervention action can be generated. In this case, when executing step S103, simulation deduction can be performed solely based on the potential state characteristics.

[0268] On the other hand, if so, it indicates that there are already intervention actions in the simulation system. The latest target decision instructions output by the industrial prediction and decision-making model can be converted into intervention action codes based on the industrial ontology and template action codes. In this case, when executing step S103, simulation deduction can be performed based on the potential state characteristics and intervention action codes, so that subsequent simulation outputs can reflect the latest control intent.

[0269] This measure ensures that the intervention action codes input to the industrial dynamic simulation model and the target decision commands output by the industrial prediction and decision-making model maintain a common source mapping relationship, reducing manual transcription and semantic bias. By adopting this approach, the model can maintain stable deduction even when target decision commands are missing, and promptly drive action updates after target decision commands are generated. This makes the connection between industrial dynamic simulation results and decision results more consistent, and improves the continuity and semantic alignment of the simulation chain.

[0270] In one possible embodiment, in the case of manual intervention for business needs, an intervention action code is generated based on the intervention result.

[0271] In this solution, in response to production demand adjustment instructions, candidate adjustment schemes can be generated based on knowledge graphs, and the candidate adjustment schemes can be converted into intervention action codes according to industrial ontology and template action codes.

[0272] In practice, production demand adjustment instructions can be triggered by order changes, capacity gaps, equipment malfunctions, or quality deviations. After receiving the production demand adjustment instructions, information such as the target production line, adjustment period, and constraints can be parsed. Then, matching process rules and equipment linkage relationships can be retrieved from the industrial knowledge graph to generate several candidate adjustment schemes.

[0273] Furthermore, each candidate adjustment scheme is encoded as an intervention action code according to the coding rules of the industrial body and template action coding instructions.

[0274] In one possible embodiment, production demand adjustment instructions can be processed through an industrial big model, candidate adjustment schemes can be generated based on a knowledge graph, and the candidate adjustment schemes can be converted into intervention action codes based on industrial ontology and template action codes.

[0275] Through the above measures, the industrial knowledge graph undertakes the semantic reasoning function from production demand to candidate solutions, while the industrial ontology and template action coding undertake the standardized conversion function from candidate solutions to intervention action coding. This enables production demand adjustment instructions to be converted into uniformly expressed intervention action information, directly serving subsequent simulation deduction. Since the candidate adjustment solutions converted from production demand adjustment instructions are input as intervention action codes into the industrial dynamic simulation model for counterfactual simulation, the feasibility of the candidate adjustment solutions can be verified, avoiding the adverse effects on the production line caused by directly issuing decision instructions containing "industrial illusions".

[0276] In one possible embodiment, after the industrial dynamic simulation model outputs a multi-round simulation time sequence, a three-dimensional simulation image of the industrial scene can be rendered and generated based on the noise-free visual features in the multi-round simulation time sequence and the potential state features, and the three-dimensional simulation image can be output and displayed.

[0277] The noise-free visual features can include the geometric features of equipment and workpieces on the production line, as well as the spatial location information of the workshop. Based on these geometric features, image rendering can be performed. The multi-round simulation time sequence can include information such as AGV scheduling time sequence and process flow time sequence. Through the multi-round simulation time sequence, the changes in the 3D simulation image as the simulation process is determined, thereby generating a continuous 3D simulation image. The time sequence values ​​can affect the rendering logic; for example, an increase in equipment temperature will change the equipment color, increased workpiece wear will show the expansion of workpiece defects, and material accumulation will update the distribution of work-in-process in the workshop.

[0278] When outputting a 3D simulation screen, the corresponding numerical indicators can be displayed simultaneously. These numerical indicators can come directly from multi-dimensional quantitative simulation indicators or be the result of further processing of multi-dimensional quantitative simulation indicators, so that managers can observe the production line status and simulation evolution results.

[0279] Adaptable adjustments can be made based on the needs of industrial scenarios. For example, if the rendering process is performed on an edge device, the rendering precision can be reduced to display only the dynamic information of key equipment. If the rendering process is performed in the cloud, a complete high-precision global scene can be rendered to meet the needs of process review. For the welding, machining, and semiconductor industries, specific defect rendering rules are applied to intuitively present the entire evolution process of scratches, deformations, and porosity throughout the generation cycle. A time-synchronization alignment mechanism ensures that the numerical indicators displayed in the output are synchronized with the 3D simulation image in time.

[0280] By combining simulation time-series information with the visually cleaned potential state for rendering, abstract time-series numerical indicators can be transformed into intuitive and visual dynamic images of workshops, production lines, and equipment. This allows managers to see the changes in equipment status, material distribution, and defects resulting from intervention actions without needing to interpret complex feature vectors. Furthermore, directly displaying 3D simulation images allows production operation platforms and process simulation applications to intuitively showcase the simulation results, lowering the barrier to entry for industrial simulation.

[0281] See Figure 8 This is a schematic diagram of the architecture of the industrial AI system provided in this application, such as... Figure 8 As shown, the industrial AI system is divided into four layers from bottom to top: equipment and resources layer, data and knowledge layer, world model layer, product and intelligent agent layer, and application scenario layer.

[0282] The equipment and resource layer includes a hardware access module, an IT system integration module, and a computing power scheduling module. The hardware access module interfaces with production line hardware such as PLCs, CNC equipment, AGVs, industrial 3D cameras, vibration sensors, temperature sensors, and pressure sensors, enabling real-time acquisition of multimodal raw data, including one-dimensional sensor time-series data, two-dimensional image data, and three-dimensional point cloud data. The IT system integration module can connect to industrial information systems such as ERP, MES, PLM, CRM, and SCADA, extracting structured or unstructured text data from these systems, including work orders, process documents, equipment ledgers, quality inspection records, and production schedules. The computing power scheduling module centrally manages multi-brand computing servers, allocating computing resources for tasks such as model pre-training and fine-tuning, real-time simulation, and edge inference.

[0283] Data acquired by the device and resource layers can be pushed up to the data and knowledge layers after standardization and cleaning via a data upload interface. Additionally, the device controls the execution of decision-making commands issued by the product and intelligent agent layers.

[0284] The data and knowledge layer serves as the input source of shared data and symbolic knowledge for all models within the world model layer. It can include industrial knowledge graphs, industrial AI datasets, industrial ontology models, industrial big data, and industrial data spaces. Industrial AI datasets can encompass specialized labeled datasets for various sub-sectors such as automotive, home appliances, machinery and equipment, semiconductors, and process chemicals. Data within these datasets can undergo processing such as multi-source data alignment, data augmentation, missing value completion, and industrial noise filtering. Industrial ontology models, including domain and business ontology, serve as a unified semantic bridge for interaction between models and agents within the industrial AI system. The industrial knowledge graph, based on the industrial ontology framework, accumulates millions of entity-related knowledge entries, including process standards, fault cases, equipment maintenance specifications, safety regulations, and historical optimization cases. Coupled with the RAG search engine, it provides real-time process knowledge retrieval support for the algorithm models within the system.

[0285] The world model layer includes the industrial world model and the large industrial model, equipped with a digital twin integrated engine to build a high-fidelity industrial virtual sandbox. It features built-in standardized bidirectional interactive interfaces that connect to the data and knowledge layer, and the product and intelligent agent layer, enabling the entire process of multimodal data representation, virtual working condition simulation, multi-objective decision optimization, and knowledge simulation compliance verification. The processing logic of each model in the world model layer is detailed in the previous descriptions and will not be repeated here.

[0286] The Product and Intelligent Agent Layer is a platform built upon the Industrial Claw Sandbox kernel and Industrial Intelligent Agent Engine, encompassing the entire process of intelligent agent development, orchestration, and scheduling. It serves as the bridge between simulation decision-making schemes and the underlying production line hardware. The intelligent agent cluster within the Product and Intelligent Agent Layer can include functional intelligent agents (e.g., indicator generation intelligent agents, data governance intelligent agents, decision analysis intelligent agents, equipment intelligent agents, quality intelligent agents, production scheduling intelligent agents, energy and carbon intelligent agents, virtual power plant intelligent agents, etc.) and digital human decision-making intelligent agents (e.g., plant manager digital human, CFO digital human, production supervisor digital human, energy station manager digital human, etc.). The Product and Intelligent Agent Layer possesses verification and scheduling capabilities, enabling accurate recognition of user natural language intent, breakdown of complex production tasks, and distributed collaborative scheduling of multiple intelligent agents. The Product and Intelligent Agent Layer can incorporate lightweight industrial large-scale models, capable of parsing the natural language optimization requirements of management and process engineers, or directly calling the industrial large-scale model from the world model layer to parse natural language optimization requirements.

[0287] The product and intelligent agent layer can receive target decision instructions output from the industrial world model and distribute them to the scheduling system of the equipment and resource layer through an automatically generated SDK (Software Development Kit) interface to complete equipment control and work order adjustment. The product and intelligent agent layer can collect real production line conditions and equipment operation data after the intelligent agent executes the data, standardize and encapsulate it, and then feed it back to the industrial dynamic simulation model of the world model layer to trigger online residual adaptive fine-tuning, forming a complete closed-loop iterative link.

[0288] At the application scenario level, standardized business modules are provided to the entire manufacturing industry chain. All functions rely on industrial world model simulation, industrial large-scale model knowledge reasoning, and intelligent agent collaboration capabilities, providing a visual interactive interface for enterprise management, process engineers, and maintenance personnel. Core industrial business modules at the application scenario level can include process simulation modules, intelligent scheduling systems, flexible manufacturing control, visual inspection platforms, full lifecycle equipment management, smart energy and carbon emissions and virtual power plants, production operation management, and safety production risk control. The user interaction interface at the application scenario level supports natural language voice or text interaction and features a 3D virtual-real twin visualization screen and an automatic simulation analysis report export module.

[0289] When users input production optimization requirements at the user interaction portal in the application scenario layer, the requirements can be analyzed through the industrial big data model to generate candidate adjustment schemes. After the candidate adjustment schemes are converted into intervention action codes, they can be handed over to the industrial world model for multiple rounds of counterfactual simulation verification. Finally, the target decision instructions are obtained and issued to the production line by the intelligent agent for execution. Simulation data and process analysis reports can be visualized, realizing the digitalization of simulation analysis, intelligent scheduling, and risk warning capabilities.

[0290] The following example, using a scenario of personalized customization in electronic SMT, illustrates the simulation decision-making process under conditions of human intervention.

[0291] First, the equipment and resource layer can collect multimodal raw data such as the vibration timing of the SMT production line's pick-and-place machine and PCB (Printed Circuit Board) defect images, and collect the production work order MES text, and push the multimodal raw data and production work order MES text to the data and knowledge layer.

[0292] The data and knowledge layer embeds text into the MES text to obtain text vectors. It then performs standardization and cleaning (e.g., multi-source data alignment, data augmentation) on the original multimodal data and text vectors to obtain standardized multimodal data. The data and knowledge layer can then pass this standardized multimodal data to the world model layer.

[0293] The industrial state representation model at the world model layer can transform standardized multimodal data into latent state features. This process can compress high-dimensional data and filter out workshop dust and sensor drift noise.

[0294] Process engineers submit production adjustment instructions to "reduce placement temperature and optimize PCB yield" through the visualization interface at the application scenario level. The industrial big data model generates two sets of candidate adjustment schemes based on the knowledge graph, and converts each candidate adjustment scheme into intervention action codes according to the industrial ontology and template action codes.

[0295] Potential state characteristics and intervention action codes are input into an industrial dynamic simulation model for a 7-day counterfactual simulation. The industrial dynamic simulation model incorporates thermodynamic and physical constraints of electronic manufacturing, simulating yield, energy consumption, and surface mount machine wear quantification indicators corresponding to three sets of intervention action codes to obtain the simulation trajectory. The simulation trajectory is then pushed to an industrial prediction and decision-making model for multi-objective optimization.

[0296] The industrial forecasting and decision-making model outputs the optimal candidate adjustment scheme through multi-objective optimization. This optimal candidate adjustment scheme is pushed to the industrial big data model, which verifies the compliance of the scheme based on the electronic process itself and safety specifications, eliminates non-compliant parameters that pose a risk of thermal deformation, and thus obtains the target decision instruction.

[0297] The target decision command is issued to the production scheduling agent and the equipment agent, which then issues the command to the underlying pick-and-place machine PLC to adjust the placement temperature parameters.

[0298] The intelligent agent can collect real production line yield, energy consumption, and equipment vibration data within 24 hours after the instruction is issued, and feed them back to the industrial dynamic simulation model to trigger an online residual adaptive mechanism to fine-tune the dynamic model parameters and reduce the deviation between virtual and real simulation.

[0299] The world model layer can push simulation comparison data and process semantic analysis reports to the visualization dashboard of the application scenario layer, enabling process engineers to view the simulation results and compare them with the real production line, and complete the closed-loop iteration of the entire chain.

[0300] Figure 9 A schematic diagram of the simulation decision-making device provided in this application is shown below. Figure 9 As shown, the simulation decision-making device 90 provided in this embodiment includes:

[0301] The encoding module 901 is used to convert process-related text into template action codes based on the industrial ontology, and then hand over the template action codes to the industrial world model; wherein, the industrial world model includes an industrial state representation model, an industrial dynamic simulation model, and an industrial prediction and decision-making model.

[0302] The mapping module 902 is used to take standardized multimodal data as input parameters of the industrial state characterization model and obtain the latent state features output by the industrial state characterization model.

[0303] The simulation module 903 is used to take the potential state features and intervention action codes as input parameters of the industrial dynamic simulation model, obtain the multi-round simulation time sequence output by the industrial dynamic simulation model, and quantize the multi-round simulation time sequence to obtain multi-dimensional quantitative simulation indicators.

[0304] The decision module 904 is used to take the multi-round simulation time sequence and multi-dimensional quantitative simulation indicators as input parameters of the industrial prediction and decision model, and obtain the target decision instructions output by the industrial prediction and decision model under the constraints of process compliance and optimization target weights.

[0305] In one possible implementation, the device further includes an acquisition module (not shown) for:

[0306] Obtain multimodal raw data from production line equipment;

[0307] Text data is obtained from an industrial information system, and text vectors are obtained by embedding the text data.

[0308] Standardized multimodal raw data and text vectors are cleaned and standardized to obtain standardized multimodal data.

[0309] In one possible implementation, the encoding module 901 is further configured to:

[0310] Semantic extraction and entity recognition are performed on process-related texts to obtain key business elements;

[0311] Threshold constraints, standardization scaling, and numerical fitting are applied to the numerical values ​​of key business elements to transform them into normalized values.

[0312] Based on the industrial ontology, constraint mapping and format unification are performed on the key business elements after numerical transformation to obtain standardized data;

[0313] Based on the pre-defined coding rules, template action codes are generated using standardized data.

[0314] In one possible implementation, the mapping module 902 is further configured to:

[0315] Separate time-series data, visual data, and text vectors from standardized multimodal data;

[0316] Steady-state baseline features and abrupt failure features are extracted from time series data, and then the steady-state baseline features and abrupt failure features are concatenated into noise-free time series features.

[0317] Visual data is encoded and processed to filter image noise and obtain noise-free visual features;

[0318] Noise-free temporal features, noise-free visual features, and text vectors are used as input parameters for a cross-modal projection encoder that has undergone contrastive learning, thereby obtaining multiple latent features output by the cross-modal projection encoder.

[0319] Adaptive weighted aggregation is performed on multiple latent features to obtain latent state features.

[0320] In one possible implementation, latent state characteristics include equipment-level latent state characteristics and production line-level latent state characteristics; intervention action codes include equipment-level intervention action codes and production line-level intervention action codes.

[0321] Industrial dynamic models include stochastic state-space models, temporal Transformers, and physical information neural networks;

[0322] Simulation module 903 is also used for:

[0323] By using device-level latent state features and device-level intervention action encoding as input parameters of a stochastic state-space model, long-term time-series prediction features of multiple devices are obtained.

[0324] By using production line-level potential state features and production line-level intervention action encodings as input parameters of the temporal Transformer, global temporal features of the entire production line can be obtained.

[0325] The long-term prediction features and global time-series features of multiple devices are input into the physical information neural network to obtain the multi-round simulation time-series sequence output by the physical information neural network.

[0326] In one possible implementation, the device further includes an analysis module (not shown) for:

[0327] Numerical analysis is performed on multi-dimensional quantitative simulation indicators to obtain structured numerical features;

[0328] By utilizing industrial knowledge graphs, causal reasoning and natural language semantic generation are performed on multi-dimensional quantitative simulation indicators carrying structured numerical features to obtain preliminary attribution conclusions for operating conditions.

[0329] Based on the process entity, the preliminary attribution conclusions of the operating conditions are revised to obtain a process semantic analysis report.

[0330] In one possible implementation, the multi-round simulation time series includes corrected long-term time series prediction features from multiple devices and corrected global time series features from multiple rounds; the decision module 904 is further configured to:

[0331] By using the corrected long-term prediction features and multi-dimensional quantitative simulation indicators of multiple devices as input parameters for industrial MPC, the candidate control parameters output by industrial MPC for each device under process compliance constraints and optimization target weight limitations are obtained.

[0332] The global time series features and multi-dimensional quantitative simulation indicators after multiple rounds of correction are used as input parameters for the deep reinforcement learning algorithm to obtain the medium- and long-term candidate scheduling schemes output by the deep reinforcement learning algorithm under the constraints of process compliance constraints, optimization target weights and candidate control parameters output by the industrial MPC for each device.

[0333] Based on the candidate control parameters of each device and the medium- and long-term candidate scheduling schemes corresponding to each set of candidate control parameters, multiple candidate decision schemes are generated.

[0334] Filter multiple candidate decision-making options based on compliance filtering strategies;

[0335] Generate the target decision instruction based on the remaining candidate decision schemes after filtering.

[0336] In one possible implementation, the decision module 904 is further configured to:

[0337] Filter multiple candidate decision schemes based on fundamental physical constraints;

[0338] The remaining candidate decision schemes after the initial filtering are then filtered a second time based on process standards, safety production regulations, and industry-specific rules.

[0339] In one possible implementation, the decision module 904 is further configured to:

[0340] The remaining candidate decision schemes after filtering are broken down according to the execution level to obtain the decision instructions for each execution level;

[0341] Construct target decision instructions based on decision instructions from multiple execution levels.

[0342] In one possible implementation, the distribution module (not shown in the figure) is used for:

[0343] The target decision instructions are handed over to the intelligent agent, which then distributes the decision instructions at each execution level of the target decision instructions to the execution object.

[0344] In one possible implementation, the adjustment module (not shown in the figure) is used for:

[0345] Obtain real production line operating data within a specified time period after executing the target decision instruction;

[0346] An online residual adaptive mechanism is triggered based on real production line operating data to adjust the model parameters of the industrial dynamic simulation model.

[0347] In one possible implementation, the mapping module 902 is further configured to:

[0348] Standardized multimodal data is used as input parameters for an industrial state representation model to obtain the latent state features output by the industrial state representation model under the guidance of semantic attention-guided features; wherein, the semantic attention-guided features are generated by a large industrial model.

[0349] In one possible implementation, the latent state features include standard latent state features and lightweight latent state features;

[0350] Mapping module 902 is also used for:

[0351] After the industrial state representation model outputs latent state features, the lightweight latent state features are handed over to the large industrial model to complete the cross-modal semantic alignment of standardized multimodal data.

[0352] In one possible implementation, the acquisition module (not shown in the figure) is further configured to:

[0353] Before encoding potential state characteristics and intervention actions as input parameters for the industrial dynamic simulation model, it is determined whether the industrial prediction and decision-making model outputs the target decision instruction.

[0354] If not, generate an intervention action code that is empty;

[0355] If so, the latest target decision instructions output by the industrial forecasting and decision-making model are converted into intervention action codes based on the industrial ontology and template action codes.

[0356] In one possible implementation, the acquisition module (not shown in the figure) is further configured to:

[0357] In response to production demand adjustment instructions, candidate adjustment schemes are generated based on knowledge graphs, and the candidate adjustment schemes are converted into intervention action codes according to industrial ontology and template action codes.

[0358] In one possible implementation, the display module (not shown in the figure) is used for:

[0359] Based on the noise-free visual features in the multi-round simulation time sequence and latent state features, a 3D simulation image of an industrial scene is rendered and displayed.

[0360] The simulation decision module provided in this embodiment can execute the method provided in the above method embodiment. Its implementation principle and technical effect are similar, and will not be described in detail here.

[0361] Figure 10 A schematic diagram of the structure of the electronic device (hereinafter referred to as the "device") provided in this application. Figure 10 As shown, the electronic device 100 provided in this embodiment includes at least one processor 1001 and a memory 1002. Optionally, the electronic device 100 further includes a communication component 1003. The processor 1001, memory 1002, and communication component 1003 are connected via a bus 1004.

[0362] In a specific implementation, at least one processor 1001 executes computer execution instructions stored in memory 1002, causing at least one processor 1001 to perform the above-described method.

[0363] The specific implementation process of processor 1001 can be found in the above method embodiments, and its implementation principle and technical effect are similar. It will not be repeated here.

[0364] In the above embodiments, it should be understood that the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor.

[0365] The memory may include random access memory (RAM) and may also include non-volatile memory (NVM), such as at least one disk storage device.

[0366] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of illustration, the buses shown in the accompanying drawings are not limited to a single bus or a single type of bus.

[0367] This application also provides a computer program product (hereinafter referred to as the "product"), including a computer program that, when executed by a processor, implements the above-described method.

[0368] This application also provides a computer-readable storage medium (hereinafter referred to as "the medium") that stores computer-executable instructions, which, when executed by a processor, implement the above-described method.

[0369] The aforementioned readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The readable storage medium can be any available medium accessible to a general-purpose or special-purpose computer.

[0370] An exemplary readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and the readable storage medium can exist as discrete components in the device.

[0371] The division of units is merely a logical functional division; in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0372] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0373] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0374] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0375] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0376] Finally, it should be noted that other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This invention is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein, and is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.

Claims

1. A simulation-based decision-making method based on an industrial world model, characterized in that, include: Based on the industrial ontology, process-related text is converted into template action codes, and the template action codes are then passed to the industrial world model; wherein, the industrial world model includes an industrial state representation model, an industrial dynamic simulation model, and an industrial prediction and decision-making model; Standardized multimodal data is used as input parameters for the industrial state characterization model to obtain the latent state features output by the industrial state characterization model. The latent state features and intervention action codes are used as input parameters of the industrial dynamic simulation model to obtain a multi-round simulation time sequence output by the industrial dynamic simulation model, and the multi-round simulation time sequence is quantized to obtain multi-dimensional quantitative simulation indicators. The multi-round simulation time sequence and the multi-dimensional quantitative simulation index are used as input parameters of the industrial prediction and decision-making model to obtain the target decision instructions output by the industrial prediction and decision-making model under the constraints of process compliance and optimization target weights.

2. The method according to claim 1, characterized in that, The steps for acquiring the standardized multimodal data include: Obtain multimodal raw data from production line equipment; Text data is obtained from an industrial information system, and text embedding is performed on the text data to obtain a text vector; The original multimodal data and the text vector are standardized and cleaned to obtain standardized multimodal data.

3. The method according to claim 1, characterized in that, The conversion of process-related text into template action codes based on industrial ontology includes: Semantic extraction and entity recognition are performed on the process-related text to obtain key business elements; Threshold constraints, standardization scaling, and numerical fitting are applied to the values ​​in the key business elements to transform the values ​​in the key business elements into normalized values. Based on the industrial ontology, constraint mapping and format unification are performed on the key business elements after numerical transformation to obtain standardized data; Template action codes are generated based on the standardized data according to the preset coding rules.

4. The method according to claim 1, characterized in that, The step of using standardized multimodal data as input parameters to the industrial state characterization model to obtain the latent state features output by the industrial state characterization model includes: Separate time-series data, visual data, and text vectors from the standardized multimodal data; Steady-state baseline features and abrupt failure features are extracted from the time series data, and the steady-state baseline features and the abrupt failure features are concatenated into noise-free time series features; The visual data is encoded to filter out image noise and obtain noise-free visual features; The noise-free temporal features, the noise-free visual features, and the text vector are used as input parameters for a cross-modal projection encoder that has undergone contrastive learning, thereby obtaining multiple latent features output by the cross-modal projection encoder. Adaptive weighted aggregation is performed on the multiple latent features to obtain latent state features.

5. The method according to claim 1, characterized in that, The latent state characteristics include equipment-level latent state characteristics and production line-level latent state characteristics; the intervention action codes include equipment-level intervention action codes and production line-level intervention action codes. The industrial dynamic simulation model includes a stochastic state-space model, a temporal Transformer, and a physical information neural network. The step of encoding the latent state features and intervention actions as input parameters to the industrial dynamic simulation model to obtain the multi-round simulation time sequence output by the industrial dynamic simulation model includes: The device-level latent state features and the device-level intervention action encoding are used as input parameters of the stochastic state-space model to obtain long-term time-series prediction features of multiple devices. The production line-level potential state features and the production line-level intervention action encoding are used as input parameters of the temporal Transformer to obtain the global temporal features of the entire production line. The long-term time-series prediction features of multiple devices and the global time-series features are input into the physical information neural network to obtain the multi-round simulation time-series sequence output by the physical information neural network.

6. The method according to claim 1 or 5, characterized in that, The method further includes: Numerical analysis is performed on the multi-dimensional quantitative simulation indicators to obtain structured numerical features; By utilizing industrial knowledge graphs, causal reasoning and natural language semantic generation are performed on multi-dimensional quantitative simulation indicators carrying structured numerical features to obtain preliminary attribution conclusions for operating conditions. Based on the process entity, the preliminary attribution conclusions of the operating conditions are revised to obtain a process semantic analysis report.

7. The method according to claim 1, characterized in that, The multi-round simulation time series includes corrected long-time prediction features of multiple devices and corrected global time series features after multiple rounds of simulation. The step of using the multi-round simulation time sequence and the multi-dimensional quantitative simulation indicators as input parameters of the industrial prediction and decision-making model to obtain the target decision instruction output by the industrial prediction and decision-making model under the constraints of process compliance and optimization target weights includes: The corrected long-term prediction features of the multiple devices and the multi-dimensional quantitative simulation indicators are used as input parameters of the industrial MPC to obtain the candidate control parameters output by the industrial MPC for each device under the process compliance constraints and the optimization target weight limit. The global time series features after multiple rounds of correction and the multi-dimensional quantized simulation indicators are used as input parameters of the deep reinforcement learning algorithm to obtain the medium- and long-term candidate scheduling schemes output by the deep reinforcement learning algorithm under the constraints of process compliance, optimization target weights and candidate control parameters output by the industrial MPC for each device. Based on the candidate control parameters of each device and the medium- and long-term candidate scheduling schemes corresponding to each set of candidate control parameters, multiple candidate decision schemes are generated. The multiple candidate decision schemes are filtered based on a compliance filtering strategy; Generate the target decision instruction based on the remaining candidate decision schemes after filtering.

8. The method according to claim 7, characterized in that, The filtering of the multiple candidate decision schemes based on the compliance filtering strategy includes: The multiple candidate decision schemes are filtered based on fundamental physical constraints. The remaining candidate decision schemes after the initial filtering are then filtered a second time based on process standards, safety production regulations, and industry-specific rules.

9. The method according to claim 7, characterized in that, The step of generating the target decision instruction based on the remaining candidate decision schemes after filtering includes: The remaining candidate decision schemes after filtering are broken down according to the execution level to obtain the decision instructions for each execution level; The target decision instruction is constructed based on decision instructions from multiple execution levels.

10. The method according to claim 1 or 9, characterized in that, The method further includes: The target decision instruction is handed over to the intelligent agent, which then distributes the decision instructions at each execution level of the target decision instruction to the execution object.

11. The method according to claim 10, characterized in that, The method further includes: Obtain real production line operating condition data within a specified time period after executing the target decision instruction; Based on the real production line operating data, an online residual adaptive mechanism is triggered to adjust the model parameters of the industrial dynamic simulation model.

12. The method according to claim 1, characterized in that, The step of using standardized multimodal data as input parameters to the industrial state characterization model to obtain the latent state features output by the industrial state characterization model includes: The standardized multimodal data is used as the input parameters of the industrial state representation model to obtain the latent state features output by the industrial state representation model under the guidance of semantic attention-guided features; wherein, the semantic attention-guided features are generated by the large industrial model.

13. The method according to claim 12, characterized in that, The latent state features include standard latent state features and lightweight latent state features; The method further includes: After the industrial state representation model outputs latent state features, the lightweight latent state features are handed over to the large industrial model to complete the cross-modal semantic alignment of the standardized multimodal data.

14. The method according to claim 1, characterized in that, The intervention action code is obtained in the following way: Before encoding the potential state features and intervention actions as input parameters of the industrial dynamic simulation model, it is determined whether the industrial prediction and decision-making model outputs the target decision instruction; If not, generate an intervention action code that is empty; If so, the latest target decision instruction output by the industrial prediction and decision-making model is converted into intervention action code based on the industrial ontology and the template action code.

15. The method according to claim 1 or 14, characterized in that, The intervention action code is obtained in the following way: In response to production demand adjustment instructions, candidate adjustment schemes are generated based on a knowledge graph, and the candidate adjustment schemes are converted into intervention action codes according to the industrial ontology and the template action codes.

16. The method according to claim 1 or 5, characterized in that, The method further includes: Based on the multi-round simulation time sequence and the noise-free visual features in the potential state features, a three-dimensional simulation image of the industrial scene is rendered and displayed.

17. An electronic device, characterized in that, include: Memory, processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory, causing the processor to perform the method as described in any one of claims 1-16.

18. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the method as described in any one of claims 1-16.

19. A computer program product, characterized in that, Includes a computer program that, when executed by a processor, implements the method described in any one of claims 1-16.

Images