A powder premixing strategy generation method and system based on physical constraints and diffusion decisions

Abstract

The present application relates to the technical field of automatic policy generation, in particular to a powder premixing policy generation method and system based on physical constraints and diffusion decision; the method comprises self-supervised multi-modal state representation based on pre-processed data, including unified process sample input and modal token construction, quality perception mask self-supervised coding, action conditional time consistency and cross-modal alignment, latent state fusion and module output; a physical constraint latent space world model is constructed based on state representation, including latent state sample input and state action decomposition, latent space state transition prediction and uncertainty estimation, physical constraint decoding and mixed uniformity response modeling, joint training and world model output; the present application effectively improves the uniformity control precision, batch consistency and engineering operation safety of the powder premixing process.

Description

Technical Field

[0001] This invention relates to the field of automated strategy generation technology, and in particular to a method and system for generating powder premixing strategies based on physical constraints and diffusion decisions. Background Technology

[0002] In the production of powder coatings, fine chemical powders, and functional materials, the premixing process directly affects the dispersion of raw materials, batch consistency, and the performance of the final product. As powder materials develop towards high performance, green manufacturing, and customization, the production process places higher demands on mixing uniformity, energy consumption control, equipment safety, and process response speed. However, the powder premixing process is influenced by multiple factors, including material particle size, density, flowability, adhesion, formulation ratio, and equipment operating conditions, exhibiting significant nonlinear, time-varying, and highly perturbed characteristics.

[0003] Existing powder premixing control methods mostly rely on fixed process parameters, manual experience, or offline sampling and testing. These methods typically determine uniformity by sampling after mixing, making it difficult to promptly reflect changes in material flow, such as abnormalities, localized agglomeration, adhesion, bridging, or insufficient dispersion during the mixing process, resulting in a lag in control. For different batches of materials or new formulations, maintaining fixed speed, angle, and mixing time cannot consistently guarantee mixing quality, easily leading to increased energy consumption, insufficient mixing, or excessive shearing.

[0004] To improve the automation level of the premixing process, some existing technologies introduce sensor data such as torque, current, vibration, vision, or spectral data, and combine them with PID control, adaptive control, model predictive control, or traditional machine learning models to adjust the mixing parameters. These methods can improve the shortcomings of single-signal monitoring and manual parameter tuning to some extent, but they typically rely on manual features, local empirical models, or fixed control laws. They struggle to accurately characterize the continuous evolution of the powder mixing state in the potential space, and also find it difficult to plan multi-step strategies balancing equipment safety, energy consumption constraints, and mixing quality objectives.

[0005] In recent years, technologies such as self-supervised multimodal characterization, world models, diffusion decision-making, and safety constraint control have provided new technical pathways for intelligent regulation of complex industrial processes. Self-supervised learning can extract a unified latent state from multi-source process data; world models can predict future mixed states under different regulatory actions; diffusion decision-making can generate multi-step candidate regulation sequences; and safety constraints can screen executable actions that meet equipment load, vibration, temperature, and process boundaries. However, in existing technologies, the aforementioned intelligent decision-making methods have not yet fully integrated the physical constraints, uniformity evolution laws, and abnormal operating condition control requirements in powder premixing processes.

[0006] Therefore, there is an urgent need for an intelligent system for online control of powder premixing uniformity, which can integrate multimodal process data such as torque, current, vibration, temperature, images, spectra, material batches and formulation ratios, construct a latent space world model with physical constraints, and use safe diffusion decision to generate control strategies that meet the requirements of mixing quality, energy consumption and equipment safety, thereby realizing online uniformity assessment, closed-loop regulation and abnormal operating condition control of the powder premixing process. Summary of the Invention

[0007] To address the challenges of real-time characterization of mixing state, inability of fixed process parameters to adapt to material and batch variations, lack of predictive capability in multi-step control strategies, and difficulty in synergistically satisfying equipment safety constraints and mixing quality objectives in existing online control processes for powder premixing uniformity, this invention provides a powder premixing strategy generation method and system based on physical constraints and diffusion decision-making. Through inputting multimodal process data of powder premixing, self-supervised multimodal state characterization, construction of a physical constraint latent space world model, generation of safe diffusion control strategies, and an online closed-loop execution and anomaly control output mechanism, online evaluation, predictive control, and safe closed-loop optimization of powder premixing uniformity are achieved, thereby improving mixing quality stability, process adaptability, and equipment operational safety.

[0008] In a first aspect, the present invention provides a powder premixing strategy generation method based on physical constraints and diffusion decisions, which adopts the following technical solution: A powder premixing strategy generation method based on physical constraints and diffusion decisions includes: Acquire multimodal powder premixing data; Data preprocessing was performed based on the acquired multimodal powder premixing data; Self-supervised multimodal state representation based on preprocessed data includes unified process sample input and modal token construction, quality-aware mask self-supervised encoding, action condition temporal consistency and cross-modal alignment, latent state fusion and module output; The physical constraint latent space world model is constructed based on state representation, including latent state sample input and state-action decomposition, latent space state transition prediction and uncertainty estimation, physical constraint decoding and mixed uniformity response modeling, joint training and world model output. Based on the constructed world model, a security diffusion control strategy is generated, including the construction of control condition states and the setting of target constraints, the generation of security diffusion candidate control sequences, rolling prediction of the world model and screening of security barriers, control utility scoring and strategy result output; Online closed-loop execution and anomaly control output are performed based on the output strategy.

[0009] Secondly, a powder premixing strategy generation system based on physical constraints and diffusion decisions includes: The data acquisition module is configured to acquire multimodal powder premixing data; The preprocessing module is configured to perform data preprocessing based on the acquired multimodal powder premixing data; The representation module is configured to perform self-supervised multimodal state representation based on preprocessed data, including unified process sample input and modal token construction, quality-aware mask self-supervised encoding, action condition temporal consistency and cross-modal alignment, latent state fusion and module output; The model building module is configured to build a physically constrained latent space world model based on state representation, including latent state sample input and state-action decomposition, latent space state transition prediction and uncertainty estimation, physical constraint decoding and mixed uniformity response modeling, joint training and world model output. The strategy generation module is configured to generate security diffusion control strategies based on the constructed world model, including constructing control condition states and setting target constraints, generating security diffusion candidate control sequences, rolling prediction of the world model and screening of security barriers, scoring control utility and outputting strategy results. The execution module is configured to perform online closed-loop execution and anomaly control output based on the output strategy.

[0010] Thirdly, the present invention provides a computer-readable storage medium storing a plurality of instructions adapted for loading and execution by a processor of a terminal device of the powder premixing strategy generation method based on physical constraints and diffusion decisions.

[0011] Fourthly, the present invention provides a terminal device, including a processor and a computer-readable storage medium, wherein the processor is used to implement various instructions; the computer-readable storage medium is used to store multiple instructions, the instructions being adapted to be loaded and executed by the processor to provide a powder premixing strategy generation method based on physical constraints and diffusion decisions.

[0012] In summary, the present invention has the following beneficial technical effects: Compared to existing powder premixing control methods, which suffer from technical bottlenecks such as difficulty in online characterization of mixing state, reliance on human experience for control parameters, weak adaptability to batch differences, and insufficient consideration of equipment safety constraints, this invention addresses the core need of "online sensing, prediction, and safe control of powder premixing uniformity." It proposes an online powder premixing uniformity control system based on a physical constraint world model and safe diffusion decision-making. This system realizes a complete processing flow from multimodal process data input, self-supervised state characterization, physical constraint world model construction, safe diffusion control strategy generation to online closed-loop execution and anomaly control, effectively improving the uniformity control accuracy, batch consistency, and engineering operation safety of the powder premixing process.

[0013] First, this invention solves the problem in traditional systems where material information, equipment information, and process observations are scattered and difficult to integrate into control decisions by uniformly modeling batch attributes, multi-source sensor observations, equipment operating status, and historical actions during the powder premixing process. Furthermore, this invention transforms asynchronous, noisy, and multi-scale process data into unified process samples through adaptive time window alignment, data quality assessment, and reliability calculation. This enables the system to maintain a stable data input foundation under conditions of sensor absence, asynchronous sampling, and fluctuating operating conditions.

[0014] Secondly, this invention utilizes a self-supervised multimodal state representation mechanism to convert multi-source process information such as current, torque, vibration, temperature, images, and spectra into latent state representations, avoiding the problem of traditional methods over-reliance on offline sampling detection or single sensor indicators. This mechanism, even with insufficient uniformity labels, learns the rules governing mixed state changes through mask reconstruction, action condition temporal consistency, and cross-modal alignment. This enables the system not only to identify the current mixed stage but also to express the impact of historical actions on subsequent mixed states, providing a more reliable state foundation for subsequent dynamic prediction.

[0015] Furthermore, this invention constructs a physically constrained latent space world model, linking potential state transition predictions with physical process quantities such as stirring power, cumulative mixing energy, shear strength, motor temperature rise, vibration amplitude, and load state. This addresses the issues of insufficient interpretability in purely data-driven models and the potential deviation of prediction results from equipment operating patterns. Through energy consistency constraints, operating boundary constraints, uniformity dynamic response constraints, and prediction uncertainty estimation, the system can simultaneously consider mixing quality improvement and equipment safety boundaries when predicting future states, thereby enhancing the reliability of the generated control strategies.

[0016] Finally, this invention forms a closed-loop control mechanism for powder premixing by generating a safe diffusion control strategy and an online closed-loop anomaly control mechanism. On the one hand, the system can generate multiple sets of future candidate action sequences based on a world model, and select the optimal control strategy through rolling prediction, safety barrier screening, and comprehensive utility scoring, avoiding sudden action changes, overmixing, or equipment overstepping limits caused by single-step greedy parameter tuning. On the other hand, the system can calculate prediction deviations based on the uniformity, power, shear, temperature, vibration, and load feedback after actual execution, identify risks such as undermixing, overmixing, overload, abnormal temperature rise, and model mismatch, and output graded control results, enabling the powder premixing process to have online correction and safety traceability capabilities.

[0017] In online control applications of powder premixing uniformity, the mean absolute error of online uniformity prediction in this invention is reduced by more than 21% compared to traditional single-modal time-series prediction methods; the coefficient of variation of final mixing uniformity is reduced by more than 18% compared to fixed-parameter mixing control; and the time to achieve target uniformity is shortened by more than 16% compared to conventional empirical parameter tuning methods. Simultaneously, the equipment safety violation rate is reduced to below 3%, and the abnormal state identification recall rate reaches over 90%. Therefore, this invention improves powder premixing uniformity, batch consistency, and control response speed while reducing the risks of overmixing, equipment overload, and abnormal shutdowns. It provides an efficient, reliable, and engineering-deployable online control technology solution for powder coatings, fine chemicals, and related powder manufacturing processes. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of a powder premixing strategy generation method based on physical constraints and diffusion decisions according to Embodiment 1 of the present invention; Figure 2 This is a schematic diagram comparing the uniformity prediction errors of various methods in Embodiment 1 of the present invention; Figure 3 This is a schematic diagram comparing the coefficients of variation of the endpoints of various methods in Embodiment 1 of the present invention; Figure 4 This is a schematic diagram comparing the safety out-of-bounds rates of various methods in Embodiment 1 of the present invention; Figure 5 This is a schematic diagram comparing the abnormal recall rates of various methods in Embodiment 1 of the present invention; Figure 6 This is a schematic diagram comparing the unit energy consumption of each method in Embodiment 1 of the present invention; Figure 7 This is a schematic diagram comparing the achievement times of various methods in Embodiment 1 of the present invention. Detailed Implementation

[0019] The present invention will be further described in detail below with reference to the accompanying drawings.

[0020] Example 1 Reference Figure 1 This embodiment of a powder premixing strategy generation method based on physical constraints and diffusion decisions includes: (1) Powder premixing multimodal process data input module This module addresses the challenges of significant batch-to-batch variations in powder premixing, heterogeneous sensor observations, and the difficulty in synchronously recording equipment status and control actions. Its goal is to unify material properties, multi-source process observations, equipment operating status, and historical actions into computable process data. To resolve issues of scattered raw data sources, inconsistent timescales, and unstable data quality, this module employs batch attribute modeling, multi-source observation acquisition, equipment status and historical action recording, adaptive time window alignment, and data reliability calculation to create a unified process data set that can be directly used by the subsequent self-supervised multimodal state characterization module.

[0021] 1) Batch attributes and multi-source raw observation input To address the challenge of unifying batch information with sensor observations, this module provides a unified representation of material-side and observation-side data in the powder premixing process through batch set definition, batch attribute vector construction, and multi-source original observation unit recording.

[0022] Let the set of powder premixed batches be: , in, This indicates a collection of powder premixed batches. Indicates the first One powder premixed batch Indicates the batch quantity.

[0023] Regarding the first For each powder batch, the material composition, particle size distribution, formulation ratio, feed quality, and environmental conditions are organized into a batch attribute vector: , in, Indicates the first Batch attribute vectors for each batch Indicates the material category and component properties. Indicates particle size distribution characteristics, Indicates the characteristics of the formula proportions. Indicates the batch feed quality. This indicates the ambient temperature, humidity, and operating parameters.

[0024] Let the set of observed modes during the powder premixing process be: , in, Represents the set of observed modes. Indicates the first Observational modes, This indicates the number of observation modes. The observation modes include one or more of the following: stirring current, torque, vibration, temperature, online images, near-infrared spectroscopy, and powder flow state signals.

[0025] To preserve the temporal attributes, sensor status, and missing data for each observation, the original observation unit is constructed: , in, Indicates the first The first batch, the first The first observation mode One original observation unit, Represents the original observation value. Indicates the sampling time. Indicates the health status of the sensor. This indicates a missing marker.

[0026] The batch attributes and multi-source original observation entry points obtained through the above processing provide the original data foundation for subsequent equipment status association and time window alignment.

[0027] 2) Data collection on equipment operating status and historical actions. To address the issue of unclear variable sources in subsequent world models for control actions, this module synchronously collects equipment operating state vectors and historical executed action vectors, incorporating the actual operating state and executed control parameters into the same batch of process records.

[0028] In order for the subsequent world model to learn the dynamic response relationship between "control action - mixed state change", the device operating state vector is first defined: , in, Indicates the first The first batch The device operating state vector at each sampling time. Indicates the actual rotational speed. Represents current. Indicates voltage. Indicates motor temperature. Indicates the vibration amplitude. Indicates the load status.

[0029] Further define the historical execution action vector: , in, Indicates the first The first batch The historical execution action vector at each sampling time, Indicates the set speed. This indicates setting the blending time or time increment. This indicates the amount of material compensation. This indicates the amount of adjustment of the stirring structure or blade attitude.

[0030] Therefore, batch attributes, multi-source raw observations, equipment operating status, and historical execution actions are uniformly organized into the first... Raw process data for each batch: , in, Indicates the first The original process data for each batch, Indicates the first The number of samples for each observation modality. Indicates the number of samples taken to indicate the device's operating status. This indicates the number of historical action samples. and These represent the sampling time for the device's operating status and the historical actions performed, respectively.

[0031] The above processing clarifies the data sources for equipment status and historical actions, providing traceable input for action variables in the subsequent latent space world model.

[0032] 3) Adaptive temporal window alignment and modal coding To address the issue that different modalities cannot directly participate in the same time slice calculation, this module employs time window partitioning, time-sensitive weight allocation, and modal coding mechanisms to generate window-level observation representations, device status representations, and historical execution action representations.

[0033] Let the first The first batch The time window for each process is: , in, Indicates the first A process time window, Indicates the window start time. Indicates the window end time. Indicates the center time of the window.

[0034] To determine the effective sampling range of each observation mode within the current time window, a set of sampling indices is defined: , in, Indicates the first The first batch, the first Observational modes in time window The set of sampling indices within.

[0035] To address asynchronous sampling and differences in sensor quality, time-sensitive weights are assigned to observation points within the window: , in, Indicates the first Each observation point within the time window Time-sensitive weights within, Indicates the first Time-scale parameters of the observation modality This represents the weighting coefficient for sensor health. This indicates the missing penalty weight coefficient.

[0036] Based on the above weights, the original observations are encoded into window-level modal representations: , in, Indicates the first Observational modes in time window The encoded observations within represent, Indicates the first The encoding function corresponding to the observation mode. This represents the encoding parameters.

[0037] To ensure that device status and historical actions are kept on the same time scale as observed modes, aggregation weights are further defined: , in, This represents the aggregate weight of the device's operating status or historical actions within a time window. Indicates the operating status of the equipment. Indicates historical actions performed. This represents the set of sampling indices for the corresponding variable within the time window. This represents the time scale parameter of the corresponding variable.

[0038] Obtain the window-level device state representation and the window-level historical execution action representation: , in, This indicates the window-level device status. This represents the device status coding function. Indicates the device status code parameter. This indicates the history of actions executed at the window level.

[0039] Through the above processing, the alignment and encoding of multi-source asynchronous data into a unified time window are achieved, providing a time-consistent process segment for subsequent self-supervised state representation.

[0040] 4) Data quality assessment and unified process sample generation To address the issue of model instability caused by directly inputting low-quality data into the state representation module, this module generates unified process samples and forms the module output by calculating observation coverage, time synchronization reliability, and comprehensive data reliability.

[0041] First calculate the... Observational modes in time window Observation coverage within: , in, Indicates the first Observation coverage of the observation modality Indicates the number of valid samples within the window. Indicates the first The observation modality satisfies the minimum number of samples required for effective observation.

[0042] Further calculation of window-level time synchronization reliability: , in, Indicates time window Time synchronization reliability. A data quality vector is constructed based on observation coverage, sensor health, missing data, and time synchronization reliability: , in, Indicates the first The first batch Data quality vectors for each time window.

[0043] To generate reliable samples that can be used by subsequent state characterization modules, the reliability of window-level aggregated data is calculated: , in, Indicates the reliability of window-level comprehensive data. This represents the Sigmoid function. , , , These represent the weight parameters corresponding to data quality, batch attributes, device status, and historical execution actions, respectively. This indicates the bias term.

[0044] Finally, window-level multimodal observations, device status, historical execution actions, and batch attributes are combined into a unified process segment: , in, Indicates the first The first batch A unified process segment within a time window.

[0045] Constructing a unified process sample: , in, Indicates the first The first batch A unified process sample for each time window.

[0046] This module ultimately outputs a dataset of multimodal powder premixing process data: , in, This represents the unified process data set output by this module. Indicates the first The number of time windows obtained from each batch division. The output of this module will serve as the input for subsequent modules.

[0047] (2) Self-supervised multimodal state representation module This module addresses the challenges of high-dimensional data, significant modal differences, and the difficulty in obtaining uniformity labels in real time during powder premixing processes. Its goal is to learn latent states from unified process data that characterize the mixing stages, equipment responses, and the relationships between actions. To overcome the limitations of traditional methods that rely on manual annotation or single sensor features, leading to insufficient state representation, this module employs quality-enhanced token construction, masked self-supervised reconstruction, action condition temporal consistency constraints, and cross-modal attention fusion to form a set of state representations that can be directly used in subsequent physically constrained latent space world models.

[0048] 1) Unified process sample input and modal token construction To address the challenge of directly inputting heterogeneous process information into a self-supervised model, this module extracts time-window process samples from unified process data and constructs modal tokens and process-wide tokens through a quality enhancement mapping mechanism.

[0049] For the For observation modalities, construct quality-enhanced modal tokens: , in, Indicates the first The first batch Within the first time window Token representation of observation modalities The layer normalization function is represented. , , , , These represent the mapping matrices corresponding to modal observations, batch attributes, device status, historical actions, and data quality, respectively. Represents the overall credibility mapping vector. This indicates the bias term.

[0050] To introduce batch-level, device-level, and action-level global contexts, a process-wide global token is further constructed: , in, Indicates the first The first batch A global token for a time window process. , , , , and This represents the corresponding mapping parameters.

[0051] Combine the global token and the tokens of each observation mode row by row to obtain the token matrix: , in, Indicates the first The first batch A multimodal token matrix with a time window. Indicates the token dimension.

[0052] After the above processing, a process token matrix for self-supervised learning is obtained, which serves as the input for subsequent mask self-supervised encoding.

[0053] 2) Quality-aware mask self-supervised coding To address the interference of low-reliability and missing data on self-supervised learning, this module dynamically adjusts the mask probability based on data reliability and observation quality, and obtains self-supervised hidden representations through quality-biased attention and mask reconstruction mechanisms.

[0054] The data quality vector is represented as: , in, Represents the observation coverage component. Indicates the sensor health component. Indicates the effective observation proportion component, This represents the time synchronization reliability component.

[0055] Based on the above mass components, the calculation of the first... Token mask probability: , in, Indicates the first The mask probability of tokens, This represents the truncation function. Represents the basic mask probability. and Indicates the weighting coefficient. and These represent the lower and upper bounds of the mask probability, respectively.

[0056] Generate the masked token based on the mask probability: , in, This represents the token behind the mask. This indicates a mask indicator variable. This represents a learnable mask token.

[0057] Combine all mask tokens as: , in, This represents the multimodal token matrix after masking.

[0058] To enable the self-supervised encoding process to simultaneously consider intermodal relationships and data quality differences, a quality-biased attention matrix is ​​constructed: , in, Represents the mass bias matrix. and This represents the mass bias coefficient. Indicates length is A vector of all 1s Represents the identity matrix.

[0059] Self-attention encoding based on the quality bias matrix: , in, This represents the self-supervised attention weight matrix. The self-supervised hidden representation matrix is ​​represented. , , , These represent the query, key, value, and output projection matrix, respectively. This represents the attention head dimension.

[0060] Reconstruct the masked token: , in, This represents the reconstructed token matrix. This represents the reconstruction mapping matrix. This represents the reconstructed bias vector.

[0061] Constructing the mask reconstruction loss: , in, Indicates the mask reconstruction loss. express The Middle A reconstruction token, This represents a constant used to prevent the denominator from being zero.

[0062] The results obtained through the above processing will serve as inputs for subsequent time consistency constraints and cross-modal fusion.

[0063] 3) Motion conditional temporal consistency and cross-modal alignment To address the problem that relying solely on static window features is insufficient to represent the evolution of the mixing process, this module introduces action condition time consistency constraints and combines them with a cross-modal alignment mechanism to enhance the consistent representation of different observation modes in the same mixing stage.

[0064] Extracting the first from the self-supervised hidden representation matrix The hidden representation of each token: , in, Indicates the first The hidden representation of a token, Indicates the first The selection vector for each position.

[0065] To ensure that the historical execution actions and the hidden representation maintain the same dimension, the window-level historical execution actions are first projected: , in, This represents the historical execution actions after projection. Represents the action projection matrix. This indicates the motion projection offset.

[0066] Based on the current global hidden representation and the projected historical execution actions, predict the global hidden representation for the next time window: , in, This represents the globally hidden representation of the predicted next time window. , , These represent the state mapping, action mapping, and state-action interaction mapping matrices, respectively. This represents element-wise multiplication. This indicates the bias term.

[0067] Constructing action condition time consistency loss: , in, This indicates the loss of consistency between action conditions and time. This represents the true global hidden representation of the next time window.

[0068] To enhance the consistent representation capability of different observation modalities under the same mixture state, the hidden representations of each modality are normalized and projected: , in, Indicates the first Normalized aligned representation of tokens, This represents the alignment projection matrix.

[0069] Calculate the cross-modal consistency weights between each observation modality and the global token: , in, Indicates the first Cross-modal consistency weights for observed modes. This represents the temperature coefficient.

[0070] Further construct cross-modal contrast loss: , in, Indicates cross-modal contrast loss. This represents the set of negative sample indices from different batches or different state stages compared to the current time window. This represents the globally aligned representation of the negative sample.

[0071] By using time consistency and cross-modal alignment constraints, the state representation can simultaneously preserve the process dynamic response and the consistency of multi-source observations.

[0072] 4) Latent state fusion and module output To address the issue that simple feature splicing is insufficient to highlight effective modalities and the impact of key actions, this module uses consistency weighting, state fusion attention, and gating fusion mechanisms to compress multi-source hidden representations into latent state samples, which then form the module output.

[0073] To enable different modalities to adaptively participate in the fusion according to their consistency and reliability, the state fusion attention score is first calculated: , in, Indicates the first State fusion attention score for observation modalities , , , , and This represents the state fusion parameters.

[0074] Combining the cross-modal consistency weights, the final modal fusion weights are obtained: , in, Indicates the first The final fusion weights for observation modalities. This represents the cross-modal consistency adjustment coefficient.

[0075] Generate fused state representation based on modality fusion weights: , in, Indicates the first The first batch The fusion state representation of each time window.

[0076] To balance global process information and multimodal fusion information, a state gating vector is introduced: , in, Represents the state gating vector. Represents the gated mapping matrix, This indicates gating bias.

[0077] Generate potential state vectors: , in, Indicates the first The first batch Potential state vectors for each time window , , , and This represents the parameters for generating the potential state.

[0078] To ensure that we can directly learn the evolutionary relationship of the latent state under the influence of historical actions, the latent state, historical actions, data confidence, and time step are encapsulated as a latent state sample: , in, Indicates the first The time step of each time window.

[0079] , in, Indicates the first The first batch Potential state samples within a time window.

[0080] This module ultimately outputs a set of self-supervised multimodal state representations: , in, This represents the set of self-supervised multimodal state representations output by this module. The output of this module will serve as the input for subsequent modules.

[0081] (3) Physically Constrained Latent Space World Model Construction Module This module addresses the nonlinear, hysteretic, and physically constrained characteristics of mixing state evolution during powder premixing. Its goal is to construct a latent space world model based on a set of state representations, capable of predicting future mixing states, physical process quantities, and homogeneity responses. To overcome the limitations of purely data-driven models in terms of interpretability and the lack of physical boundary constraints in control predictions, this module employs state-action decomposition, latent space transition prediction, uncertainty estimation, physical decoding, and homogeneity response constraints to form a world model that can be invoked for subsequent safe diffusion control strategies.

[0082] 1) Latent state sample input and state-action decomposition To address the issues of unclear sources of action variables and ambiguous state-action interactions during world model training, this module extracts potential mixed states, historical executed actions, data credibility, and time steps from the state representation set and constructs the world model input vector.

[0083] For the The first batch Given potential state samples within a time window, define a selection mapping: , in, Indicates a potential mixed state. Indicates historical actions performed. Indicates data credibility. Indicates the time step. , , , These represent the corresponding selection matrix or selection vector, respectively.

[0084] To enable action variables to interact with the latent state for computation, a dimensional projection is performed on the historically executed actions: , in, This represents the action representation after projection. Represents the action projection matrix. This indicates the motion projection offset.

[0085] Construct the world model input vector based on the latent state, projected action, state-action interaction term, and time step: , in, Represents the world model input vector. This indicates element-wise multiplication.

[0086] The resulting world model input vector, obtained after the above processing, will be used for subsequent latent space state transition prediction.

[0087] 2) Latent space state transition prediction and uncertainty estimation To address the problem that fixed dynamic models are difficult to adapt to complex hybrid processes, this module adopts state-action interaction, transition gating, and variance estimation mechanisms to establish a confidence-weighted latent space state transition prediction model.

[0088] To describe the moderating effect of the executed action on the magnitude of the state change, a transition gating vector is constructed: , in, Represents the latent space state transition gating vector. This represents the Sigmoid function. Represents the gated mapping matrix, This indicates gating bias.

[0089] Further generate candidate state increments: , in, This represents the predicted potential state increment. Represents the state increment mapping matrix, This indicates the state increment bias.

[0090] By combining the time step and transition gating, the potential state prediction for the next time window is obtained: , in, This represents the predicted potential state for the next time window.

[0091] To reduce over-deterministic predictions in situations with high noise or low observation confidence, a state prediction variance is constructed: , in, This represents the variance of the latent state prediction. Represents a smooth non-negative activation function. Represents the variance mapping matrix. Indicates variance bias. This represents a constant that prevents the value from being zero.

[0092] Based on the predicted state and the actual next state, construct a confidence-weighted state prediction loss: , in, This represents the state prediction loss. Represents the potential state dimension. and The first and second represent the actual potential state and the predicted potential state, respectively. Each component.

[0093] The potential state and potential state prediction variance obtained through the above processing will be used for subsequent physical constraint decoding and multi-step rolling prediction.

[0094] 3) Physical constraint decoding and hybrid uniformity response modeling To address the issue that model predictions may violate the actual operating rules of equipment, this module improves the interpretability and controllability of the world model through physical decoders, energy consistency constraints, operating boundary constraints, and uniformity dynamic response constraints.

[0095] To recover physical process quantities from potential states and actions, a physical decoder is defined: , in, Represents a vector of physical process quantities. Indicates the physical decoder. Indicates the physical decoder parameters. Indicates the stirring power. Indicates the cumulative mixed energy. Indicates shear strength. Indicates motor temperature. Indicates the vibration amplitude. Indicates the load status.

[0096] Perform the same physical decoding to predict the next state: , in, This represents the physical process quantity of the next time window obtained from the decoding of the predicted latent state.

[0097] To constrain the relationship between accumulated mixed energy and power input, an energy consistency loss is constructed: , in, Indicates energy consistency loss, This indicates the cumulative mixed energy predicted for the next time window. Indicates the first The mixed energy utilization coefficient corresponding to each batch.

[0098] To avoid the model generating physical states that do not conform to the device's safety boundaries, a runtime boundary constraint loss is constructed: , in, This represents the loss due to running boundary constraints. and These represent the lower and upper limits of shear strength, respectively. Indicates the upper limit of motor temperature. Indicates the upper limit of vibration amplitude. Indicates the maximum load. This represents the linear rectifier function.

[0099] Further construct a uniformity response predictor: , in, Indicates the first The first batch Predicted mixing uniformity for each time window, This represents a uniformity response predictor. This indicates its parameters.

[0100] To express the combined effects of mixing energy, shearing action, and overmixing risk on uniformity variation, a dynamic uniformity consistency constraint is constructed: , in, Over-mixed risk weights, , , , This indicates the parameters for calculating mixed risks. This represents the response term to changes in uniformity. , , This represents the weighting coefficients corresponding to power, shear, and load. Indicates the safe shear threshold. This represents the loss of dynamic uniformity.

[0101] The physical decoder and uniformity response predictor obtained through the above processing can be used in the latent space world model to output future state prediction results with engineering constraints.

[0102] 4) Joint training and world model output To address the challenge of balancing prediction accuracy and safety constraints in training with a single loss method, this module integrates state prediction loss, energy consistency loss, boundary constraint loss, and uniformity response loss into the training objective. It also encapsulates a state transition predictor, a physical decoder, a safety barrier function, and an uncertainty estimator to form the module output.

[0103] The state prediction loss, energy consistency loss, boundary constraint loss, and uniformity dynamic consistency loss are combined into the world model training objective: , in, This represents the joint training loss of the world model. , , , These represent energy consistency, boundary constraints, uniformity dynamic consistency, and parameter regularization weights, respectively. This represents all the parameters of the world model that need to be trained.

[0104] The optimal parameters are obtained by minimizing the joint training objective: , in, This represents the optimal parameters of the world model after training.

[0105] To enable direct safety assessment of candidate action sequences in the future, a physical safety barrier function is defined: , in, This represents the physical safety barrier function. The smaller its output value, the closer the corresponding physical state is to the device's operating boundary.

[0106] This module ultimately outputs a physically constrained latent space world model: , in, This represents a physically constrained latent space world model. This represents a latent space state transition predictor. Indicates the physical decoder. This represents a uniformity response predictor. Represents the physical security barrier function. This represents a forecast uncertainty estimator. This represents the optimal parameters of the world model. The output of this module will serve as the input for subsequent modules.

[0107] (4) Module for generating safe diffusion control strategies This module addresses the challenges of high continuity of control actions, strict equipment safety boundaries, and the difficulty of single-step parameter tuning in considering future mixing trends during powder premixing. Its goal is to generate safe, smooth, and homogenization-enhancing candidate control strategies based on a world model. To overcome the limitations of traditional fixed-rule or greedy control methods in multi-step prediction and optimization, this module constructs control conditions, generates conditional diffusion candidate sequences, performs rolling predictions using a world model, screens safety barriers, and scores comprehensive utility results, forming a set of control strategies that can be executed online.

[0108] 1) Construction of control conditions and setting of target constraints To address the lack of state context and objective constraints in the control strategy, this module calls the physics decoder and uniformity predictor in the world model to generate the control condition vector and define the current control objective.

[0109] The physically constrained latent space world model output by the previous module is represented as follows: , in, This represents a physically constrained latent space world model. This represents a latent space state transition predictor. This represents a uniformity response predictor. Indicates the physical decoder. Represents the physical security barrier function. This represents a forecast uncertainty estimator. This represents the optimal parameters of the world model.

[0110] To determine the current starting point for regulation, samples of potential states were used. Extract the current potential mixed state and the previous executed action: , in, Indicates the current potential mixing state. This indicates the action performed in the previous time window. and These represent the state selection matrix and the action selection matrix, respectively.

[0111] Based on the physics decoder and uniformity response predictor in the world model, calculate the current physical state and the current uniformity state: , in, This represents the vector of current physical process quantities. This represents the optimal parameters of the physical decoder. This indicates the current predicted mixing uniformity. This represents the optimal parameters for the uniformity response predictor.

[0112] To enable the diffusion control model to simultaneously perceive the current state, target deviation, equipment state, and previous action, a control condition vector is constructed: , in, Represents the control condition vector. Indicates the first The target mixing uniformity for each batch , , , These represent the mapping matrices corresponding to the potential state, physical state, previous action, and uniformity deviation, respectively. This indicates the bias term.

[0113] The resulting control condition vector, obtained through the above processing, will serve as the condition input for generating subsequent security diffusion candidate actions.

[0114] 2) Generation of candidate regulatory sequences for safe diffusion To address the issue that single-step control cannot adequately consider future mixing trends in advance, this module employs a conditional diffusion generation mechanism, which generates candidate action sequences for multiple future time windows under the guidance of regulatory condition vectors.

[0115] Let the first Group candidate action sequences in the initial diffusion step The normalized noise at that time is expressed as: , in, Indicates the first Initial representation of the diffusion of candidate action sequences. This represents the total number of backward diffusion iterations. Indicates the length of the predictive control time domain. Indicates the action dimension. Indicates the number of candidate action sequences. This represents the unit covariance matrix.

[0116] To incorporate the current regulation objective and safety preference into the denoising process, a conditional noise estimate is constructed: , in, Indicates the first Conditional noise estimation for each diffusion step. This represents a conditional noise prediction network. Indicate its parameters, Indicates the first Candidate sequence representation for each diffusion step.

[0117] To ensure that candidate action sequences move in a direction that improves uniformity and reduces control mutations during the generation phase, a diffusion guidance function is defined: , in, This represents the diffusion guiding function. express The Middle Normalized action vectors for a future time window, This represents the time discount factor. This indicates the weighting of uniformity deviation. Indicates the first The action direction weight vector for each future time window.

[0118] Based on conditional noise estimation and diffusion guidance function, perform inverse denoising update: , in, Indicates the first The candidate sequence representation for each diffusion step, and Represents the diffusion scheduling coefficient. Indicates the intensity of diffused noise. Indicates standard Gaussian noise. This represents the guiding gradient weights.

[0119] When diffusion iterates in reverse... Then, the normalized actions are mapped to the range of actions that the device can execute: , in, Indicates the first The first candidate sequence in the group Candidate actions to be performed within a future time window. and They represent the first The lower limit and upper limit of actions corresponding to each batch. This represents the normalized action vector after diffusion is complete.

[0120] Get the first Group candidate regulatory action sequence: , in, Indicates the first Group candidate regulatory action sequence.

[0121] The resulting set of candidate action sequences, obtained through the above processing, is used for subsequent rolling predictions and safety screening in the world model.

[0122] 3) World model rolling prediction and security barrier screening To address the lack of safety feasibility assessment for candidate actions, this module calls upon the world model for multi-step rolling prediction and filters candidate sequences through physical safety barriers, action change rate constraints, and prediction uncertainty penalties.

[0123] First, the current potential state is used as the initial state for rolling prediction: , in, Indicates the first The initial predicted potential state of the candidate sequence group.

[0124] Based on candidate execution actions Step-by-step space state rolling prediction: , in, Indicates the first Group candidate sequences in the 1st Predicted potential states for a future time window Indicates the current time step.

[0125] For each step of predicting the potential state, perform physical decoding and uniformity response prediction: , in, Indicates the first Predicted physical process quantities for a future time window This indicates the corresponding predicted mixing uniformity.

[0126] To account for model prediction uncertainty in the safety screening process, the state prediction uncertainty under the candidate sequence is calculated: , in, Indicates the first Uncertainty regarding the potential state prediction for each future time window.

[0127] Based on the physical safety barrier function and the rate of change of action constraint, the robust safety margin of the candidate sequence is calculated: , in, Indicates the first Safety margin for the rate of change of action within a future time window Indicates the maximum permissible variation of the action. Represents the infinite norm, Indicates the first Robust safety margin of candidate action sequences This represents the penalty weight for uncertainty.

[0128] Map robust safety margins to candidate policy safety gating values: , in, Indicates the first Safety gating values ​​for candidate action sequences. This represents the slope coefficient of the safety gate.

[0129] The predicted uniformity, predicted physical process quantity, and safety gating value obtained through the above processing are used as the basis for scoring candidate strategies.

[0130] 4) Regulation utility score and strategy outcome output To address the potential risks to equipment or sudden changes in operation that may result from solely pursuing uniformity improvements, this module constructs a joint scoring mechanism that considers quality benefits, operation costs, risk costs, and safety gating, and generates the module's output.

[0131] To measure the improvement effect of candidate sequences on the mixing uniformity objective, a quality gain function is defined: , in, Indicates the first The quality gain of a group of candidate action sequences Indicates the mixed risk threshold. This indicates a mixed penalty weight.

[0132] To suppress excessive action mutations and inefficient energy consumption, the action cost function is defined as follows: , in, Indicates the cost of an action. This represents the action amplitude weight matrix. This represents the weight matrix of the rate of change of action.

[0133] To further penalize candidate sequences that are close to the device boundary and have high prediction uncertainty, a risk cost function is defined: , in, Indicates the risk and cost. This represents the weight of the cost of prediction uncertainty.

[0134] Taking into account the quality benefits, action costs, risk costs, and safety threshold values, a candidate sequence regulation score is obtained: , in, Indicates the first The comprehensive regulatory score of the candidate action sequence. , , These represent the weighting coefficients for quality benefits, action costs, and risk costs, respectively.

[0135] Select the candidate sequence with the highest comprehensive regulation score: , in, Indicates the optimal candidate sequence number. This indicates the final selected sequence of safety control actions.

[0136] To enable subsequent execution of the initial action and online feedback correction, the optimal action sequence, predicted uniformity trajectory, predicted physical state trajectory, safety margin, and scoring results are encapsulated into a control strategy unit: , in, Indicates the first The first batch A time window-based control strategy unit.

[0137] This module ultimately outputs a set of secure diffusion control strategies: , in, This represents the set of safe diffusion control strategies output by this module. The output of this module will serve as the input for subsequent modules.

[0138] (5) Online closed-loop execution and anomaly control output module This module addresses the need to translate safety control strategies into executable equipment commands, and to perform deviation correction and anomaly control based on real-world feedback after execution. Its goal is to achieve closed-loop execution of online control of powder premixing uniformity. To solve the problem that control strategies only rely on offline prediction and cannot handle on-site disturbances and model mismatches, this module uses strategy parsing, equipment command generation, online feedback acquisition, prediction deviation calculation, anomaly identification, and hierarchical control to form the final closed-loop control result.

[0139] 1) Control strategy analysis and equipment instruction generation To address the inconsistency between model actions and device command boundaries and rate of change constraints, this module extracts the optimal strategy unit from the set of safety diffusion control strategies and generates actual issued commands through action selection, command mapping, rate of change limits, and execution gating.

[0140] To uniformly represent the strategy unit output by the previous module, it is written as: , in, Indicates the first The first batch A time-window control strategy unit, This represents the optimal sequence of safety control actions. This represents the corresponding predicted feedback trajectory. This represents the robust safety margin of the optimal sequence. The comprehensive regulation score represents the optimal sequence. This represents the control condition vector.

[0141] To extract the first action to be executed in the current cycle from a multi-step action sequence, a selection vector is defined and the action to be executed is calculated: , in, This indicates the first action to be performed within the current time window. Indicates length is The first choice vector. express 3D identity matrix This indicates matrix vectorization operations.

[0142] To translate model actions into instructions that the device controller can execute, initial device instructions are calculated: , in, Indicates the initial device instruction. and These represent the lower and upper limits of the device command, respectively. , , and Indicates instruction mapping parameters.

[0143] To avoid equipment shock caused by sudden changes in control commands, a rate of change constraint is imposed on the initial equipment commands: , in, This indicates the equipment instruction after the rate of change constraint has been applied. This indicates the device commands from the previous time window. Indicates the maximum allowed instruction variation.

[0144] To implement conservative control when safety margins are insufficient or regulation scores are abnormal, an execution gating is constructed: , in, Indicates the execution gate value, and This represents the gating slope coefficient. This indicates the minimum control score threshold. Indicates a command to maintain or conservatively execute.

[0145] After the above processing, the execution command is obtained as the actual control command issued in the current time window.

[0146] 2) Online feedback collection and prediction deviation calculation To address the lack of closed-loop correction data after control execution, this module collects feedback data such as uniformity, power, shear, temperature, vibration, and load in the next time window, and calculates the normalized prediction deviation, deviation intensity, and feedback compensation amount.

[0147] After execution, construct the online feedback observation vector: , in, This represents the online feedback observation vector for the next time window. This indicates online estimation of mixing uniformity. This indicates the actual stirring power. Indicates the actual shear strength. Indicates motor temperature. Indicates the vibration amplitude. Indicates the load status.

[0148] To obtain the first-step prediction quantity corresponding to the feedback observation, the first-step prediction feedback is extracted from the prediction feedback trajectory: , in, This represents the first-step prediction feedback vector. This indicates the dimension of the feedback observation.

[0149] Calculate the normalized bias between online feedback and prediction results: , in, Represents the feedback prediction bias vector. This represents the normalized feedback prediction bias vector. This represents the diagonal normalized matrix composed of the scaling parameters of each feedback variable.

[0150] To comprehensively assess the degree of prediction bias and insufficient safety margin, the closed-loop bias strength is calculated: , in, Indicates the strength of the closed-loop deviation. Represents the bias weight matrix. This indicates the penalty weight for safety margin.

[0151] Feedback compensation is generated based on the closed-loop deviation strength. , in, This indicates the feedback compensation threshold value. This indicates the deviation threshold that triggers compensation. Indicates the amount of feedback compensation. This represents the equipment instruction compensation scale matrix. , and This indicates the feedback compensation parameter.

[0152] The candidate execution instructions after compensation are obtained: , in, This indicates the candidate execution instruction after compensation.

[0153] The feedback deviations and compensation instructions obtained through the above processing are used for subsequent abnormal risk identification and hierarchical control.

[0154] 3) Abnormal status identification and hierarchical control To address the issue that single threshold alarms cannot distinguish between anomaly types and handling methods, this module integrates feedback observation, prediction deviation, safety margin, control score, and compensation instructions to calculate anomaly probability, risk level, and safety control instructions.

[0155] To standardize the representation of anomaly detection input, anomaly discrimination features are constructed: , in, Indicates anomaly detection features, , , , , and This represents the parameters for anomaly detection.

[0156] Let the set of exception categories be: , in, Represents a set of exception categories. Indicates the first Class of abnormal states, Indicates the number of exception categories.

[0157] Calculate the probability of each anomaly category: , in, Indicates the first The probability of an abnormal state. , and This represents the anomaly classification parameters.

[0158] Determine the current primary anomaly category: , in, This indicates the current primary anomaly category number.

[0159] To unify anomaly probability, prediction bias, and equipment boundary violation into a single risk intensity, the anomaly risk value is calculated: , in, Indicates an abnormal risk value. , , , , , and This represents the parameters used in risk calculation.

[0160] Risk level is determined based on abnormal risk values: , in, Indicates the risk level. Indicates an indicator function, Indicates the number of risk level thresholds. Indicates the first One risk threshold.

[0161] To select control actions based on anomaly type and risk level, the probability of control actions is calculated: , in, Indicates the first The probability of selecting a class of control actions One-hot encoding representing the main anomaly category, Indicates the number of control action categories. , , , and This indicates the parameters for selecting control actions.

[0162] Generate security control instructions: , in, This indicates the safety instructions following the tiered control measures. Indicates the first Pre-defined control commands.

[0163] After the above processing, the anomaly category, risk level, and safety control instructions are obtained for the final closed-loop control result encapsulation.

[0164] 4) Encapsulation and Module Output of Closed-Loop Control Results To address the issue of scattered storage of execution instructions, predicted trajectories, actual feedback, abnormal risks, and control effects, this module unifies and encapsulates control actions, equipment instructions, feedback observations, prediction deviations, anomaly categories, risk levels, and control effectiveness, forming the final output.

[0165] To evaluate the effectiveness of the current control cycle, the improvement in uniformity target is calculated: , in, This represents the target improvement in uniformity during the current control cycle. This represents the online estimated mixing uniformity for the previous time window.

[0166] By combining uniformity improvement, feedback deviation, safety margin, and anomaly risk, the closed-loop control effectiveness score is calculated: , in, This indicates the score for the effectiveness of closed-loop regulation. , , , , and This represents the scoring parameters.

[0167] To form a complete closed-loop control record, a single-cycle output result is constructed: , in, Indicates the first The first batch The results of closed-loop regulation within a time window.

[0168] This module ultimately outputs a set of online closed-loop execution and anomaly control results: , in, This represents the set of online closed-loop execution and anomaly control results output by this module. The output includes safe execution instructions, actual feedback observations, predicted deviations, anomaly risk levels, and closed-loop control effectiveness results, used to support the traceability, verification, and continuous optimization of the online control process for powder premixing uniformity.

[0169] Experimental verification To verify the effectiveness and reliability of the online powder premixing uniformity control system based on a physically constrained world model and safe diffusion decision proposed in this invention in terms of multimodal process perception, potential state characterization, mixing state prediction, safe control strategy generation, and anomaly closed-loop management, a powder premixing multimodal process control dataset was constructed. This dataset originates from typical production scenarios such as powder coatings, fine chemical powders, and functional filler premixing. It includes formulation composition, particle size distribution, feed mass, stirring current, torque, vibration, temperature, load, online images, near-infrared spectroscopy, historical control actions, endpoint offline uniformity detection results, and abnormal operation records, covering typical samples under different powder systems, batch sizes, equipment loads, and mixing stages.

[0170] ① Powder coating samples mainly include premixing process data for resin powder, pigments, fillers, additives, and functional powders, used to verify the system's ability to predict the uniformity of multi-component powders under conditions of particle size and density differences; ② Fine chemical powder samples mainly include mixing process data for highly dispersible powders, easily agglomerated powders, and low-content additives, used to verify the system's ability to identify risks of local agglomeration, undermixing, and overmixing; ③ Functional filler samples mainly include premixing data for conductive fillers, flame-retardant fillers, and composite modified powders, used to verify the system's ability to safely regulate under equipment load changes, shear strength changes, and temperature rise constraints. All samples have undergone production record alignment, sensor anomaly removal, and offline detection result verification to form labeled data containing batch attributes, process observations, control actions, uniformity results, anomaly types, and safety boundary states.

[0171] The multimodal process data in the dataset were first processed through time window alignment and unified state characterization, and then divided into three task levels according to the online powder premixing control task: ① Process state characterization task, including multi-source sensor observation alignment, data credibility assessment, modal token construction, and latent state encoding; ② Mixing state prediction task, including latent space state transition prediction, mixing uniformity prediction, physical process quantity decoding, and prediction uncertainty estimation; ③ Online safety control task, including candidate action sequence generation, safety barrier screening, target uniformity achievement, abnormal state identification, and hierarchical control output. A total of 1350 powder premixing batch samples were constructed, of which 900 were used for model training and parameter adjustment, and 450 were used for validation and testing. The test samples covered different formulation ratios, different initial dispersion states, different sensor missing ratios, different load levels, and different anomaly types.

[0172] To comprehensively evaluate system performance, four comparison methods were selected: ① FixedM, a fixed process parameter mixing control method, sets stirring speed, mixing time, and feeding rhythm based on experience, without dynamic adjustment based on online status; ② PIDT, a PID closed-loop control method based on torque and current feedback, can adjust some control parameters according to equipment load changes, but lacks multimodal state understanding and future process prediction; ③ LSTMM, a control method based on long short-term memory network prediction and model predictive control, can use time-series data to predict uniformity changes and perform rolling optimization, but does not explicitly introduce physical safety barriers and diffusion candidate strategy generation; ④ DiffSafe, a diffusion control method without physical safety constraints, can generate multi-step candidate action sequences, but lacks energy, shear, temperature rise, vibration, and load boundary constraints; ⑤ The system of this invention achieves online control of powder premixing uniformity through multimodal process data input, self-supervised state characterization, physical constraint world model, safe diffusion decision-making, and online anomaly control mechanism.

[0173] All methods were evaluated under the same test samples, initial operating conditions, and uniform safety boundary conditions. To verify the stability of the system under real production disturbance conditions, three types of complexity conditions were set: ① Low complexity samples, with good powder flowability, small differences in formulation components, and complete sensor signals, mainly used to verify the ability to control the uniformity of conventional mixing; ② Medium complexity samples, with differences in particle size distribution, slight agglomeration, equipment load fluctuations, and some sensor noise, mainly used to verify the ability to characterize multimodal states and predict world models; ③ High complexity samples, with difficulties in dispersing low-content components, powder adhesion, sudden load changes, abnormal temperature rise, abnormal vibration, and local sensor loss, mainly used to verify the ability to control safe diffusion and manage abnormal closed loops.

[0174] The following evaluation metrics are used: ① Uniformity prediction error (UE) measures the average absolute deviation between the system's online predicted mixing uniformity and the offline detection results; ② Endpoint coefficient of variation (CV) measures the degree of fluctuation in the endpoint mixing uniformity of different batches of the same formulation; ③ Time to reach target (TT) measures the average mixing time required for the system to reach the target uniformity; ④ Safety deviation rate (OR) measures the proportion of overload, overheating, over-vibration, or shear deviation that occur during the control process; ⑤ Anomaly recall rate (AR) measures the completeness of the system's identification of abnormal states such as undermixing, overmixing, equipment malfunction, and model mismatch; ⑥ Unit energy consumption (EC) measures the normalized energy consumption required to complete a single batch of premixing.

[0175] Experimental results show that the system of this invention exhibits good overall performance in the online control of powder premixing uniformity. The FixedM method is stable and simple to implement, but lacks responsiveness to batch differences and mid-process disturbances, with a final coefficient of variation of 6.9% and a time to reach the target of 31.4 minutes. The PIDT method can make local adjustments based on torque and current changes, shortening the time to reach the target to 25.2 minutes, but its anomaly recall rate is only 68.5%, and it is insufficient in identifying undermixing and sensor drift. The LSTMM method improves the uniformity prediction capability, reducing the uniformity prediction error to 0.061 and the safe out-of-bounds rate to 5.2%, but it still relies on strong model stability under high load and overmixing risk scenarios. The DiffSafe method can generate more aggressive multi-step control actions, shortening the time to reach the target to 20.6 minutes, but due to the lack of physical safety constraints, the safe out-of-bounds rate increases to 8.4%. In comparison, the uniformity prediction error of the system of the present invention is reduced to 0.045, the coefficient of variation of the endpoint is reduced to 3.7%, the time to reach the target is shortened to 18.9 minutes, the safety out-of-bounds rate is reduced to 2.8%, the anomaly recall rate reaches 90.6%, and the unit energy consumption is reduced to 0.86, which is significantly better than the comparative methods such as FixedM, PIDT, LSTMM and DiffSafe.

[0176] Under high-complexity sample testing conditions, the system of this invention maintains good stability. Because this invention unifies batch attributes, multimodal process observations, equipment status, and historical execution actions into a potential state, and uses a physical constraint world model to predict future states and screen safety barriers for candidate actions, it can reduce invalid parameter tuning and equipment out-of-bounds behavior under conditions of low-content, difficult-to-disperse components, sudden load changes, abnormal temperature rise, and local sensor deficiencies. Test results show that the system of this invention exhibits a less than 7.4% increase in the endpoint coefficient of variation in high-complexity samples, while the increases for FixedM, PIDT, LSTMM, and DiffSafe are approximately 18.6%, 16.3%, 12.9%, and 15.7%, respectively. Simultaneously, the safety out-of-bounds rate of the system of this invention in high-complexity samples remains below 3.6%, lower than FixedM's 9.8%, PIDT's 8.7%, LSTMM's 6.4%, and DiffSafe's 10.5%, indicating that the system of this invention has stronger robustness and safety control capabilities in complex powder premixing scenarios.

[0177] In summary, the experimental results verify the effectiveness of the system of this invention in sensing multimodal processes of powder premixing, self-supervised state characterization, prediction using a physical constraint world model, safe diffusion control, and anomaly closed-loop management. This system not only improves the accuracy of mixing uniformity prediction and batch consistency, but also shortens the time to achieve target uniformity, reduces equipment safety violation rates and unit energy consumption, providing a stable, interpretable, and engineering-approved online control technology solution for powder coatings, fine chemicals, and functional powder manufacturing processes.

[0178] Table 1. Comparison of data from different methods under six major indicators. In performance evaluation, the anomaly recall rate (AR) is a positive indicator, used to measure the completeness of the system's identification of abnormal states such as undermixing, overmixing, equipment malfunction, and model mismatch. The uniformity prediction error (UE), the endpoint coefficient of variation (CV), the time to reach the target (TT), the safety out-of-bounds rate (OR), and the unit energy consumption (EC) are negative indicators. UE measures the deviation of online uniformity prediction, CV measures the fluctuation of endpoint uniformity across different batches, TT measures the efficiency of achieving the target uniformity, OR measures the safety of equipment operation, and EC measures the energy consumption required to complete a single batch of premixing.

[0179] The experimental results are shown in Table 1. Figure 2 , Figure 3 , Figure 4 , Figure 5 , Figure 6 , Figure 7 As shown, in the task of online control of powder premixing uniformity, FixedM can ensure the stability of the basic production process, but it is difficult to adapt to changes in powder composition, particle size distribution, and environmental conditions; PIDT can improve the control response under local load fluctuations, but it lacks in-depth characterization of mixing uniformity, overmixing risk, and anomaly types; LSTMM can improve uniformity control capability by utilizing time-series prediction, but it lacks physical process constraints and safety barrier mechanisms; DiffSafe can generate more aggressive candidate action sequences, but it is prone to introducing safety out-of-bounds violations under high load and temperature rise disturbance conditions. In contrast, the system of this invention, by introducing multimodal process data input, self-supervised state characterization, physical constraint world model, safety diffusion decision-making, and online anomaly control mechanism, realizes a complete processing flow from process state perception, future mixing prediction, candidate action generation to safe closed-loop execution, and has better uniformity control capability and equipment safety stability in complex powder premixing scenarios.

[0180] The results show that the uniformity prediction error (UE) of the system of the present invention is reduced to 0.045, which is approximately 45.1% lower than PIDT, approximately 26.2% lower than LSTMM, and approximately 21.1% lower than DiffSafe; the endpoint coefficient of variation (CV) is reduced to 3.7%, which is approximately 46.4% lower than FixedM, approximately 36.2% lower than PIDT, and approximately 19.6% lower than LSTMM; the time to reach the target (TT) is shortened to 18.9 minutes, which is approximately 39.8% shorter than FixedM, approximately 25.0% shorter than PIDTorque, and approximately 21.1% shorter than DiffSafe. The safety margin is shortened by approximately 8.3%; the safety deviation rate (OR) is reduced to 2.8%, which is approximately 59.4% lower than PIDT, approximately 46.2% lower than LSTMM, and approximately 66.7% lower than DiffSafe; the anomaly recall rate (AR) reaches 90.6%, which is approximately 32.3% higher than PIDT, approximately 18.6% higher than LSTMM, and approximately 24.3% higher than DiffSafe; and the unit energy consumption (EC) is reduced to 0.86, which is approximately 14.0% lower than FixedM, approximately 8.5% lower than PIDT, and approximately 5.5% lower than LSTMM. These results demonstrate that the system of this invention has significant advantages in powder premixing uniformity prediction, batch consistency control, target achievement efficiency, equipment safety constraints, anomaly identification, and energy consumption control, verifying its engineering feasibility and practical application potential in intelligent manufacturing and continuous quality control of powder premixing.

[0181] Example 2 This embodiment provides a powder premixing strategy generation system based on physical constraints and diffusion decisions.

[0182] A computer-readable storage medium storing a plurality of instructions adapted for loading and execution by a processor of an end device of the powder premixing strategy generation method based on physical constraints and diffusion decisions.

[0183] A terminal device includes a processor and a computer-readable storage medium, the processor being configured to implement various instructions; the computer-readable storage medium being configured to store multiple instructions adapted for loading and execution by the processor of the powder premixing strategy generation method based on physical constraints and diffusion decisions.

[0184] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for generating powder premixing strategies based on physical constraints and diffusion decisions, characterized in that, include: Acquire multimodal powder premixing data; Data preprocessing was performed based on the acquired multimodal powder premixing data; Self-supervised multimodal state representation based on preprocessed data includes unified process sample input and modal token construction, quality-aware mask self-supervised encoding, action condition temporal consistency and cross-modal alignment, latent state fusion and module output; The physical constraint latent space world model is constructed based on state representation, including latent state sample input and state-action decomposition, latent space state transition prediction and uncertainty estimation, physical constraint decoding and mixed uniformity response modeling, joint training and world model output. Based on the constructed world model, a security diffusion control strategy is generated, including the construction of control condition states and the setting of target constraints, the generation of security diffusion candidate control sequences, rolling prediction of the world model and screening of security barriers, control utility scoring and strategy result output; Online closed-loop execution and anomaly control output are performed based on the output strategy.

2. The powder premixing strategy generation method based on physical constraints and diffusion decisions according to claim 1, characterized in that, The data preprocessing based on the acquired multimodal powder premixing data includes: to address the difficulty in unifying batch information and sensor observations, a unified representation of the material-side and observation-side data of the powder premixing process is achieved through batch set definition, batch attribute vector construction, and multi-source original observation unit recording; and to retain the time attribute, sensor status, and missing information of each observation value, original observation units are constructed. The processed batch attributes and multi-source original observation entry points provide the raw data foundation for subsequent equipment state association and time window alignment. To address the issue of unclear variable sources in subsequent world models of control actions, the actual operating state and executed control parameters are incorporated into the same batch process record through synchronous acquisition of equipment operating state vectors and historical executed action vectors. Furthermore, to resolve the problem of different modalities not being able to directly participate in the same time slice calculation, a time window partitioning, time-sensitive weight allocation, and modal encoding mechanism are employed to generate window-level observation representations, equipment state representations, and historical executed action representations. To ensure that equipment states and historical executed actions maintain the same time scale as the observation modalities, aggregation weights are further defined to obtain window-level equipment state representations and window-level historical executed action representations, achieving alignment and encoding of multi-source asynchronous data into a unified time window. Finally, to address the problem of model instability caused by low-quality data directly entering the state representation module, unified process samples are generated through observation coverage, time synchronization reliability, and comprehensive data reliability calculations, forming the module output.

3. The powder premixing strategy generation method based on physical constraints and diffusion decisions according to claim 2, characterized in that, The unified process sample input and modal token construction include extracting time window process samples from unified process data and constructing modal tokens and process global tokens through a quality enhancement mapping mechanism. Specifically, for the first... For observation modalities, quality-enhanced modal tokens are constructed, and process-wide tokens are further constructed to introduce batch-level, device-level, and action-level global contexts. Then, the global tokens and tokens from each observation modality are combined row-wise to obtain a token matrix. in, Indicates the first The first batch A multimodal token matrix with a time window. The token dimension is represented by the quality-aware mask self-supervised encoding, which includes dynamically adjusting the mask probability based on data credibility and observation quality, and obtaining a self-supervised hidden representation through quality-biased attention and mask reconstruction mechanisms. The data quality vector is represented as: in, Represents the observation coverage component. Indicates the sensor health component. Indicates the effective observation proportion component, Represents the time synchronization reliability component; and calculates the first [component] based on the quality component. The masking probabilities of tokens are calculated; then, masked tokens are generated based on these probabilities. All masked tokens are combined to construct a quality bias attention matrix so that the self-supervised encoding process simultaneously considers intermodal relationships and data quality differences. in, Represents the mass bias matrix. and This represents the mass bias coefficient. Indicates length is A vector of all 1s The identity matrix is ​​represented, and self-attention encoding is performed based on the quality bias matrix; finally, the masked token is reconstructed: in, This represents the reconstructed token matrix. This represents the reconstruction mapping matrix. Representing the reconstructed bias vector, constructing the masked reconstruction loss: in, Indicates the mask reconstruction loss. express The Middle A reconstruction token, This represents a constant used to prevent the denominator from being zero.

4. The powder premixing strategy generation method based on physical constraints and diffusion decisions according to claim 3, characterized in that, The action condition temporal consistency and cross-modal alignment include introducing action condition temporal consistency constraints and combining them with a cross-modal alignment mechanism to enhance the consistent representation capability of different observation modalities in the same mixing stage. Specifically, the first action condition is extracted from the self-supervised hidden representation matrix. To ensure that the historical execution actions and their hidden representations maintain the same dimension, the window-level historical execution actions are first projected: in, This represents the historical execution actions after projection. Represents the action projection matrix. The action projection bias is represented; based on the current global hidden representation and the historical executed actions after projection, the global hidden representation for the next time window is predicted, and then the action conditional temporal consistency loss is constructed; to enhance the consistent representation capability of different observation modalities in the same mixed state, the hidden representations of each modality are normalized and projected: in, Indicates the first Normalized aligned representation of tokens, The alignment projection matrix is ​​represented; finally, the cross-modal consistency weights between each observed modality and the global token are calculated, and a cross-modal contrastive loss is further constructed; the latent state fusion and module output include compressing the multi-source hidden representation into latent state samples through consistency weighting, state fusion attention, and gated fusion mechanisms, and forming the module output. To enable different modalities to adaptively participate in fusion according to their consistency and reliability, the state fusion attention score is first calculated, and combined with the cross-modal consistency weights, the final modal fusion weights are obtained. A fused state representation is generated based on these modal fusion weights. Simultaneously, to balance global process information and multimodal fusion information, a state gating vector is introduced: in, Represents the state gating vector. Represents the gated mapping matrix, This represents the gating bias, which in turn generates the potential state vector: in, Indicates the first The first batch Potential state vectors for each time window , , , and The parameters for generating latent states are represented. To ensure that the evolution relationship of latent states under the influence of historical actions can be directly learned, the latent states, historical actions, data credibility, and time steps are encapsulated as latent state samples, and finally, a set of self-supervised multimodal state representations is output.

5. The powder premixing strategy generation method based on physical constraints and diffusion decision-making according to claim 4, characterized in that, The latent state sample input and state-action decomposition include, to address the issues of unclear sources of action variables and ambiguous state-action interactions in world model training, extracting latent mixed states, historical executed actions, data reliability, and time steps from the state representation set, and constructing a world model input vector, where for the ... The first batch Given potential state samples within a time window, define a selection mapping: in, Indicates a potential mixed state. Indicates historical actions performed. Indicates data credibility. Indicates the time step. , , , These represent the corresponding selection matrix or selection vector, respectively; to enable the action variables to interact with the latent state for computation, a dimensional projection is performed on the historical actions: in, This represents the action representation after projection. Represents the action projection matrix. This represents the action projection bias; then, based on the latent state, projected action, state-action interaction term, and time step, the world model input vector is constructed: in, Represents the world model input vector. This represents element-wise multiplication; the latent space state transition prediction and uncertainty estimation include constructing a transition gating vector to describe the moderating effect of the executed action on the magnitude of state changes: in, Represents the latent space state transition gating vector. This represents the Sigmoid function. Represents the gated mapping matrix, This indicates the gating bias, which in turn generates candidate state increments: in, This represents the predicted potential state increment. Represents the state increment mapping matrix, This represents the state increment bias; combined with the time step and transition gating, the potential state prediction for the next time window is obtained: in, The predicted potential state for the next time window is represented; to reduce over-deterministic prediction, a state prediction variance is constructed; and a confidence-weighted state prediction loss is constructed based on the predicted state and the actual next state.

6. The powder premixing strategy generation method based on physical constraints and diffusion decisions according to claim 5, characterized in that, The physical constraint decoding and hybrid uniformity response modeling include defining a physical decoder to recover physical process quantities from potential states and actions, performing the same physical decoding on the predicted next state, and constructing an energy consistency loss to constrain the relationship between accumulated hybrid energy and power input. in, Indicates energy consistency loss, This indicates the cumulative mixed energy predicted for the next time window. Indicates the first The hybrid energy utilization coefficients corresponding to each batch; to avoid the model generating physical states that do not conform to the equipment safety boundaries, an operational boundary constraint loss is constructed: in, This represents the loss due to running boundary constraints. and These represent the lower and upper limits of shear strength, respectively. Indicates the upper limit of motor temperature. Indicates the upper limit of vibration amplitude. Indicates the maximum load. Represent the linear rectification function; further construct a uniformity response predictor: in, Indicates the first The first batch Predicted mixing uniformity for each time window, This represents a uniformity response predictor. The parameters are represented; finally, to express the combined effects of mixed energy, shearing action, and overmixing risk on uniformity changes, a dynamic uniformity consistency constraint is constructed; the joint training and world model output include, to address the problem that single-loss training cannot simultaneously consider prediction accuracy and safety constraints, combining state prediction loss, energy consistency loss, boundary constraint loss, and dynamic uniformity consistency loss into a world model training objective: in, This represents the joint training loss of the world model. , , , These represent energy consistency, boundary constraints, uniformity dynamic consistency, and parameter regularization weights, respectively. This represents all the parameters to be trained in the world model; then, the optimal parameters are obtained by minimizing the joint training objective. in, The parameters represent the optimal parameters of the world model after training; a physical safety barrier function is defined to facilitate subsequent safety assessment of candidate action sequences. in, Represents the physical security barrier function; the final output is a physically constrained latent space world model: in, This represents a physically constrained latent space world model. This represents a latent space state transition predictor. Indicates the physical decoder. This represents a uniformity response predictor. Represents the physical security barrier function. This represents a forecast uncertainty estimator. This represents the optimal parameters of the world model.

7. The powder premixing strategy generation method based on physical constraints and diffusion decisions according to claim 6, characterized in that, The construction of the control condition state and the setting of the target constraint include, to address the problem of the lack of state context and target constraints in the control strategy, calling the physical decoder and uniformity predictor in the world model to generate the control condition vector and clarify the current control target. First, the physical constraint latent space world model is represented as follows: , in, This represents a physically constrained latent space world model. This represents a latent space state transition predictor. This represents a uniformity response predictor. Indicates the physical decoder. Represents the physical security barrier function. This represents a forecast uncertainty estimator. This represents the optimal parameters of the world model; to determine the current starting point for regulation, it is derived from potential state samples. Extract the current potential mixed state and the previous executed action: in, Indicates the current potential mixing state. This indicates the action performed in the previous time window. and Let these represent the state selection matrix and action selection matrix, respectively; and based on the physics decoder and uniformity response predictor in the world model, calculate the current physical state and the current uniformity state: in, This represents the vector of current physical process quantities. This represents the optimal parameters of the physical decoder. This indicates the current predicted mixing uniformity. The parameters represent the optimal parameters of the uniformity response predictor; to enable the diffusion control model to simultaneously perceive the current state, target deviation, equipment state, and previous action, a control condition vector is constructed: in, Represents the control condition vector. Indicates the first The target mixing uniformity for each batch , , , These represent the mapping matrices corresponding to the potential state, physical state, previous action, and uniformity deviation, respectively. The bias term is represented; the generation of the safe diffusion candidate control sequence includes using a conditional diffusion generation mechanism to generate candidate action sequences for multiple future time windows under the guidance of the control condition vector; let the first... Group candidate action sequences in the initial diffusion step The normalized noise at that time is expressed as: in, Indicates the first Initial representation of the diffusion of candidate action sequences. This represents the total number of backward diffusion iterations. Indicates the length of the predictive control time domain. Indicates the action dimension. Indicates the number of candidate action sequences. Represent the unit covariance matrix; to incorporate the current regulatory objective and safety preference during the denoising process, a conditional noise estimate is constructed: in, Indicates the first Conditional noise estimation for each diffusion step. This represents a conditional noise prediction network. Indicate its parameters, Indicates the first The candidate sequence is represented by a diffusion step; to ensure that the candidate action sequence moves in a direction that improves uniformity and reduces control mutations during the generation stage, a diffusion guidance function is defined: in, This represents the diffusion guiding function. express The Middle Normalized action vectors for a future time window, This represents the time discount factor. This indicates the weighting of uniformity deviation. Indicates the first Action direction weight vectors for each future time window; inverse denoising update is performed based on conditional noise estimation and diffusion guidance function: in, Indicates the first The candidate sequence representation for each diffusion step, and Represents the diffusion scheduling coefficient. Indicates the intensity of diffused noise. Indicates standard Gaussian noise. This represents the guiding gradient weights; when the diffusion iterates in reverse to... Then, the normalized actions are mapped to the range of actions that the device can execute: in, Indicates the first The first candidate sequence in the group Candidate actions to be performed within a future time window. and They represent the first The lower limit and upper limit of actions corresponding to each batch. This represents the normalized action vector after diffusion is complete, yielding the first... Group candidate regulatory action sequence: in, Indicates the first Group candidate regulatory action sequence.

8. The powder premixing strategy generation method based on physical constraints and diffusion decisions according to claim 7, characterized in that, The world model rolling prediction and safety barrier screening includes calling the world model to perform multi-step rolling prediction, and screening candidate sequences through physical safety barriers, action change rate constraints, and prediction uncertainty penalties. The current potential state is first used as the initial state for rolling prediction. in, Indicates the first The initial predicted potential state of the candidate sequence is determined, and the candidate execution action is then performed. The process involves rolling prediction of potential space states at each step; physical decoding and uniformity response prediction for each predicted potential state; and calculating the state prediction uncertainty under the candidate sequence to account for model prediction uncertainty in the safety screening. in, Indicates the first The potential state prediction uncertainty for each future time window is considered, and then the robust safety margin of the candidate sequence is calculated based on the physical safety barrier function and the rate of change of action constraint. in, Indicates the first Safety margin for the rate of change of action within a future time window Indicates the maximum permissible variation of the action. Represents the infinite norm, Indicates the first Robust safety margin of candidate action sequences The uncertainty penalty weights are represented, and finally, the robust safety margin is mapped to the candidate policy safety gating value: in, Indicates the first Safety gating values ​​for candidate action sequences. The slope coefficient represents the safety gating. The regulation utility score and strategy result output include: defining a quality gain function to measure the improvement effect of candidate sequences on the mixing uniformity target; defining an action cost function to suppress excessive action mutations and ineffective energy consumption; defining a risk cost function to further penalize candidate sequences that are close to the equipment boundary and have high prediction uncertainty; and then combining the quality gain, action cost, risk cost, and safety gating value to obtain the candidate sequence regulation score. in, Indicates the first The comprehensive regulatory score of the candidate action sequence. , , These represent the weighting coefficients for quality benefits, action costs, and risk costs, respectively; finally, the candidate sequence with the highest comprehensive control score is selected. in, Indicates the optimal candidate sequence number. This represents the final selected sequence of safe control actions; to execute the first action and perform online feedback correction, the optimal action sequence, predicted uniformity trajectory, predicted physical state trajectory, safety margin, and scoring results are encapsulated into a control strategy unit: in, Indicates the first The first batch The control strategy unit is divided into several time windows, and the final output is a set of safe diffusion control strategies.

9. The powder premixing strategy generation method based on physical constraints and diffusion decisions according to claim 8, characterized in that, The online closed-loop execution and anomaly control output based on the output strategy includes: to address the inconsistency between model actions and equipment command boundaries and rate of change constraints, the optimal strategy unit is extracted from the set of safety diffusion control strategies, and actual commands are generated through action selection, command mapping, rate of change limits, and execution gating; to address the lack of closed-loop correction basis after control execution, feedback data such as uniformity, power, shear, temperature, vibration, and load in the next time window are collected, and normalized prediction deviation, deviation intensity, and feedback compensation amount are calculated; then, to address the difficulty of distinguishing anomaly types and handling methods with single threshold alarms, feedback observation, prediction deviation, safety margin, control score, and compensation commands are integrated to calculate anomaly probability, risk level, and safety control commands; finally, control actions, equipment commands, feedback observation, prediction deviation, anomaly category, risk level, and control effectiveness are uniformly encapsulated to form the final output.

10. A powder premixing strategy generation system based on physical constraints and diffusion decisions, executing the powder premixing strategy generation method based on physical constraints and diffusion decisions as described in claim 1, characterized in that, include: The data acquisition module is configured to acquire multimodal powder premixing data; The preprocessing module is configured to perform data preprocessing based on the acquired multimodal powder premixing data; The representation module is configured to perform self-supervised multimodal state representation based on preprocessed data, including unified process sample input and modal token construction, quality-aware mask self-supervised encoding, action condition temporal consistency and cross-modal alignment, latent state fusion and module output; The model building module is configured to build a physically constrained latent space world model based on state representation, including latent state sample input and state-action decomposition, latent space state transition prediction and uncertainty estimation, physical constraint decoding and mixed uniformity response modeling, joint training and world model output. The strategy generation module is configured to generate security diffusion control strategies based on the constructed world model, including constructing control condition states and setting target constraints, generating security diffusion candidate control sequences, rolling prediction of the world model and screening of security barriers, scoring control utility and outputting strategy results. The execution module is configured to perform online closed-loop execution and anomaly control output based on the output strategy.

Images