Time-aware artificial intelligence is reshaping dynamic control methods and systems in industrial manufacturing.

Abstract

This application provides a time-aware artificial intelligence-based method and system for dynamic control in industrial manufacturing. The method includes: collecting a set of raw data from the industrial manufacturing process every n minutes and performing valid interval verification to continuously accumulate and form a time-series dataset; extracting multiple sets of features from the latest accumulated time-series data every m hours based on expertise in industrial manufacturing; dynamically adjusting the complexity of the time-series-aware model and using the latest extracted features for a new round of training; and using the newly trained time-series-aware model to predict the optimal execution point and acceptable fluctuation range of other parameters for the next m hours. This application employs a rolling cycle mechanism every m hours, enabling autonomous recovery after sensor drift or sudden disturbances. The decision time of this method can be controlled within 60 seconds, which is far more efficient than traditional CFD simulation or dynamic simulation, meeting the real-time control requirements in industrial manufacturing processes.

Description

Technical Field

[0001] This application relates to the field of industrial manufacturing, and more specifically, to methods and systems for dynamic control in industrial manufacturing based on time-aware artificial intelligence. Background Technology

[0002] Fermentation technology, as one of the core pillars of modern biomanufacturing, is widely used in the large-scale production of various biological products such as antibiotics, enzymes, vaccines, amino acids, vitamins, alcohol, dairy products, and organic acids. It holds an irreplaceable position in many industries, including pharmaceuticals, food, agriculture, and chemicals.

[0003] However, despite the significant advantages of fermentation, it is fundamentally a highly complex, strongly coupled, and dynamically changing nonlinear biological system. The fermentation process is influenced by a synergistic effect of numerous factors, such as fluctuations in the composition and quality of raw materials, changes in the metabolic state of microbial strains, adjustments to process equipment operating parameters (e.g., stirring, aeration, and feeding), and external disturbances such as environmental temperature, humidity, and power fluctuations. This results in significant uncertainties and dynamic evolutionary characteristics in both spatial and temporal dimensions, greatly increasing the difficulty of precise control.

[0004] Current industrial fermentation control models are still dominated by fixed formulas and human experience, mainly relying on operators to set static parameters (such as stirring speed, aeration rate, and feeding interval) for process regulation. This traditional "rule-driven, manual intervention" control model has many problems: it cannot cope with dynamic changes in the process, because the requirements for process parameters are significantly different at different stages of fermentation (such as lag phase, logarithmic growth phase, stationary phase, and decline phase), and fixed strategies are difficult to adapt dynamically; it is highly dependent on experience and personnel capabilities, and the control effect is limited by the knowledge level and responsiveness of the operators, which can easily lead to increased fluctuations between different batches; it lacks a feedback closed-loop mechanism, resulting in lag in process regulation, inability to respond promptly to changes in microbial metabolism, and easy loss of adjustment windows; and it is difficult to establish a unified model, as there are conditional heterogeneities between different strains, process routes, and production workshops, which limits the universality and transferability of traditional modeling methods. Summary of the Invention

[0005] In view of the shortcomings of the prior art, the purpose of this application is to provide a time-aware artificial intelligence-based method and system for reshaping dynamic control in industrial manufacturing.

[0006] The first aspect of this application provides a method for dynamic control in industrial manufacturing based on time-aware artificial intelligence, comprising:

[0007] Every n minutes, a set of raw data from the industrial manufacturing process is collected and valid intervals are checked based on first-order principle constraints, and the data is continuously accumulated to form a time-series dataset.

[0008] Every m hours, multiple sets of features are extracted from the latest accumulated time-series dataset;

[0009] Every m hours, based on the current accumulated total amount of raw data, the complexity of the time-series awareness model is dynamically adjusted, and a new round of training is performed using the latest extracted features. The training process includes:

[0010] The output of the time-series-aware model is interpreted as parameters of a probability distribution. The parameters of the time-series-aware model are updated by optimizing the log-likelihood function with respect to the real observation data and adding a penalty term for the first-order principle constraint, thus completing this round of training.

[0011] Using the latest trained time-aware model, predict the optimal execution point and the acceptable fluctuation range of other parameters for the next m hours;

[0012] After the predicted output, new raw data continues to accumulate, and the feature extraction and time-series-aware model rolling training enters the next m-hour cycle.

[0013] Optionally, every n minutes, a set of raw data from the industrial manufacturing process is collected and continuously accumulated to form a time-series dataset, including:

[0014] Every n minutes, taking the batch stage of each reactor as the basic unit, the raw log is sliced ​​into time windows, and the parameters within the time slice are extracted to form a set of raw data.

[0015] A time-series dataset is composed of multiple consecutively collected n-minute raw data points arranged in timestamp order.

[0016] Optionally, the valid interval verification includes:

[0017] Trim all raw data parameters to the device's safe range;

[0018] If any parameter in the original data contains a hard outlier, the original data set will be labeled "early stop".

[0019] For each parameter in each set of raw data, a triplet is set, including: set value, measured value, and deviation;

[0020] Manual operations involving replenishing or removing materials are recorded using timestamps.

[0021] Optionally, if the deviation exceeds the limit twice consecutively, the diagnostic logic is triggered, requiring sensor verification or manual inspection.

[0022] Optionally, every m hours, multiple sets of features are extracted from the latest accumulated time-series data, including:

[0023] Based on the knowledge of growth curve kinetics, the fermentation stage (bio_phase) of each set of raw data is determined, and the stage information is added to the corresponding raw data in the form of a unique heat vector.

[0024] The static encoder reads invariant metadata from the original data and compresses it into static latent variables; the dynamic encoder reads real-time changing data from the original data and compresses it into dynamic latent variables; the static latent variables and the dynamic latent variables are then concatenated.

[0025] The flow rate variables in the raw data are scaled according to the corresponding instantaneous working volume and a derivative channel is added to capture leakage or evaporation.

[0026] The recorded data of the manual operation is masked and binary-labeled to filter out sudden fluctuations triggered by human intervention for the time-series-aware model.

[0027] Optionally, the time-aware model adopts a Transformer-based encoder-decoder architecture, including:

[0028] Input processing layer: It contains a static encoder and a dynamic encoder. The static encoder is responsible for processing metadata that does not change over time and compressing it into static latent variables. The dynamic encoder is responsible for processing real-time time series data and compressing it into dynamic latent variables. The two are concatenated and output.

[0029] Encoder: Composed of multiple layers of Transformer encoder modules stacked together; each Transformer encoder module uses its core self-attention mechanism to capture the intrinsic correlation and long-range dependency between all time points in the sequence output by the input processing layer, generating a context vector that can characterize the state of the entire historical production process.

[0030] Decoder: Composed of multiple stacked Transformer decoder modules; The decoder receives the context vector output by the encoder and generates a serialized prediction result for the next prediction period in an autoregressive manner.

[0031] Output layer: This is a probability distribution layer; for each prediction result, this layer outputs the parameters of a probability distribution as the optimal setting point and acceptable fluctuation range.

[0032] A second aspect of this application provides a time-aware artificial intelligence-based dynamic control system for reshaping industrial manufacturing, comprising:

[0033] Data acquisition module: Every n minutes, it collects a set of raw data from the industrial manufacturing process and performs a valid interval check based on first-order principle constraints, continuously accumulating to form a time-series dataset;

[0034] Feature generation module: Every m hours, extract multiple sets of features from the latest accumulated time-series dataset;

[0035] Model training module: Every m hours, based on the current accumulated total amount of raw data, the complexity of the time-series awareness model is dynamically adjusted, and a new round of training is performed using the latest extracted features. The training process includes:

[0036] The output of the time-series-aware model is interpreted as parameters of a probability distribution. The parameters of the time-series-aware model are updated by optimizing the log-likelihood function with respect to the real observation data and adding a penalty term for the first-order principle constraint, thus completing this round of training.

[0037] Model prediction module: Using the latest trained time-aware model, predict the optimal execution point and the acceptable fluctuation range of other parameters for the next m hours;

[0038] The looping module continues to accumulate new raw data after the prediction output, and then enters the feature extraction and time-series-aware model rolling training for the next m-hour cycle.

[0039] A third aspect of this application provides an industrial manufacturing system, including a perception layer, a decision-making layer, and an execution layer;

[0040] The perception layer serves as the data entry point, continuously receiving time-series datasets and features, and synchronously transmitting them to the decision layer;

[0041] The decision-making layer will adopt any of the time-aware AI-based dynamic control methods for reshaping industrial manufacturing, or the time-aware model trained by the time-aware AI-based dynamic control system for reshaping industrial manufacturing, and encapsulate it as a gRPC microservice. Based on the data input from the perception layer, the decision-making layer will start the gRPC microservice for analysis and reasoning to generate the optimal control strategy for the current production state. At the same time, it will refer to the SOP to ensure the compliance of the optimal control strategy, and then send the control command to the execution layer.

[0042] After receiving control commands from the decision-making layer, the execution layer follows the operating procedures in accordance with the SOP, drives the PLC script to be transformed into specific equipment actions, and feeds back the execution results to the perception layer.

[0043] A fourth aspect of this application provides a terminal, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it can be used to execute any of the time-aware artificial intelligence-based dynamic control methods for reshaping industrial manufacturing, or to run any of the time-aware artificial intelligence-based dynamic control systems or industrial manufacturing systems.

[0044] A fifth aspect of this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, can be used to perform any of the time-aware artificial intelligence-based dynamic control methods for reshaping industrial manufacturing, or to operate any of the time-aware artificial intelligence-based dynamic control systems or industrial manufacturing systems.

[0045] This application presents a time-aware artificial intelligence-based method for reshaping dynamic control in industrial manufacturing. Employing a rolling cycle mechanism every m hours, it can autonomously recover after sensor drift or sudden disturbances. First-order principle constraints are embedded into the time-aware model's inference process, ensuring control safety without sacrificing computational speed. The decision-making time of this method can be controlled within 60 seconds, an efficiency far exceeding traditional CFD or dynamic simulations, better meeting the real-time control needs of industrial manufacturing processes.

[0046] Other technical effects resulting from the additional features will be further illustrated in the corresponding embodiments. Attached Figure Description

[0047] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0048] Figure 1 A flowchart illustrating a time-aware artificial intelligence-based dynamic control method for reshaping industrial manufacturing, according to an exemplary embodiment;

[0049] Figure 2 for Figure 1 A structural diagram illustrating a time-aware artificial intelligence-based dynamic control system in industrial manufacturing, according to an exemplary embodiment;

[0050] Figure 3 A comparison of pH values—optimal, predicted, and measured—is shown according to an exemplary embodiment.

[0051] Figure 4 A comparison of the optimal value, predicted value, and measured value of variable s3 according to an exemplary embodiment;

[0052] Figure 5 A comparison of the optimal value, predicted value, and measured value of variable s4 according to an exemplary embodiment;

[0053] Figure 6 This is a comparison of the optimal value, predicted value, and measured value of variable s2 according to an exemplary embodiment;

[0054] Figure 7 A comparison of the optimal value, predicted value, and measured value of variable s6 according to an exemplary embodiment;

[0055] Figure 8 A comparison of the optimal value, predicted value, and measured value of variable s7 according to an exemplary embodiment;

[0056] Figure 9 This is a comparison of the optimal value, predicted value, and measured value of variable p1 according to an exemplary embodiment.

[0057] Figure 10 A comparison of the optimal value, predicted value, and measured value of variable s5 according to an exemplary embodiment;

[0058] Figure 11 This is a comparison of the optimal value, predicted value, and measured value of variable s1 according to an exemplary embodiment;

[0059] Figure 12 A comparison of the optimal value, predicted value, and measured value of variable s2 (optimization perspective) according to an exemplary embodiment;

[0060] Figure 13 For example, a comparison of the optimal value, predicted value and measured value of variable s3 (optimization perspective) according to an exemplary embodiment;

[0061] Figure 14 This is a comparison of the optimal value, predicted value, and measured value of variable y4 according to an exemplary embodiment. Detailed Implementation

[0062] The present application will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present application, but do not limit the present application in any way. It should be noted that those skilled in the art can make several modifications and improvements without departing from the concept of the present application, and these all fall within the protection scope of the present application. Parts not described in detail in the following embodiments can be implemented using existing technology.

[0063] Terminology Explanation:

[0064] Optimal Setpoint: In this application, "optimal setpoint" refers to the mean (μ) of the probability distribution (e.g., Gaussian distribution) predicted by the output layer of the time-aware model. It represents the ideal value that a certain process parameter is most likely to reach at a future time point, as predicted by the time-aware model.

[0065] Acceptable Fluctuation Band: This refers to the range of values ​​generated based on the standard deviation (σ) of the probability distribution output by the time-series-aware model. The standard deviation (σ) is used to quantify the uncertainty of the time-series-aware model's predictions.

[0066] SOP: Standard Operating Procedure, refers to a document that formalizes the standard steps, requirements, and precautions for a process or operation, guiding personnel to perform tasks according to unified standards. Its core function is to ensure consistency, standardization, and repeatability of operations, reduce human error, improve work efficiency and quality stability, and is also an important tool for ensuring safety and compliance (such as industry regulatory requirements).

[0067] First-order biochemical principles refer to the self-evident fundamental laws and constraints derived from basic physics, chemistry, and biology that must be strictly followed in the manufacturing process. These principles are not learned from data, but rather serve as prior knowledge to guide and constrain artificial intelligence models, ensuring that their predictions and decisions are scientifically feasible and safe.

[0068] In this application, the first-order biochemical principle is specifically embodied in the following points:

[0069] The law of conservation of mass states that the total mass of materials input into a system must equal the sum of the masses of output materials, remaining materials, and consumed materials. For example, the consumption rate of feed (such as glucose) must be related to transformation processes such as cell growth and product synthesis.

[0070] The law of conservation of energy states that the energy input (such as heating) and output (such as heat dissipation and heat of chemical reaction) of a system must be in balance and consistent with the temperature change of the system.

[0071] Fundamental constraint of reaction kinetics: The rate of consumption or formation of a substance cannot be negative. This includes the carbon source injection rate or the nitrogen matrix consumption rate.

[0072] Physical feasibility boundary: Process parameters must be within their physically possible range. For example, pressure cannot be negative; the luminescence index cannot be negative.

[0073] In this application, the first-order biochemical principle is explicitly applied at the following levels to ensure the scientific rationality of system decisions:

[0074] Model Training: During model training, first-order principle constraints (such as "non-negative oxygen uptake rate") are added as a soft penalty term to the composite loss function. If the prediction results of the time-series-aware model violate these basic principles, its loss value will increase, thereby guiding the parameters of the time-series-aware model to optimize in a direction that conforms to scientific principles during backpropagation.

[0075] Predictive filtering: After the time-aware model outputs its predictions, a first-order principle is used as a "physics-based guard" mechanism to filter out settings that are impossible to achieve in reality. For example, the system ensures that the predicted oxygen uptake rate and glucose uptake rate are non-negative.

[0076] Data validation: Starting from the data validity range validation step at the very beginning of the process, the system design is anchored to these first-order principles to ensure that the data input into the time-series-aware model is physically reasonable.

[0077] In this way, the first-order biochemical principle solidifies domain knowledge into the system, enabling it not only to learn the correlations in the data, but also to make safe, reliable judgments that conform to common scientific sense, thereby ensuring the domain consistency and stability of the entire regulation process.

[0078] Different fermentation stages (such as the lag phase, logarithmic growth phase, stationary phase, and decline phase) have significantly different requirements for process parameters, making it difficult for fixed strategies to adapt dynamically. The process is highly dependent on experience and personnel capabilities, with control effectiveness limited by the operator's knowledge and responsiveness, easily leading to increased fluctuations between different batches. The lack of a feedback loop mechanism results in lagging process control, failing to respond promptly to changes in microbial metabolism and easily missing adjustment windows. Furthermore, establishing a unified model is difficult, as conditional heterogeneity exists between different strains, process routes, and production workshops, limiting the universality and transferability of traditional modeling methods. Based on these problems, this application provides a time-aware artificial intelligence-based dynamic control method for reshaping industrial manufacturing to address these issues.

[0079] Reference Figure 1 As shown, a time-aware artificial intelligence-based method for reshaping dynamic control in industrial manufacturing can be implemented using the following steps:

[0080] S100 collects a set of raw data from the industrial manufacturing process every n minutes and performs a valid interval check based on first-order principle constraints, continuously accumulating to form a time-series dataset.

[0081] S200, every m hours, extracts multiple sets of features from the latest accumulated time series dataset;

[0082] S300, every m hours, dynamically adjusts the complexity of the time-series awareness model based on the current accumulated total amount of raw data, and uses the latest extracted features for a new round of training. The training process includes:

[0083] The output of the time-aware model is interpreted as the parameters of a probability distribution. The parameters of the time-aware model are updated by optimizing the log-likelihood function with respect to the real observation data and adding a penalty term for the first-order principle constraint, thus completing this round of training.

[0084] S400 uses a state-of-the-art time-aware model to predict the optimal execution point and acceptable fluctuation range of other parameters for the next m hours.

[0085] After the S500 predicts the output, it continues to accumulate new raw data and enters the next m-hour cycle of feature extraction and time-aware model rolling training.

[0086] Specifically, n can be an integer or a decimal (e.g., 1.5 minutes, 3 minutes), and m can also be an integer or a decimal (e.g., 0.5 hours, 2 hours). To ensure logical matching between the data acquisition frequency and the model training / feature extraction frequency (i.e., each m-cycle can fully contain several n-cycle data acquisitions), the following must be satisfied: when the units of n and m are uniformly converted to "seconds", the number of seconds corresponding to m must be a positive integer multiple of the number of seconds corresponding to n.

[0087] For example: Suppose that in an industrial manufacturing scenario, the data acquisition interval n is set to 1.5 minutes, and the model training and feature extraction interval m is set to 0.5 hours.

[0088] n = 1.5 minutes = 1.5 × 60 = 90 seconds;

[0089] m = 0.5 hours = 0.5 × 3600 = 1800 seconds.

[0090] Calculate the ratio of m to n: 1800 seconds ÷ 90 seconds = 20, the result is a positive integer. That is to say, under this setting, 20 sets of raw data will be collected completely within each 0.5-hour m-cycle (one set is collected every 1.5 minutes), which can ensure that subsequent feature extraction can be based on "complete and continuous time segments" and avoid the model training effect being affected by missing or fragmented data.

[0091] For example: n is 30 minutes, m is 4 hours. Under this setting, within each 4-hour m-cycle, 8 sets of raw data will be collected completely at a frequency of once every 30 minutes, ensuring that subsequent feature extraction and model training can be based on continuous and complete time-series segments, avoiding the impact of missing or fragmented data on the model training effect.

[0092] The constraint loss in S100 to S300 of the above embodiments of this application is anchored layer by layer to the first-order principles of biochemistry, rather than general ML (machine learning) experience. This design differs from general machine learning, which relies solely on data patterns and empirical fitting. Instead, it ensures that the entire process of the time-series-aware model, from input data collection to training and optimization, is constrained by scientific rules in the field of biochemistry, ensuring that the final result conforms to data patterns without violating basic scientific principles.

[0093] To achieve preprocessing and integrity control of raw data, in some specific embodiments of this application, in S100, every n minutes, a set of raw data from the industrial manufacturing process is collected and valid intervals are checked, and the data is continuously accumulated to form a time-series dataset. This can be achieved by the following steps:

[0094] S101, Multi-reactor stratification: The raw log is segmented by reactor ID × batch stage to avoid signal crosstalk between reactors with different designs.

[0095] Specifically, the raw data is stored and processed separately according to "reactor number" and "which batch stage the reactor is currently in" to prevent data from different reactor designs from being mixed together and interfering with the time-series sensing model's learning of the specific stage patterns within a single reactor.

[0096] For example: Reactor A is a 500L stirred reactor and Reactor B is a 1000L tower reactor. Their operating signals (such as temperature and pressure changes) have different patterns. By further segmenting them according to their respective batch stages (such as A feeding and B reacting), the data can be made "pure" and mutual interference can be avoided.

[0097] S102, Validation of valid intervals and marking "early termination": All continuous variables are clipped to the safe interval of the equipment; if a hard outlier occurs, the entire time slice is marked with an Early-Termination label.

[0098] Specifically, a safe range refers to a value that falls within the design tolerance range. A hard outlier refers to an abnormal value that significantly exceeds the reasonable range, such as a temperature suddenly exceeding the equipment's tolerance limit.

[0099] Specifically, "early termination" refers to "Early-Termination." This is a specific label used for data validation, and its meaning is as follows:

[0100] Triggering condition: If, during data acquisition, any parameter value in a set of raw data (or within a specific time slice) is identified as a hard outlier, then that entire set of data will be labeled "early stop." Hard outliers refer to abnormal values ​​that clearly exceed reasonable ranges or the equipment's safe operating range. For example, a temperature reading that suddenly exceeds the equipment's maximum tolerance limit is a hard outlier.

[0101] This label is used to mark data for a specific time period as abnormal or unreliable. By adding this label, problematic data can be identified and filtered out during subsequent training and analysis of the time-series awareness model, thereby ensuring the stability and accuracy of the model's predictions and preventing erroneous data from interfering with production control decisions.

[0102] S103, Setpoint-Measured Value Triplet: Saves a (target, measured, Δ) triplet for each signal (parameter) so that subsequent processes can quantify and control quality.

[0103] S104, Manual Operation SOP Marking: Manual operations such as material replenishment and material extraction are unified into the same set of SOPs and timestamps are added to generate category masks for downstream time-series awareness models.

[0104] It should be noted that in other embodiments, the execution order of steps S102-S104 is arbitrary. The execution order can be S104, S103, S102, or S103, S104, S102, etc., and is not limited to one.

[0105] In the above implementation of this application, S101 effectively avoids signal crosstalk between reactors with different designs (such as different volumes and structures) by segmenting the data according to "reactor ID × batch stage". This ensures that the time-aware model trained for each reactor and its specific process stage has high specificity and accuracy.

[0106] S102 ensures data quality and security. Pruning parameters to the equipment's safe range prevents the time-series-aware model from learning dangerous or impossible operating conditions. Labeling data containing hard outliers as "early stop" establishes a clear abnormal data handling mechanism, guaranteeing the reliability of the training dataset.

[0107] S103 stores a (target, measured, Δ) triple for each parameter, and its core function is to quantify control quality. This allows the time-aware model to learn not only the current state of the process (measured value), but also the system's execution capability and response characteristics (deviation Δ), providing key features for subsequent precise control and fault diagnosis.

[0108] By uniformly labeling manual operations such as material replenishment and extraction, S104 allows the time-series awareness model to clearly distinguish which process fluctuations are caused by external human intervention and which are spontaneous changes within the system. This prevents the time-series awareness model from mistakenly learning manual operations as inherent process patterns, thus improving the interpretability and prediction accuracy of the time-series awareness model.

[0109] After processing through four steps (S101-S104), the raw, mixed industrial logs are transformed into a phase-aligned, reactor-specific, high-quality feature tensor. This tensor has an extremely low missing data rate (<0.5%) and all breakpoints and anomaly handling are clearly traceable. This provides a clean, reliable, standardized, and information-rich "foundation" for subsequent feature engineering and time-aware model training, and is the fundamental guarantee for the successful operation of the entire system.

[0110] The absence of these data preprocessing steps will lead to a series of serious problems, ultimately causing the entire intelligent control system to fail, for example:

[0111] Without S101, data confusion will occur, and the time-aware model will become inaccurate. That is, data from different reactors will be mixed for training, and the time-aware model will try to find an "average" pattern that does not apply to any reactor, resulting in inaccurate predictions for all reactors.

[0112] The absence of S102 can lead to model contamination and dangerous decision-making. Data containing extreme error values ​​(hard outliers) will contaminate the training set, potentially causing the time-series-aware model to learn completely wrong or even dangerous relationships, thus outputting unsafe and illogical control commands.

[0113] Without S103, crucial information will be lost, hindering precise control. The sequential awareness model will be unable to directly acquire information about the quality of control, significantly reducing its learning efficiency and accuracy. Simultaneously, the system will lose its ability to quickly diagnose anomalies based on deviation (Δ).

[0114] The absence of S104 can lead to confusion between "human" and "natural" factors, resulting in incorrect predictions. In other words, the sequence-aware model may misinterpret the transient effects of human intervention as intrinsic characteristics of the process. For example, it might incorrectly assume that concentration fluctuations caused by manual feeding are normal metabolic behaviors of microorganisms, thus making incorrect predictions even without human intervention.

[0115] In summary, the absence of steps S101-S104 leads to a typical "garbage in, garbage out" problem. The AI ​​model will be trained on a low-quality dataset full of noise, errors, and obfuscated information, and its final output will inevitably be unreliable, inaccurate, and insecure. This will prevent the realization of the core advantages designed in this application, such as dynamic regulation, autonomous recovery, and real-time correction.

[0116] After obtaining the phase-aligned, reactor-specific tensor in step S100, further interpretation of the data and integration with biochemical mechanisms and workshop practices are required. In some specific embodiments of this application, in step S200, multiple sets of features are extracted from the latest accumulated time-series data every m hours, which can be achieved through the following steps:

[0117] S201, explicit phase encoding: Each row of raw data is given a bio_phase unique heat vector derived from the growth curve kinetics to avoid the time-aware model from confusing the early oxygen peak with the accumulation of later products.

[0118] Specifically, a "growth curve" is a curve used to describe the growth and reproduction patterns of microorganisms (such as bacteria and yeast) in a specific environment. In biomanufacturing processes such as fermentation, this curve is usually divided into four classic stages:

[0119] Lag phase (or adaptation phase): Microorganisms have just entered a new environment and are adapting, so their growth is very slow.

[0120] Logarithmic growth phase (or exponential phase): Microorganisms are fully adapted to the environment and reproduce rapidly at an exponential rate, during which the number of cells increases dramatically.

[0121] Stationary phase: Due to nutrient consumption and metabolic waste accumulation, the proliferation and death rates of microorganisms reach equilibrium, and the total bacterial count reaches a peak plateau. Many target products (such as antibiotics) are synthesized in large quantities during this stage.

[0122] Decline phase: The environment deteriorates further, the mortality rate exceeds the reproduction rate, and the number of live bacteria begins to decline.

[0123] This application mentions these stages and uses their patterns to guide time-aware models.

[0124] Specifically, “bio_phase” refers to the “biological phase”, and the process of deriving its one-heat vector is as follows:

[0125] Calculate the cell growth rate: First, the cell growth rate within the current window is calculated based on real-time collected data (such as cell concentration, exhaust carbon dioxide concentration, and other key parameters).

[0126] Determine the current stage: Based on the calculated growth rate, the system can infer the current fermentation stage.

[0127] For example, an extremely high and positive growth rate usually indicates that the process is in the "logarithmic growth phase".

[0128] A growth rate close to zero or beginning to decline usually indicates that the process has entered a "stabilization / production period".

[0129] One-hot vector encoding is performed: Once the phase is determined (e.g., "lag phase," "log phase," "stationary phase"), the system converts it into a one-hot vector. This is a standard machine learning encoding method used to transform categorical information into a numerical format that time-series-aware models can process. For example:

[0130] If the stage is a "delay period", the vector could be [1,0,0].

[0131] If the phase is the "logarithmic phase", the vector could be [0, 1, 0].

[0132] If the phase is a "stable period", the vector may be [0,0,1].

[0133] The generated vector will then be added to the corresponding original data row.

[0134] Specifically, if the time-aware model cannot distinguish these stages, it may incorrectly associate the characteristics of early oxygen peaks with the synthesis of later products, leading to prediction bias. S201's explicit stage encoding, by introducing biological stage labels, forces the time-aware model to learn the independent patterns of different stages.

[0135] For example:

[0136] During fermentation, microbial metabolic activities exhibit significant stage-specific kinetic characteristics: the logarithmic growth phase, characterized by rapid cell proliferation and intense respiration, is marked by a significant oxygen consumption peak (dissolved oxygen trough), during which energy metabolism primarily serves biomass accumulation; the stationary phase, on the other hand, is characterized by a slowdown in cell proliferation, with metabolic flux shifting towards target product synthesis, resulting in a relative decrease in oxygen consumption and a product concentration accumulation phase. Without bio-phase identifiers, only superficial phenomena such as "low dissolved oxygen concentration" can be captured, but the underlying biological driving mechanisms cannot be analyzed. This easily leads to confusion between the essential difference between the high oxygen consumption dominated by growth in the logarithmic phase and the low oxygen consumption dominated by product synthesis in the stationary phase, resulting in misjudgments of process states and deviations in regulatory strategies. By introducing bio-phase one-heat encoding vectors (e.g., [0,1,0] for the logarithmic phase, [0,0,1] for the stationary phase), clear stage-specific prior information can be provided for the time-series-aware model, enabling it to learn the specific kinetic laws and optimize objective functions for different physiological stages, thereby achieving mechanistic understanding and precise regulation of the fermentation process.

[0137] S202, dual encoder structure: the static encoder reads invariant metadata, and the dynamic encoder processes the real-time signal from step S100; the two are latently spliced ​​together to capture the interaction.

[0138] Specifically, static encoders handle fundamental information that remains unchanged throughout the entire process (immutable metadata), such as the type of bacterial strain used, the size and shape of the reactor, and the materials used in the equipment. This information constitutes the "inherent conditions" of fermentation, determining the basic characteristics of the process, such as the temperature sensitivity of a particular bacterial strain and the heat dissipation efficiency of a specific reactor.

[0139] Dynamic encoder: Focuses on processing real-time data collected in step S100, such as real-time changes in temperature, pH, and dissolved oxygen. This data represents the "real-time status" of fermentation, reflecting dynamic changes in the process, such as oxygen consumption caused by cell growth and pH fluctuations after feeding.

[0140] Latent space splicing and interactive capture: The two encoders transform the information they process into latent features that the time-aware model can understand, and then merge these two types of latent features. In this way, the time-aware model can learn the relationship between "fixed conditions" and "real-time state"—for example, "under the conditions of this strain + 500L reactor, when the temperature rises to 37°C, dissolved oxygen will drop rapidly," thus more accurately grasping the laws of the fermentation process.

[0141] S203, Dynamic Volume Normalization: All flow variables are automatically proportionally adjusted based on the instantaneous working volume recorded in S100 to ensure the comparability of flow data under different volumes; at the same time, a new volume change rate over time (dV / dt) is added as an independent data channel to monitor and capture abnormal volume changes caused by leakage or evaporation in real time.

[0142] Specifically, automatic scaling refers to scaling the value of the flow rate variable proportionally based on the current instantaneous working volume of the reactor.

[0143] For example, when the liquid volume in the reactor is 500L, the feed flow rate is 10L / h, and the relative ratio is 10 / 500=0.02 (that is, the flow rate is 0.02L / h per liter of volume). If the volume is subsequently changed to 1000L, the same absolute flow rate of 10L / h will be adjusted to 10 / 1000=0.01, which is expressed in a uniform "flow rate per unit volume".

[0144] This adjustment eliminates the interference of volume changes on the actual impact of flow rate, allowing flow rate data at different times and volumes to have a consistent physical meaning, which facilitates the time-series sensing model to accurately capture the correlation between flow rate and reaction process.

[0145] S204, Intervention Mask: A binary label generated in step S104 that allows the time-aware model to filter out sudden fluctuations triggered by the operator.

[0146] For example, when an operator manually adds feed, the concentration of the fermentation broth may suddenly increase; this change is not a result of natural microbial metabolism. By labeling these intervention events in the data (1 indicates intervention, 0 indicates no intervention), the time-aware model can determine which fluctuations are caused by human intervention, thus focusing attention on the natural laws of the system itself and avoiding the erroneous attribution of the effects of human intervention to the inherent characteristics of the fermentation process.

[0147] It should be noted that in other embodiments, the execution order of steps S202 to S204 is arbitrary, and S204, S203, and S202, or S203, S202, and S204 are not uniquely limited.

[0148] In the above embodiment S201 of this application, by adding a bio_phase one-heat vector derived from growth curve kinetics to each row of raw data, the time-aware model can be forced to learn the independent patterns of different biological stages. This can effectively prevent the time-aware model from confusing the high oxygen consumption peak caused by rapid cell proliferation in the early stage with the conditions of the later product accumulation stage.

[0149] S202 employs a dual encoder structure, with static and dynamic encoders to process unchanging metadata and real-time changing data, respectively. By concatenating the latent variables compressed from both, the time-series awareness model can effectively capture the interaction between fixed "inherent conditions" (such as reactor size) and "real-time states" (such as temperature fluctuations), thereby more accurately grasping the fermentation process.

[0150] S203 scales all flow variables according to the instantaneous working volume, ensuring the comparability of flow data under different volumes. Simultaneously, the newly added dV / dt data channel can be used to monitor and capture abnormal volume changes caused by leakage or evaporation in real time.

[0151] S204 utilizes manually generated binary tags, allowing the time-series-aware model to effectively filter out sudden fluctuations triggered by operators. This prevents the time-series-aware model from incorrectly attributing the impact of human intervention to inherent characteristics of the fermentation process.

[0152] In the embodiments described above, through steps S201-204, the system further processes the preprocessed data, incorporating biochemical mechanisms and practical workshop experience, ultimately generating an interpretable and information-rich feature tensor. This high-quality feature tensor enables the time-series-aware model to gain a deeper understanding of process dynamics, forming the foundation for achieving high-precision prediction and reliable control.

[0153] The absence of feature engineering steps S201-S204 will directly lead to a significant decrease in the performance and reliability of the time-series-aware model, for example:

[0154] Without S201, the model will confuse the process stages, meaning the time-series sensing model will be unable to distinguish between different fermentation stages, potentially confusing early oxygen peaks with later product accumulation characteristics, leading to incorrect judgments at critical production stages.

[0155] Without S202, the interaction effects cannot be understood, meaning the sequential awareness model will be unable to learn the relationship between static metadata and dynamic real-time signals. For example, it will have difficulty understanding why the same temperature change will have different effects in reactors of different specifications.

[0156] Without S203, flow data becomes less comparable and anomalies cannot be detected. Flow data at different volumes will lose consistent physical meaning, interfering with the learning of the time-series sensing model. Simultaneously, the system will lose an effective channel for capturing volume anomalies such as leaks or evaporation.

[0157] Without S204, it is impossible to distinguish between human-induced and natural fluctuations. In other words, the time-series perception model will mistakenly identify sudden fluctuations triggered by human intervention (such as manual material replenishment) as natural changes in the process, thus learning incorrect causal relationships and affecting the accuracy of its predictions.

[0158] As can be seen, S201-S204 constitute a complete setup. If any of these feature engineering steps are missing, the time-series-aware model may learn based on raw time-series data that has not been deeply processed and lacks domain knowledge background. This will result in the time-series-aware model's understanding of complex industrial processes being superficial, and its prediction accuracy, generalization ability, and final control effect will be greatly reduced, failing to meet the needs of reliable and efficient control in actual production.

[0159] Based on the interpretable feature tensor obtained in step S200, which incorporates biochemical mechanisms and workshop practices, a time-series-aware model is constructed. Step S300 can specifically employ the following steps:

[0160] S301, the Transformer-based encoder-decoder architecture first determines the core architecture of the time-aware model, specifically including:

[0161] Input processing layer: It contains a static encoder and a dynamic encoder. The static encoder is responsible for processing metadata that does not change over time and compressing it into static latent variables. The dynamic encoder is responsible for processing real-time time series data and compressing it into dynamic latent variables. The two are concatenated and used as the input of the main body of the time series-aware model.

[0162] Encoder: Composed of multiple layers of Transformer encoder modules stacked together; each Transformer encoder module uses its core self-attention mechanism to capture the intrinsic correlations and long-range dependencies between all time points in the sequence output by the input processing layer, generating a context vector that can characterize the state of the entire historical production process.

[0163] Decoder: Composed of multiple stacked Transformer decoder modules; The decoder receives the context vector output by the encoder and generates a serialized prediction result for the next prediction period in an autoregressive manner.

[0164] Output layer: This is a probability distribution layer. For each prediction result, this layer outputs parameters of a probability distribution, such as the mean (μ) and standard deviation (σ) of a Gaussian distribution. The predicted mean (μ) serves as the "optimal setting" for this parameter, while the standard deviation (σ) is used to quantify the uncertainty of the prediction, thereby generating an "acceptable fluctuation band".

[0165] S302, Adaptive Capacity Control

[0166] After determining the architecture, the parameters of the Transformer model, such as the number of encoder and decoder layers, the number of attention heads, and the model embedding dimension, are dynamically adjusted based on the amount of raw data obtained in step S100. This is done to balance the expressive power of the temporal awareness model with the risk of overfitting and to ensure that the network size matches the amount of data.

[0167] For example: "Total raw data" is measured in terms of the number of completed production batches:

[0168] Phase 1 (Data volume < 10 batches): In the initial stage of system deployment, there is very little available training data. To prioritize preventing overfitting of the time-aware model, the system automatically configures the time-aware model to a lightweight version:

[0169] Number of encoder / decoder layers: 1 layer each

[0170] Number of heads to focus on: 2

[0171] Model embedding dimension: 64

[0172] Phase Two (10 batches ≤ Data Volume < 50 batches): As the data volume increases, the time-series-aware model requires stronger expressive power to learn complex production patterns. At this point, the system automatically upgrades the time-series-aware model to the standard version to seek a balance between fitting ability and generalization ability.

[0173] Number of encoder / decoder layers: 2 layers each

[0174] Number of heads to focus on: 4

[0175] Model embedding dimension: 128

[0176] Phase 3 (Data volume ≥ 50 batches): After the system has accumulated a large amount of historical data, it becomes capable of supporting deeper and wider network structures to capture more subtle patterns in the data. The system automatically adjusts the time-series-aware model to an enhanced version:

[0177] Encoder / Decoder Layers: 4 layers each

[0178] Number of heads to focus on: 8

[0179] Model embedding dimension: 256

[0180] By employing this phased, automated capacity control strategy based on the amount of raw data, we can ensure that the time-series-aware model matches the amount of data at different stages of data accumulation, thereby maximizing the learning and expressive capabilities of the time-series-aware model while avoiding overfitting.

[0181] S303, the probabilistic loss function defines the training objective.

[0182] The training objective is defined as minimizing a composite loss function. This function consists of two parts: a negative log-likelihood (NLL) loss term and a soft penalty term constrained by the first principle. The time-aware model no longer predicts just a single value, but instead predicts the probability distribution of each parameter at future time points (e.g., the mean μ and standard deviation σ of a Gaussian distribution).

[0183] Specifically, the composite loss function (Ltotal) can be expressed by the following formula:

[0184] ;

[0185] Where: L total It is the total loss value that needs to be minimized.

[0186] The first term is the negative log-likelihood (NLL) loss: N is the total number of data points. i It is the actual observed value of the i-th data point. i and σ i These are the mean and standard deviation of the Gaussian distribution predicted by the time-aware model for the i-th data point, respectively.

[0187] The second term is the first-order principle constraint penalty term: λ(lambda) is a hyperparameter used to balance the importance of the NLL loss and the penalty term. P represents a set of predicted values ​​output by the time-aware model. C k (P) is a function used to evaluate whether the predicted value P of the time-aware model violates the k-th first-order principle constraint. If it violates the constraint (e.g., the predicted oxygen uptake is negative), C k The value of (P) is positive, thus incurring a penalty; if no violation occurs, its value is zero or negative, max(0,C) k If (P) is zero, no penalty is imposed.

[0188] Traditional loss functions (such as mean squared error, MSE) can only enable time-series-aware models to learn and predict a single, deterministic value (mean μ), and cannot assess the confidence level of the prediction. This application, however, employs the aforementioned composite loss function for training the time-series-aware model. By minimizing the mean loss function (NLL), it drives the time-series-aware model to simultaneously learn and predict both the mean (μ) and standard deviation (σ), thereby quantifying uncertainty and improving control robustness.

[0189] Meanwhile, in the composite loss function described above in this application, the standard deviation σ directly quantifies the degree of uncertainty of the time-series sensing model for each predicted value. This enables the system to provide not only the "optimal setpoint" (μ) but also a dynamic "acceptable fluctuation band" (determined by σ), making field control more flexible and robust, and able to tolerate normal fluctuations without affecting the quality of the final product.

[0190] Existing AI models that rely solely on data-driven approaches may learn spurious correlations that contradict physical or biochemical laws, leading to unrealistic or even dangerous control commands (such as predicting negative material concentrations or reaction rates). The composite loss function described in this application, by superimposing a penalty term based on first-order principle constraints, penalizes any predictions that violate fundamental scientific laws (such as conservation of mass, energy, and non-negativity of reaction rates) during training. This design embeds domain knowledge into the time-series-aware model, fundamentally ensuring that the decisions output by the model possess physical and biochemical rationality. This significantly improves the reliability and safety of the entire dynamic control system, thereby guaranteeing the physical rationality of the prediction results and enhancing system security.

[0191] Furthermore, the composite loss function described above in this application organically combines data-driven (NLL term) and mechanism-constrained (penalty term). The NLL term ensures that the time-series-aware model can fit the complex dynamic patterns in the real observation data to the greatest extent, while the penalty term defines the scientific boundaries for it. The time-series-aware model trained in the end can accurately capture subtle changes in the production process without making judgments that violate basic scientific common sense. It achieves an effective balance between the flexibility of data-driven and the stability of the mechanism-based time-series-aware model, thus taking into account both the accuracy of data-driven and the reliability of the mechanism-based time-series-aware model.

[0192] S304, Phase-consistent course learning.

[0193] This step involves implementing a training strategy, which involves sampling data in windows according to the stage labels in step S201, training sequentially according to the fermentation stages, accelerating the convergence of the time-series-aware model through a "from easy to difficult" learning approach, and enhancing its generalization ability for long-term predictions.

[0194] In practical applications, the time-series sensing model trained according to the above embodiments of this application maintains an error of less than 8% between its predictions and actual measurements during the fermentation process of new batches that were not trained. This demonstrates that the time-series sensing model has strong generalization ability and can stably adapt to different batch production scenarios, rather than simply overfitting to historical data. Furthermore, the inference speed of the time-series sensing model is at least 1000 times that of traditional physical simulators. This means that while traditional simulators require a considerable amount of time (e.g., several minutes or even longer) to complete a prediction, the time-series sensing model can output results in milliseconds. This combination of high precision and high speed enables the time-series sensing model to provide reliable predictions for the dynamic control of the fermentation process and to meet the real-time decision-making needs of industrial production. It effectively solves the speed bottleneck of traditional simulators and lays the foundation for efficient and precise production control.

[0195] Based on the same technical concept, in some specific embodiments of this application, a time-aware artificial intelligence-based dynamic control system 100 for reshaping industrial manufacturing is described, such as... Figure 2 As shown, it includes:

[0196] Data acquisition module 110: Every n minutes, it collects a set of raw data from the industrial manufacturing process and performs a valid interval check based on first-order principle constraints, continuously accumulating to form a time-series dataset;

[0197] Feature generation module 120: Every m hours, based on the expertise of industrial manufacturing, it extracts multiple sets of features from the latest accumulated time series data;

[0198] Model Training Module 130: Every m hours, based on the current accumulated total amount of raw data, dynamically adjust the complexity of the time-series awareness model and use the latest extracted features for a new round of training; the training process includes:

[0199] The output of the time-aware model is interpreted as parameters of a probability distribution. The parameters of the time-aware model are updated by optimizing the log-likelihood function with respect to the real observation data and adding a penalty term of the first-order principle constraint, thus completing this round of training.

[0200] Model prediction module 140: Using the latest trained time-aware model, predict the optimal execution point and the acceptable fluctuation range of other parameters for the next m hours;

[0201] Circular rolling module 150: After the prediction output, it continues to accumulate new data and enters the next m-hour cycle of feature extraction and time-series-aware model rolling training.

[0202] The specific implementation techniques of each module / unit in the above embodiments of this application can be referred to the steps of the dynamic control method for reshaping industrial manufacturing based on time-aware artificial intelligence in the above embodiments, which will not be repeated here.

[0203] In some specific embodiments, the aforementioned time-aware artificial intelligence-based dynamic control methods or systems for reshaping industrial manufacturing are deployed and autonomously controlled to form a specific industrial manufacturing system, employing a three-layer runtime stack:

[0204] The perception layer, as the data entry point, continuously receives time-series datasets and features that have passed the valid interval verification; and synchronously transmits this data to the decision layer.

[0205] The decision layer, based on the data input from the perception layer, calls a trained time-series awareness model encapsulated as a gRPC microservice (gRPC microservices refer to a microservice architecture implemented based on the gRPC framework (using Protobuf to define interface contracts and HTTP / 2 to achieve high-performance transmission), to analyze and reason, generate the optimal control strategy for the current production state, and ensure the compliance of the strategy by referring to the SOP, and then issue control instructions to the execution layer;

[0206] After receiving control commands from the decision-making layer, the execution layer strictly follows the operating procedures in accordance with the SOPs to drive the PLC scripts (PLC scripts are program codes that run in a programmable logic controller (PLC) to define and execute specific industrial control logic and automation processes) into specific equipment actions, and feeds back the execution results to the perception layer, forming a cycle of "perception → decision → execution → re-perception" to achieve dynamic closed-loop control of the production process.

[0207] In some specific implementations, the industrial manufacturing system is designed with two API interfaces: REST and GraphQL, for external MES (Manufacturing Execution System) or LIMS (Laboratory Information Management System) to call. External systems can achieve three types of interactions through these interfaces: calling the / predict interface to obtain the predicted results of the production process, calling the / optimize interface to obtain the optimized control scheme, or subscribing to onControlUpdate to receive real-time updates of system control commands.

[0208] The response latency of these interfaces can be controlled within 200 milliseconds, supporting real-time HMI (Human-Machine Interface) display, ensuring fast and synchronous information transmission between external systems and this system, and meeting the needs of real-time monitoring and collaborative management in the production process.

[0209] In some specific implementations, a self-healing stabilization cycle mechanism is also set up. Specifically, if the difference between the set value and the measured value in the data obtained by the sensing layer exceeds the normal range for two consecutive cycles, the system will automatically trigger the diagnostic logic defined in step S102 and then issue a suggestion: either check whether the sensor is faulty or arrange for manual on-site inspection, so as to promptly identify abnormalities and maintain the stability of the production process.

[0210] This mechanism forms a self-healing cycle that can quickly respond to anomalies by monitoring parameter deviations in real time and automatically initiating diagnostic processes, reducing production fluctuations caused by sensor failures or on-site problems.

[0211] In some specific implementations, each completed production batch, after data cleaning, is automatically stored in the database to continuously enrich the training samples. During non-production hours at night, the system uses this new data to retrain the time-series awareness model in step S300 and generates a "canary version" for testing. Only when the key performance indicators (KPIs) of this new version of the time-series awareness model fluctuate by no more than 0.25% will it be officially upgraded to the production environment version. This continuous learning pipeline ensures that the time-series awareness model can continuously absorb new production data for iterative optimization, while a strict performance verification mechanism avoids the impact of upgrades on production stability, achieving continuous improvement and safe iteration of the time-series awareness model's capabilities.

[0212] In the embodiments described above, the industrial manufacturing system is a closed-loop service system built on cloud-native technology. This system integrates the core achievements of the three previous steps: leveraging the data cleanliness ensured by S100 (removing noise, outliers, etc., to ensure data reliability), combining the domain characteristics extracted by S200 (such as fermentation stage, equipment parameters, and other professional information), and the efficient constrained prediction capability achieved by S300 (accurately predicting production status and conforming to process principle limitations), ultimately achieving stable operation and high-yield targets in the production process.

[0213] The embodiments described above in this application exhibit domain consistency: from valid interval verification to constraint loss, each layer is anchored to the first-order principles of biochemistry, rather than relying on general machine learning experience. They also demonstrate small-sample resilience: staged learning and a dual-encoder design enable the time-series-aware model to maintain high performance with only about 5% of the usual data volume. Furthermore, they offer sub-minute real-time correction: the Δ triplet logic combined with a self-healing loop results in end-to-end latency of less than one minute. Finally, they ensure regulatory compliance: SOP marking and a dual-interface design guarantee seamless implementation under my country's GMP (Good Manufacturing Practice) framework.

[0214] The preferred features in the above embodiments can be used individually in any embodiment, or in any combination thereof, provided they do not conflict with each other. Furthermore, parts not described in detail in the embodiments can be implemented using existing technologies.

[0215] The following examples and comparative examples will be used to further illustrate this application in order to better understand the above-mentioned technical solutions. It should be understood that the following are only some examples and are not intended to limit this application.

[0216] Application Example 1:

[0217] The program used in this application example was written in Python 3.6 and ran on a Linux server (CPU: Intel Xeon E5-2685 v3, 24 cores; 100GB memory). All comparative tests were conducted under the same hardware conditions.

[0218] This embodiment specifically describes how ManuDrive, based on time-aware artificial intelligence, reshapes the dynamic control process in industrial manufacturing. The following example illustrates how ManuDrive executes a fully automated, four-hour cycle inference loop in a 5L penicillin fed fermentation process. Specifically, it includes the following steps (1)-(4):

[0219] (1) Data preprocessing and integrity control

[0220] Real-time data collection: The Python OPC-UA client subscribes to the plant's historical database and writes approximately 16 tags to a memory buffer every 30 minutes. The 16 tags are shown in Table 1.

[0221]

[0222] Among them, items 1-5 are process control variables, i.e. the execution quantities that will be of concern later; items 6-10 are key status variables; items 11-13 are cumulative consumption; and items 14-16 are derived / calculated features.

[0223] Validation of valid intervals and early termination marking: All new values ​​were soft-pruned and hard-outlier triggered by the validated physical and chemical range (e.g., light flux intensity 1000–5000 µmol / m² / s, culture medium pH 6.5–8.5, metabolic heat output >0 W) to trigger the Early-Termination marking of the 30-minute fragment.

[0224] Point setting - actual measurement - deviation triplet: Save for each label The four directly quantifiable parameters (eddy current rate, carbon source injection rate, luminous flux intensity, and nitrogen matrix replenishment rate) are highlighted.

[0225] SOP labeling: For example, manual operations such as replenishing catalyst A and calibrating photon sensors are intercepted by MESAPI and written into a category mask to help the time-series awareness model distinguish between human intent and process drift.

[0226] Through the above steps (1), we can obtain a phase-aligned time series buffer that is updated every 30 minutes, and the error processing is deterministic and traceable.

[0227] (2) Automatic generation of domain features

[0228] In this embodiment, the automatic generation of domain features is set to be triggered once every 4 hours, specifically including the following operations:

[0229] Window extraction: At times 0h, 4h, 8h... the system extracts the most recent eight rows (4 hours) of data.

[0230] Explicit phase coding: Calculate the unique heat vector of the culture growth phase (stagnation / exponential / production) based on the rate of change of optical density within the window.

[0231] Volume dynamic normalization: All flow variables are divided by the instantaneous fermentation broth volume and dV / dt is added; this can promptly expose leaks or evaporations.

[0232] Static-Dynamic Dual Encoding Concatenation: Static metadata (strain, impeller, working volume) is merged with dynamic tensors to form inference input.

[0233] Through the above step (2), a feature tensor of 4 hours can be obtained.

[0234] (3) Temporal model reasoning

[0235] Model loading: The registry retrieves a Transformer model containing two encoder layers and two decoder layers. In this embodiment, the model dimension is 128, and each layer has four attention heads.

[0236] Prediction content: The time-aware model outputs the optimal setpoints for four execution quantities in the next 4 hours, as well as the acceptable fluctuation bands for all other tags. The four execution quantities include eddy current rate (rpm), carbon source injection rate (L / h), light flux intensity (µmol / m² / s), and nitrogen matrix replenishment rate (mL / h).

[0237] Physical protection: Filter out infeasible settings using first-order principles, such as quantum yield efficiency not being negative and carbon source injection rate not being negative.

[0238] Latency: 42 milliseconds from feature generation to inference completion, running on the edge box. In this embodiment, the edge box is a 4-core N5105.

[0239] Through the above step (3), the timestamped setting point and acceptable range can be obtained for the next step of execution.

[0240] (4) Cyber-physical execution closed loop

[0241] Point-to-point distribution: The inference results are sent to the field PLC via the ManuDrive control API.

[0242] PLC Action: The PLC writes the new setting point into the actuator PID module and limits other variables to the prediction range.

[0243] Feedback monitoring: Real-time sensor data will be continuously added to the historical database for the next four hours; if Δ(set point, actual measurement) exceeds the limit for two consecutive cycles, the automatic diagnostic program will prompt sensor calibration or manual inspection.

[0244] The loop continues: when the next four hours arrive, the process returns to step (2) until the batch ends.

[0245] Step (4) outputs a completely unattended rolling closed loop, with a final luminescence index increase of 5.9%, a batch time reduction of 6.3%, and a total sensor-to-execution delay of <55 seconds (more than 1000 times faster than CFD / kinetics combined calculation).

[0246] like Figures 3-14 The figure shown is a data graph of predicted and actual values ​​during the process described in the above embodiments of this application. The three data lines in the figure represent different information during the process, as detailed below:

[0247] The red data line (actual data) represents the real-time measured values ​​of various parameters collected by physical sensors (such as thermometers, pH meters, etc.) during the production process. This line represents objective facts and serves as the "benchmark" or "factual basis" for evaluating the accuracy of the time-series sensing model's predictions.

[0248] The blue data line (prediction data 1) represents the preliminary prediction results made by the time-series-aware model based on historical data for a future period. It mainly demonstrates the time-series-aware model's understanding and learning ability of the basic trends of the process, but has not yet undergone full optimization and calibration of the system in this application.

[0249] The green data line (optimized prediction data 2) represents the final "optimal control strategy" after complete processing by the method of this application embodiment. It is an optimized setpoint or acceptable range generated based on the initial prediction (blue data line), further incorporating first-order biochemical principle constraints and SOP compliance checks. This green line is the target trajectory that the system recommends the execution layer to actually operate on, and it is also the core embodiment of the value of this application.

[0250] In addition, Table 2 shows the performance evaluation results of the time-aware model on several independent production batches.

[0251]

[0252] In Table 2 above, each row is identified by a filename (e.g., A403240535), representing a complete and independent production batch data record from a specific reactor. The columns are hierarchical, showing the performance metrics of the time-aware model at different prediction time points. The top-level time columns (24h, 48h, 96h, 120h) represent checkpoints for predicting the future at different time lengths of production.

[0253] The meanings of each item in the secondary indicator column are as follows:

[0254] NLL_total (Total Negative Log-Likelihood): This is the main loss function metric optimized during the training of time-aware models. It measures how well the probability distribution predicted by the time-aware model matches the actual data. The lower the NLL value, the closer the predicted distribution of the time-aware model is to the actual situation, and the better the performance of the time-aware model.

[0255] RMSE_total (Root Mean Square Error): This is a supplementary and more intuitive performance metric. It measures the average error between the mean of the distribution predicted by the time-aware model and the actual measured value. Although the time-aware model is optimized on an NLL basis, RMSE can intuitively reflect how accurate the predicted center point is.

[0256] Table 2 illustrates the performance levels achieved by the time-series-aware model after training on historical data from different (selected) reactors, with the "training loss" directly represented by the NLL_total metric. The table also provides RMSE as a supplement, together forming a quantitative evaluation of the training effectiveness of the time-series-aware model.

[0257] Based on the same technical concept, in other embodiments of this application, a terminal is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it can be used to execute the dynamic control method for reshaping industrial manufacturing based on time-aware artificial intelligence as described in any of the above embodiments, or to run the dynamic control system for reshaping industrial manufacturing based on time-aware artificial intelligence as described in any of the above embodiments.

[0258] Based on the same technical concept, in other embodiments of this application, a computer-readable storage medium is provided, on which a computer program is stored. When executed by a processor, the program can be used to execute the above-described dynamic control method for reshaping industrial manufacturing based on time-aware artificial intelligence, or to run the above-described dynamic control system for reshaping industrial manufacturing based on time-aware artificial intelligence.

[0259] Optionally, the memory is used to store programs; the memory may include volatile memory, such as random-access memory (RAM), such as static random-access memory (SRAM), double data rate synchronous dynamic random-access memory (DDR SDRAM), etc.; the memory may also include non-volatile memory, such as flash memory. The memory is used to store computer programs (such as application programs, functional modules, etc. that implement the above methods), computer instructions, etc., and the aforementioned computer programs, computer instructions, etc., can be partitioned and stored in one or more memories. Furthermore, the aforementioned computer programs, computer instructions, data, etc., can be accessed by the processor.

[0260] The aforementioned computer programs, computer instructions, etc., can be stored in partitions within one or more memory locations. Furthermore, the aforementioned computer programs, computer instructions, data, etc., can be accessed by a processor.

[0261] A processor is used to execute a computer program stored in memory to implement the various steps of the methods involved in the above embodiments. For details, please refer to the relevant descriptions in the preceding method embodiments.

[0262] The processor and memory can be separate structures or integrated structures. When the processor and memory are separate structures, they can be coupled together via a bus.

[0263] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0264] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0265] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0266] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0267] The foregoing has described some specific embodiments of this application. It should be understood that this application is not limited to the specific embodiments described above, and those skilled in the art can make various modifications or variations within the scope of the claims, which do not affect the substantive content of this application. The above-described preferred features can be used in any combination without conflict.

Claims

1. A time-aware artificial intelligence-based method for reshaping dynamic control in industrial manufacturing, characterized in that, include: Every n minutes, a set of raw data from the industrial manufacturing process is collected and valid intervals are checked based on first-order principle constraints, and the data is continuously accumulated to form a time-series dataset. Every m hours, multiple sets of features are extracted from the latest accumulated time-series dataset; Every m hours, based on the current accumulated total amount of raw data, the complexity of the time-series awareness model is dynamically adjusted, and a new round of training is performed using the latest extracted features. The training process includes: The output of the time-series-aware model is interpreted as parameters of a probability distribution. The parameters of the time-series-aware model are updated by optimizing the log-likelihood function with respect to the real observation data and adding a penalty term for the first-order principle constraint, thus completing this round of training. Using the latest trained time-aware model, predict the optimal execution point and the acceptable fluctuation range of other parameters for the next m hours; After predicting the output, new raw data continues to accumulate, and the feature extraction and time-series-aware model rolling training enters the next m-hour cycle. Every m hours, multiple sets of features are extracted from the latest accumulated time-series dataset, including: Based on the knowledge of growth curve kinetics, the fermentation stage (bio_phase) of each set of raw data is determined, and the stage information is added to the corresponding raw data in the form of a unique heat vector. The static encoder reads invariant metadata from the original data and compresses it into static latent variables; the dynamic encoder reads real-time changing data from the original data and compresses it into dynamic latent variables; the static latent variables and the dynamic latent variables are then concatenated. The flow rate variables in the raw data are scaled according to the corresponding instantaneous working volume and a derivative channel is added to capture leakage or evaporation. Manually recorded data is masked and binary-labeled to filter out sudden fluctuations triggered by human intervention for the time-series-aware model.

2. The method for dynamic control in industrial manufacturing based on time-aware artificial intelligence, as described in claim 1, is characterized in that... Every n minutes, a set of raw data from the industrial manufacturing process is collected and continuously accumulated to form a time-series dataset, including: Every n minutes, taking the batch stage of each reactor as the basic unit, the raw log is sliced ​​into time windows, and the parameters within the window slice are extracted to form a set of raw data. A time-series dataset is composed of multiple consecutively collected n-minute raw data points arranged in timestamp order.

3. The method for dynamic control in industrial manufacturing based on time-aware artificial intelligence, as described in claim 2, is characterized in that... The valid interval verification based on first-order principle constraints includes: Trim all raw data parameters to the device's safe range; If any parameter in the original data contains a hard outlier, the original data set will be labeled "early stop". For each parameter in each set of raw data, a triplet is set, including: set value, measured value, and deviation; Manual operations involving replenishing or removing materials are recorded using timestamps.

4. The method for dynamic control in industrial manufacturing based on time-aware artificial intelligence, as described in claim 3, is characterized in that... If the deviation exceeds the limit twice consecutively, the diagnostic logic will be triggered, requiring sensor verification or manual inspection.

5. The method for dynamic control in industrial manufacturing based on time-aware artificial intelligence, as described in claim 1, is characterized in that... The time-aware model adopts a Transformer-based encoder-decoder architecture, including: Input processing layer: It contains a static encoder and a dynamic encoder. The static encoder is responsible for processing metadata that does not change over time and compressing it into static latent variables. The dynamic encoder is responsible for processing real-time time series data and compressing it into dynamic latent variables. The two are concatenated and output. Encoder: Composed of multiple layers of Transformer encoder modules stacked together; each Transformer encoder module uses its core self-attention mechanism to capture the intrinsic correlation and long-range dependency between all time points in the sequence output by the input processing layer, generating a context vector that can characterize the state of the entire historical production process. Decoder: Composed of multiple stacked Transformer decoder modules; The decoder receives the context vector output by the encoder and generates a serialized prediction result for the next prediction period in an autoregressive manner. Output layer: This is a probability distribution layer; for each prediction result, this layer outputs the parameters of a probability distribution as the optimal setting point and acceptable fluctuation range.

6. A time-aware artificial intelligence-based dynamic control system for reshaping industrial manufacturing, characterized in that, include: Data acquisition module: Every n minutes, it collects a set of raw data from the industrial manufacturing process and performs a valid interval check based on first-order principle constraints, continuously accumulating to form a time-series dataset; Feature generation module: Every m hours, extract multiple sets of features from the latest accumulated time-series dataset; Model training module: Every m hours, based on the current accumulated total amount of raw data, the complexity of the time-series awareness model is dynamically adjusted, and a new round of training is performed using the latest extracted features. The training process includes: The output of the time-series-aware model is interpreted as parameters of a probability distribution. The parameters of the time-series-aware model are updated by optimizing the log-likelihood function with respect to the real observation data and adding a penalty term for the first-order principle constraint, thus completing this round of training. Model prediction module: Using the latest trained time-aware model, predict the optimal execution point and the acceptable fluctuation range of other parameters for the next m hours; The looping module continues to accumulate new raw data after the prediction output, and then enters the feature extraction and time-series-aware model rolling training for the next m-hour cycle. Every m hours, multiple sets of features are extracted from the latest accumulated time-series dataset, including: Based on the knowledge of growth curve kinetics, the fermentation stage (bio_phase) of each set of raw data is determined, and the stage information is added to the corresponding raw data in the form of a unique heat vector. The static encoder reads invariant metadata from the original data and compresses it into static latent variables; the dynamic encoder reads real-time changing data from the original data and compresses it into dynamic latent variables; the static latent variables and the dynamic latent variables are then concatenated. The flow rate variables in the raw data are scaled according to the corresponding instantaneous working volume and a derivative channel is added to capture leakage or evaporation. Manually recorded data is masked and binary-labeled to filter out sudden fluctuations triggered by human intervention for the time-series-aware model.

7. An industrial manufacturing system, characterized in that, It includes the perception layer, decision-making layer, and execution layer; The perception layer serves as the data entry point, continuously receiving time-series datasets and features, and synchronously transmitting them to the decision layer; The decision-making layer will adopt the time-aware AI-based dynamic control method for reshaping industrial manufacturing as described in any one of claims 1-5, or the time-aware model trained by the time-aware AI-based dynamic control system for reshaping industrial manufacturing as described in claim 6, and encapsulate it as a gRPC microservice. Based on the data input from the perception layer, the decision layer initiates gRPC microservices to perform analysis and reasoning, generating the optimal control strategy for the current production state. At the same time, it refers to standard operating procedures to ensure the compliance of the optimal control strategy, and then sends the control instructions to the execution layer. After receiving control commands from the decision-making layer, the execution layer follows the standard operating procedures to drive the PLC script into specific equipment actions and feeds back the execution results to the perception layer.

8. A terminal, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it can be used to perform the method of any one of claims 1-5, or to run the system of any one of claims 6-7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program can be used to perform the method of any one of claims 1-5, or to run the system of any one of claims 6-7.

Images