Multimodal time series data fusion modeling method, process state prediction method and device

Abstract

The application discloses a kind of multi-modal time series data fusion modeling method, process state prediction method and device, modeling method includes: obtaining and preprocessing the historical multi-source heterogeneous batch data of target distillation equipment, obtains training data set on unified time axis;Mechanism-data hybrid driven model is constructed, mechanism-data hybrid driven model includes mechanism skeleton layer and AI parameter correction layer;Training data set is used to train AI parameter correction layer, to make mechanism skeleton layer introduce AI parameter correction layer updated to be inverted parameter in training process Calculation of predicted process state;In the error between predicted process state and actual process state in training data set reaches minimum condition, generate pre-trained mechanism-data hybrid driven model.Therefore, by the embodiment of the application, the prediction accuracy and robustness of the model in the whole life cycle are significantly improved.At the same time, the generalization ability and decision reliability of the model under extreme conditions are greatly enhanced.

Description

Technical Field

[0001] This application relates to the field of computer modeling and optimization control technology for chemical processes, and in particular to a multimodal time series data fusion modeling method, a process state prediction method and apparatus. Background Technology

[0002] In chemical distillation production, the process relies on batch or semi-batch distillation operations, which require extremely precise control of temperature, pressure, and component separation. A distributed control system (DCS) is deployed on-site to collect multi-source, heterogeneous, time-series process data, including reactor temperature, jacket temperature, vacuum level, steam regulating valve opening, and receiving tank weighing. Simultaneously, a laboratory information management system (LIMS) is used to obtain low-frequency batch quality index data. Operators must adjust the heat and mass transfer efficiency during the distillation process based on this massive amount of data to ensure the unit operates safely and stably.

[0003] Among related technologies, a pure mechanistic modeling method based on the principles of chemical thermodynamics and transport processes is employed. This method simulates the complex behavior within a distillation column by establishing a set of differential equations describing mass conservation, energy conservation, and phase equilibrium. Secondly, a purely data-driven modeling method has emerged in recent years. This method utilizes machine learning algorithms to directly mine the nonlinear mapping relationships between input variables (such as operating conditions) and output variables (such as product quality and critical states) in historical production data to achieve process prediction or soft measurement.

[0004] However, the prediction accuracy of pure mechanistic models heavily depends on the accuracy of their model parameters. In real industrial environments, key parameters such as the overall heat transfer coefficient can drift due to time-varying factors such as fouling of heat exchanger tubes, changes in material viscosity, and catalyst activity decay. Mechanistic models struggle to correct for these time-varying characteristics in real time, leading to insufficient accuracy. On the other hand, while pure data-driven models may exhibit high fitting accuracy under specific operating conditions, they are essentially black-box models, lacking physical mechanism constraints. This can cause their predictions to seriously violate the laws of mass conservation or energy conservation, posing significant safety risks and insufficient reliability when used for real-time control or optimization decision-making. Summary of the Invention

[0005] This application provides a multimodal time series data fusion modeling method, a process state prediction method, and an apparatus. To provide a basic understanding of some aspects of the disclosed embodiments, a brief summary is given below. This summary is not intended as a general commentary, nor is it intended to identify key / important components or describe the scope of protection of these embodiments. Its sole purpose is to present some concepts in a simple form as a prelude to the detailed description that follows.

[0006] In a first aspect, embodiments of this application provide a method for multimodal time-series data fusion modeling, the method comprising:

[0007] Acquire and preprocess historical multi-source heterogeneous batch data of the target distillation equipment to obtain a training dataset on a unified time axis; A mechanism-data hybrid driven model is constructed, which includes a mechanism skeleton layer and an AI parameter correction layer. The AI ​​parameter correction layer is used to take observable process features as input and learn the real-time correction amount relative to the basic reference value of the parameter to be inverted. The mechanism skeleton layer is used to dynamically update the basic reference value of the parameter to be inverted by calling the real-time correction amount output by the AI ​​parameter correction layer during calculation, and predict the process state at the next moment based on the updated parameter to be inverted. The AI ​​parameter correction layer is trained using the training dataset so that the mechanism skeleton layer can incorporate the updated parameters to be inverted during the training process of the AI ​​parameter correction layer to calculate the predicted process state. A pre-trained mechanism-data hybrid driven model is generated when the error between the predicted process state and the actual process state in the training dataset is minimized.

[0008] Secondly, embodiments of this application provide a method for predicting process states, including: During the online operation of the target distillation equipment, the latest observation data within a preset period is acquired in real time; The pre-trained mechanism-data hybrid driven model is invoked; wherein, the pre-trained parameter correction model is obtained by training the multimodal time series data fusion modeling method described above; The observable process features within the current time and historical time window are extracted from the latest observation data, and the observable process features are input into the AI ​​parameter correction layer in the pre-trained mechanism-data hybrid driven model. The AI ​​parameter correction layer outputs the target correction amount relative to the basic reference value of the parameter to be inverted, and the mechanism skeleton layer in the mechanism-data hybrid driving model uses the target correction amount to dynamically update the basic reference value to obtain the updated target parameter to be inverted. Substitute the target parameters to be inverted into the differential equation of the mechanism framework layer, and use a numerical integration algorithm to forward calculate and predict the target process state at the next moment. Output the target process state for the next moment to guide operators in adjusting process parameters or as the setpoint input for the closed-loop control system.

[0009] Thirdly, embodiments of this application provide a multimodal time-series data fusion modeling apparatus, the apparatus comprising: The historical multi-source heterogeneous batch data acquisition module is used to acquire and preprocess the historical multi-source heterogeneous batch data of the target distillation equipment to obtain a training dataset on a unified time axis. The model building module is used to construct a mechanism-data hybrid driven model, which includes a mechanism skeleton layer and an AI parameter correction layer. The AI ​​parameter correction layer is used to take observable process features as input and learn the real-time correction amount relative to the basic reference value of the parameter to be inverted. The mechanism skeleton layer is used to dynamically update the basic reference value of the parameter to be inverted by calling the real-time correction amount output by the AI ​​parameter correction layer during calculation, and predict the process state at the next moment based on the updated parameter to be inverted. The training prediction module is used to train the AI ​​parameter correction layer using the training dataset, so that the mechanism skeleton layer can introduce the updated parameters to be inverted during the training process of the AI ​​parameter correction layer to calculate the predicted process state. The model generation module is used to generate a pre-trained mechanism-data hybrid driven model while minimizing the error between the predicted process state and the actual process state in the training dataset.

[0010] Fourthly, embodiments of this application provide a process state prediction device, the device comprising: The latest observation data acquisition module is used to acquire the latest observation data within a preset period in real time during the online operation of the target distillation equipment; The model invocation module is used to invoke the pre-trained mechanism-data hybrid driven model; wherein, the pre-trained parameter correction model is trained through the multimodal time series data fusion modeling method described above; The feature processing module is used to extract observable process features from the latest observation data within the current time and historical time windows, and input the observable process features into the AI ​​parameter correction layer in the pre-trained mechanism-data hybrid driven model. The basic reference value update module is used to output the target correction amount relative to the basic reference value of the parameter to be inverted through the AI ​​parameter correction layer, and the mechanism skeleton layer in the mechanism-data hybrid driving model uses the target correction amount to dynamically update the basic reference value to obtain the updated target parameter to be inverted. The process state prediction module is used to substitute the target parameters to be inverted into the differential equation of the mechanism skeleton layer, and predict the target process state at the next moment through forward calculation using a numerical integration algorithm. The process status output module is used to output the target process status at the next moment to guide operators in adjusting process parameters or as the setpoint input for the closed-loop control system.

[0011] The technical solutions provided in this application embodiment may include the following beneficial effects: In this application embodiment, on the one hand, this application introduces an AI parameter correction layer to dynamically update the parameters to be inverted in the model, and uses a data-driven approach to capture and compensate for parameter drift caused by time-varying factors in real time, thereby effectively overcoming the accuracy decay problem caused by the inability of traditional models to adaptively adjust, and significantly improving the prediction accuracy and robustness of the model throughout its entire life cycle; on the other hand, this application constructs a mechanism skeleton layer with differential equations as the core, which forcibly constrains the prediction logic within the framework of physical laws, allowing the AI ​​model to learn the residual correction of physical parameters rather than directly fitting the final result, thereby ensuring that the output at each moment conforms to objective physical laws while retaining the flexibility of data-driven approaches, and greatly enhancing the model's generalization ability and decision reliability under extreme conditions.

[0012] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0013] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0014] Figure 1 This is a flowchart illustrating a multimodal time-series data fusion modeling method provided in an embodiment of this application; Figure 2 This is an information diagram of derived feature variables in a specific scenario provided in this application embodiment; Figure 3 This is a schematic diagram illustrating a training dataset generation process provided in an embodiment of this application; Figure 4 This is an information diagram of a mechanism-data hybrid driven model provided in an embodiment of this application; Figure 5 This is a schematic diagram illustrating a hybrid-driven model construction process provided in an embodiment of this application; Figure 6 This is an infographic illustrating model training and state prediction provided in an embodiment of this application; Figure 7 This is a schematic diagram of a model training process provided in an embodiment of this application; Figure 8 This is a flowchart illustrating a process state prediction method provided in an embodiment of this application. Figure 9 This is a schematic diagram of the structure of a process state prediction device provided in an embodiment of this application; Figure 10 This is a schematic diagram of the structure of a process state prediction device provided in an embodiment of this application; Figure 11 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0015] The following description and accompanying drawings fully illustrate specific embodiments of this application to enable those skilled in the art to practice them.

[0016] It should be understood that the described embodiments are merely some, not all, of the embodiments in this application. All other embodiments obtained by those skilled in the art based on the embodiments in this application without inventive effort are within the scope of protection of this application.

[0017] In the following description, when referring to the accompanying drawings, the same numbers in different drawings denote the same or similar elements unless otherwise indicated. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0018] In the description of this application, it should be understood that the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances. Furthermore, in the description of this application, unless otherwise stated, "multiple" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship.

[0019] This application provides a multimodal time-series data fusion modeling method, a process state prediction method, and an apparatus to address the problems existing in the aforementioned related technologies. In the embodiments of this application, on the one hand, this application introduces an AI parameter correction layer to dynamically update the parameters to be inverted in the model, using a data-driven approach to capture and compensate for parameter drift caused by time-varying factors in real time. This effectively overcomes the accuracy decay problem caused by the inability of traditional models to adaptively adjust, significantly improving the prediction accuracy and robustness of the model throughout its entire lifecycle. On the other hand, this application constructs a mechanism skeleton layer with differential equations as its core, forcibly constraining the prediction logic within the framework of physical laws. It only allows the AI ​​model to learn the residual correction of physical parameters rather than directly fitting the final result. This ensures that the output at each moment conforms to objective physical laws while retaining the flexibility of data-driven approaches, greatly enhancing the model's generalization ability and decision reliability under extreme conditions. Exemplary embodiments are described in detail below.

[0020] This application provides a multimodal time series data fusion modeling method, a process state prediction method, and an apparatus to solve the problems existing in the aforementioned related technologies. The following will be discussed in conjunction with the appendix... Figure 1 -Appendix Figure 8 This application provides a detailed description of the multimodal time-series data fusion modeling method provided in its embodiments. This method can be implemented using a computer program and can run on a multimodal time-series data fusion modeling device based on the von Neumann architecture. The computer program can be integrated into an application or run as a standalone utility application.

[0021] Please see Figure 1 This is a flowchart illustrating a multimodal time-series data fusion modeling method provided in this application embodiment. Figure 1 As shown, the method in this application embodiment may include the following steps: S101, acquire and preprocess historical multi-source heterogeneous batch data of the target distillation equipment to obtain a training dataset on a unified time axis; The target distillation equipment refers to the reaction vessel, distillation column, and its associated condenser and receiving devices with distillation capabilities in the industrial setting; it is the physical carrier of the process physical process. Historical multi-source heterogeneous batch data refers to the entire process time-series data collected from multiple independent sources such as distributed control systems (DCS) and laboratory information management systems (LIMS) during the historical production process of the target distillation equipment. These data differ in data structure, sampling frequency, and physical dimensions. The training dataset refers to the preprocessed, standardized data sample set containing feature sequences (such as temperature, pressure, and derived variables) arranged along a unified time axis and their corresponding batch quality labels (such as fraction purity and yield), used to train the AI ​​parameter correction layer.

[0022] In some embodiments of this application, the specific process of acquiring and preprocessing historical multi-source heterogeneous batch data of the target distillation equipment to obtain a training dataset on a unified time axis includes: extracting time-series process data of the target distillation equipment from the distributed control system and batch quality data of the target distillation equipment from the laboratory information management system; converting and mapping the time-series process data and batch quality data to the same time axis using the start time of the production order as the base time to obtain historical multi-source heterogeneous batch data; cleaning and reconstructing the historical multi-source heterogeneous batch data to obtain a complete time-series sequence; extracting derived feature variables to characterize the process state of the reactor from the complete time-series sequence; normalizing the time-series process data, batch quality data, and derived feature variables respectively to eliminate the influence of dimensions to obtain all normalized data; organizing all normalized data into samples by batch to form a training dataset, where each sample contains a feature sequence and its label arranged along the time axis.

[0023] Specifically, the process of extracting derived characteristic variables to characterize the process state of the reactor from the complete time series includes: calculating the real-time difference between the reactor jacket temperature and the reactor internal temperature in the complete time series as the real-time jacket temperature difference; performing first derivative calculations on the weight data of the isopropanol receiving tank in the complete time series to calculate the real-time distillation rate; automatically identifying and labeling the heating stage, initial distillation stage, and later distillation stage of the distillation process based on the change curves of the reactor internal temperature and pressure in the complete time series, combined with the preset distillation process stage division rules; and using the real-time jacket temperature difference, real-time distillation rate, and labeled heating stage, initial distillation stage, and later distillation stage as derived characteristic variables to characterize the process state of the reactor.

[0024] In one possible implementation, the plant's data storage server is first accessed to extract all time-series data related to the distillation process in reactor #21 within a specified time period (e.g., October 2025 to December 2025) from the historical database of the Distributed Control System (DCS). This data includes, but is not limited to: reactor internal temperature (PV value), reactor jacket temperature, reactor vacuum, steam regulating valve opening (OP value), agitator motor speed, first / second stage condenser outlet temperature, and isopropanol receiving tank weighing data. The raw sampling frequency for this data is 1 time / second. From the Laboratory Information Management System (LIMS), batch quality data for all completed batches corresponding to the aforementioned time period is extracted. This data typically includes the final product purity, acid value, and moisture content of the batch; only 1-2 records are entered for each batch at the end of production, representing low-frequency data. The system iterates through each production batch, reading the production order start time in the batch control system, and uses this time point as the baseline time origin (T0). Subsequently, the system converts the timestamps of the high-frequency time-series data extracted from the DCS and the timestamps of the low-frequency quality data extracted from the LIMS into a relative time starting from T0 (such as "150 minutes after the start of the batch"). This maps the heterogeneous data with different sampling frequencies and clock references onto the same continuous and aligned unified time axis, forming preliminary historical multi-source heterogeneous batch data.

[0025] Then, the system performs automated cleaning and reconstruction: The system detects a sudden jump in the reactor temperature reading to 999℃ (far exceeding physical limits) at a certain moment, identifying it as a transient sensor fault. Based on the linear trend of adjacent time points, this point is removed or corrected. For the 5-second data gap caused by network fluctuations, the system uses linear interpolation based on windows of 10 data points before and after the gap to smoothly fill in the missing segments, ensuring the continuity of data for subsequent analysis and ultimately obtaining a complete time series.

[0026] Secondly, the system reads the reactor jacket temperature and reactor interior temperature in real time at each moment, calculates the difference between them, and uses this difference as a real-time jacket temperature difference feature sequence characterizing the heat transfer driving force. The system acquires the weighing data of the isopropanol receiving tank, performs a first-order derivative operation on it (i.e., calculates the change in weight between adjacent time points divided by the time interval), and uses the calculation result as a real-time distillation rate feature sequence characterizing the current evaporation intensity. The system analyzes the change curves of the reactor interior temperature and pressure. Based on preset distillation process stage division rules (e.g., rapid temperature rise and no significant pressure drop as "heating stage"; relatively stable temperature and pressure and high distillation rate as "initial distillation"; continuous temperature rise but significant decrease in distillation rate as "late distillation"), the system automatically scans the time axis of the entire batch using a pattern recognition algorithm and labels the corresponding time period as heating stage, initial distillation, or late distillation. The system uses the calculated real-time jacket temperature difference, real-time distillation rate, and process stage labels as new derived feature variables to characterize the reactor process state. For example, a derived feature variable in one scenario... Figure 2 As shown.

[0027] Finally, the system performs Min-Max Normalization on the original DCS time-series process data, LIMS batch quality data, and newly extracted derived feature variables. This involves linearly scaling all data to the [0, 1] interval to eliminate the influence of different physical units (such as ℃, kPa, kg / h) on model training. The system then packages all normalized data into independent samples based on production batch IDs. After traversing all historical batches, the final training dataset is formed.

[0028] For example Figure 3 As shown, Figure 3 This is a schematic diagram of the training dataset generation process provided in this application. First, high-frequency time-series process data is extracted from the Distributed Control System (DCS) and low-frequency batch quality data is extracted from the Laboratory Information Management System (LIMS). Next, based on the production order start time, these two types of heterogeneous data are transformed and mapped to the same time axis to achieve time alignment. Subsequently, the aligned data is cleaned and reconstructed to generate a complete time-series sequence. On this basis, derived feature variables (such as real-time jacket temperature difference, real-time distillation rate, and process stage labeling) for characterizing the reactor process state are further extracted. Then, all original and derived feature data are normalized to eliminate the influence of dimensions. Finally, the normalized data is organized into feature sequences and their labels by batch to form a standardized dataset that can be used for model training.

[0029] It should be noted that by extracting and unifying multi-source heterogeneous data from DCS and LIMS, and then extracting derived features such as real-time temperature difference, distillation rate, and process stage, the information dimensions characterizing the equipment status are significantly enriched, making up for the deficiencies of the original sensor data. Data cleaning, reconstruction, and normalization operations effectively eliminate noise and dimensional differences, which not only greatly improves data quality but also provides standardized, high signal-to-noise ratio input for subsequent AI models, enabling more stable and accurate parameter inversion and process prediction.

[0030] S102, Construct a mechanism-data hybrid driven model, which includes a mechanism skeleton layer and an AI parameter correction layer. The AI ​​parameter correction layer is used to take observable process features as input and learn the real-time correction amount relative to the basic reference value of the parameter to be inverted. The mechanism skeleton layer is used to dynamically update the basic reference value of the parameter to be inverted by calling the real-time correction amount output by the AI ​​parameter correction layer during calculation, and predict the process state at the next moment based on the updated parameter to be inverted. The parameters to be inverted are key physical parameters in the differential equations of the mechanistic framework layer, used to characterize heat transfer efficiency (e.g., overall heat transfer coefficient U), mass transfer efficiency (e.g., mass transfer coefficient K), or reaction kinetic characteristics (e.g., reaction rate constant k). These parameters are preset to change dynamically with operating conditions and require real-time inversion correction through the AI ​​parameter correction layer. The base reference value is the initial reference value of the parameters to be inverted under ideal or design conditions (e.g., the nominal value provided in the equipment design manual). This value serves as the correction benchmark and is superimposed with the real-time correction amount output by the AI ​​parameter correction layer to obtain the updated parameter estimate. The real-time correction amount is used to dynamically compensate for the deviation between actual operating conditions and the ideal model. Observable process characteristics are a sequence of physical quantities (e.g., reactor temperature, jacket temperature difference, distillation rate, process stage labels, etc.) that characterize the current process state, obtained directly from sensors or the control system of the target distillation equipment, or indirectly calculated. Process status refers to the key physical quantities output from the mechanistic framework layer that describe the operating status of the distillation equipment (such as the temperature inside the reactor, the pressure inside the reactor, the cumulative weight of the isopropanol receiving tank, and the liquid level of the material inside the reactor).

[0031] In some embodiments of this application, the specific process of constructing a mechanism-data hybrid driven model includes: generating at least one differential equation based on the law of conservation of mass or the law of conservation of energy; determining at least one key parameter in the differential equation used to characterize heat transfer efficiency or mass transfer efficiency as a parameter to be inverted that dynamically changes with operating conditions, and setting a basic reference value for the parameter to be inverted; generating a mechanism skeleton layer through the differential equation with the basic reference value set; constructing an AI parameter correction layer through a machine learning algorithm; and connecting the mechanism skeleton layer and the AI ​​parameter correction layer to obtain the mechanism-data hybrid driven model. The mechanism-data hybrid driven model is, for example... Figure 4As shown.

[0032] It should be noted that by generating differential equations based on the law of conservation of mass / energy and establishing a mechanistic framework layer, the constraint that the model follows objective physical laws is fundamentally established. This eliminates the possibility of purely data-driven models making meaningless predictions that violate the laws of thermodynamics, thus ensuring the scientific validity and generalization reliability of the model. Simultaneously, by defining key parameters characterizing heat / mass transfer efficiency as parameters to be inverted that dynamically change with operating conditions and setting basic reference values, the traditional fixed-parameter mechanistic model can be transformed into a dynamic framework with adaptive capabilities. This allows the model to capture complex time-varying characteristics such as equipment fouling and material fluctuations in real time through an AI parameter correction layer. Finally, the organic connection between the AI ​​parameter correction layer constructed through machine learning algorithms and the mechanistic framework layer achieves a two-way empowerment of physical mechanisms guiding data learning and data-driven correction of physical deviations. This preserves the interpretability of the mechanistic model while endowing the data model with adaptive capabilities.

[0033] Specifically, the process of generating at least one differential equation based on the law of conservation of mass or the law of conservation of energy includes: for the reactor in the target distillation equipment, based on the law of conservation of mass, constructing a material balance differential equation to describe the liquid level of the material in the reactor or the cumulative amount of distilled products. The material balance differential equation is used to characterize the weight change rate of the isopropanol receiving tank as a function of the condensation rate of the vapor distillate, and the condensation rate of the vapor distillate is related to the parameters to be inverted and the temperature and pressure inside the reactor; or, for the reactor in the target distillation equipment, based on the law of conservation of energy, constructing an energy balance differential equation to describe the dynamic change of temperature inside the reactor. The energy balance differential equation is used to characterize the rate of change of temperature inside the reactor over time as a function of the difference between the input heat and the output heat. The input heat is calculated from the steam flow rate and the latent heat of steam determined by the opening of the steam regulating valve, and the output heat is determined by the heat transfer rate between the reactor jacket and the material inside the reactor, and the heat transfer rate is positively correlated with the real-time jacket temperature difference and the parameters to be inverted.

[0034] Specifically, the process of generating the mechanism framework layer by setting differential equations with basic reference values ​​includes: substituting the basic reference values ​​into the differential equations as static model parameters; configuring the calculation logic; combining the static model parameters and the calculation logic into the mechanism framework layer; wherein, the calculation logic performs the following operations in each control cycle: receiving real-time correction values ​​from the AI ​​parameter correction layer, superimposing the real-time correction values ​​onto the basic reference values ​​to calculate the updated parameters to be inverted; substituting the updated parameters to be inverted into the differential equations, and using a numerical integration algorithm to forward calculate and predict the process state at the next moment; wherein, the predicted process state includes at least the temperature inside the reactor, the pressure inside the reactor, and the cumulative weight of the isopropanol receiving tank.

[0035] In one possible implementation, the target distillation equipment is, for example, reactor #21 and its associated distillation condenser receiving system in a silicone oil production line of a chemical enterprise. The energy balance differential equation is constructed as follows: For the reactor body, the system constructs an energy balance differential equation based on the law of conservation of energy. This equation represents the actual temperature inside the reactor. rate of change over time It is a function representing the difference between input heat and output heat.

[0036] Input heat The real-time steam flow rate is calculated by converting the steam regulating valve opening signal acquired by the DCS system into the actual flow rate, and then multiplying it by the latent heat of the steam at the current pressure. Its dimension is power. .

[0037] Output heat (heat transfer rate) The heat transfer rate between the reactor jacket and the material inside the reactor is determined by the heat transfer rate. According to Newton's law of cooling, this heat transfer rate is related to the real-time temperature difference in the jacket. ) and the parameters to be inverted (here, the overall heat transfer coefficient) They are positively correlated, that is .in, The effective heat exchange area of ​​the reactor jacket is considered a constant. The overall heat transfer coefficient is a physical quantity used to comprehensively reflect the intensity of the heat transfer process; its dimensions are... .

[0038] Therefore, the system can establish the following correct energy conservation equation, which describes the dynamic evolution of the actual temperature:

[0039] in, Specific heat capacity is the amount of heat required to raise the temperature of a unit mass of material by one degree Celsius. Let be the total mass of the material inside the vessel. The physical meaning of this equation is that the rate of change of the internal energy of the material inside the vessel is equal to the net heat flow input.

[0040] Construction of the boiling point (equilibrium temperature) evolution equation: To describe the boiling point rise caused by changes in material composition during distillation, the system defines an independent variable based on the principle of phase equilibrium: theoretical maximum temperature / equilibrium temperature. ).

[0041] definition: It refers to the equilibrium temperature required for the material to boil under the current pressure and composition inside the vessel.

[0042] Evolutionary logic: As distillation proceeds, lighter components (such as isopropanol) are distilled off, increasing the proportion of heavier components in the still, leading to a rise in boiling point. Therefore, It is a quantity that changes over time.

[0043] Correlation equations: System establishment The relationship with the distillation process. rate of change The primary driver is the change in material composition, which in turn is related to the distillation rate. therefore:

[0044] in, This represents the system pressure. In the simplified model, it can be obtained by fitting experimental data. The empirical relationship between the distillation rate and the distillation rate. This equation does not directly accept... This serves as input, thus avoiding confusion in physical logic.

[0045] Construction of the differential equation for mass equilibrium: For the distillation product collection end, the system constructs a mass balance differential equation based on the law of conservation of mass. This equation accounts for the weight change rate of the isopropanol receiving tank. Characterized as a function of the condensation rate of the gaseous distillate. This refers to the cumulative mass of distillate collected in the isopropanol receiving tank. It is the cumulative variable in the mass balance equation and a direct indicator of distillation progress and yield. This refers to the increase in the weight of the receiving vessel per unit time, i.e., the distillation rate. The parameters to be retrieved: This condensation rate is related to the temperature and pressure inside the vessel, as well as the parameters to be retrieved (here, these are expressed as the condensation efficiency coefficient or the gas-liquid equilibrium constant). It is directly related to the phase change efficiency of the gas phase in the condenser.

[0046] Through the above steps, the system generates a set of differential equations based on conservation laws.

[0047] Among them, the determination of the parameters to be inverted and the setting of the basic reference values ​​are crucial. In the model constructed above, the system identifies two key dimensions of the parameters to be inverted: Energy-side parameters to be inverted: Overall heat transfer coefficient in the energy balance equation It is affected by factors such as scaling on heat exchange tubes and changes in material viscosity.

[0048] Parameters to be retrieved on the mass / material transfer side: Correction coefficients in the mass balance equation related to condensation efficiency or gas-liquid equilibrium. .

[0049] At this time, the system performs the following operations: Will and The parameters to be inverted are jointly determined to be dynamically changing with operating conditions. This is done by consulting the equipment design manual or based on cold-state commissioning data. Set basic reference value (For example, Based on historical best operating data or design specifications, for Set basic reference value These baseline reference values ​​represent the device's benchmark performance under ideal conditions.

[0050] In one possible implementation, during the generation of the mechanistic framework layer, the system utilizes a system of differential equations with a set of basic reference values ​​to generate the mechanistic framework layer. Parameter initialization: Set the base reference value and Substitute the corresponding values ​​into the equation to obtain the initial values ​​for the static model parameters.

[0051] Computational logic configuration: The system configures the core computational logic, which is set to perform the following operations in each control cycle (e.g., per second): Receive correction values: Receive real-time correction values ​​from the AI ​​parameter correction layer via a preset interface. and .

[0052] Dynamic parameter update: Perform superposition operation to calculate the updated parameters to be inverted.

[0053]

[0054] Forward prediction is as follows: Substituting into the energy balance differential equation, and combining it with the actual process conditions collected at the current moment (such as...) The reactor temperature at the next moment is predicted using a numerical integration algorithm (such as the fourth-order Runge-Kutta method) to perform forward calculations. and the pressure inside the reactor.

[0055] Will Substituting into the mass balance differential equation and considering the current temperature and pressure, calculate the predicted cumulative weight of the isopropanol receiving tank at the next moment. At this point, a mechanistic framework layer with multi-parameter dynamic update capability has been constructed.

[0056] In one possible implementation, an AI parameter correction layer is constructed: the system employs a Long Short-Term Memory (LSTM) network as the machine learning algorithm to construct the AI ​​parameter correction layer. This network is designed to receive multi-source observable process features as input, including but not limited to: reactor temperature, jacket temperature difference, steam valve opening, stirring speed, real-time distillation rate, and process stage labels. Its output layer contains two neurons, each used to output parameters relative to the current time and a past time window. Real-time correction amount and relative to Real-time correction amount Model Connection: The system connects the output of the AI ​​parameter correction layer to the input of the mechanistic skeleton layer, which receives the correction values, via a data flow interface. This connection forms a tightly coupled closed-loop structure: the AI ​​layer is responsible for sensing the operating conditions and outputting multiple maintenance correction values, while the mechanistic layer is responsible for calling these correction values ​​to perform physical calculations.

[0057] For example Figure 5 As shown, Figure 5 This is a schematic diagram illustrating the process of constructing a hybrid-driven model provided in this application. First, a differential equation is constructed based on the law of conservation of mass or energy, and the parameters to be inverted are determined from it. Then, basic reference values ​​for these parameters are set in the differential equation, thereby generating a mechanistic skeleton layer. At the same time, an AI parameter correction layer is independently constructed using machine learning algorithms. Finally, the mechanistic skeleton layer and the AI ​​parameter correction layer are connected and fused to obtain a hybrid-driven model that combines physical mechanisms and data intelligence.

[0058] S103, use the training dataset to train the AI ​​parameter correction layer so that the mechanism skeleton layer can introduce the updated parameters to be inverted during the training process of the AI ​​parameter correction layer to calculate the predicted process state. In some embodiments of this application, the specific process of training an AI parameter correction layer using a training dataset so that the mechanism skeleton layer can incorporate the updated parameters to be inverted during the training process of the AI ​​parameter correction layer to calculate the predicted process state includes: initializing the model parameters of the AI ​​parameter correction layer and loading the preset basic reference values ​​of the parameters to be inverted into the mechanism skeleton layer; reading the current sample from the training dataset in batches, inputting the observable process features arranged along the time axis in the current sample into the AI ​​parameter correction layer, and outputting the real-time correction amount corresponding to the current sample relative to the basic reference value; superimposing the real-time correction amount onto the basic reference value to obtain the updated parameters to be inverted; substituting the updated parameters to be inverted into the differential equation of the mechanism skeleton layer, and using a numerical integration algorithm to forward calculate and predict the process state at the next moment to obtain the predicted process state.

[0059] It should be noted that by initializing the AI ​​parameter correction layer and loading the basic reference values ​​of the parameters to be inverted, an initial benchmark for the fusion of mechanism and data was established, ensuring the physical rationality of the model training starting point. During training, samples were traversed in batches, and observable process features were used to drive the AI ​​layer to output real-time correction values, realizing dynamic calibration of the basic reference values. This cleverly transformed parameter drift under complex working conditions into a data-driven learning task. Furthermore, the updated parameters to be inverted were substituted into the differential equations of the mechanism skeleton layer, and the process state at the next moment was predicted forward using numerical integration algorithms. This not only forced the AI ​​model to be trained under the constraints of physical conservation laws, effectively avoiding the risk of violating objective laws common in pure black box models, but also enabled the AI ​​parameter correction layer to accurately learn the optimal strategy to compensate for the bias of the mechanism model through the prediction-feedback closed-loop mechanism. This fundamentally improved the model's adaptability to the time-varying characteristics of industrial processes and its long-term prediction accuracy.

[0060] Specifically, the process of substituting the updated parameters to be inverted into the differential equation of the mechanistic framework layer and using a numerical integration algorithm to forward calculate and predict the process state at the next moment includes: substituting the updated parameters to be inverted into the differential equation of the mechanistic framework layer; discretizing and solving the differential equation based on the numerical integration algorithm to calculate the change in process state from the current moment to the next moment; wherein, the process state includes at least one of the following: reactor temperature, reactor pressure, and material level in the reactor; and determining the predicted process state at the next moment based on the change in process state and the actual process state at the current moment.

[0061] Further, the loss function value between the predicted process state and the actual process state in the training dataset is calculated; this loss function value is used as the error between the predicted process state and the actual process state in the training dataset; if the error has not reached its minimum, the gradient is calculated using the backpropagation algorithm based on the loss function value, and the model parameters of the AI ​​parameter correction layer are updated; the step of batch-reading the current sample from the training dataset continues until the error between the predicted process state and the actual process state in the training dataset reaches its minimum. Model training and state prediction are, for example... Figure 6 As shown.

[0062] S104 generates a pre-trained mechanism-data hybrid driven model when the error between the predicted process state and the actual process state in the training dataset is minimized.

[0063] For example Figure 7 As shown, Figure 7This is a schematic diagram of a model training process provided in this application. First, the parameters are initialized and the basic reference values ​​of the parameters to be inverted are loaded. Then, the training samples are read and the features are input into the AI ​​layer, and the real-time correction amount relative to the basic reference value is output. This correction amount is passed to the mechanism skeleton layer to update the parameters, and the updated parameters are substituted into the differential equation for discrete solution and forward calculation, thereby predicting the process state at the next moment. Then, the error between the prediction result and the true value is calculated by combining the loss function. If the loss is not minimized, the error is fed back to the input end through the backpropagation algorithm for iterative optimization until the loss function reaches the minimum value. Finally, the pre-trained hybrid driving model is output.

[0064] In this application embodiment, on the one hand, this application introduces an AI parameter correction layer to dynamically update the parameters to be inverted in the model, and uses a data-driven approach to capture and compensate for parameter drift caused by time-varying factors in real time, thereby effectively overcoming the accuracy decay problem caused by the inability of traditional models to adaptively adjust, and significantly improving the prediction accuracy and robustness of the model throughout its entire life cycle; on the other hand, this application constructs a mechanism skeleton layer with differential equations as the core, which forcibly constrains the prediction logic within the framework of physical laws, allowing the AI ​​model to learn the residual correction of physical parameters rather than directly fitting the final result, thereby ensuring that the output at each moment conforms to objective physical laws while retaining the flexibility of data-driven approaches, and greatly enhancing the model's generalization ability and decision reliability under extreme conditions.

[0065] Please see Figure 8 This document provides a flowchart illustrating process state prediction in an embodiment of this application. Figure 8 As shown, the method in this application embodiment may include the following steps: S201: During the online operation of the target distillation equipment, the latest observation data within a preset period is acquired in real time; S202, invoke the pre-trained mechanism-data hybrid driven model; wherein, the pre-trained parameter correction model is obtained by training the multimodal time series data fusion modeling method described above; S203 extracts observable process features from the latest observation data within the current time and historical time window, and inputs the observable process features into the AI ​​parameter correction layer in the pre-trained mechanism-data hybrid driven model; S204, the AI ​​parameter correction layer outputs the target correction amount relative to the basic reference value of the parameter to be inverted, and the mechanism skeleton layer in the mechanism-data hybrid driving model uses the target correction amount to dynamically update the basic reference value to obtain the updated target parameter to be inverted; S205, substitute the target parameters to be inverted into the differential equation of the mechanism framework layer, and predict the target process state at the next moment through forward calculation using a numerical integration algorithm; S206 outputs the target process state for the next moment to guide operators in adjusting process parameters or as a setpoint input for the closed-loop control system.

[0066] In this application embodiment, on the one hand, this application introduces an AI parameter correction layer to dynamically update the parameters to be inverted in the model, and uses a data-driven approach to capture and compensate for parameter drift caused by time-varying factors in real time, thereby effectively overcoming the accuracy decay problem caused by the inability of traditional models to adaptively adjust, and significantly improving the prediction accuracy and robustness of the model throughout its entire life cycle; on the other hand, this application constructs a mechanism skeleton layer with differential equations as the core, which forcibly constrains the prediction logic within the framework of physical laws, allowing the AI ​​model to learn the residual correction of physical parameters rather than directly fitting the final result, thereby ensuring that the output at each moment conforms to objective physical laws while retaining the flexibility of data-driven approaches, and greatly enhancing the model's generalization ability and decision reliability under extreme conditions.

[0067] The following are embodiments of the apparatus described in this application, which can be used to execute the embodiments of the method described in this application. For details not disclosed in the apparatus embodiments of this application, please refer to the embodiments of the method described in this application.

[0068] Please see Figure 9 This illustration shows a schematic diagram of a multimodal time-series data fusion modeling apparatus provided in an exemplary embodiment of this application. This multimodal time-series data fusion modeling apparatus can be implemented as all or part of an electronic device through software, hardware, or a combination of both. The apparatus 1 includes an employing module 10, a feature extraction module 20, a training module 30, and a generation module 40.

[0069] The historical multi-source heterogeneous batch data acquisition module 10 is used to acquire and preprocess the historical multi-source heterogeneous batch data of the target distillation equipment to obtain a training dataset on a unified time axis. The model building module 20 is used to build a mechanism-data hybrid driven model, which includes a mechanism skeleton layer and an AI parameter correction layer. The AI ​​parameter correction layer is used to take observable process features as input and learn the real-time correction amount relative to the basic reference value of the parameter to be inverted. The mechanism skeleton layer is used to dynamically update the basic reference value of the parameter to be inverted by calling the real-time correction amount output by the AI ​​parameter correction layer during calculation, and predict the process state at the next moment based on the updated parameter to be inverted. The training prediction module 30 is used to train the AI ​​parameter correction layer using the training dataset, so that the mechanism skeleton layer can use the updated parameters to be inverted during the training process of the AI ​​parameter correction layer to calculate the predicted process state. The model generation module 40 is used to generate a pre-trained mechanism-data hybrid driven model when the error between the predicted process state and the actual process state in the training dataset is minimized.

[0070] It should be noted that the multimodal time-series data fusion modeling device provided in the above embodiments is only illustrated by the division of the above functional modules when executing the multimodal time-series data fusion modeling method. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the multimodal time-series data fusion modeling device and the multimodal time-series data fusion modeling method embodiments provided in the above embodiments belong to the same concept, and the implementation process is detailed in the method embodiments, which will not be repeated here.

[0071] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0072] In this application embodiment, on the one hand, this application introduces an AI parameter correction layer to dynamically update the parameters to be inverted in the model, and uses a data-driven approach to capture and compensate for parameter drift caused by time-varying factors in real time, thereby effectively overcoming the accuracy decay problem caused by the inability of traditional models to adaptively adjust, and significantly improving the prediction accuracy and robustness of the model throughout its entire life cycle; on the other hand, this application constructs a mechanism skeleton layer with differential equations as the core, which forcibly constrains the prediction logic within the framework of physical laws, allowing the AI ​​model to learn the residual correction of physical parameters rather than directly fitting the final result, thereby ensuring that the output at each moment conforms to objective physical laws while retaining the flexibility of data-driven approaches, and greatly enhancing the model's generalization ability and decision reliability under extreme conditions.

[0073] Please see Figure 10 This illustration shows a schematic diagram of a process state prediction device provided in an exemplary embodiment of this application. This multimodal time-series data fusion modeling device can be implemented as all or part of an electronic device through software, hardware, or a combination of both. The device 2 includes a latest observation data acquisition module 50, a model invocation module 60, a feature processing module 70, a basic reference value update module 80, a process state prediction module 90, and a process state output module 100.

[0074] The latest observation data acquisition module 50 is used to acquire the latest observation data within a preset period in real time during the online operation of the target distillation equipment; The model invocation module 60 is used to invoke the pre-trained mechanism-data hybrid driven model; wherein, the pre-trained parameter correction model is trained through the multimodal time series data fusion modeling method described above; The feature processing module 70 is used to extract observable process features from the latest observation data within the current time and historical time window, and input the observable process features into the AI ​​parameter correction layer in the pre-trained mechanism-data hybrid driven model. The basic reference value update module 80 is used to output the target correction amount relative to the basic reference value of the parameter to be inverted through the AI ​​parameter correction layer, and the mechanism skeleton layer in the mechanism-data hybrid driving model uses the target correction amount to dynamically update the basic reference value to obtain the updated target parameter to be inverted. The process state prediction module 90 is used to substitute the target parameters to be inverted into the differential equation of the mechanism skeleton layer, and predict the target process state at the next moment through forward calculation using a numerical integration algorithm. The process status output module 100 is used to output the target process status at the next moment to guide the operator to adjust the process parameters or as the set value input of the closed-loop control system.

[0075] It should be noted that the multimodal time-series data fusion modeling device provided in the above embodiments is only illustrated by the division of the above functional modules when executing the multimodal time-series data fusion modeling method. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the multimodal time-series data fusion modeling device and the multimodal time-series data fusion modeling method embodiments provided in the above embodiments belong to the same concept, and the implementation process is detailed in the method embodiments, which will not be repeated here.

[0076] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0077] In this application embodiment, on the one hand, this application introduces an AI parameter correction layer to dynamically update the parameters to be inverted in the model, and uses a data-driven approach to capture and compensate for parameter drift caused by time-varying factors in real time, thereby effectively overcoming the accuracy decay problem caused by the inability of traditional models to adaptively adjust, and significantly improving the prediction accuracy and robustness of the model throughout its entire life cycle; on the other hand, this application constructs a mechanism skeleton layer with differential equations as the core, which forcibly constrains the prediction logic within the framework of physical laws, allowing the AI ​​model to learn the residual correction of physical parameters rather than directly fitting the final result, thereby ensuring that the output at each moment conforms to objective physical laws while retaining the flexibility of data-driven approaches, and greatly enhancing the model's generalization ability and decision reliability under extreme conditions.

[0078] This application also provides a computer-readable medium having program instructions stored thereon, which, when executed by a processor, implement the multimodal time-series data fusion modeling method provided in the above-described method embodiments.

[0079] This application also provides a computer program product containing instructions that, when run on a computer, causes the computer to execute the multimodal time-series data fusion modeling method of the above-described method embodiments.

[0080] Please see Figure 11 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 11 As shown, the electronic device 1000 may include: at least one processor 1001, at least one network interface 1004, a user interface 1003, a memory 1005, and at least one communication bus 1002.

[0081] The communication bus 1002 is used to realize the connection and communication between these components.

[0082] The user interface 1003 may include a display screen and a camera. Optionally, the user interface 1003 may also include a standard wired interface and a wireless interface.

[0083] The network interface 1004 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface).

[0084] The processor 1001 may include one or more processing cores. The processor 1001 connects to various parts within the electronic device 1000 using various interfaces and lines. It executes various functions and processes data by running or executing instructions, programs, code sets, or instruction sets stored in the memory 1005, and by calling data stored in the memory 1005. Optionally, the processor 1001 may be implemented using at least one hardware form selected from Digital Signal Processing (DSP), Field-Programmable Gate Array (FPGA), and Programmable Logic Array (PLA). The processor 1001 may integrate one or more of the following: a Central Processing Unit (CPU), a Graphics Processing Unit (GPU), and a modem. The CPU primarily handles the operating system, user interface, and applications; the GPU is responsible for rendering and drawing the content to be displayed on the screen; and the modem handles wireless communication. It is understood that the modem may also be implemented as a separate chip, without being integrated into the processor 1001.

[0085] The memory 1005 may include random access memory (RAM) or read-only memory. Optionally, the memory 1005 may include a non-transitory computer-readable storage medium. The memory 1005 can be used to store instructions, programs, code, code sets, or instruction sets. The memory 1005 may include a program storage area and a data storage area, wherein the program storage area may store instructions for implementing an operating system, instructions for at least one function (such as touch function, sound playback function, image playback function, etc.), instructions for implementing the above-described method embodiments, etc.; the data storage area may store data involved in the above-described method embodiments, etc. Optionally, the memory 1005 may also be at least one storage system located remotely from the aforementioned processor 1001. Figure 11 As shown, the memory 1005, which serves as a computer storage medium, may include an operating system, a network communication module, a user interface module, and a multimodal time-series data fusion modeling application.

[0086] exist Figure 11In the illustrated electronic device 1000, the user interface 1003 is mainly used to provide an input interface for the user and to acquire user input data; while the processor 1001 can be used to call the multimodal time series data fusion modeling application stored in the memory 1005 and specifically perform the following operations: Acquire and preprocess historical multi-source heterogeneous batch data of the target distillation equipment to obtain a training dataset on a unified time axis; A mechanism-data hybrid driven model is constructed, which includes a mechanism skeleton layer and an AI parameter correction layer. The AI ​​parameter correction layer is used to take observable process features as input and learn the real-time correction amount relative to the basic reference value of the parameter to be inverted. The mechanism skeleton layer is used to dynamically update the basic reference value of the parameter to be inverted by calling the real-time correction amount output by the AI ​​parameter correction layer during calculation, and predict the process state at the next moment based on the updated parameter to be inverted. The AI ​​parameter correction layer is trained using the training dataset so that the mechanism skeleton layer can incorporate the updated parameters to be inverted during the training process of the AI ​​parameter correction layer to calculate the predicted process state. A pre-trained mechanism-data hybrid driven model is generated when the error between the predicted process state and the actual process state in the training dataset is minimized.

[0087] In one embodiment, the processor 1001, when executing the mechanism-data hybrid driven model, specifically performs the following operations: Generate at least one differential equation based on the law of conservation of mass or the law of conservation of energy; At least one key parameter in the differential equation used to characterize heat transfer efficiency or mass transfer efficiency is determined as a parameter to be inverted that dynamically changes with operating conditions, and a basic reference value is set for the parameter to be inverted. A mechanistic framework layer is generated by setting differential equations based on fundamental reference values. An AI parameter correction layer is constructed using machine learning algorithms; By connecting the mechanism framework layer with the AI ​​parameter correction layer, a mechanism-data hybrid driven model is obtained.

[0088] In one embodiment, when the processor 1001 generates at least one differential equation based on the law of conservation of mass or the law of conservation of energy, it specifically performs the following operations: For the reaction vessel in the target distillation equipment, based on the law of conservation of mass, a material balance differential equation is constructed to describe the liquid level or the cumulative amount of distilled products within the vessel. This material balance differential equation characterizes the weight change rate of the isopropanol receiving tank as a function of the condensation rate of the vapor distillate, and the condensation rate of the vapor distillate is related to the parameters to be inverted and the temperature and pressure within the reaction vessel; or, For the reaction vessel in the target distillation equipment, based on the law of conservation of energy, an energy balance differential equation is constructed to describe the dynamic temperature change within the reaction vessel. This energy balance differential equation characterizes the rate of temperature change over time as a function of the difference between the input and output heat. The input heat is calculated from the steam flow rate and latent heat of steam determined by the opening of the steam regulating valve. The output heat is determined by the heat transfer rate between the reactor jacket and the material inside the reactor, and the heat transfer rate is positively correlated with the real-time jacket temperature difference and the parameters to be inverted.

[0089] In one embodiment, when the processor 1001 generates the mechanistic skeleton layer by executing a differential equation with a base reference value set, it specifically performs the following operations: Substitute the basic reference values ​​into the differential equation as static model parameters; Configure the computation logic; The static model parameters and computational logic are combined into a mechanistic skeleton layer; the computational logic performs the following operations within each control cycle: Receive real-time correction values ​​from the AI ​​parameter correction layer, and superimpose the real-time correction values ​​onto the base reference values ​​to calculate the updated parameters to be inverted; The updated parameters to be inverted are substituted into the differential equation, and the process state at the next moment is predicted by forward calculation using a numerical integration algorithm. The predicted process state includes at least the temperature inside the reactor, the pressure inside the reactor, and the cumulative weight of the isopropanol receiving tank.

[0090] In one embodiment, when the processor 1001 executes the following operations to calculate the predicted process state by training an AI parameter correction layer using a training dataset so that the mechanism skeleton layer incorporates the updated parameters to be inverted during the training process of the AI ​​parameter correction layer: Initialize the model parameters of the AI ​​parameter correction layer and load the preset basic reference values ​​of the parameters to be inverted into the mechanism skeleton layer; The current sample is read from the training dataset in batches, and the observable process features arranged along the time axis in the current sample are input into the AI ​​parameter correction layer. The real-time correction amount of the current sample relative to the basic reference value is output. The real-time correction is superimposed on the base reference value to obtain the updated parameters to be inverted; The updated parameters to be inverted are substituted into the differential equation of the mechanism framework layer, and the process state at the next moment is predicted by forward calculation using a numerical integration algorithm, thus obtaining the predicted process state.

[0091] In one embodiment, the processor 1001 also performs the following operations: Calculate the loss function value between the predicted process state and the actual process state in the training dataset; The loss function value is used as the error between the predicted process state and the actual process state in the training dataset. If the error has not reached its minimum, the gradient is calculated using the backpropagation algorithm based on the loss function value, and the model parameters of the AI ​​parameter correction layer are updated. Continue executing the step of iterating through the current sample in batches from the training dataset until the error between the predicted process state and the actual process state in the training dataset reaches its minimum.

[0092] In one embodiment, when the processor 1001 performs the following operations to obtain the predicted process state by substituting the updated parameters to be inverted into the differential equation of the mechanism skeleton layer and forward calculating using a numerical integration algorithm to predict the process state at the next moment: Substitute the updated parameters to be inverted into the differential equation of the mechanistic framework layer; The differential equation is discretized and solved using a numerical integration algorithm to calculate the change in process state from the current time to the next time; wherein the process state includes at least one of the following: reactor temperature, reactor pressure, and material level in the reactor. Based on the changes in process status and the actual process status at the current moment, the predicted process status for the next moment is determined.

[0093] In one embodiment, when processor 1001 acquires and preprocesses historical multi-source heterogeneous batch data of the target distillation equipment to obtain a training dataset on a unified timeline, it specifically performs the following operations: Extract time-series process data of the target distillation equipment from the distributed control system, and extract batch quality data of the target distillation equipment from the laboratory information management system; Using the start time of the production order as the base time, the time-series process data and batch quality data are converted and mapped to the same time axis to obtain historical multi-source heterogeneous batch data; The historical multi-source heterogeneous batch data is cleaned and reconstructed to obtain a complete time series sequence. From the complete time series, derived feature variables are extracted to characterize the process state of the reactor; The time-series process data, batch quality data, and derived characteristic variables were normalized to eliminate the influence of dimensions and obtain all normalized data. All normalized data are organized into samples in batches to form a training dataset. Each sample contains a feature sequence arranged along the time axis and its label.

[0094] In one embodiment, when the processor 1001 extracts derived feature variables characterizing the process state of the reactor from the complete time sequence, it specifically performs the following operations: Calculate the real-time difference between the reactor jacket temperature and the reactor internal temperature in the complete time series, and use it as the real-time jacket temperature difference; The first derivative is performed on the weight data of the isopropanol receiving tank in the complete time series to calculate the real-time distillation rate; Based on the temperature and pressure change curves inside the reactor in the complete time sequence, and combined with the preset distillation process stage division rules, the heating stage, the initial stage of distillation, and the later stage of distillation are automatically identified and marked. Real-time jacket temperature difference, real-time distillation rate, marked heating stage, initial distillation stage, and later distillation stage are used as derived characteristic variables to characterize the process state of the reactor.

[0095] In this application embodiment, on the one hand, this application introduces an AI parameter correction layer to dynamically update the parameters to be inverted in the model, and uses a data-driven approach to capture and compensate for parameter drift caused by time-varying factors in real time, thereby effectively overcoming the accuracy decay problem caused by the inability of traditional models to adaptively adjust, and significantly improving the prediction accuracy and robustness of the model throughout its entire life cycle; on the other hand, this application constructs a mechanism skeleton layer with differential equations as the core, which forcibly constrains the prediction logic within the framework of physical laws, allowing the AI ​​model to learn the residual correction of physical parameters rather than directly fitting the final result, thereby ensuring that the output at each moment conforms to objective physical laws while retaining the flexibility of data-driven approaches, and greatly enhancing the model's generalization ability and decision reliability under extreme conditions.

[0096] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program for multimodal time-series data fusion modeling can be stored in a computer-readable storage medium. When executed, the program can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory, or random access memory, etc.

[0097] The above-disclosed embodiments are merely preferred embodiments of this application and should not be construed as limiting the scope of this application. Therefore, any equivalent variations made in accordance with the claims of this application shall still fall within the scope of this application.

Claims

1. A method for multimodal time series data fusion modeling, characterized in that, The method includes: Acquire and preprocess historical multi-source heterogeneous batch data of the target distillation equipment to obtain a training dataset on a unified time axis; A mechanism-data hybrid driven model is constructed, comprising a mechanism skeleton layer and an AI parameter correction layer. The AI ​​parameter correction layer is used to take observable process features as input and learn real-time correction amounts relative to the base reference values ​​of the parameters to be inverted. The mechanism skeleton layer is used to dynamically update the base reference values ​​of the parameters to be inverted by calling the real-time correction amounts output by the AI ​​parameter correction layer during calculation, and predict the process state at the next moment based on the updated parameters to be inverted. The AI ​​parameter correction layer is trained using the training dataset so that the mechanism skeleton layer incorporates the parameters to be inverted updated during the training of the AI ​​parameter correction layer to calculate the predicted process state. When the error between the predicted process state and the actual process state in the training dataset is minimized, a pre-trained mechanism-data hybrid driven model is generated. The mechanism-data hybrid driven model includes: generating at least one differential equation based on the law of conservation of mass or the law of conservation of energy; determining at least one key parameter in the differential equation used to characterize heat transfer efficiency or mass transfer efficiency as a parameter to be inverted that dynamically changes with operating conditions, and setting a basic reference value for the parameter to be inverted; generating a mechanism skeleton layer through the differential equation with the basic reference value set; constructing an AI parameter correction layer through a machine learning algorithm; and connecting the mechanism skeleton layer and the AI ​​parameter correction layer to obtain the mechanism-data hybrid driven model. The generation of at least one differential equation based on the law of conservation of mass or the law of conservation of energy includes: for the reactor in the target distillation equipment, based on the law of conservation of mass, constructing a material balance differential equation to describe the liquid level of the material in the reactor or the cumulative amount of distilled products, wherein the material balance differential equation is used to characterize the weight change rate of the isopropanol receiving tank as a function of the condensation rate of the vapor distillate, and the condensation rate of the vapor distillate is related to the parameters to be inverted and the temperature and pressure in the reactor; for the reactor in the target distillation equipment, based on the law of conservation of energy, constructing an energy balance differential equation to describe the dynamic change of temperature in the reactor, wherein the energy balance differential equation is used to characterize the rate of change of temperature in the reactor over time as a function of the difference between the input heat and the output heat; wherein the input heat is calculated from the steam flow rate and the latent heat of steam determined by the opening of the steam regulating valve, and the output heat is determined by the heat transfer rate between the reactor jacket and the material in the reactor, and the heat transfer rate is positively correlated with the real-time jacket temperature difference and the parameters to be inverted; The step of generating a mechanistic skeleton layer by setting a differential equation with basic reference values ​​includes: substituting the basic reference values ​​into the differential equation as static model parameters; configuring computational logic; and combining the static model parameters and the computational logic into a mechanistic skeleton layer. The computational logic performs the following operations in each control cycle: receiving real-time correction values ​​from the AI ​​parameter correction layer and superimposing these real-time correction values ​​onto the basic reference values ​​to calculate updated parameters to be inverted; substituting the updated parameters to be inverted into the differential equation and using a numerical integration algorithm to forward calculate and predict the process state at the next moment; wherein the predicted process state includes at least the reactor temperature, reactor pressure, and the cumulative weight of the isopropanol receiving tank.

2. The method according to claim 1, characterized in that, The step of training the AI ​​parameter correction layer using the training dataset, so that the mechanism skeleton layer incorporates the parameters to be inverted updated during the training of the AI ​​parameter correction layer to calculate the predicted process state, includes: Initialize the model parameters of the AI ​​parameter correction layer, and load the preset basic reference values ​​of the parameters to be inverted into the mechanism skeleton layer; The current sample is read from the training dataset in batches, and the observable process features arranged along the time axis in the current sample are input to the AI ​​parameter correction layer. The real-time correction amount corresponding to the current sample relative to the basic reference value is output. The real-time correction is superimposed on the base reference value to obtain the updated inversion parameters; The updated parameters to be inverted are substituted into the differential equation of the mechanism framework layer, and the process state at the next moment is predicted by forward calculation using a numerical integration algorithm to obtain the predicted process state.

3. The method according to claim 2, characterized in that, The method further includes: Calculate the loss function value between the predicted process state and the actual process state in the training dataset; The loss function value is used as the error between the predicted process state and the actual process state in the training dataset. If the error does not reach its minimum, the gradient is calculated using the backpropagation algorithm based on the loss function value, and the model parameters of the AI ​​parameter correction layer are updated. Continue executing the step of iterating through and reading the current sample from the training dataset in batches until the error between the predicted process state and the actual process state in the training dataset reaches its minimum.

4. The method according to claim 2, characterized in that, The step of substituting the updated parameters to be inverted into the differential equation of the mechanism framework layer, and using a numerical integration algorithm to forward calculate and predict the process state at the next moment to obtain the predicted process state includes: Substitute the updated parameters to be inverted into the differential equation of the mechanism framework layer; The differential equation is discretized and solved using a numerical integration algorithm to calculate the change in process state from the current time to the next time; wherein the process state includes at least one of the following: reactor temperature, reactor pressure, and material level in the reactor. Based on the change in process state and the actual process state at the current moment, the predicted process state for the next moment is determined.

5. The method according to claim 1, characterized in that, The acquisition and preprocessing of historical multi-source heterogeneous batch data of the target distillation equipment to obtain a training dataset on a unified time axis includes: Extract the time-series process data of the target distillation equipment from the distributed control system, and extract the batch quality data of the target distillation equipment from the laboratory information management system; Using the start time of the production order as the base time, the time-series process data and the batch quality data are converted and mapped to the same time axis to obtain historical multi-source heterogeneous batch data; The historical multi-source heterogeneous batch data is cleaned and reconstructed to obtain a complete time series sequence. From the complete time series, derived feature variables are extracted to characterize the process state of the reactor; The time-series process data, the batch quality data, and the derived feature variables are normalized to eliminate the influence of dimensions and obtain all normalized data. The normalized data is organized into samples in batches to form the training dataset, and each sample contains a feature sequence and its label arranged along the time axis.

6. The method according to claim 5, characterized in that, The step of extracting derived feature variables from the complete time series to characterize the process state of the reactor includes: Calculate the real-time difference between the reactor jacket temperature and the reactor interior temperature in the complete time sequence, and use it as the real-time jacket temperature difference; The first derivative operation is performed on the weight data of the isopropanol receiving tank in the complete time series to calculate the real-time distillation rate; Based on the temperature and pressure change curves inside the reactor in the complete time sequence, and combined with the preset distillation process stage division rules, the heating stage, the initial stage of distillation, and the later stage of distillation are automatically identified and marked. The real-time jacket temperature difference, the real-time distillation rate, and the marked heating stages, initial distillation stage, and later distillation stage are used as derived characteristic variables to characterize the process state of the reactor.

7. A process state prediction method, characterized in that, The method includes: During the online operation of the target distillation equipment, the latest observation data within a preset period is acquired in real time; The pre-trained mechanism-data hybrid driven model is invoked; wherein the pre-trained parameter correction model is trained by the multimodal time series data fusion modeling method described in any one of claims 1-6; The observable process features within the current time and historical time window are extracted from the latest observation data, and the observable process features are input into the AI ​​parameter correction layer in the pre-trained mechanism-data hybrid driven model. The AI ​​parameter correction layer outputs a target correction amount relative to the basic reference value of the parameter to be inverted, and the mechanism skeleton layer in the mechanism-data hybrid driving model uses the target correction amount to dynamically update the basic reference value to obtain the updated target parameter to be inverted. Substitute the target parameters to be inverted into the differential equation of the mechanism framework layer, and predict the target process state at the next moment by forward calculation using a numerical integration algorithm; The target process state at the next moment is output to guide operators in adjusting process parameters or as a setpoint input for the closed-loop control system.

8. A multimodal time-series data fusion modeling apparatus implemented using the method described in any one of claims 1-6, characterized in that, The device includes: The historical multi-source heterogeneous batch data acquisition module is used to acquire and preprocess the historical multi-source heterogeneous batch data of the target distillation equipment to obtain a training dataset on a unified time axis. The model building module is used to construct a mechanism-data hybrid driven model, which includes a mechanism skeleton layer and an AI parameter correction layer. The AI ​​parameter correction layer is used to take observable process features as input and learn the real-time correction amount relative to the basic reference value of the parameter to be inverted. The mechanism skeleton layer is used to dynamically update the basic reference value of the parameter to be inverted by calling the real-time correction amount output by the AI ​​parameter correction layer during calculation, and predict the process state at the next moment based on the updated parameter to be inverted. The training prediction module is used to train the AI ​​parameter correction layer using the training dataset, so that the mechanism skeleton layer can use the parameters to be inverted updated during the training of the AI ​​parameter correction layer to calculate the predicted process state. The model generation module is used to generate a pre-trained mechanism-data hybrid driven model when the error between the predicted process state and the actual process state in the training dataset is minimized.

9. A process state prediction device, characterized in that, The device includes: The latest observation data acquisition module is used to acquire the latest observation data within a preset period in real time during the online operation of the target distillation equipment; The model invocation module is used to invoke a pre-trained mechanism-data hybrid driven model; wherein the pre-trained parameter correction model is trained by the multimodal temporal data fusion modeling method described in any one of claims 1-6; The feature processing module is used to extract observable process features from the latest observation data within the current time and historical time window, and input the observable process features into the AI ​​parameter correction layer in the pre-trained mechanism-data hybrid driven model. The basic reference value update module is used to output a target correction amount relative to the basic reference value of the parameter to be inverted through the AI ​​parameter correction layer, and the mechanism skeleton layer in the mechanism-data hybrid driving model uses the target correction amount to dynamically update the basic reference value to obtain the updated target parameter to be inverted. The process state prediction module is used to substitute the target parameters to be inverted into the differential equation of the mechanism skeleton layer, and predict the target process state at the next moment through forward calculation using a numerical integration algorithm. The process status output module is used to output the target process status at the next moment to guide operators in adjusting process parameters or as the setpoint input for the closed-loop control system.

Images