Industrial parameter prediction methods, apparatus, computer equipment and readable storage media

Abstract

Embodiments of the present application provide an industrial parameter prediction method and device, computer equipment and a readable storage medium. The method comprises: obtaining reactor basic data at multiple time points; for each time point, performing material balance calculation according to the reactor basic data to obtain reaction kinetics characteristics, and sequentially constructing a first feature sequence based on multiple reaction kinetics characteristics corresponding to the multiple time points; for each time point, performing solution cumulative volume calculation according to the reactor basic data to obtain time cumulative effect characteristics, and sequentially constructing a second feature sequence based on multiple time cumulative effect characteristics corresponding to the multiple time points; obtaining original sensor data, and inputting a mechanism fusion vector obtained by fusing the original sensor data, the first feature sequence and the second feature sequence into a target model to obtain an industrial parameter prediction result corresponding to the current time point. In this way, the accuracy of industrial parameter prediction can be improved.

Description

Technical Field

[0001] This application relates to the field of industrial parameter measurement technology, and in particular to an industrial parameter prediction method, apparatus, computer equipment, and readable storage medium. Background Technology

[0002] In modern intelligent manufacturing, enterprises rely on real-time and accurate sensing of key process parameters to achieve efficient, stable, and intelligent production control. For example, in industries such as chemical engineering, metallurgy, and pharmaceuticals, parameters like pH value and reaction conversion rate are core indicators determining product quality and production safety. However, these key parameters are often difficult to obtain directly due to environmental limitations and measurement lags. Therefore, accurate prediction of these parameters is necessary to ensure the stability of the production process, the consistency of product quality, and the safety of equipment operation.

[0003] In related technologies, purely data-driven models (such as neural networks and support vector machines) are generally used to predict industrial parameters. Specifically, this method collects raw measurement data from the production process, uses it as input features to directly train the model, and then uses the trained model to predict unknown operating conditions. However, the performance of purely data-driven models is highly dependent on the coverage of the training data. When actual operating conditions deviate (such as equipment aging, changes in raw material batches, or fluctuations in environmental parameters), the model struggles to adapt to the new scenario, leading to inaccurate predictions. Summary of the Invention

[0004] This application proposes an industrial parameter prediction method, apparatus, computer equipment, and readable storage medium, which can improve the accuracy of industrial parameter prediction.

[0005] To achieve the above objectives, a first aspect of this application proposes an industrial parameter prediction method, the method comprising:

[0006] Acquire basic reactor data corresponding to multiple time points, wherein the multiple time points include the current time point and multiple historical time points preceding the current time point;

[0007] For each time point, material balance calculations are performed based on the corresponding reactor basic data to obtain reaction kinetic characteristics, and a first feature sequence is constructed sequentially based on the multiple reaction kinetic characteristics corresponding to the multiple time points.

[0008] For each time point, the cumulative volume of the solution is calculated based on the corresponding reactor basic data to obtain the time cumulative effect characteristics, and a second feature sequence is constructed sequentially based on the multiple time cumulative effect characteristics corresponding to the multiple time points;

[0009] The raw sensor data is acquired, and the mechanism fusion vector obtained by fusing the raw sensor data, the first feature sequence, and the second feature sequence is input into the target model to obtain the industrial parameter prediction result corresponding to the current time point;

[0010] The target model is obtained by performing mechanism learning on a preset model. The mechanism learning is based on the difference between the sample industrial parameter prediction results output by the preset model at the prediction time point and the sample industrial parameters, the first mechanism parameter value, and the second mechanism parameter value, respectively. The preset model is subjected to difference minimization learning. The first mechanism parameter value is calculated based on the sample reaction kinetic characteristics corresponding to the prediction time point, and the second mechanism parameter value is calculated based on the sample time cumulative effect characteristics corresponding to the prediction time point.

[0011] Accordingly, a second aspect of the embodiments of this application provides an industrial parameter prediction device, the device comprising:

[0012] The acquisition module is used to acquire basic reactor data corresponding to multiple time points, wherein the multiple time points include the current time point and multiple historical time points preceding the current time point;

[0013] The calculation module is used to perform material balance calculations based on the corresponding reactor basic data for each time point, obtain reaction kinetic characteristics, and construct a first feature sequence based on the multiple reaction kinetic characteristics corresponding to the multiple time points in sequence.

[0014] The construction module is used to calculate the cumulative volume of solution based on the corresponding reactor basic data for each time point, obtain the time cumulative effect characteristics, and construct a second feature sequence based on the multiple time cumulative effect characteristics corresponding to the multiple time points in sequence.

[0015] The input module is used to acquire raw sensor data and input the mechanism fusion vector obtained by fusing the raw sensor data, the first feature sequence, and the second feature sequence into the target model to obtain the industrial parameter prediction result corresponding to the current time point;

[0016] The target model is obtained by performing mechanism learning on a preset model. The mechanism learning is based on the difference between the sample industrial parameter prediction results output by the preset model at the prediction time point and the sample industrial parameters, the first mechanism parameter value, and the second mechanism parameter value, respectively. The preset model is subjected to difference minimization learning. The first mechanism parameter value is calculated based on the sample reaction kinetic characteristics corresponding to the prediction time point, and the second mechanism parameter value is calculated based on the sample time cumulative effect characteristics corresponding to the prediction time point.

[0017] In some embodiments, the industrial parameter prediction device further includes a training module for:

[0018] Acquire basic data of the sample reactor corresponding to multiple sample time points, wherein the multiple sample time points include the predicted time point and multiple sample historical time points preceding the predicted time point.

[0019] For each sample time point, material balance calculations are performed based on the corresponding sample reactor basic data to obtain the sample reaction kinetic characteristics, and the first feature sequence of the sample is constructed sequentially based on the multiple sample reaction kinetic characteristics corresponding to the multiple sample time points.

[0020] For each sample time point, the cumulative volume of the solution is calculated based on the corresponding sample reactor basic data to obtain the sample time cumulative effect characteristics, and a sample second feature sequence is constructed sequentially based on the multiple sample time cumulative effect characteristics corresponding to the multiple sample time points.

[0021] The original sensor data of the sample is acquired, and the sample mechanism fusion vector obtained by fusing the original sensor data of the sample, the first feature sequence of the sample, and the second feature sequence of the sample is input into the preset model to obtain the sample industrial parameter prediction result corresponding to the prediction time point;

[0022] The first mechanism parameter value is calculated based on the sample reaction kinetic characteristics corresponding to the predicted time point, and the second mechanism parameter value is calculated based on the sample time cumulative effect characteristics corresponding to the predicted time point;

[0023] Obtain the sample industrial parameters corresponding to the prediction time point, and construct a first sub-loss based on the difference between the prediction result of the sample industrial parameters and the sample industrial parameters; construct a second sub-loss based on the difference between the prediction result of the sample industrial parameters and the value of the first mechanism parameter; construct a third sub-loss based on the difference between the prediction result of the sample industrial parameters and the value of the second mechanism parameter; and construct a target loss based on the first sub-loss, the second sub-loss and the third sub-loss.

[0024] Based on the target loss, the parameters of the preset model are adjusted to obtain the target model.

[0025] In some implementations, the training module is further configured to:

[0026] Based on the average of multiple historical first sub-losses corresponding to the preset model, a first average is determined; based on the average of multiple historical second sub-losses, a second average is determined; and based on the average of multiple historical third sub-losses, a third average is determined.

[0027] A first fluctuation reference value is determined based on the difference between the first sub-loss and the first mean; a second fluctuation reference value is determined based on the difference between the second sub-loss and the second mean; and a third fluctuation reference value is determined based on the difference between the third sub-loss and the third mean.

[0028] The reference sum is determined based on the sum of the first fluctuation reference value, the second fluctuation reference value, and the third fluctuation reference value;

[0029] A first weight is determined based on the ratio of the first fluctuation reference value to the reference sum; a second weight is determined based on the ratio of the second fluctuation reference value to the reference sum; and a third weight is determined based on the ratio of the third fluctuation reference value to the reference sum.

[0030] The first sub-loss is adjusted based on the first weight to obtain the target first sub-loss; the second sub-loss is adjusted based on the second weight to obtain the target second sub-loss; and the third sub-loss is adjusted based on the third weight to obtain the target third sub-loss.

[0031] The target loss is constructed based on the sum of the first sub-loss, the second sub-loss, and the third sub-loss.

[0032] In some implementations, the training module is further configured to:

[0033] Obtain the historical loss of the previous training round of the preset model, and calculate the difference between the historical loss and the target loss to obtain the loss decrease rate value;

[0034] Obtain a preset learning speed benchmark value, and based on the relationship between the loss decrease rate value and the learning speed benchmark value, adjust the weight matrices of the memory unit layer and the fully connected layer of the preset model to obtain the adjusted memory unit layer and the fully connected layer.

[0035] The model parameters of the preset model are adjusted to obtain the updated preset model;

[0036] The parameters of the preset model are repeatedly adjusted based on the target loss until the preset number of training iterations is reached, thereby obtaining the target model.

[0037] In some implementations, the weight matrix includes a first weight matrix corresponding to the memory unit layer and a second weight matrix corresponding to the fully connected layer. The training module is further configured to:

[0038] Based on the relationship between the loss decrease rate value and the learning speed benchmark value, the adjustment rank value corresponding to the target loss is determined, wherein when the loss decrease rate value is greater than the learning speed benchmark value, the adjustment rank value is less than the historical adjustment rank value corresponding to the historical loss, or when the loss decrease rate value is less than the learning speed benchmark value, the adjustment rank value is greater than the historical adjustment rank value.

[0039] The first low-rank matrix corresponding to the first weight matrix is ​​adjusted based on the adjusted rank value, and the memory cell layer is updated according to the adjusted first low-rank matrix to obtain the updated memory cell layer.

[0040] The second low-rank matrix corresponding to the second weight matrix is ​​adjusted based on the adjusted rank value, and the fully connected layer is updated according to the adjusted second low-rank matrix to obtain the updated fully connected layer.

[0041] In some embodiments, the reactor basic data includes inlet flow rate data, outlet flow rate data, temperature data, input concentration data, reactor internal concentration data, and reactor volume data. The calculation module is further used for:

[0042] For each time point, a first parameter is determined based on the ratio of the inlet flow rate data to the reactor volume data;

[0043] Obtain a preset reaction rate constant, and determine a second parameter based on the ratio of the outlet flow rate data to the reactor volume data. Then, determine a target second parameter based on the sum of the second parameter and the preset reaction rate constant.

[0044] Based on the product of the first parameter and the input concentration data, a first product is obtained, and based on the ratio of the first product to the target second parameter, a first ratio is obtained;

[0045] Obtain the initial time point, and determine the third parameter based on the initial time point, each time point, and the target second parameter;

[0046] The first material balance term is obtained based on the product of the first ratio and the third parameter;

[0047] The second material balance term is obtained by multiplying the concentration data in the reactor with the third parameter.

[0048] The reaction kinetic characteristics at each time point are obtained based on the sum of the first material balance term and the second material balance term.

[0049] In some implementations, the reactor basic data includes inlet flow rate data, outlet flow rate data, and reactor volume data. The construction module is further configured to:

[0050] For each time point, residence flow data is calculated based on the inlet flow data and outlet flow data, and the average residence time is obtained based on the ratio of the reactor volume data to the residence flow data.

[0051] Based on the average residence time and each time point, the residence time distribution description information of the material in the reactor is determined;

[0052] The cumulative amount description information is obtained by multiplying the dwell time distribution description information and the import flow data;

[0053] The accumulated amount description information is integrated over time to obtain the time-cumulative effect characteristics.

[0054] Accordingly, a third aspect of the present application provides a computer device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the industrial parameter prediction method of any one of the embodiments of the first aspect of the present application.

[0055] Accordingly, a fourth aspect of the embodiments of this application proposes a computer-readable storage medium storing a computer program that, when executed by a processor, implements the industrial parameter prediction method of any one of the embodiments of the first aspect of this application.

[0056] This application embodiment acquires reactor basic data corresponding to multiple time points, including the current time point and multiple historical time points preceding the current time point. For each time point, material balance calculations are performed based on the corresponding reactor basic data to obtain reaction kinetic characteristics, and a first feature sequence is constructed sequentially based on the multiple reaction kinetic characteristics corresponding to the multiple time points. For each time point, solution cumulative volume calculations are performed based on the corresponding reactor basic data to obtain time cumulative effect characteristics, and a second feature sequence is constructed sequentially based on the multiple time cumulative effect characteristics corresponding to the multiple time points. Raw sensor data is acquired, and the mechanism fusion vector obtained by fusing the raw sensor data, the first feature sequence, and the second feature sequence is input into the target model to obtain the industrial parameter prediction result corresponding to the current time point. The target model is obtained by performing mechanism learning on a preset model. Mechanism learning is based on minimizing the differences between the sample industrial parameter prediction result output by the preset model for the prediction time point and the sample industrial parameter, the first mechanism parameter value, and the second mechanism parameter value, respectively. The first mechanism parameter value is calculated based on the sample reaction kinetic characteristics corresponding to the prediction time point, and the second mechanism parameter value is calculated based on the sample time cumulative effect characteristics corresponding to the prediction time point. In this way, by calculating reaction kinetic characteristics (such as material balance) and time-cumulative effect characteristics (such as solution cumulative volume), the dynamic relationships implicit in the process mechanism (such as reaction rate and equipment aging trend) can be transformed into a quantifiable feature sequence, which is then concatenated with the original sensor data to form a mechanism fusion vector. This overcomes the deficiency of purely data-driven models in perceiving complex physical processes, ensuring that the input features not only include apparent data but also embed the essential laws of the process, enhancing the model's sensitivity to changes in key parameters. Furthermore, the target model is trained through mechanism learning, forcing the model output to not only match the sample industrial parameters but also minimize the differences from mechanistic parameters such as reaction kinetics and time-cumulative effects. That is, physical constraints are introduced into the loss function, enabling the model to learn the causal mechanisms behind the data, rather than shallow statistical correlations. Therefore, this application, through the explicit construction of mechanism features, enhances the model's robustness to changes in operating conditions, reduces the impact of insufficient data coverage, and improves the accuracy of industrial parameter predictions. Attached Figure Description

[0057] Figure 1 This is a schematic diagram of the architecture of the industrial parameter prediction system provided in the embodiments of this application;

[0058] Figure 2 This is a flowchart of the industrial parameter prediction method provided in the embodiments of this application;

[0059] Figure 3 This is a model structure diagram provided in the embodiments of this application;

[0060] Figure 4 This is a framework diagram of the overall industrial parameter prediction method provided in the embodiments of this application;

[0061] Figure 5 This is a comparison chart of the lightweight performance of the models provided in the embodiments of this application;

[0062] Figure 6 This is a comparison chart of the prediction results of multiple models provided in the embodiments of this application;

[0063] Figure 7 This is a comparison chart of the accuracy of predictions from multiple models provided in the embodiments of this application;

[0064] Figure 8 This is a schematic diagram of the functional modules of the industrial parameter prediction device provided in the embodiments of this application;

[0065] Figure 9 This is a schematic diagram of the hardware structure of the computer device provided in the embodiments of this application. Detailed Implementation

[0066] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0067] It should be noted that although functional modules are divided in the device schematic diagram and a logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than the module division in the device or the order in the flowchart. The terms "first," "second," etc., in the specification, claims, and the aforementioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0068] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0069] In modern intelligent manufacturing, enterprises rely on real-time and accurate sensing of key process parameters to achieve efficient, stable, and intelligent production control. For example, in industries such as chemical engineering, metallurgy, and pharmaceuticals, parameters like pH value and reaction conversion rate are core indicators determining product quality and production safety. However, these key parameters are often difficult to obtain directly due to environmental limitations and measurement lags. Therefore, accurate prediction of these parameters is necessary to ensure the stability of the production process, the consistency of product quality, and the safety of equipment operation.

[0070] In related technologies, purely data-driven models (such as neural networks and support vector machines) are generally used to predict industrial parameters. Specifically, this method collects raw measurement data from the production process, uses it as input features to directly train the model, and then uses the trained model to predict unknown operating conditions. However, the performance of purely data-driven models is highly dependent on the coverage of the training data. When actual operating conditions deviate (such as equipment aging, changes in raw material batches, or fluctuations in environmental parameters), the model struggles to adapt to the new scenario, leading to inaccurate predictions.

[0071] Based on this, embodiments of this application provide an industrial parameter prediction method, apparatus, computer equipment, and readable storage medium, which can improve the accuracy of industrial parameter prediction.

[0072] The industrial parameter prediction method, apparatus, computer equipment, and readable storage medium provided in this application are specifically described through the following embodiments. First, the industrial parameter prediction system in the embodiments of this application is described.

[0073] Please refer to Figure 1 In some embodiments, this application provides an industrial parameter prediction system, including a terminal 11 and a server 12.

[0074] For example, terminal 11, as a core edge device in the industrial field, is mainly responsible for real-time data acquisition, lightweight model inference, and local control. Specifically, terminal 11 can be used to acquire sensor data (such as flow rate, temperature, hydrogen ion concentration, etc.) in the industrial field in real time, perform data preprocessing (such as filtering, interpolation, and feature fusion), and run a lightweight target model (such as a KD-LLSTM model) for online prediction to obtain prediction results. Alternatively, terminal 11 can also send the preprocessed data to server 12 for prediction, receive the prediction results returned by server 12, and directly integrate the prediction results into the industrial control system to achieve real-time feedback control (such as adjusting valve opening), ensuring the stability and continuity of the production process while avoiding the high maintenance costs of relying on physical sensors. For example, the prediction results can be high-precision real-time estimates of key parameters (such as pH value), which can be directly used for local control decisions.

[0075] Furthermore, the terminal 11 can be an embedded module of an industrial-grade programmable logic controller (PLC), a distributed control system (DCS), an edge computing gateway, or an industrial control computer, etc. The specific device of the terminal 11 needs to meet the harsh environment of the industrial site (such as high temperature, strong electromagnetic interference), low power consumption requirements, and real-time requirements.

[0076] For example, server 12 acts as the backend computing hub, focusing on model building, optimization, and historical data analysis to ensure the overall performance of the soft measurement system. Specifically, server 12 can perform the construction, training, and compression of lightweight pre-defined models, such as training models using sample data and ensuring that the prediction results conform to physical laws (such as reaction kinetics and material balance) through a composite loss function.

[0077] Furthermore, server 12 can be a local industrial server, such as a rack server equipped with a graphics processing unit (GPU) accelerator card, a private cloud platform, or a high-performance computing cluster (such as a multi-node GPU cluster) in a data center, etc.

[0078] In some implementations, terminal 11 uploads raw sensor data and prediction results to server 12 in real time for model performance monitoring, data drift detection, and iterative training. Server 12 can send the trained model to terminal 11 for deployment to adapt to its hardware computing power limitations. Based on the data fed back by terminal 11 and changes in operating conditions (such as changes in reaction rate caused by temperature fluctuations), server 12 can trigger model retraining or parameter adjustment, and redeploy the updated model to terminal 11 to achieve adaptive closed-loop optimization of the industrial parameter prediction system.

[0079] The industrial parameter prediction method in this application can be illustrated through the following embodiments.

[0080] It should be noted that in all specific embodiments of this application, when processing data related to user identity or characteristics, such as user information, user behavior data, user historical data, and user location information, user permission or consent will be obtained first. Furthermore, the collection, use, and processing of this data will comply with relevant laws, regulations, and standards. In addition, when embodiments of this application require access to sensitive personal information of users, separate permission or consent from the user will be obtained through pop-ups or redirects to confirmation pages. Only after obtaining the user's separate permission or consent will the necessary user-related data for the normal operation of the embodiments of this application be obtained.

[0081] In this embodiment, the description will focus on an industrial parameter prediction device, which can be integrated into a computer device. See [link to relevant documentation]. Figure 2 , Figure 2 This is a flowchart illustrating the steps of the industrial parameter prediction method provided in this application embodiment. Taking the industrial parameter prediction device specifically integrated into a terminal or server as an example, the specific process when the processor on the terminal or server executes the program instructions corresponding to the industrial parameter prediction method is as follows:

[0082] Step 101: Obtain basic reactor data corresponding to multiple time points, including the current time point and multiple historical time points preceding the current time point.

[0083] In some implementations, in order to achieve accurate modeling of the dynamic characteristics of industrial processes, reactor baseline data containing current and historical time points can be collected to construct an input dataset with temporal correlation, thereby improving the model's prediction accuracy and physical consistency.

[0084] In this context, multiple time points can be time series nodes in an industrial process divided according to a fixed sampling period (such as per second or per minute) or an event-triggered mechanism. These multiple time points typically include the current time point for which industrial parameters need to be predicted, as well as several preceding time points, so that the target model can construct a time series based on these multiple time points. For example, if the current time point is tn, then the multiple time points can include t1, t2, ..., tn, etc., which can be obtained through real-time sensor data acquisition or by backtracking from a historical database.

[0085] A reactor can be a device used to realize chemical reactions, physical changes, or nuclear reaction processes, and can be applied in chemical, oil refining, and metallurgical industries.

[0086] Among them, the basic data of the reactor can be physical quantities that describe the operating status of the reactor, such as inlet flow rate data, outlet flow rate data, temperature data, hydrogen ion inlet concentration, etc., which are obtained in real time by sensors (such as electromagnetic flow meters and temperature probes) installed in the reactor pipes and chambers.

[0087] The current time point can be the latest data moment received and processed by the system in real time, which can be determined by synchronous acquisition of data from the clock signal and sensors of the industrial control system. The current time point is also the time point at which industrial parameter prediction is required.

[0088] Among them, historical time points can be multiple recorded timestamps (such as tn-1, tn-2, ... tn-1, tn-2, ...) before the current time point (such as tn), which can be obtained by extracting past process parameters stored in industrial data acquisition systems or historical databases.

[0089] In some implementations, sensors can be installed at key locations in each reactor (such as inlet, outlet, specific internal points, etc.) to monitor parameters such as temperature, pressure, and flow rate. The sensors are then connected to a data acquisition system, which automatically records and stores data at each point in time to form time-series data.

[0090] For example, the basic data of a reactor may include temperature data, pressure data, input flow rate data, output flow rate data, hydrogen ion concentration data, chemical component concentration data, etc., and relevant sensor data can be selected as the basic data of the reactor according to the actual situation.

[0091] By using the above methods, basic reactor data with temporal continuity and physical correlation can be obtained to characterize the dynamic evolution of industrial processes. This provides high-fidelity time-series input data for the subsequent construction of reaction kinetics characteristics and time cumulative effect characteristics, thereby facilitating the improvement of the model's prediction accuracy and stability for key parameters under complex operating conditions.

[0092] Step 102: For each time point, perform material balance calculations based on the corresponding reactor basic data to obtain reaction kinetic characteristics, and construct a first feature sequence based on multiple reaction kinetic characteristics corresponding to multiple time points in sequence.

[0093] In some implementations, in order to obtain the dynamic change characteristics of the reaction process within multiple time points corresponding to the current time point, the kinetic characteristics of the reaction process can be accurately obtained by performing material balance calculations on the reactor basic data at each time point, and a first feature sequence can be constructed based on multiple reaction kinetic characteristics corresponding to multiple time points to obtain the trend of reactor basic data changing over time, thereby achieving high-precision prediction of key parameters of industrial processes.

[0094] Among them, reaction kinetic characteristics can be parameters that describe the changes in chemical reaction rate, reactant and product concentrations over time, reflecting the kinetic behavior of the reaction process in the reactor.

[0095] The first feature sequence can be an ordered set of reaction kinetic features corresponding to multiple time points arranged in chronological order. It is formed by integrating the reaction kinetic feature values ​​of each time point according to the temporal logic of the process flow, and is used to characterize the dynamic evolution law of the reaction process.

[0096] Understandably, since the definition of pH directly depends on the activity or concentration of hydrogen ions, the core of calculating pH lies in measuring the hydrogen ion concentration in the solution. The concentration of hydrogen ions can be measured to quickly and accurately determine the acidity or alkalinity of a solution, providing crucial data support for scientific experiments and industrial production.

[0097] In some implementations, it can be based on the collected reactor baseline data (including inlet flow data). Export flow data Temperature data T, input concentration data (hydrogen ion inlet concentration) Concentration data in the reactor and reactor volume data Based on the principle of mass conservation and the laws of reaction kinetics, the dynamic changes of hydrogen ion concentration are calculated at each time point, reaction kinetic characteristics are extracted, and the first characteristic sequence is constructed.

[0098] Specifically, based on the principle of mass conservation in the reactor, the change in hydrogen ion concentration can be determined by the amount of hydrogen ions brought in by the inlet flow rate data. The amount of hydrogen ions discharged in the export flow data The amount of hydrogen ions consumed or generated during the reaction The value is determined by the reaction rate (r) and reactor volume (V). Therefore, the rate of change of hydrogen ion concentration over time can be expressed as:

[0099] (1)

[0100] Furthermore, since the reaction follows a first-order kinetic model, the reaction rate is related to the concentration data (hydrogen ion concentration) within the reactor. The relationship is:

[0101] (2)

[0102] Where k is the reaction rate constant, determined by the Arrhenius equation:

[0103] (3)

[0104] Where A is the frequency factor. Here, R is the activation energy, R is the gas constant, and T is the temperature.

[0105] Therefore, we can substitute formula (2) into formula (1) and rearrange to obtain the differential equation:

[0106] (4)

[0107] Furthermore, to simplify the calculation, we can let , Substituting into formula (4), we get:

[0108] (5)

[0109] Furthermore, with initial conditions (For example Starting with the measured concentration at time ( ), we solve formula (5) to obtain the analytical solution for the hydrogen ion concentration, which is also the reaction kinetic characteristic:

[0110] (6)

[0111] The above formula describes the dynamic change of hydrogen ion concentration over time, and its reaction kinetics are reflected in the decay term. This can reflect the effect of the reaction rate constant k and temperature T on the concentration change. Larger concentrations (i.e., faster reaction rates or higher temperatures) lead to faster concentration decay; and a steady-state term is also present. , which represents the hydrogen ion concentration after the system reaches equilibrium.

[0112] In some implementations, parameters can be identified using a particle swarm optimization algorithm at each time point. and The corresponding data is injected into the above formula (6) to calculate the corresponding reaction kinetic characteristics. The reaction kinetic characteristics of multiple time points are arranged in chronological order to obtain the first characteristic sequence, which is used to reflect the evolution of hydrogen ion concentration over time (such as rapid decrease, slow stabilization, etc.) and to indirectly characterize the changes in key process parameters such as reaction rate, temperature effect, and material flow rate.

[0113] By using the above methods, the basic reactor data obtained from the original sensor data can be transformed into reaction kinetic characteristics with clear physical meaning, and then into quantifiable feature sequences. This enhances the model's ability to model long-term time-series dependencies and dynamic characteristics, thereby significantly improving the stability and accuracy of the prediction results.

[0114] In some implementations, to achieve high-precision modeling of the reaction kinetics of complex industrial processes, the reaction kinetic characteristics at each time point can be calculated by constructing material balance equations to quantify the dynamic evolution of hydrogen ion concentration, fully consider the physical mechanisms, and effectively avoid model prediction distortion caused by neglecting the physical mechanisms. For example, the reactor basic data may include inlet flow rate data, outlet flow rate data, temperature data, input concentration data, reactor internal concentration data, and reactor volume data. Step 102, "for each time point, perform material balance calculations based on the corresponding reactor basic data to obtain reaction kinetic characteristics," may include:

[0115] (102.1) For each time point, determine the first parameter based on the ratio of inlet flow rate data to reactor volume data;

[0116] (102.2) Obtain the preset reaction rate constant, and determine the second parameter based on the ratio of the outlet flow rate data to the reactor volume data, and determine the target second parameter based on the sum of the second parameter and the preset reaction rate constant;

[0117] (102.3) Based on the product of the first parameter and the input concentration data, the first product is obtained, and based on the ratio of the first product to the target second parameter, the first ratio is obtained;

[0118] (102.4) Obtain the initial time point, and determine the third parameter based on the initial time point, each time point, and the target second parameter;

[0119] (102.5) Based on the product of the first ratio and the third parameter, the first material balance term is obtained;

[0120] (102.6) Based on the product of the concentration data in the reactor and the third parameter, the second material balance term is obtained;

[0121] (102.7) Based on the sum of the first material balance term and the second material balance term, the reaction kinetic characteristics at each time point are obtained.

[0122] The inlet flow rate data can be the volume of material flowing through the reactor inlet per unit time, which can be obtained in real time by an electromagnetic flow meter installed at the pipeline inlet.

[0123] The reactor volume data can be the geometric volume of the reactor cavity (e.g., V=330 L), which can be obtained through equipment nameplate parameters or 3D modeling measurements at the industrial site.

[0124] The first parameter can be the ratio of the inlet flow rate to the reactor volume, which can be used to characterize the amount of material entering the reactor per unit volume per unit time, that is, the contribution rate of the inlet flow rate to the material concentration in a unit volume reactor.

[0125] Among them, the reaction rate constant can be a temperature-dependent parameter describing the rate of a chemical reaction, which can be preset through experimental calibration or literature data.

[0126] Among them, the outlet flow rate data can be the volume of material flowing through the reactor outlet per unit time, which can be obtained in real time by a vortex flow meter installed at the pipeline outlet of the reactor.

[0127] The second parameter can be the ratio of the outlet flow rate to the reactor volume. It can be used to characterize the sum of the material output rate (outlet flow rate data) and the chemical reaction rate per unit volume per unit time, and to take into account the dual effects of natural material outflow and chemical reaction consumption on the concentration in the reactor.

[0128] The target second parameter can be the sum of the second parameter and the reaction rate constant.

[0129] The first product can be the product of the first parameter and the input hydrogen ion concentration, which is used to characterize the total amount of hydrogen ions brought in by the imported material per unit volume of reactor per unit time, that is, to quantify the degree to which the external material input drives the change in the hydrogen ion concentration of the reaction system.

[0130] The first ratio can be the ratio of the first product to the target second parameter, which is used to characterize the steady-state hydrogen ion concentration contributed by the imported material when the system reaches dynamic equilibrium under the current operating conditions. In other words, it is used to reflect the hydrogen ion concentration level that the system will eventually approach under the current feed and reaction conditions.

[0131] The initial time point can be the reactor start-up time or the start timestamp of data acquisition. It can be obtained through the time synchronization signal of the industrial control system.

[0132] The third parameter can be a time-cumulative effect factor calculated based on the initial time point, the current time point, and the second parameter. It can be used to characterize the degree of decay of the initial influence during the evolution of the system from the initial state to the current state.

[0133] The first material balance term can be the product of the first ratio and the third parameter.

[0134] The second material balance term can be the product of the hydrogen ion concentration in the reactor and the third parameter.

[0135] In some implementations, for each time point, based on import flow data Reactor volume data The ratio determines the first parameter. The formula can be expressed as follows:

[0136] ;

[0137] Understandably, the first parameter reflects the relationship between the rate at which material enters the reactor and the reactor size.

[0138] Furthermore, it can be based on export flow data. Reactor volume data The ratio determines the second parameter. The target second parameter is determined based on the sum of the second parameter and the preset reaction rate constant k. The reaction rate constant k is determined by the Arrhenius equation and reflects the relationship between the reaction rate and temperature. Calculating the second objective parameter can synthesize the effects of material outflow and chemical reaction on the material concentration within the reactor.

[0139] For example, it can be based on the first parameter Compared with input concentration data The product of these two factors yields the first product, which is the product of the amount of material entering per unit volume per unit time and its concentration. This first product is then used in conjunction with the target second parameter. The ratio of the two values ​​is used to obtain the first ratio. This is to determine the relative relationship between the incoming material and the outflow and reaction rate.

[0140] Furthermore, it can be based on the initial time point At each time point t and the target's second parameter The third parameter is determined to be The dynamic characteristics of the concentration decay over time during the reaction of materials in the reactor are determined by using a third parameter.

[0141] Furthermore, the first material balance term can be obtained based on the product of the first ratio and the third parameter: Therefore, the cumulative effect of material changes over time in a dynamic equilibrium state can be described by the first material balance term.

[0142] For example, it can be based on concentration data within the reactor. Product with the third parameter The second material balance term is obtained: Therefore, the influence of the initial state on the subsequent reaction process can be reflected through the second material balance term.

[0143] Finally, the reaction kinetics at the current time point t can be calculated based on the sum of the first and second material balance terms. To obtain the material concentration in the reactor at each time point:

[0144] ;

[0145] In some implementations, the reaction kinetic characteristics at each time point can be directly calculated by directly inputting the reactor's basic data (such as inlet flow rate data, outlet flow rate data, temperature data, input concentration data, reactor internal concentration data, and reactor volume data) at each time point into the formula for the reaction kinetic characteristics derived above.

[0146] In this way, not only can the physicochemical laws of industrial processes (such as reaction kinetic equations and material balance principles) be transformed into quantifiable time-series characteristics, but also high-value mechanism-driven inputs can be provided for subsequent soft sensor models. This reduces model complexity while enhancing the prediction accuracy and generalization performance of reaction process state changes.

[0147] Step 103: For each time point, calculate the cumulative volume of the solution based on the corresponding reactor basic data to obtain the time cumulative effect characteristics, and construct a second feature sequence based on multiple time cumulative effect characteristics corresponding to multiple time points in sequence.

[0148] In some implementations, in order to achieve accurate modeling of the dynamic characteristics of material residence time in industrial processes, the characteristics of time accumulation effect can be calculated based on reactor basic data, and a second feature sequence can be formed in sequence to enhance the model's ability to characterize long time delays and time accumulation effects, and avoid prediction bias caused by ignoring residence time distribution.

[0149] Among them, the time accumulation effect characteristic can be a dynamic characteristic calculated based on the reactor residence time distribution function and the material accumulation amount, which is used to characterize the dynamic behavior of material accumulation in the reactor over time.

[0150] The second feature sequence can be an ordered set of time cumulative effect features corresponding to multiple time points arranged in chronological order. It can be formed by integrating the cumulative effect feature values ​​of each time point according to the temporal logic of the process flow, and is used to characterize the evolution law of the material residence time distribution.

[0151] In some implementations, the residence time of materials in the reactor plays a decisive role in the reaction effect. Therefore, the cumulative effect characteristics at each time point can be calculated to enable the model to understand the dynamic influence of the distribution of the residence time of materials in the reactor on the industrial parameters (such as pH value) that need to be predicted.

[0152] Furthermore, for a single reactor (or a reactor system formed by multiple reactors in series), this allows us to determine the probability density of the material's residence time t in the reactor, for example, when t equals When E(t) decreases to This indicates that the residence time of most materials is concentrated in Nearby. Specifically, the description of the dwell time distribution (i.e., the dwell time distribution function) is as follows:

[0153] ;

[0154] in, The average stay time is calculated using the following formula:

[0155] ;

[0156] Where V is the reactor volume, and F can be the average of the inlet and outlet flow rates extracted from the reactor's basic data, i.e., (inlet flow rate + outlet flow rate) / 2. This eliminates the influence of instantaneous fluctuations and reflects the stable flow characteristics of the material within the reactor.

[0157] In some implementations, for a reactor system consisting of three reactors connected in series, the overall residence time distribution information can be used for calculation, specifically:

[0158] (7)

[0159] Where t represents each time point in the current calculation. The calculation logic for the overall residence time distribution description of multiple reactors in series is the same as that for the overall residence time distribution description of three reactors in series, and will not be elaborated here.

[0160] In some implementations, the cumulative amount of material in the reactor can be calculated using the residence time distribution. Taking reactor No. 1 as an example, the cumulative volume of the hydrogen ion-containing solution is:

[0161] (8)

[0162] Substituting formula (7) into formula (8), we can obtain the time cumulative effect characteristics for each time point that needs to be calculated:

[0163] ;

[0164] in, This is import flow data.

[0165] Furthermore, by sequentially summarizing and storing the cumulative effect features corresponding to multiple time points into the corresponding sequence R, a second feature sequence corresponding to a single reactor or multiple reactor systems can be obtained, such as... .

[0166] By using the above methods, the physical laws governing the distribution of material residence time in industrial processes can be transformed into quantifiable temporal features, thereby enhancing the model's ability to model long time delays and dynamic cumulative effects. This provides input features with both physical consistency and dynamic characteristics for the subsequent training of lightweight neural networks, improving the stability and generalization of key parameter predictions.

[0167] In some implementations, to achieve accurate quantification of the material residence time distribution characteristics in industrial processes, residence flow rate and average residence time can be calculated based on flow rate data and reactor volume, and further, residence time cumulative effect characteristics can be constructed to enhance the model's ability to model the physical laws governing the dynamic behavior of materials, effectively solving the prediction bias problem caused by neglecting the time-cumulative effect. For example, the reactor basic data includes inlet flow rate data, outlet flow rate data, and reactor volume data. Step 103, "for each time point, calculate the cumulative solution volume based on the corresponding reactor basic data to obtain the time-cumulative effect characteristics," may include:

[0168] (103.1) For each time point, calculate the residence flow rate based on the inlet flow rate data and the outlet flow rate data, and obtain the average residence time based on the ratio of the reactor volume data to the residence flow rate data;

[0169] (103.2) Determine the residence time distribution description information of the material in the reactor based on the average residence time and each time point;

[0170] (103.3) The cumulative volume description information is obtained by multiplying the residence time distribution description information and the import flow data;

[0171] (103.4) Perform time integration on the cumulative amount description information to obtain the time cumulative effect characteristics.

[0172] The residence flow rate data can be the difference between the reactor inlet flow rate and the reactor outlet flow rate, which is used to characterize the stability of material flow within the reactor.

[0173] The average residence time can be the ratio of reactor volume to residence flow rate, used to characterize the average residence time of materials in the reactor.

[0174] The dwell time distribution description information can be based on the dwell time distribution function of the average dwell time and the current time point, that is, the probability distribution used to describe the dwell time of materials.

[0175] The cumulative quantity description information can be the product of the residence time distribution description information and the inlet flow data, which is used to quantify the contribution of dynamic input to the reaction process.

[0176] In some implementations, dwell flow data can be used to eliminate the impact of instantaneous flow fluctuations, ensuring the stability of subsequent calculations. For example, dwell flow data... It can be based on import flow data and export flow data The arithmetic mean was calculated as follows:

[0177] ;

[0178] Furthermore, average stay time The calculation formula is:

[0179] ;

[0180] Where V is the volume of a single reactor. By calculating the average residence time, the dynamic retention characteristics of the material within the reactor can be quantified.

[0181] In some implementations, the residence time distribution description information (i.e., residence time distribution function) for a single reactor is as follows:

[0182] ;

[0183] In some implementations, for a reactor system consisting of three reactors connected in series, the overall residence time distribution is used to describe the information. The calculation can be specifically as follows:

[0184] ;

[0185] By calculating the residence time distribution description information, the flow model within the reactor (such as plug flow, completely mixed flow, etc.) can be accurately identified.

[0186] Furthermore, for a reactor system consisting of n reactors in series, the expression for its overall residence time distribution is as follows:

[0187] ;

[0188] The specific expression can be adjusted according to the actual situation.

[0189] Further, cumulative description information The calculation formula is:

[0190] ;

[0191] in, To input traffic data, This provides information on the residence time distribution. By calculating the cumulative information, the flux of material carried by the portion of the material entering the reactor per unit time with a residence time of t can be determined.

[0192] By integrating the cumulative information over time from 0 to t, the following formula can be obtained:

[0193] ;

[0194] Descriptive information on the distribution of dwell time Substituting the expression into the above formula, we can obtain the time cumulative effect characteristics for each time point that needs to be calculated as follows:

[0195] ;

[0196] in, This is import flow data. Therefore, by integrating over time, the cumulative amount of material at each point in time from the initial moment to the present can be quantified.

[0197] By using the above methods, the physical laws governing the distribution of material residence time in industrial processes can be transformed into quantifiable temporal characteristics, thereby enhancing the model's ability to model long time delays and dynamic cumulative effects. This significantly improves the process optimization capability and the model's prediction accuracy for industrial parameters, and has broad industrial application value.

[0198] Step 104: Obtain the original sensor data, and input the mechanism fusion vector obtained by fusing the original sensor data, the first feature sequence, and the second feature sequence into the target model to obtain the industrial parameter prediction result corresponding to the current time point;

[0199] The target model is obtained by performing mechanism learning on the preset model. Mechanism learning is based on the difference between the predicted industrial parameters of the sample output at the prediction time point and the sample industrial parameters, the first mechanism parameter value, and the second mechanism parameter value, respectively. The preset model is subjected to difference minimization learning. The first mechanism parameter value is calculated based on the sample reaction kinetic characteristics corresponding to the prediction time point, and the second mechanism parameter value is calculated based on the sample time cumulative effect characteristics corresponding to the prediction time point.

[0200] In some implementations, in order to achieve high-precision prediction of key parameters of industrial processes, a mechanism fusion vector with strong physical interpretability and data-driven advantages can be constructed by fusing mechanistic features (first feature sequence and second feature sequence) that reflect physical laws with the actual observed raw sensor data. This vector can then be input into the target model to fully explore the nonlinear relationship between process variables, thereby improving the prediction accuracy and robustness of key industrial parameters at the current time point and realizing real-time, accurate soft measurement and dynamic detection of complex industrial process states.

[0201] The raw sensor data can be data that can be obtained in real time and accurately at the industrial site, such as electrical parameters and material composition.

[0202] The mechanism fusion vector can be a high-dimensional input vector composed of the original sensor data, the first feature sequence (reaction kinetic characteristics), the second feature sequence (time cumulative effect characteristics), and the basic reactor data, such as... Mechanism fusion vectors can effectively guide the knowledge of the model.

[0203] The target model can be a lightweight neural network model optimized through mechanism learning (such as the KD-LLSTM model), that is, a model obtained by learning and optimizing parameters through machine learning algorithms, whose prediction process can fully follow known physical laws.

[0204] Among them, the industrial parameter prediction results can be the key parameter estimates (such as pH value prediction) output by the target model at the current time point, which are generated by inputting the mechanism fusion vector into the target model and performing forward inference.

[0205] The preset model can be an initial model without physical constraints, such as the KD-LLSTM model.

[0206] The prediction time point can be the target moment when industrial parameters need to be predicted during the model training or inference phase.

[0207] Among them, the sample industrial parameter prediction results can be intermediate prediction values ​​output by the preset model in the training state at the prediction time point.

[0208] Among them, the industrial parameters of the sample can be actual measured and verified true key parameter values ​​(such as the true value of pH).

[0209] The first mechanism parameter value can be the pH value in the reactor calculated based on the reaction kinetic characteristics of the sample at the predicted time point, using the reaction kinetic equation.

[0210] The second mechanism parameter value can be the pH value in the reactor calculated based on the time cumulative effect equation and the sample time cumulative effect characteristics corresponding to the predicted time point.

[0211] Among them, the sample reaction kinetic characteristics can be parameters that describe the changes in chemical reaction rate, reactant and product concentrations over time at the predicted time point, reflecting the kinetic behavior of the reaction process in the reactor.

[0212] Among them, the cumulative effect feature of sample time can be the cumulative amount of material calculated for the prediction time point.

[0213] In some implementations, the raw sensor data (e.g., relevant material data, electrical parameters, etc.), the first feature sequence (reaction kinetic characteristics), the second feature sequence (time cumulative effect characteristics), and the reactor basic data (inlet flow rate data, outlet flow rate data, temperature data, input concentration data, reactor internal concentration data, and reactor volume data) can be directly concatenated to obtain a mechanism fusion vector, such as... Then, the mechanism fusion vector is directly input into the trained target model, and the target model can output the corresponding pH value, such as a pH value of 7.2.

[0214] In some implementations, the sample response kinetics characteristics and sample time cumulative effect characteristics corresponding to the predicted time point can be obtained. The calculation method for the sample response kinetics characteristics at the predicted time point is similar to that for calculating the response kinetics characteristics at each time point as described above. The calculation method for the sample time cumulative effect characteristics at the predicted time point is similar to that for calculating the time cumulative effect characteristics at each time point as described above. The calculation method has been described above and will not be repeated here.

[0215] For example, the concentration of hydrogen ions at the current predicted time point can be determined based on the sample reaction kinetic characteristics corresponding to the predicted time point. Then, using the calculated concentration of hydrogen ions, the pH value is defined as the negative logarithm of the hydrogen ion concentration. Therefore, the negative logarithm of the hydrogen ion concentration needs to be taken to obtain the pH value, which is the first mechanism parameter value.

[0216] In some implementations, the cumulative consumption of reactants (such as acids) in an acid-base neutralization reaction affects the concentration of hydrogen ions in the solution. Therefore, the concentration of hydrogen ions can be calculated based on the cumulative effect characteristics of the sample time corresponding to the predicted time point by combining other chemical equilibrium equations or reaction kinetic equations, and then the pH value can be calculated. The pH value here is the second mechanism parameter value.

[0217] Furthermore, based on the pH value predicted by the preset model (the predicted result of the sample industrial parameters), the differences between these values ​​and the actual sample industrial parameters, the first mechanism parameter value, and the second mechanism parameter value can be compared. The first sub-loss, the second sub-loss, and the third sub-loss can be calculated using the mean square error formula (or other error formulas). Finally, the first sub-loss, the second sub-loss, and the third sub-loss are added together to construct the target loss. By minimizing the target loss, the preset model is subjected to mechanism learning. This ensures that the preset model not only fits the actual observation data during the mechanism learning process but also follows the physicochemical laws corresponding to the reaction kinetic equation and the time cumulative effect equation. This achieves the task of knowledge-guided model training and can significantly improve the physical consistency and accuracy of the pH prediction by the trained target model.

[0218] In some implementations, besides predicting pH, this application can be applied to different industrial scenarios. For example, the industrial parameter prediction method can cover various key parameters (such as concentration, conversion rate, energy consumption, etc.) in fields such as chemistry, metallurgy, chemical engineering, and materials. The raw sensor data and corresponding mechanistic characteristics can be dynamically adjusted according to the physical and / or chemical laws of the predicted industrial parameters, but the model design and optimization framework remain consistent. Specific concepts can be found in the embodiments of this application regarding pH prediction as an industrial parameter, which will not be elaborated upon here.

[0219] This application embodiment acquires reactor basic data corresponding to multiple time points, including the current time point and multiple historical time points preceding the current time point. For each time point, material balance calculations are performed based on the corresponding reactor basic data to obtain reaction kinetic characteristics, and a first feature sequence is constructed sequentially based on the multiple reaction kinetic characteristics corresponding to the multiple time points. For each time point, solution cumulative volume calculations are performed based on the corresponding reactor basic data to obtain time cumulative effect characteristics, and a second feature sequence is constructed sequentially based on the multiple time cumulative effect characteristics corresponding to the multiple time points. Raw sensor data is acquired, and the mechanism fusion vector obtained by fusing the raw sensor data, the first feature sequence, and the second feature sequence is input into the target model to obtain the industrial parameter prediction result corresponding to the current time point. The target model is obtained by performing mechanism learning on a preset model. Mechanism learning is based on minimizing the differences between the sample industrial parameter prediction result output by the preset model for the prediction time point and the sample industrial parameter, the first mechanism parameter value, and the second mechanism parameter value, respectively. The first mechanism parameter value is calculated based on the sample reaction kinetic characteristics corresponding to the prediction time point, and the second mechanism parameter value is calculated based on the sample time cumulative effect characteristics corresponding to the prediction time point. In this way, by calculating reaction kinetic characteristics (such as material balance) and time-cumulative effect characteristics (such as solution cumulative volume), the dynamic relationships implicit in the process mechanism (such as reaction rate and equipment aging trend) can be transformed into a quantifiable feature sequence, which is then concatenated with the original sensor data to form a mechanism fusion vector. This overcomes the deficiency of purely data-driven models in perceiving complex physical processes, ensuring that the input features not only include apparent data but also embed the essential laws of the process, enhancing the model's sensitivity to changes in key parameters. Furthermore, the target model is trained through mechanism learning, forcing the model output to not only match the sample industrial parameters but also minimize the differences from mechanistic parameters such as reaction kinetics and time-cumulative effects. That is, physical constraints are introduced into the loss function, enabling the model to learn the causal mechanisms behind the data, rather than shallow statistical correlations. Therefore, this application, through the explicit construction of mechanism features, enhances the model's robustness to changes in operating conditions, reduces the impact of insufficient data coverage, and improves the accuracy of industrial parameter predictions.

[0220] In some implementations, to ensure the target model maintains stable and reliable predictive performance under varying operating conditions or data noise interference, an input vector fusing physical mechanism features and data-driven features can be constructed based on multi-source data at sample time points. The preset model is then optimized by introducing a triple-constraint loss function comprising reaction kinetics (first mechanism parameter value), time accumulation effect (second mechanism parameter value), and true values. This enhances the model's physical consistency and generalization ability, enabling the preset model to learn not only data-driven pattern recognition capabilities during training but also the physical and chemical mechanisms of chemical processes, thereby improving the model's prediction accuracy, generalization ability, and robustness. For example, the target model can be trained in the following manner:

[0221] (A.1) Obtain basic data of the sample reactor corresponding to multiple sample time points, wherein the multiple sample time points include the predicted time point and multiple historical sample time points preceding the predicted time point;

[0222] (A.2) For each sample time point, material balance calculation is performed based on the corresponding sample reactor basic data to obtain the sample reaction kinetic characteristics, and the first feature sequence of the sample is constructed sequentially based on the multiple sample reaction kinetic characteristics corresponding to multiple sample time points.

[0223] (A.3) For each sample time point, the cumulative volume of solution is calculated based on the corresponding sample reactor basic data to obtain the sample time cumulative effect characteristics, and the sample second feature sequence is constructed sequentially based on the multiple sample time cumulative effect characteristics corresponding to multiple sample time points.

[0224] (A.4) Obtain the original sensor data of the sample, and input the sample mechanism fusion vector obtained by fusing the original sensor data of the sample, the first feature sequence of the sample, and the second feature sequence of the sample into the preset model to obtain the sample industrial parameter prediction result corresponding to the prediction time point;

[0225] (A.5) The first mechanism parameter value is calculated based on the sample reaction kinetic characteristics corresponding to the predicted time point, and the second mechanism parameter value is calculated based on the sample time cumulative effect characteristics corresponding to the predicted time point;

[0226] (A.6) Obtain the sample industrial parameters corresponding to the prediction time point, and construct the first sub-loss based on the difference between the prediction result of the sample industrial parameters and the sample industrial parameters; construct the second sub-loss based on the difference between the prediction result of the sample industrial parameters and the value of the first mechanism parameter; construct the third sub-loss based on the difference between the prediction result of the sample industrial parameters and the value of the second mechanism parameter; and construct the target loss based on the first sub-loss, the second sub-loss and the third sub-loss.

[0227] (A.7) Based on the target loss, the parameters of the preset model are adjusted to obtain the target model.

[0228] In this model, multiple sample time points can be time series nodes in an industrial process divided according to a fixed sampling period (such as per second or per minute) or an event-triggered mechanism. These multiple sample time points typically include the prediction time point for the industrial parameter to be predicted, as well as multiple preceding sample time points, so that the pre-defined model can assemble a sample time series based on these multiple sample time points. For example, if the sample time point is tn, then multiple sample time points can include t1, t2, ..., tn, etc., which can be obtained from a historical database or a pre-defined sample dataset.

[0229] Among them, the basic data of the sample reactor can be physical quantities that describe the operating status of the reactor, such as inlet flow rate data, outlet flow rate data, temperature data, hydrogen ion inlet concentration, etc., which can be obtained by sensors (such as electromagnetic flow meters and temperature probes) installed in the reactor pipes and chambers, or obtained from historical databases and sample datasets.

[0230] Among them, the historical time points of the sample can be multiple recorded timestamps (such as tn-1, tn-2, ... tn-1, tn-2, ...) before the predicted time point (such as tn), which can be obtained by extracting past process parameters stored in the industrial data acquisition system or historical database.

[0231] Among them, the sample reaction kinetic characteristics can be parameters that describe the changes in chemical reaction rate, reactant and product concentrations over time, reflecting the kinetic behavior of the reaction process in the reactor.

[0232] The first feature sequence of the sample can be an ordered set of sample reaction kinetic features corresponding to multiple sample time points arranged in chronological order. It is formed by integrating the sample reaction kinetic feature values ​​of each sample time point according to the temporal logic of the process flow, and is used to characterize the dynamic evolution law of the reaction process.

[0233] Among them, the sample time cumulative effect feature can be a dynamic feature calculated based on the reactor residence time distribution function and the material accumulation amount, which is used to characterize the dynamic behavior of material accumulation in the reactor over time.

[0234] The second feature sequence of the sample can be an ordered set of cumulative effect features of sample time corresponding to multiple sample time points arranged in chronological order. It is formed by integrating the cumulative effect feature values ​​of each sample time point according to the temporal logic of the process flow, and is used to characterize the evolution law of the material residence time distribution.

[0235] The sample mechanism fusion vector can be a high-dimensional input vector composed of the original sensor data of the sample, the first feature sequence of the sample (reaction kinetic characteristics), the second feature sequence of the sample (time cumulative effect characteristics), and the basic data of the sample reactor, such as... The sample mechanism fusion vector can effectively guide the model with knowledge.

[0236] Among them, the sample industrial parameter prediction results can be the key parameter estimates (such as pH value prediction) output by the preset model at the current time point, which are generated by inputting the sample mechanism fusion vector into the preset model and performing forward inference.

[0237] The first sub-loss can be the root mean square error between the predicted industrial parameters of the sample and the industrial parameters of the sample.

[0238] The second sub-loss can be the root mean square error between the predicted industrial parameters of the sample and the value of the first mechanism parameter.

[0239] The third sub-loss can be the root mean square error between the predicted industrial parameters of the sample and the value of the second mechanism parameter.

[0240] The target loss can be a weighted sum of the first sub-loss, the second sub-loss, and the third sub-loss. The weights of the first, second, and third sub-losses can be set according to the actual situation to achieve joint constraint optimization of physical laws and data laws.

[0241] In some implementations, after acquiring basic reactor data corresponding to multiple sample time points, relevant calculations are performed to obtain a first feature sequence and a second feature sequence of the sample. The sample mechanism fusion vector obtained by fusing the original sensor data, the first feature sequence, and the second feature sequence is then input into a preset model to obtain the predicted industrial parameters of the sample corresponding to the predicted time point. This process is the same as the process described above, where basic reactor data corresponding to multiple time points are acquired, relevant calculations are performed to obtain a first feature sequence and a second feature sequence, and the mechanism fusion vector obtained by fusing the original sensor data, the first feature sequence, and the second feature sequence is input into a target model to obtain the predicted industrial parameters of the current time point. The only difference is the model state (one is a preset model to be trained, and the other is a trained target model) and the data (real data and sample data). The calculation concept and process can be referred to above and will not be repeated here.

[0242] For example, the concentration of hydrogen ions at the current predicted time point can be determined based on the sample reaction kinetic characteristics corresponding to the predicted time point. Then, using the calculated hydrogen ion concentration, the pH value is defined as the negative logarithm of the hydrogen ion concentration. Therefore, the negative logarithm of the hydrogen ion concentration is needed to obtain the pH value, which is the first mechanistic parameter value. For example, if the hydrogen ion concentration at the predicted time point... The concentration is 0.02 mol / L, while the pH value (i.e., the first mechanism parameter value) is the negative logarithm of the hydrogen ion concentration. It is approximately 1.7.

[0243] In some implementations, during acid-base neutralization reactions, the cumulative consumption of reactants (such as acids) (i.e., the cumulative effect characteristic over sample time) affects the concentration of hydrogen ions in the solution. Therefore, the hydrogen ion concentration can be calculated based on the cumulative effect characteristic over sample time at the predicted time point by combining other chemical equilibrium equations or reaction kinetic equations, and then the pH value can be calculated. Here, the pH value is the second mechanism parameter value. For example, if the reactor volume is 5 liters, and the calculated cumulative effect characteristic over sample time shows that NaOH has consumed 0.9 mol of hydrogen ions, and the total amount of remaining hydrogen ions is 1.0 - 0.9 = 0.1 mol, then the current hydrogen ion concentration in the reactor is 0.1 mol / 5 = 0.02 mol. This can be calculated using the formula... The value of the second mechanism parameter can be calculated to be 1.7.

[0244] It should be noted that the methods for calculating hydrogen ions and pH may differ in different scenarios. The appropriate method should be used according to the specific circumstances.

[0245] In some implementations, accurate pH values, i.e., sample industrial parameters, can be obtained from historical databases, sample datasets, or directly measured. Then, the predicted results of the sample industrial parameters are used to calculate the differences between these values ​​and the actual sample industrial parameters, the first mechanism parameter value calculated using formulas, and the second mechanism parameter value, respectively, to obtain the corresponding first sub-loss, second sub-loss, and third sub-loss. Furthermore, the differences can be calculated using root mean square error, mean absolute error, or mean absolute percentage error methods; this application does not limit the specific calculation method.

[0246] In some implementations, the first sub-loss It can be calculated in the following way:

[0247] ;

[0248] Where N represents the total number of samples, Indicates the industrial parameters of the sample. This represents the predicted results of the sample industrial parameters. Therefore, by minimizing the predicted value (the predicted results of the sample industrial parameters),... Corresponding true values ​​(sample industrial parameters) The squared error maximizes the model's ability to fit the training data.

[0249] In some implementations, the second sub-loss It can be calculated in the following way:

[0250] ;

[0251] in, This represents the value of the first mechanistic parameter corresponding to the i-th sample. This represents the predicted results of industrial parameters for the sample. Based on this, the theoretical values ​​(i.e., the first mechanism parameter values) calculated by chemical reaction kinetic models (such as the Arrhenius equation) can be obtained. ) and the corresponding sample industrial parameter prediction results By comparison, the output of the pre-defined model is constrained to conform to physical laws.

[0252] In some implementations, the third sub-loss It can be calculated in the following way:

[0253] ;

[0254] in, This represents the value of the second mechanistic parameter corresponding to the i-th sample. This represents the predicted results of the sample industrial parameters; thus, the values ​​of the second mechanism parameters calculated based on the material residence time distribution can be used. Corresponding sample industrial parameter prediction results By comparing and constraining the model to capture the cumulative effect over time, we can ensure that the preset model captures long-term dynamic characteristics.

[0255] In some implementations, the predicted industrial parameters of the i-th sample in the current batch can be compared one by one with the predicted industrial parameters, the first mechanism parameter value, and the second mechanism parameter value. For each sub-loss calculation dimension (e.g., the first sub-loss calculation dimension), the differences corresponding to all samples are summed, and then divided by the total number of samples N to obtain the corresponding sub-loss. In this way, the impact of outliers or noise on training can be effectively reduced, ensuring that the model focuses on the overall trend rather than local noise.

[0256] In some implementations, to balance the contributions of the first sub-loss, the second sub-loss, and the third sub-loss, corresponding weights can be assigned to each of them, as follows:

[0257] ;

[0258] Furthermore, based on the actual situation, corresponding weights can be assigned to the first sub-loss, the second sub-loss, and the third sub-loss. Specifically, the first sub-loss can be assigned the first weight. For example, 0.4, assigning a second weight to the second sub-loss. For example, 0.3, assigning a third weight to the third sub-loss. For example, 0.3, etc. After adjusting the first sub-loss, second sub-loss and third sub-loss by weight, the final target loss can be obtained by adding them together.

[0259] In some implementations, the parameters of a preset model can be adjusted using a gradient descent-type optimization algorithm based on the target loss until a preset number of training iterations is reached, such as 200, or the target loss is lower than the preset value for multiple consecutive iterations (e.g., 5), at which point training can be stopped and the target model can be obtained.

[0260] By employing the above methods, the physicochemical laws of industrial processes (such as reaction kinetic equations and time-cumulative effect models) can be deeply integrated with data-driven prediction models. Through a composite loss function with triple constraints, the model can be optimized from the perspectives of data fitting, physical consistency, and long-term modeling. This guides the model to learn data patterns while following the basic principles of chemical reactions, thereby significantly improving the prediction accuracy, robustness, and generalization ability of industrial parameters (such as pH value). Especially under complex operating conditions or when the data contains noise, it can still maintain stable and reliable prediction performance, providing highly reliable decision support for real-time detection and intelligent control of industrial processes.

[0261] In some implementations, to address the model training imbalance caused by changes in operating conditions in a fixed-weight loss function, the fluctuation characteristics of each sub-loss can be dynamically sensed and weights allocated accordingly. This allows the model to more efficiently coordinate multi-objective constraints during training, significantly shortening convergence time and improving the model's adaptability to complex industrial processes and prediction accuracy. For example, "constructing the target loss based on the first, second, and third sub-losses" in (A.6) can include:

[0262] (A.6.1) Determine the first mean based on the mean of multiple historical first sub-losses corresponding to the preset model, determine the second mean based on the mean of multiple historical second sub-losses, and determine the third mean based on the mean of multiple historical third sub-losses;

[0263] (A.6.2) Based on the difference between the first sub-loss and the first mean, a first fluctuation reference value is determined; based on the difference between the second sub-loss and the second mean, a second fluctuation reference value is determined; and based on the difference between the third sub-loss and the third mean, a third fluctuation reference value is determined.

[0264] (A.6.3) Determine the reference sum based on the sum of the first fluctuation reference value, the second fluctuation reference value, and the third fluctuation reference value;

[0265] (A.6.4) Determine the first weight based on the ratio of the first fluctuation reference value to the reference sum, determine the second weight based on the ratio of the second fluctuation reference value to the reference sum, and determine the third weight based on the ratio of the third fluctuation reference value to the reference sum;

[0266] (A.6.5) Adjust the first sub-loss based on the first weight to obtain the target first sub-loss, adjust the second sub-loss based on the second weight to obtain the target second sub-loss, and adjust the third sub-loss based on the third weight to obtain the target third sub-loss;

[0267] (A.6.6) Construct the target loss based on the sum of the first sub-loss, the second sub-loss, and the third sub-loss of the target.

[0268] Among them, the historical first sub-loss can be the historical value of the first sub-loss recorded by the preset model during the training process.

[0269] The first mean can be the arithmetic mean of multiple historical first sub-losses.

[0270] Among them, the historical second sub-loss can be the historical value of the second sub-loss recorded by the preset model during the training process.

[0271] The second mean can be the arithmetic mean of multiple historical second sub-losses.

[0272] Among them, the historical third sub-loss can be the historical value of the third sub-loss recorded by the preset model during training.

[0273] The third mean can be the arithmetic mean of multiple historical third sub-losses.

[0274] The first fluctuation reference value can be the difference between the current first sub-loss and the first mean, which is used to characterize the degree of deviation between the current first sub-loss and the first mean.

[0275] The second fluctuation reference value can be the difference between the current second sub-loss and the second mean, which is used to characterize the degree of deviation between the current second sub-loss and the second mean.

[0276] The third fluctuation reference value can be the difference between the current third sub-loss and the third mean, which is used to characterize the degree of deviation between the current third sub-loss and the third mean.

[0277] The reference sum can be the sum of the first fluctuation reference value, the second fluctuation reference value, and the third fluctuation reference value, which is obtained by accumulating the three types of fluctuation reference values.

[0278] The first weight can be the proportion of the first fluctuation reference value to the total reference value, used to quantify the relative importance of the first fluctuation reference value in the total fluctuation. The larger the first weight, second weight, or third weight, the more significant the current fluctuation of the sub-loss is and the further it deviates from historical stability. The preset model will prioritize optimizing this loss term in this training step to accelerate its convergence and suppress instability.

[0279] The second weight can be the proportion of the second fluctuation reference value to the total reference value, used to quantify the relative importance of the second fluctuation reference value in the total fluctuation.

[0280] The third weight can be the proportion of the third fluctuation reference value to the total reference value, used to quantify the relative importance of the second fluctuation reference value in the total fluctuation.

[0281] The first sub-loss of the target can be the product of the first sub-loss and the first weight. The first sub-loss, second sub-loss, and third sub-loss of the target are used to characterize the actual contribution of the corresponding sub-losses after dynamic weight adjustment to the target loss.

[0282] The target second sub-loss can be the product of the second sub-loss and the second weight.

[0283] The target third sub-loss can be the product of the third sub-loss and the third weight.

[0284] In some implementations, the first mean The calculation formula is as follows:

[0285] ;

[0286] Where T represents the training epoch, which can be selected according to the actual situation, such as selecting data from the first 10 or first 15 training epochs of the current training epoch. Let represent the historical first sub-loss in the t-th training round.

[0287] Similarly, the second mean is calculated. Third mean The process is the same as the calculation formula for the first mean, and will not be elaborated here. By introducing the historical mean, a stability benchmark is established, allowing the system to determine whether the current loss deviates from the normal range. If a sub-loss consistently exceeds its historical mean, it indicates that this part has not yet converged or has been disturbed, requiring attention.

[0288] For example, in the current training epoch, the corresponding first sub-loss, second sub-loss, and third sub-loss can be obtained, and their absolute differences from the corresponding historical averages can be calculated as fluctuation reference values. Taking the first sub-loss as an example, the corresponding first fluctuation reference value... Through the first sub-loss Compared with the first mean The calculation yields the following formula: .

[0289] Furthermore, the first fluctuation reference value Second fluctuation reference value Third fluctuation reference value By adding them together, we can obtain the reference sum S.

[0290] Furthermore, the corresponding weights can be calculated based on the proportion of each fluctuation reference value in the total. This is used to calculate the first weight corresponding to the first sub-loss. For example, the calculation formula is as follows:

[0291] ;

[0292] in, This represents the first fluctuation reference value; S represents the reference sum.

[0293] Similarly, the second and third weights can also be calculated in the same way as above, and will not be elaborated here.

[0294] For example, the original sub-loss can be weighted and scaled using the calculated first, second, and third weights to obtain the adjusted target first sub-loss, target second sub-loss, and target third sub-loss. The specific process is as follows:

[0295] ;

[0296] in, Indicates the first weight. Indicates the loss of the first child; Indicates the second weight. Indicates the loss of the second child; Indicates the third weight. L represents the third sub-loss; L represents the final target loss.

[0297] By using the above methods, the optimization imbalance caused by fixed weights can be effectively avoided, the model convergence speed can be significantly accelerated (especially in the initial training or sudden change of working conditions), and the robustness of the model to noise and anomalies can be enhanced. This ensures the synergistic optimization among data fitting, physical laws and dynamic characteristics, thereby improving the prediction accuracy, stability and generalization ability of the model in complex industrial scenarios.

[0298] In some implementations, to enhance the model's adaptability to data noise and changing operating conditions in complex industrial environments, and to achieve lightweight deployment of the model, the difference between historical loss and target loss can be calculated to obtain the loss descent rate value. Based on the relationship between the loss descent rate value and the learning speed benchmark value, the weight matrices of the preset model's memory unit layer and fully connected layer can be adjusted. This balances model compression ratio and prediction accuracy, significantly reducing the model's parameter count and computational complexity while ensuring high-precision prediction capabilities, enabling the model to run stably and in real-time even on embedded devices or in resource-constrained environments. For example, (A.7) may include:

[0299] (A.7.1) Obtain the historical loss of the previous training round of the preset model, and calculate the difference between the historical loss and the target loss to obtain the loss descent rate value;

[0300] (A.7.2) Obtain the preset learning speed benchmark value, and adjust the weight matrix of the memory unit layer and the fully connected layer of the preset model based on the relationship between the loss descent rate value and the learning speed benchmark value, so as to obtain the adjusted memory unit layer and the fully connected layer.

[0301] (A.7.3) Adjust the model parameters of the preset model to obtain the updated preset model;

[0302] (A.7.4) Repeatedly adjust the parameters of the preset model based on the target loss until the preset number of training iterations is reached to obtain the target model.

[0303] The historical loss can be the target loss calculated by the preset model in the previous training round of the current training round.

[0304] The loss reduction rate value can be the difference between the target loss in the current training round and the historical loss, which reflects the current learning speed of the preset model.

[0305] The learning speed benchmark can be a preset threshold used to measure the convergence efficiency of the preset model, which can be determined through empirical settings or cross-validation.

[0306] Among them, the memory unit layer can be the core module in the preset model responsible for storing and updating temporal features, and it can achieve dynamic parameter compression through low-rank singular value decomposition.

[0307] Among them, the fully connected layer can be a linear transformation module in the output layer of the preset model that maps the hidden state to the final predicted value. It can achieve dynamic parameter compression through low-rank singular value decomposition.

[0308] The weight matrix can be the parameter matrix used for feature mapping in the memory cell layer and the fully connected layer. It can be reconstructed into a low-rank form through adaptive rank optimization singular value decomposition to reduce the number of parameters and retain the main information.

[0309] In some implementations, since the memory cell layer is the core of the preset model for capturing the long-term time dependence and dynamic evolution characteristics of industrial processes, its parameter updates directly affect the model's ability to model physical laws such as time accumulation effects and reaction dynamics; the fully connected layer is the last link that maps hidden features to predicted values ​​of key parameters (such as pH value), and its weights directly affect the output accuracy and stability; these two layers contain the most and most critical learnable parameters in the preset model and are most sensitive to changes in the loss function; therefore, the rank parameters of the weight matrices of the memory cell layer and the fully connected layer can be dynamically adjusted to ensure training efficiency while improving the model's adaptability and reliability in complex industrial scenarios.

[0310] In some implementations, while ensuring model compression efficiency, it is necessary to design a reasonable optimal update mechanism. Firstly, the historical loss of the preset model in the previous training round can be obtained. And based on historical losses and the target losses of the current training round The difference is used to calculate the rate of loss decrease. The specific calculation process is as follows:

[0311] ;

[0312] Furthermore, in order to dynamically adjust the model's compression level and learning efficiency, a learning speed baseline value can be set. This allows for lightweight design while maintaining model accuracy. The baseline learning speed can be determined based on experimental data or the operator's experience.

[0313] In some implementations, if Greater than This indicates that the preset model learns quickly, and the rank k of the weight matrices in the memory unit layer and fully connected layer can be reduced to further compress the model, accelerate model convergence, and save computational resources; if Less than When the loss decreases and then plateaus, it indicates that the learning speed of the preset model has slowed down. At this point, the rank of the weight matrix of the memory unit layer and the fully connected layer can be increased to improve the expressive power of the preset model, help the model capture more complex patterns, and avoid underfitting.

[0314] Specifically, the weight matrix of the preset model may include the input gate, forget gate, output gate, and cell state matrix.

[0315] Furthermore, after adjusting the weight matrices of the memory unit layer and the fully connected layer, the learnable parameters of the entire preset model can be updated to continuously optimize the performance of the preset model.

[0316] In some implementations, other model parameters of the preset model are adjusted to obtain an updated preset model. Then, the next training round can be started, and a new target loss is calculated in the next training round. Then, the process from (A.7.1) to (A.7.3) is repeated to adjust and obtain an updated preset model. This process is repeated until the model converges and the target model is obtained.

[0317] By using the above methods, the rank parameter of the weight matrix of the core layer of the model can be dynamically adjusted based on the rate of loss descent, so as to achieve an adaptive balance between compression efficiency and prediction accuracy during the model training process. This provides a lightweight model with high accuracy, low latency and strong generalization ability for subsequent real-time soft measurement deployment in industrial sites, significantly reducing hardware resource consumption and improving the stability and reliability of the model under complex working conditions.

[0318] In some implementations, to achieve a dynamic balance between model parameter compression efficiency and prediction accuracy, the rank value can be adjusted based on the comparison between the loss descent rate value and the learning speed benchmark value, and the low-rank representation of the weight matrix can be updated accordingly to adaptively optimize the model structure and improve the model's convergence speed and generalization ability. For example, the weight matrix includes a first weight matrix corresponding to the memory unit layer and a second weight matrix corresponding to the fully connected layer, wherein the first weight matrix includes an input gate, a forget gate, an output gate, and a cell state matrix, or may only include the cell state matrix; (A.7.2) "Adjusting the weight matrices of the memory unit layer and the fully connected layer of the preset model based on the magnitude relationship between the loss descent rate value and the learning speed benchmark value to obtain the adjusted memory unit layer and the fully connected layer" can include:

[0319] (A.7.2.1) Based on the relationship between the loss descent rate value and the learning rate benchmark value, determine the adjustment rank value corresponding to the target loss. Wherein, when the loss descent rate value is greater than the learning rate benchmark value, the adjustment rank value is less than the historical adjustment rank value corresponding to the historical loss, or when the loss descent rate value is less than the learning rate benchmark value, the adjustment rank value is greater than the historical adjustment rank value.

[0320] (A.7.2.2) Adjust the first low-rank matrix corresponding to the first weight matrix based on the adjustment rank value, and update the memory cell layer according to the adjusted first low-rank matrix to obtain the updated memory cell layer;

[0321] (A.7.2.3) Adjust the second low-rank matrix corresponding to the second weight matrix based on the adjusted rank value, and update the fully connected layer according to the adjusted second low-rank matrix to obtain the updated fully connected layer.

[0322] The adjustment rank value can be a matrix rank parameter that is dynamically adjusted based on the comparison between the current loss rate and the learning rate benchmark value. It is obtained by increasing or decreasing the historical adjustment rank value after determining whether the loss rate exceeds the benchmark value.

[0323] Among them, the historical adjustment rank value can be the rank parameter used by the preset model in the previous training round.

[0324] The first weight matrix can be the weight matrix used for feature mapping in the memory unit layer.

[0325] The first low-rank matrix can be the low-rank form of the first weight matrix after being reconstructed by adaptive rank optimization singular value decomposition, which can be generated by low-rank approximation of the first weight matrix.

[0326] The second weight matrix can be the weight matrix used for output mapping in the fully connected layer.

[0327] The second low-rank matrix can be the low-rank form of the second weight matrix after being reconstructed by adaptive rank optimization singular value decomposition, which can be generated by low-rank approximation of the second weight matrix.

[0328] In some implementations, while ensuring model compression efficiency, it is necessary to design a reasonable optimal update mechanism. Firstly, the historical loss of the preset model in the previous training round can be obtained. And based on historical losses and the target losses of the current training round The difference is used to calculate the rate of loss decrease. The specific calculation process is as follows:

[0329] ;

[0330] Furthermore, in order to dynamically adjust the model's compression level and learning efficiency, a learning speed baseline value can be set. This allows for lightweight design while maintaining model accuracy. The baseline learning speed can be determined based on experimental data or the operator's experience.

[0331] In some implementations, if Greater than This indicates that the preset model learns quickly, which can reduce the rank k of the weight matrix of the memory unit layer and the fully connected layer, i.e., set the adjustment rank value for the current round. The historical adjusted rank value corresponding to the historical loss of the previous training round is less than the historical loss of the previous training round. This further compresses the model, accelerates model convergence, and saves computational resources; if Less than When the loss decreases and plateaus, it indicates that the learning speed of the preset model has slowed down. At this point, the rank of the weight matrix of the memory unit layer and the fully connected layer can be increased, that is, the adjustment rank value for the current round can be set. The historical adjusted rank value corresponding to the historical loss of the previous training round is greater than that of the previous training round. This is to enhance the expressive power of the preset model, help the model capture more complex patterns, and avoid underfitting.

[0332] In some implementations, the core structure of the pre-defined model includes a memory unit layer (including an input gate, a forget gate, an output gate, and cell states) and a fully connected layer. The memory unit layer... For example, the process of performing singular value decomposition on its corresponding first weight matrix is ​​as follows:

[0333] ;

[0334] Where d is the feature dimension of the input mechanism fusion vector, and h is the number of neurons in the hidden layer. It is a left singular matrix. It is a singular value matrix. It is the transpose of a right singular matrix. This represents the first low-rank matrix corresponding to the first weight matrix.

[0335] Adjust the rank value determined in the current training round. Applying this to the first weight matrix yields the updated first weight matrix. :

[0336] ;

[0337] in, It is the rank of the matrix. It is the updated low-rank left singular matrix. It is the updated low-rank singular value matrix. It is the transpose of the updated low-rank right singular matrix.

[0338] In some implementations, it is possible to Replace the original weight matrix Then, the updated memory cell layer is obtained, and the fully connected layer is updated (reconstructed) using the adjusted second address matrix to obtain the updated fully connected layer. During reconstruction, settings are... To ensure that k is much smaller than the input dimension d and the hidden layer dimension h, the number of parameters is significantly reduced while ensuring that the main information in the matrix is ​​not lost. Then, the learnable parameters of the overall preset model are adjusted and updated according to the target loss to obtain the updated target model.

[0339] By using the loss descent rate as a feedback signal to dynamically adjust the low-rank representation of the weight matrices of the memory unit layer and fully connected layer in the preset model, a synergistic optimization of model compression and learning efficiency is achieved. This not only improves the training stability and convergence speed of the preset model but also significantly reduces the number of model parameters and inference latency, making it more suitable for edge computing environments in industrial settings.

[0340] Please refer to Figure 3 and Figure 4 , Figure 3 This is a model structure diagram of an embodiment of this application. Figure 4 This is a framework diagram of the overall industrial parameter prediction method according to an embodiment of this application.

[0341] For example, during the training of the preset model, basic data of the sample reactor (such as inlet flow rate data, outlet flow rate data, temperature data, concentration and volume data, etc.) corresponding to multiple sample time points can be acquired. Material balance calculations are performed for each sample time point to extract the sample reaction kinetic characteristics, and cumulative solution volume calculations are performed to extract the sample time-cumulative effect characteristics. Then, based on the multiple sample reaction kinetic characteristics corresponding to multiple sample time points, a first sample feature sequence is constructed, and based on the multiple sample time-cumulative effect characteristics corresponding to multiple sample time points, a second sample feature sequence is constructed. At the same time, the original sensor data of the samples (such as flow rate, temperature, etc.) are acquired.

[0342] Furthermore, such as Figure 3 As shown, the original sensor data, the first feature sequence, and the second feature sequence of the sample can be fused into a sample mechanism fusion vector. This vector integrates the dynamic relationships implicit in the physical mechanism (such as reaction rate and cumulative effect) with the observation data. Then, the sample mechanism fusion vector can be input into a preset model to output the predicted industrial parameters (such as pH value) of the sample at the predicted time point.

[0343] Furthermore, a target loss (including data loss and physical constraint loss) can be constructed based on the differences between the predicted results of the sample industrial parameters and the sample industrial parameters, the values ​​of the first mechanism parameter (calculated from reaction kinetic characteristics), and the values ​​of the second mechanism parameter (calculated from time cumulative effect characteristics). Based on the target loss of the current training epoch and the historical loss of the previous training epoch, the corresponding loss descent rate is determined. Then, the rank parameter is adjusted based on the loss descent rate to achieve dynamic low-rank compression of the weight matrices of the memory cell layer and the fully connected layer of the preset model, resulting in compressed memory cell layers and fully connected layers. Thus, in the current training epoch, the parameters of the preset model can be adjusted based on the target loss to obtain a preset model with updated parameters. This process continues until the preset convergence condition is met, at which point the target model is obtained.

[0344] In some implementations, after training, the target model undergoes 8-bit integer (INT8) quantization optimization and format conversion (such as ONNX format) before being deployed to industrial edge devices, such as Programmable Logic Controllers (PLCs) and Distributed Control Systems (DCS), to achieve low-latency, high-frequency real-time inference. For example, the actual generated mechanism fusion vector can be directly input into the target model to predict accurate industrial parameters at the current time point. This allows for real-time and accurate parameter estimation to address the problems of short lifespan, high maintenance costs, and measurement lag associated with physical sensors in harsh environments such as high temperatures and strong corrosion. It not only replaces high-cost sensors and reduces maintenance expenses but also improves the model's adaptability and robustness to complex operating conditions (such as long time delays and nonlinear dynamics) by fusing physical mechanisms with data-driven models, ensuring the stability and continuity of the production process. Ultimately, this achieves process optimization, quality control, and intelligent decision-making, driving manufacturing enterprises towards high efficiency, safety, and sustainability.

[0345] This invention achieves a breakthrough in the industrial applicability of the model by deeply integrating physical mechanisms with innovative lightweight technology. Specific performance indicators are reflected in lightweight performance and key parameter prediction performance, which will be described in detail below.

[0346] Please refer to Figure 5In some implementations, the target model achieves significant computational efficiency improvements by employing adaptive rank-optimized singular value decomposition and INT8 quantization optimization. Compared to the standard Long Short-Term Memory (LSTM) model, the target model (KD-LLSTM model) proposed in this invention achieves a substantial reduction in the number of parameters, expected to be reduced by 70%, while maintaining accuracy. The target model size is reduced to 30% of that of LSTM. The single inference time is significantly shortened when deployed on industrial edge devices, reduced to 50% of that of LSTM, thus meeting the stringent real-time control cycle requirements of industrial environments.

[0347] The number of parameters in a standard LSTM model is the baseline value (assumed to be 100%), while the target model in this application compresses the number of parameters to 30% of that of an LSTM (i.e., a reduction of 70%) by dynamically adjusting the rank of the weight matrix. This compression is based on a low-rank approximation reconstruction of the core layers of the model (memory unit layer and fully connected layer). During training, the rank value is automatically adjusted according to the rate of loss decrease, balancing model accuracy and compression rate, preserving the main information while significantly reducing redundant parameters.

[0348] Furthermore, after deploying the target model on industrial edge devices (such as PLCs or DCS), the model parameters are converted from 32-bit floating-point numbers (FP32) to 8-bit integers (INT8), further reducing computational complexity and memory consumption, and adapting to resource-constrained industrial edge devices. The single inference time of the target model is only 50% of that of the LSTM model. This application has a significant advantage in meeting industrial real-time control cycle requirements (such as millisecond-level response).

[0349] Please refer to Figure 6 and Figure 7 This application significantly improves the prediction accuracy and physical consistency of the target model by constructing mechanistic features and designing a knowledge-guided target loss. Taking pH prediction as an example, by constructing mechanistic features and designing a knowledge-guided target loss for training, the prediction accuracy and physical consistency of the target model in this application are substantially improved.

[0350] exist Figure 6 The image shows the prediction results of the target model under fluctuating operating conditions (such as the pH value over time). Figure 6 It can be seen that the predicted values ​​of the target model are closer to the actual values, and the fluctuation range is smaller. From Figure 7As can be seen, compared with BP neural networks or standard LSTM models that rely solely on data, this application demonstrates stronger stability and greater accuracy in capturing pH time-varying trends when facing fluctuations in operating conditions. Its root mean square error can be reduced by 20% to 30%, effectively avoiding prediction distortion caused by a lack of physical constraints. In scenarios involving changes in raw material batches and temperature fluctuations, the prediction stability of the target model in this application is significantly better than that of purely data-driven models, exhibiting good generalization performance.

[0351] Please see Figure 8 This application also provides an industrial parameter prediction device that can implement the above-described industrial parameter prediction method. The industrial parameter prediction device includes:

[0352] The acquisition module 81 is used to acquire basic reactor data corresponding to multiple time points, including the current time point and multiple historical time points preceding the current time point.

[0353] The calculation module 82 is used to perform material balance calculations based on the corresponding reactor basic data for each time point, obtain reaction kinetic characteristics, and construct a first feature sequence based on multiple reaction kinetic characteristics corresponding to multiple time points in sequence.

[0354] Module 83 is used to calculate the cumulative volume of solution for each time point based on the corresponding reactor basic data, obtain the time cumulative effect characteristics, and sequentially construct a second feature sequence based on multiple time cumulative effect characteristics corresponding to multiple time points.

[0355] The input module 84 is used to acquire raw sensor data and input the mechanism fusion vector obtained by fusing the raw sensor data, the first feature sequence, and the second feature sequence into the target model to obtain the industrial parameter prediction result corresponding to the current time point;

[0356] The target model is obtained by performing mechanism learning on the preset model. Mechanism learning is based on the difference between the predicted industrial parameters of the sample output at the prediction time point and the sample industrial parameters, the first mechanism parameter value, and the second mechanism parameter value, respectively. The preset model is subjected to difference minimization learning. The first mechanism parameter value is calculated based on the sample reaction kinetic characteristics corresponding to the prediction time point, and the second mechanism parameter value is calculated based on the sample time cumulative effect characteristics corresponding to the prediction time point.

[0357] The specific implementation of this industrial parameter prediction device is basically the same as the specific embodiment of the above-described industrial parameter prediction method, and will not be repeated here. Subject to meeting the requirements of the embodiments of this application, the industrial parameter prediction device may also be equipped with other functional modules to implement the industrial parameter prediction method in the above embodiments.

[0358] This application also provides a computer device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-described industrial parameter prediction method. This computer device can be any smart terminal, including tablet computers, in-vehicle computers, etc.

[0359] Please see Figure 9 , Figure 9 The hardware structure of a computer device according to another embodiment is illustrated. The computer device includes:

[0360] The processor 91 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this application.

[0361] The memory 92 can be implemented in the form of read-only memory (ROM), static storage device, dynamic storage device, or random access memory (RAM). The memory 92 can store the operating system and other application programs. When the technical solutions provided in the embodiments of this specification are implemented by software or firmware, the relevant program code is stored in the memory 92 and is called and executed by the processor 91 to execute the industrial parameter prediction method of the embodiments of this application.

[0362] Input / output interface 93 is used to implement information input and output;

[0363] Communication interface 94 is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, network cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0364] Bus 95 transmits information between various components of the device (e.g., processor 91, memory 92, input / output interface 93, and communication interface 94);

[0365] The processor 91, memory 92, input / output interface 93, and communication interface 94 are interconnected within the device via bus 95.

[0366] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described industrial parameter prediction method.

[0367] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0368] The embodiments described in this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided by the embodiments of this application. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by the embodiments of this application are also applicable to similar technical problems.

[0369] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of this application, and may include more or fewer steps than shown, or combine certain steps, or different steps.

[0370] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0371] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0372] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0373] It should be understood that in this application, "at least one" and "several" refer to one or more, and "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 three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0374] In the embodiments provided in this application, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for instance, the division of the units described above is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0375] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0376] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0377] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes multiple instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing programs, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0378] The preferred embodiments of the present application have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of the embodiments of the present application shall be within the scope of the claims of the present application.

Claims

1. An industrial parameter prediction method characterized by, The method comprises: obtaining reactor basic data corresponding to a plurality of time points, wherein the plurality of time points include a current time point and a plurality of historical time points before the current time point; for each time point, performing material balance calculation according to the corresponding reactor basic data to obtain reaction kinetics characteristics, and sequentially constructing a first feature sequence based on a plurality of reaction kinetics characteristics corresponding to the plurality of time points; for each time point, performing solution cumulative volume calculation according to the corresponding reactor basic data to obtain time cumulative effect characteristics, and sequentially constructing a second feature sequence based on a plurality of time cumulative effect characteristics corresponding to the plurality of time points; obtaining original sensor data, and inputting a mechanism fusion vector obtained by fusing the original sensor data, the first feature sequence and the second feature sequence into a target model to obtain an industrial parameter prediction result corresponding to the current time point; wherein the target model is obtained by mechanism learning on a preset model, the mechanism learning is difference minimization learning on the preset model based on differences between sample industrial parameter prediction results output by the preset model for a prediction time point and sample industrial parameters, first mechanism parameter values and second mechanism parameter values, the first mechanism parameter values are calculated according to sample reaction kinetics characteristics corresponding to the prediction time point, and the second mechanism parameter values are calculated according to sample time cumulative effect characteristics corresponding to the prediction time point; The target model is trained by: obtaining sample reactor basic data corresponding to a plurality of sample time points, wherein the plurality of sample time points include a prediction time point and a plurality of sample historical time points before the prediction time point; for each sample time point, performing material balance calculation according to the corresponding sample reactor basic data to obtain sample reaction kinetics characteristics, and sequentially constructing a sample first feature sequence based on a plurality of sample reaction kinetics characteristics corresponding to the plurality of sample time points; for each sample time point, performing solution cumulative volume calculation according to the corresponding sample reactor basic data to obtain sample time cumulative effect characteristics, and sequentially constructing a sample second feature sequence based on a plurality of sample time cumulative effect characteristics corresponding to the plurality of sample time points; obtaining sample original sensor data, and inputting a sample mechanism fusion vector obtained by fusing the sample original sensor data, the sample first feature sequence and the sample second feature sequence into a preset model to obtain a sample industrial parameter prediction result corresponding to the prediction time point; calculating a first mechanism parameter value according to the sample reaction kinetics characteristics corresponding to the prediction time point, and calculating a second mechanism parameter value according to the sample time cumulative effect characteristics corresponding to the prediction time point; obtaining a sample industrial parameter corresponding to the prediction time point, and constructing a first sub-loss according to the difference between the sample industrial parameter prediction result and the sample industrial parameter, a second sub-loss according to the difference between the sample industrial parameter prediction result and the first mechanism parameter value, and a third sub-loss according to the difference between the sample industrial parameter prediction result and the second mechanism parameter value, and constructing a target loss based on the first sub-loss, the second sub-loss and the third sub-loss; adjusting the parameters of the preset model based on the target loss to obtain a target model.

2. The industrial parameter prediction method of claim 1, wherein, The target loss is constructed based on the first sub-loss, the second sub-loss and the third sub-loss, including: determining a first average value based on the average value of a plurality of historical first sub-losses corresponding to the preset model, determining a second average value based on the average value of a plurality of historical second sub-losses, and determining a third average value based on the average value of a plurality of historical third sub-losses; determining a first fluctuation reference value based on the difference between the first sub-loss and the first average value, determining a second fluctuation reference value based on the difference between the second sub-loss and the second average value, and determining a third fluctuation reference value based on the difference between the third sub-loss and the third average value; determining a reference sum based on the sum of the first fluctuation reference value, the second fluctuation reference value and the third fluctuation reference value; determining a first weight based on the ratio of the first fluctuation reference value to the reference sum, determining a second weight based on the ratio of the second fluctuation reference value to the reference sum, and determining a third weight based on the ratio of the third fluctuation reference value to the reference sum; adjusting the first sub-loss based on the first weight to obtain a target first sub-loss, adjusting the second sub-loss based on the second weight to obtain a target second sub-loss, and adjusting the third sub-loss based on the third weight to obtain a target third sub-loss; constructing a target loss based on a sum of the target first sub-loss, the target second sub-loss, and the target third sub-loss.

3. The industrial parameter prediction method of claim 1, wherein, The adjusting the parameters of the preset model based on the target loss to obtain a target model comprises: obtaining a historical loss of a previous training round of the preset model, and calculating a difference between the historical loss and the target loss to obtain a loss drop rate value; obtaining a preset learning speed reference value, and adjusting weight matrices of a memory cell layer and a fully connected layer of the preset model based on a size relationship between the loss drop rate value and the learning speed reference value to obtain the memory cell layer and the fully connected layer after adjustment; adjusting the model parameters of the preset model to obtain an updated preset model; repeating the adjusting the parameters of the preset model based on the target loss until a preset training number is reached to obtain a target model.

4. The industrial parameter prediction method of claim 3, wherein, The weight matrices comprise a first weight matrix corresponding to the memory cell layer and a second weight matrix corresponding to the fully connected layer, and the adjusting the weight matrices of the memory cell layer and the fully connected layer of the preset model based on the size relationship between the loss drop rate value and the learning speed reference value to obtain the memory cell layer and the fully connected layer after adjustment comprises: determining an adjustment rank value corresponding to the target loss based on the size relationship between the loss drop rate value and the learning speed reference value, wherein when the loss drop rate value is greater than the learning speed reference value, the adjustment rank value is less than a historical adjustment rank value corresponding to the historical loss, or when the loss drop rate value is less than the learning speed reference value, the adjustment rank value is greater than the historical adjustment rank value; adjusting a first low-rank matrix corresponding to the first weight matrix based on the adjustment rank value, updating the memory cell layer according to the first low-rank matrix after adjustment to obtain an updated memory cell layer; adjusting a second low-rank matrix corresponding to the second weight matrix based on the adjustment rank value, updating the fully connected layer according to the second low-rank matrix after adjustment to obtain an updated fully connected layer.

5. The industrial parameter prediction method of claim 1, wherein, The reactor basic data comprise inlet flow data, outlet flow data, temperature data, input concentration data, reactor internal concentration data, and reactor volume data, and the reaction kinetics characteristics are obtained by material balance calculation based on the corresponding reactor basic data for each time point, comprising: determining a first parameter based on a ratio of the inlet flow data to the reactor volume data for each time point; obtaining a preset reaction rate constant, determining a second parameter based on a ratio of the outlet flow data to the reactor volume data, and determining a target second parameter based on a sum of the second parameter and the preset reaction rate constant; a first product based on a product of the first parameter and the input concentration data, and a first ratio based on a ratio of the first product and the target second parameter; an initial time point is obtained, and a third parameter is determined according to the initial time point, the each time point, and the target second parameter; a first material balance term is obtained based on a product of the first ratio and the third parameter; a second material balance term is obtained based on a product of the reactor internal concentration data and the third parameter; a reaction kinetic characteristic of the each time point is obtained based on a sum of the first material balance term and the second material balance term.

6. The industrial parameter prediction method of claim 1, wherein, The reactor basic data includes inlet flow data, outlet flow data, and reactor volume data, and the time cumulative effect characteristic is obtained by calculating the solution cumulative volume according to the corresponding reactor basic data for the each time point, including: For the each time point, the residence flow data is calculated according to the inlet flow data and outlet flow data, and the average residence time is obtained based on a ratio of the reactor volume data and the residence flow data; The residence time distribution description information of the material in the reactor is determined based on the average residence time and the each time point; The cumulative amount description information is obtained according to a product of the residence time distribution description information and the inlet flow data; The time cumulative effect characteristic is obtained by time integration calculation on the cumulative amount description information.

7. An industrial parameter prediction device characterized by comprising: The device includes: An acquisition module is configured to acquire reactor basic data corresponding to a plurality of time points, wherein the plurality of time points include a current time point and a plurality of historical time points before the current time point; A calculation module is configured to, for each time point, perform material balance calculation according to the corresponding reactor basic data to obtain a reaction kinetic characteristic, and sequentially construct a first feature sequence based on a plurality of reaction kinetic characteristics corresponding to the plurality of time points; A construction module is configured to, for the each time point, perform solution cumulative volume calculation according to the corresponding reactor basic data to obtain a time cumulative effect characteristic, and sequentially construct a second feature sequence based on a plurality of time cumulative effect characteristics corresponding to the plurality of time points; An input module is configured to acquire original sensor data, and input a mechanism fusion vector obtained by fusing the original sensor data, the first feature sequence, and the second feature sequence into a target model to obtain an industrial parameter prediction result corresponding to the current time point; The target model is obtained by mechanism learning on a preset model, and the mechanism learning is a difference minimization learning on the preset model based on differences between sample industrial parameter prediction results output by the preset model for a prediction time point and sample industrial parameters, first mechanism parameter values, and second mechanism parameter values, the first mechanism parameter values are calculated according to sample reaction kinetic characteristics corresponding to the prediction time point, and the second mechanism parameter values are calculated according to sample time cumulative effect characteristics corresponding to the prediction time point; The target model is trained by: obtaining sample reactor basic data corresponding to a plurality of sample time points, wherein the plurality of sample time points include a prediction time point and a plurality of sample historical time points before the prediction time point; for each sample time point, performing material balance calculation according to the corresponding sample reactor basic data to obtain sample reaction kinetics characteristics, and sequentially constructing a sample first feature sequence based on a plurality of sample reaction kinetics characteristics corresponding to the plurality of sample time points; for each sample time point, performing solution cumulative volume calculation according to the corresponding sample reactor basic data to obtain sample time cumulative effect characteristics, and sequentially constructing a sample second feature sequence based on a plurality of sample time cumulative effect characteristics corresponding to the plurality of sample time points; obtaining sample original sensor data, and inputting a sample mechanism fusion vector obtained by fusing the sample original sensor data, the sample first feature sequence and the sample second feature sequence into a preset model to obtain a sample industrial parameter prediction result corresponding to the prediction time point; calculating a first mechanism parameter value according to the sample reaction kinetics characteristics corresponding to the prediction time point, and calculating a second mechanism parameter value according to the sample time cumulative effect characteristics corresponding to the prediction time point; obtaining a sample industrial parameter corresponding to the prediction time point, and constructing a first sub-loss according to the difference between the sample industrial parameter prediction result and the sample industrial parameter, a second sub-loss according to the difference between the sample industrial parameter prediction result and the first mechanism parameter value, and a third sub-loss according to the difference between the sample industrial parameter prediction result and the second mechanism parameter value, and constructing a target loss based on the first sub-loss, the second sub-loss and the third sub-loss; adjusting parameters of the preset model based on the target loss to obtain a target model.

8. A computer device, comprising: The computer device comprises a memory and a processor, the memory stores a computer program, and the processor realizes the industrial parameter prediction method of any one of claims 1 to 6 when executing the computer program.

9. A computer readable storage medium, the storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to realize the industrial parameter prediction method of any one of claims 1 to 6.

Images