Intelligent and automated small-scale synthesis reaction control method

Abstract

This invention provides an intelligent and automated small-scale synthetic reaction control method, belonging to the field of synthetic reaction control technology. This invention suppresses pH oscillations caused by peristaltic pump pulse feeding through a pulse denoising median filter and a dead-zone nonlinear link. Multivariable soft-measurement values ​​replace direct measurement when the pH probe drifts due to scaling. A hardware-level emergency cut-off circuit, independent of the computer control program, intervenes first when the temperature slope inside the reactor exceeds the temperature rise rate threshold. After the reaction, an optimal stop-time drying endpoint determination algorithm dynamically determines the drying endpoint based on the drying device temperature slope sequence and heating power ratio sequence. A robotic arm and slide rail automatically transfer the first layer of filter plates in the filtration device. The entire process is uniformly scheduled by computer via RS485 communication, solving the problem that small-scale synthetic reactors cannot accurately identify reaction kinetic parameters online under sparse sensor measurement conditions.

Description

Technical Field

[0001] This invention belongs to the field of synthetic reaction control technology, and more specifically, relates to an intelligent and automated method for controlling small-scale synthetic reactions. Background Technology

[0002] In existing technologies, automated synthesis reactors are widely used in chemical, pharmaceutical, and hazardous chemical synthesis fields, achieving automatic regulation of the reaction process through the integration of temperature sensors, pH sensors, and programmable controllers. For small reactors, due to their low heat capacity, the temperature rises extremely rapidly during exothermic reactions. Traditional PID control, relying on fixed parameters, struggles to adapt to the dynamic changes in reaction kinetics. While model predictive control can anticipate temperature rise trends, the accuracy of its prediction equations is highly dependent on reaction kinetic parameters such as activation energy and frequency factor. These parameters drift with changes in material batches, concentrations, and temperature programs during actual production. Existing methods typically employ offline identification to pre-calibrate kinetic parameters or rely on purely data-driven black-box models for online estimation. The former fails to reflect batch-to-batch parameter drift, while the latter suffers from significantly reduced identification accuracy under sparse measurement conditions due to a lack of physical constraints. This leads to a persistent deviation between the model predictive control's prediction equations and the actual reaction behavior, ultimately causing control lag. In other words, existing technologies have a technical problem: small synthesis reactors cannot accurately identify reaction kinetic parameters online under sparse sensor measurement conditions, which leads to temperature control lag due to parameter mismatch in model predictive control and causes safety risks. Summary of the Invention

[0003] In view of this, the present invention provides an intelligent and automated small-scale synthesis reaction control method, which can solve the technical problem in the prior art that small-scale synthesis reactors cannot perform online high-precision identification of reaction kinetic parameters under sparse sensor measurement conditions, resulting in temperature control lag due to parameter mismatch in model predictive control and causing safety risks.

[0004] This invention is implemented as follows: This invention provides an intelligent and automated method for controlling small-scale synthetic reactions, comprising the following steps:

[0005] The raw materials are loaded into raw material tank 1 and raw material tank 2 respectively. The peristaltic pump 1 and peristaltic pump 2 are started by computer, the feed rate and feed flow rate are set, and the mechanical stirring motor is started simultaneously and the stirring speed and stirring duration are set.

[0006] Start the temperature control equipment, set the program heating and cooling rate and the duration of constant temperature. The PT100 temperature probe collects the temperature inside the vessel in real time and uploads it to the computer. The pH meter probe collects the pH value inside the vessel in real time and uploads it to the computer. All data is uploaded to the computer and recorded via RS485 communication line.

[0007] The computer inputs the internal temperature of the vessel collected by the PT100 temperature probe, the internal pH value collected by the pH meter probe, and the feed flow rates of peristaltic pump 1 and peristaltic pump 2 into the physical information dynamics identification model. The physical information dynamics identification model outputs the current activation energy estimate and frequency factor estimate. The computer updates the model prediction control parameters based on the activation energy estimate and frequency factor estimate, dynamically adjusts the rotation speed of peristaltic pump 1 and peristaltic pump 2 and the power of the temperature control equipment, and at the same time, the pulse noise denoising median filter filters the internal pH value collected by the pH meter probe, and the dead zone nonlinear element suppresses the limit cycle oscillation.

[0008] The computer continuously monitors the pH value and temperature slope inside the reactor. When the pH value deviation exceeds the pH deviation threshold, it switches to a multivariate soft measurement value based on conductivity, flow rate of the automatic discharge device flow meter, and feed flow rate, and triggers the automatic cleaning device valve to open for a short cleaning cycle of the pH meter probe. When the temperature slope inside the reactor exceeds the temperature rise rate threshold, the hardware-level emergency cut-off circuit intervenes independently of the computer control program and opens the cooling channel.

[0009] After the reaction is completed, the computer sequentially closes the valves of raw material tank 1 and raw material tank 2, and opens the valve of the automatic feeding device. The flow meter of the automatic feeding device monitors the discharge flow rate in real time and feeds back to control the opening of the valve of the automatic feeding device. The material flows into the filter device evenly. The liquid flows into the liquid recovery tank through the second layer side pipe of the filter device, and the solid is retained in the first layer filter plate of the filter device.

[0010] The computer opens the valve of the automatic cleaning device, and the cleaning liquid is used to clean the inner wall of the vessel, the top cover, and the stirring paddle 360° through the inlet pipe and the multi-directional rotating nozzle of the automatic cleaning device. The cleaning liquid flows out through the valve of the automatic discharge device and rinses the solids on the first layer of the filter plate of the filter device. The computer controls the slide rail servo motor to drive the robotic arm. The robotic arm moves the first layer of the filter plate and the solid material to the top of the drying device through the suction cup. The drying device determines the drying endpoint according to the optimal stop time drying endpoint judgment algorithm and automatically stops.

[0011] Specifically, the structure of the physical information dynamics identification model is as follows: it uses a multi-layer feedforward neural network as the backbone, the input layer receives the temperature inside the vessel, the pH value inside the vessel, the feed flow rate and time, the output layer outputs the estimated value of activation energy and the estimated value of frequency factor, and the loss function consists of two parts: the data fitting residual term and the physical equation residual term.

[0012] The physical equation residuals embed the Arrhenius rate equation and the material balance differential equation in the form of soft constraints, requiring the network output to satisfy the reaction rate equation. and material balance constraint equations .

[0013] Specifically, the training dataset for the physical information dynamics identification model is established by: collecting historical experimental data under different temperature programs and different feed flow rates, using the process quantity sequence at each sampling time as the input sample, and using the activation energy and frequency factor identified offline for the corresponding batch as labels. At the same time, virtual samples that satisfy the reaction rate equation and material balance constraint equation are generated through numerical simulation to expand the training set.

[0014] Specifically, the training of the physical information dynamics identification model involves: using the Adam optimizer to minimize the total loss function, which is the weighted sum of the mean square error of the data fitting residuals and the physical equation residuals. The weight coefficients of the physical equation residuals are adaptively adjusted with each training round. After training, the input sequence is updated in real time using a sliding window method, and the current activation energy estimate and frequency factor estimate are output.

[0015] The physical information dynamic identification model also includes a dynamic identification confidence evaluation function. This function calculates the confidence index value based on a weighted average of three normalized indices: the variance of the temperature slope within the vessel, the drift rate of the pH value within the vessel, and the fluctuation rate of the feed flow rate. ,when When using a high learning rate, When using a moderate learning rate, The parameters are frozen and updated, and the output is switched to multivariate soft measurement values.

[0016] Among them, the confidence index value The calculation formula is ,in , , The weighting coefficients and , It is a dimensionless quantity.

[0017] Specifically, the pulse denoising median filter takes the median value of the original sampling sequence of the pH meter probe within a fixed-length sliding window as the filtered output. The dead-zone nonlinear element is specifically introduced into the model predictive control output with a zero-centered insensitive zone. When the absolute value of the control deviation is less than the dead-zone width, the control output remains unchanged.

[0018] Specifically, the multivariate soft measurement value is a substitute output value for estimating the current pH value in the reactor based on the conductivity, the flow rate of the automatic discharge device flow meter, and the feed flow rate through a pre-established regression relationship.

[0019] Specifically, the hardware-level emergency shut-off circuit is a safety protection circuit that is independent of the computer control program, directly monitors the temperature inside the vessel by a hardware comparison circuit, and immediately shuts off the heating circuit and opens the cooling channel when the temperature rise rate threshold is exceeded. The unit of the temperature rise rate threshold is ℃ / s.

[0020] The optimal stop-drying endpoint decision algorithm models the drying process as a Markov chain with random perturbations, making a decision to continue heating or stop heating at each sampling time, and defining the cost function as the weighted expectation of over-drying loss and under-drying loss.

[0021] The optimal stop-drying endpoint determination algorithm is based on past... Temperature slope sequence of drying device at each sampling point Ratio sequence of heating power Calculate the expected cost gradient ,when Once the drying endpoint is determined, the computer sends a shutdown command to the drying unit.

[0022] Wherein, the expected cost gradient The calculation formula is , It is a dimensionless quantity.

[0023] Specifically, the step of the automatic feeding device flow meter monitoring the discharge flow rate in real time and feeding back to control the opening of the automatic feeding device valve involves the automatic feeding device flow meter uploading the real-time flow signal to the computer via an RS485 communication line, and the computer adjusting the opening of the automatic feeding device valve according to the flow deviation to maintain uniform material flow.

[0024] The reactor consists of a vessel body, a vessel cover, a mechanical stirring motor, a stirring paddle, a pH meter probe, a PT100 temperature probe, a multi-directional rotating nozzle for an automatic cleaning device, a reactor inlet, and a reactor outlet. The reactor control system consists of a mobile platform, raw material tank 1, raw material tank 2, peristaltic pump 1, peristaltic pump 2, the reactor, a temperature control device, valves for the automatic cleaning device, valves for the automatic discharging device, a flow meter for the automatic discharging device, a liquid recovery tank, a filter device, a robotic arm, a slide rail, a drying device, and a computer. The computer is connected to each component via an RS485 communication line.

[0025] This invention introduces a physical information kinetic identification model into the control method of small-scale synthetic reactions. The Arrhenius rate equation and material balance differential equation are embedded into a neural network loss function in the form of soft constraints. This allows the network to minimize the residuals of the reaction kinetic physical equations while fitting sparse measurement data. This enables high-precision online identification of activation energy and frequency factor estimates even under conditions of long sensor sampling intervals or high noise. The identification results update the prediction equation parameters of the model predictive control in real time, ensuring that the controller always predicts temperature rise trends and adjusts cooling power based on the actual kinetic characteristics of the current batch. This eliminates control lag caused by parameter mismatch and solves the technical problem that small-scale synthetic reactors cannot achieve high-precision online identification of reaction kinetic parameters under sparse sensor measurement conditions. In summary, this invention solves the technical problem mentioned in the background art where small-scale synthetic reactors cannot achieve high-precision online identification of reaction kinetic parameters under sparse sensor measurement conditions, leading to temperature control lag due to parameter mismatch in model predictive control and resulting in safety risks. Attached Figure Description

[0026] Figure 1 This is a flowchart of the method of the present invention.

[0027] Figure 2 This is a structural diagram of a small-scale automated synthesis reactor.

[0028] Figure 3 This is a schematic diagram of the composition of the reactor control system.

[0029] In the attached drawings, the reference numerals are explained as follows: 1. Reactor; 2. Temperature control equipment; 3. Peristaltic pump 1; 4. Peristaltic pump 2; 5. Raw material tank 1; 6. Raw material tank 2; 7. Automatic cleaning device valve; 8. Automatic discharging device valve; 9. Automatic discharging device flow meter; 10. Mechanical stirring motor; 11. Liquid recovery tank; 12. Robotic arm; 13. Slide rail; 14. Drying device; 15. Computer; 16. Mobile platform; 17. pH meter probe; 18. PT100 temperature probe; 19. Automatic cleaning device multi-rotary nozzle; 20. Stirring paddle; 21. Reactor inlet; 22. Reactor outlet; 23. Filtration device. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below.

[0031] like Figure 1 The diagram shown is a flowchart of an intelligent and automated small-scale synthetic reaction control method provided by the present invention. This method includes the following steps:

[0032] S01. Load the raw materials into raw material tank 1 and raw material tank 2 respectively. Start peristaltic pump 1 and peristaltic pump 2 through computer, set the feed amount and feed flow rate, and start the mechanical stirring motor simultaneously and set the stirring speed and stirring duration.

[0033] S02. Start the temperature control equipment, set the program heating and cooling rate and the constant temperature duration. The PT100 temperature probe collects the temperature inside the vessel in real time and uploads it to the computer. The pH meter probe collects the pH value inside the vessel in real time and uploads it to the computer. All data is uploaded to the computer and recorded via RS485 communication line.

[0034] S03. The computer inputs the internal temperature of the vessel collected by the PT100 temperature probe, the internal pH value collected by the pH meter probe, and the feed flow rate of peristaltic pump 1 and peristaltic pump 2 into the physical information dynamics identification model. The physical information dynamics identification model outputs the current activation energy estimate and frequency factor estimate. The computer updates the model prediction control parameters based on the activation energy estimate and frequency factor estimate, dynamically adjusts the rotation speed of peristaltic pump 1 and peristaltic pump 2 and the power of the temperature control equipment, and at the same time, the pulse noise reduction median filter filters the internal pH value collected by the pH meter probe, and the dead zone nonlinear element suppresses the limit cycle oscillation.

[0035] S04. The computer continuously monitors the pH value and temperature slope inside the reactor. When the pH value deviation exceeds the pH deviation threshold, it switches to a multivariate soft measurement value based on conductivity, flow rate of the automatic discharge device flow meter, and feed flow rate, and triggers the automatic cleaning device valve to open for a short cleaning cycle of the pH meter probe. When the temperature slope inside the reactor exceeds the temperature rise rate threshold, the hardware-level emergency cut-off circuit intervenes independently of the computer control program and opens the cooling channel.

[0036] S05. After the reaction is completed, the computer will close the valves of raw material tank 1 and raw material tank 2 in sequence, and open the valve of the automatic feeding device. The flow meter of the automatic feeding device will monitor the discharge flow rate in real time and control the opening of the valve of the automatic feeding device. The material will flow into the filter device evenly. The liquid will flow into the liquid recovery tank through the second layer side pipe of the filter device, and the solid will remain in the first layer filter plate of the filter device.

[0037] S06. The computer opens the valve of the automatic cleaning device. The cleaning liquid is discharged through the inlet pipe and the multi-directional rotating nozzle of the automatic cleaning device covers the inner wall of the vessel, the top cover, and the stirring paddle 360°. The cleaning liquid flows out through the valve of the automatic discharge device and rinses the solids on the first layer of the filter plate of the filter device. The computer controls the slide rail servo motor to drive the robotic arm. The robotic arm moves the first layer of the filter plate and the solid material to the top of the drying device through the suction cup. The drying device determines the drying endpoint according to the optimal stop time drying endpoint judgment algorithm and automatically stops.

[0038] The specific structure of the physical information dynamics identification model is as follows: It uses a multi-layer feedforward neural network as its backbone. The input layer receives the in-vessel temperature, pH value, feed flow rate, and time. The output layer outputs the activation energy estimate and frequency factor estimate. The loss function consists of two parts: a data fitting residual term and a physical equation residual term. The physical equation residual term embeds the Arrhenius rate equation and the material balance differential equation in a soft-constraint form, requiring the network output to satisfy the reaction rate equation. ;in For the reaction rate, For reference reaction rate, For frequency factor estimates, As a reference frequency factor, This is an estimated activation energy value. For reference activation energy, The temperature inside the vessel. The reference temperature is used; the material balance constraint equation is as follows: ;in For component concentration, For reference concentration, The steps for establishing the training dataset of the physical information dynamics identification model specifically include: collecting historical experimental data under multiple different temperature programs and different feed flow rates, recording the temperature change curves, pH change curves, and feed flow rate change curves within the reactor over time; using the process quantity sequence at each sampling moment as input samples, and the activation energy and frequency factor identified offline for the corresponding batch as labels, constructing a training set; and simultaneously expanding the training set by generating virtual samples that satisfy the reaction rate equation and material balance constraint equation through numerical simulation. The steps for training the physical information dynamics identification model specifically include: using the Adam optimizer to minimize the total loss function, which is the weighted sum of the mean square error of the data fitting residuals and the physical equation residuals; the weight coefficients of the physical equation residuals are adaptively adjusted with each training round; after training, deploying the physical information dynamics identification model on a computer; and updating the input sequence in real time using a sliding window during online inference and outputting the current activation energy estimate and frequency factor estimate to update the prediction equation parameters for model predictive control.

[0039] The physical information dynamics identification model embeds the reaction rate equation and material balance constraint equation into the loss function with soft constraints, enabling the network to simultaneously satisfy data fitting and physical mechanism consistency under sparse measurement conditions. This allows for high-precision identification of activation energy and frequency factor estimates even when sensor noise is high or sampling intervals are long. The identification results update the prediction equations of the model predictive control in real time, enabling the computer to predict the temperature rise trend inside the vessel in advance and adjust the cooling power ahead of time. This avoids control lag caused by the decrease in accuracy of the pure data-driven model during operating condition drift, thus improving the initiative and robustness of the temperature control inside the vessel. Especially in the case of small-volume vessels with low heat capacity and fast temperature rise rate, the physical constraints significantly suppress the uncertainty of parameter identification, keeping the prediction accuracy of the model predictive control at a high level.

[0040] The physical information dynamics identification model also includes a dynamics identification confidence evaluation function. This function is calculated by weighting three normalized indices within the current sampling window: the variance of the temperature slope within the vessel, the drift rate of the pH value within the vessel, and the fluctuation rate of the feed flow rate, to obtain the confidence index value. ; The calculation formula is expressed as follows: ;in The variance of the slope of the temperature inside the vessel. This represents the maximum permissible value for the variance of the temperature slope inside the vessel. This represents the absolute value of the pH drift rate within the vessel. This represents the maximum allowable pH drift rate within the reactor. For feed flow rate fluctuation rate, This represents the maximum allowable value for feed flow rate fluctuation. , , The weighting coefficients and , It is a dimensionless quantity; when At that time, a high learning rate is used to accelerate the online updating of the physical information dynamics identification model parameters; when At that time, the online parameters of the physical information dynamics identification model are updated normally using a moderate learning rate; when At that time, the online parameters of the frozen physical information dynamics identification model are updated and switched to multivariable soft measurement value output, pending... Updates will resume once the value drops to 0.6 or below.

[0041] The optimal stop-drying endpoint decision algorithm models the drying process of the drying device as a Markov chain with random perturbations. At each sampling time, it makes a decision to continue heating or stop heating. The cost function is defined as the weighted expectation of over-drying loss and under-drying loss, based on past... Temperature slope sequence of drying device at each sampling point Ratio sequence of heating power Calculate the expected cost gradient in the current state. ; The calculation formula is expressed as follows: ;in This is the maximum reference value for the temperature slope of the drying device. This is the maximum reference value for the heating power ratio. and These are the weighting coefficients. The estimated cost at the stopping time. As a cost, normalized reference value, It is a dimensionless quantity; when When the drying endpoint is determined, the computer sends a shutdown command to the drying unit.

[0042] The optimal drying endpoint determination algorithm models the drying process as a stochastic decision problem, replacing fixed time parameters with dynamically calculated marginal costs. This allows the drying device to determine whether to continue heating at each sampling moment based on the actual evolution trend of the ratio of the drying device's temperature slope to its heating power. Heating is immediately terminated when the marginal cost of continuing heating first exceeds the cost of stopping heating. This adaptively finds the optimal drying endpoint even when there are differences in moisture content, ambient temperature, and particle size among different batches of material. It avoids both excessively short fixed times leading to high product moisture content and excessively long fixed times leading to over-drying or energy waste, thus improving the drying device's adaptability to batch-to-batch differences and ensuring consistent product quality.

[0043] The pulse denoising median filter is a digital filtering device that takes the median value of the original sampling sequence of the pH meter probe within a fixed-length sliding window as the filtered output. It is used to suppress the periodic fluctuations in pH value inside the reactor caused by the periodic pulse feed flow rates of peristaltic pumps 1 and 2. The dead-zone nonlinear element is an insensitive zone centered at zero introduced at the model predictive control output. When the absolute value of the control deviation is less than the dead-zone width, the control output remains unchanged. This is used to suppress the control loop that eliminates limit cycle oscillations caused by small fluctuations in pH value inside the reactor by frequent reverse adjustments. The multivariable soft measurement value is an alternative output value that estimates the current pH value inside the reactor based on conductivity, flow rate of the automatic discharge device flow meter, and feed flow rate through a pre-established regression relationship when the pH value deviation inside the reactor exceeds the pH deviation threshold. The model predictive control is a predictive control strategy that predicts the trend of temperature change inside the reactor within multiple control steps forward based on the reaction rate equation and material balance constraint equation. Under the premise of satisfying the temperature constraint inside the reactor, it solves the optimal control sequence and executes the first step. The hardware-level emergency shut-off circuit is a safety protection circuit that is independent of the computer control program, directly monitored by the hardware comparison circuit, and immediately shuts off the heating circuit and opens the cooling channel when the temperature rise rate threshold is exceeded. The sliding window is a dynamic data sequence consisting of a fixed number of sampling points taken from the current sampling time towards the historical direction, used for state estimation in the online reasoning of the physical information dynamics identification model and the optimal stop-time drying endpoint determination algorithm. The temperature rise rate threshold is the upper limit of the temperature rise rate inside the vessel, pre-calculated based on the vessel's heat capacity and the upper limit of the reaction exothermic rate, and fixed in the hardware-level emergency shut-off circuit and the computer control program, in °C / s. The pH deviation threshold is the maximum allowable deviation between the measured pH value inside the vessel and the multivariable soft measurement value, pre-set and stored in the computer. The first layer filter plate of the filtration device is a sintered metal filter plate with uniform pores and high compressive strength, used to retain solid materials. The heating power ratio is the ratio of the current heating power of the drying device to the rated power of the drying device, and is a dimensionless quantity.

[0044] The intelligent and automated small-scale synthesis reactor and reactor control system consist of the following components.

[0045] The reactor comprises: a vessel body, a vessel lid, a mechanical stirring motor, a stirring paddle, a pH meter probe, a PT100 temperature probe, a multi-directional rotating nozzle for an automatic cleaning device, a reactor inlet, and a reactor outlet. The vessel lid is located on top of the vessel body, and the mechanical stirring motor is located on top of the vessel lid. The output shaft of the mechanical stirring motor is connected to one end of the stirring paddle, and the other end of the stirring paddle extends through the vessel lid into the interior of the vessel body. Two feed inlets are provided on one side of the top of the vessel lid, and the vessel lid also has openings for the insertion of the PT100 temperature probe, the pH meter probe, and the multi-directional rotating nozzle for the automatic cleaning device. A discharge outlet is provided at the bottom of the vessel body, and a reactor inlet and a reactor outlet are respectively provided on both sides of the vessel body. The mechanical stirring motor is connected to a computer via an RS485 communication line.

[0046] The reactor control system comprises: a mobile platform, raw material tank 1, raw material tank 2, peristaltic pump 1, peristaltic pump 2, reactor, temperature control equipment, automatic cleaning device valves, automatic discharging device valves, automatic discharging device flow meter, liquid recovery tank, filtration device, robotic arm, slide rail, drying device, and computer. The mobile platform is made of aluminum alloy and has casters at its bottom. Raw material tank 1, raw material tank 2, peristaltic pump 1, peristaltic pump 2, reactor, temperature control equipment, automatic cleaning device valves, automatic discharging device valves, automatic discharging device flow meter, liquid recovery tank, filtration device, robotic arm, slide rail, and drying device are all mounted on the surface of the mobile platform. Raw material tank 1 and raw material tank 2 are sealed tanks, each with a solenoid valve at the bottom. Peristaltic pump 1 and peristaltic pump 2 are connected to the computer via RS485 communication lines. One end of the feed pipe 1 connects to the pagoda head of raw material tank 1, and the other end, after passing through peristaltic pump 1, is inserted into the reactor body through the feed inlet on the reactor lid. Internally; one end of the feed pipe 2 is connected to the pagoda head of the raw material tank 2, and the other end is inserted into the reactor body through the feed inlet of the reactor lid after passing through the peristaltic pump 2; the temperature control device is connected to the reactor inlet and outlet through two flexible hoses respectively, using liquid as the transmission medium to provide the energy required for heating or cooling the reactor, and the temperature control device is connected to the computer through an RS485 communication line; the flow meter of the automatic discharge device is installed at the reactor outlet, and the valve of the automatic discharge device is installed below the flow meter of the automatic discharge device, and both the flow meter and the valve of the automatic discharge device are connected to the computer through an RS485 communication line; the filter device is set directly below the reactor outlet, the first layer of the filter device is a sintered metal filter plate, the second layer of the filter device is a metal plate, and the side opening of the second layer of the filter device is connected to a pipe to export the filtered liquid to the liquid recovery tank, which is set below the filter device; the drying device is adjacent to the filter device, the material is a cast aluminum heating plate, and multiple parts are embedded on the plate surface. The drying device includes a platinum resistance thermometer and a high-precision PID temperature controller connected to a computer via an RS485 communication line. The slide rail is driven by a servo motor and also connected to the computer via an RS485 communication line. One end of the robotic arm is mounted on the slide rail, and the other end has a suction cup for connecting to the first layer of the filter plate in the filtration device, used to move the first layer of the filter plate and solid materials above the drying device. The valve of the automatic cleaning device is installed on the water inlet pipe and connected to the computer via an RS485 communication line. One end of the water inlet pipe is connected to a cleaning liquid switch, and the other end enters the reactor body through the reactor lid insertion hole and connects to the multi-directional rotating nozzle of the automatic cleaning device. The computer is connected via RS485 communication lines to peristaltic pump 1, peristaltic pump 2, mechanical stirring motor, temperature control equipment, automatic cleaning device valve, automatic discharging device valve, automatic discharging device flow meter, high-precision PID temperature controller, and slide rail servo motor, respectively, to achieve real-time monitoring and control of the entire reaction process through integrated software.

[0047] The specific implementation of step S01 is as follows: First, the technicians inject the liquid raw materials into raw material tank 1 and raw material tank 2 respectively and confirm that they are sealed. Through the integrated software interface of the computer, the feed rate and feed flow rate of peristaltic pump 1 and peristaltic pump 2 are set respectively. The computer sends a start command to peristaltic pump 1 and peristaltic pump 2 through the RS485 communication line. Peristaltic pump 1 transports the raw materials in raw material tank 1 to the inside of the reactor body through feed pipe 1. Peristaltic pump 2 transports the raw materials in raw material tank 2 to the inside of the reactor body through feed pipe 2. At the same time, the computer sends a start command to the mechanical stirring motor and sets the stirring speed and stirring duration. The mechanical stirring motor drives the stirring paddle to mix evenly inside the reactor body, ensuring that the two raw materials are mixed evenly immediately after entering the reactor body, providing a stable initial process state for subsequent reaction kinetic identification.

[0048] The specific implementation of step S02 is as follows: The computer sends program heating and cooling commands to the temperature control device via an RS485 communication line, sets multiple heating rates, constant temperature durations, and cooling rates. The temperature control device provides or removes heat to the reactor body through a circulation loop formed by the liquid medium through the reactor inlet and outlet. The PT100 temperature probe is inserted into the reactor body through the opening in the reactor lid, and collects the reactor temperature in real time at a sampling frequency of not less than 1 time / s and uploads it to the computer via the RS485 communication line. The pH meter probe simultaneously collects the pH value inside the reactor and uploads it to the computer. The computer timestamps and stores all uploaded data to form a complete process time series, which serves as the input data source for the physical information dynamics identification model in subsequent steps.

[0049] The specific implementation of step S03 is as follows: The computer uses a sliding window to extract a fixed-length sequence of in-vessel temperature, pH value, and feed flow rate from the current moment to the historical direction. This sequence is input into the physical information dynamics identification model deployed on the computer. The model outputs the current activation energy estimate and frequency factor estimate through a multi-layer feedforward neural network. The computer substitutes these estimates into the prediction equation of the model's predictive control to solve for the optimal control sequence within the current control cycle. The first-step control quantity is then used to send speed adjustment commands to peristaltic pump 1 and peristaltic pump 2, and a power adjustment command to the temperature control device. Simultaneously, the original sampling sequence from the pH meter probe is processed by a pulse noise reduction median filter. It is recommended that the filter window length be 5–15 sampling points to effectively suppress periodic pH fluctuations caused by the peristaltic pump pulse feed. The dead zone width is recommended to be set to 5%–10% of the pH value deviation to ensure that small fluctuations do not trigger frequent reverse adjustments, thereby eliminating limit cycle oscillations. The kinetic identification confidence evaluation function calculates the confidence index value in each control cycle. ,when High learning rates are used to accelerate online updates. Normal updates are performed using a moderate learning rate. The online parameter update is frozen, and the system switches to multivariate soft measurement mode. Updates will resume once the value drops to 0.6 or below to ensure the stability of the identification process.

[0050] The specific implementation of step S04 is as follows: In each sampling cycle, the computer compares the current pH value inside the vessel with the multivariate soft measurement value from the previous moment. The recommended pH deviation threshold is 0.3–0.5 pH units. When the deviation exceeds the pH deviation threshold, the computer determines that the pH meter probe may be experiencing scaling and drift. It then switches the control loop to a multivariate soft measurement value established based on conductivity, the flow rate of the automatic discharge device flow meter, and the feed velocity. An opening command is sent to the automatic cleaning device valve, allowing the cleaning fluid to briefly flush the pH meter probe. The recommended cleaning cycle is 30–60 seconds. Simultaneously, the computer estimates the temperature slope inside the vessel using the first-order difference of the temperature sequence. The recommended temperature rise rate threshold is 3–5 °C / s. When the temperature slope inside the vessel exceeds the temperature rise rate threshold, the hardware-level emergency cut-off circuit is directly triggered by the hardware comparison circuit, cutting off the heating circuit and opening the cooling channel. This action is independent of the computer control program, with a response delay less than the software control loop, thus achieving the fastest safe intervention in the early stages of a rapid temperature rise inside the vessel.

[0051] The specific implementation of step S05 is as follows: The computer detects the end of the reaction based on the stable pH value in the reactor and the zero feed flow rate. Then, it sends a closing command to the valves of raw material tank 1 and raw material tank 2 in sequence, and sends an opening command to the valve of the automatic discharge device. The flow meter of the automatic discharge device transmits the real-time flow signal to the computer via the RS485 communication line. The computer controls the opening of the valve of the automatic discharge device according to the flow deviation through proportional adjustment, so that the material flows into the filter device located directly below the outlet of the reactor at a uniform flow rate. The first layer of the sintered metal filter plate of the filter device intercepts the solid material, and the liquid flows into the liquid recovery tank by gravity through the side pipe of the second layer of the filter device. After the reading of the flow meter of the automatic discharge device returns to zero, the mechanical stirring motor is turned off, and the solid-liquid separation is completed.

[0052] The specific implementation of step S06 is as follows: The computer sends an opening command to the valve of the automatic cleaning device. The cleaning liquid is ejected at high pressure from the nozzle of the multi-directional rotating nozzle of the automatic cleaning device through the water inlet pipe. The reaction force drives the multi-directional rotating nozzle to rotate, achieving 360° coverage and rinsing of the inner wall of the vessel, the top cover, and the stirring paddle. The rinsing waste liquid is discharged through the valve of the automatic discharge device and simultaneously washes the solids on the first layer of the filter plate of the filter device. After cleaning, the valve of the automatic cleaning device is closed. Subsequently, the computer sends a movement command to the slide rail servo motor. The slide rail drives the robotic arm to move horizontally to directly above the filter device. The suction cup at the end of the robotic arm attracts the first layer of the filter plate of the filter device. The servo motor drives the slide rail in the opposite direction to move the first layer of the filter device and the solid material as a whole horizontally to the top of the drying device and place it. The computer sends a start command to the high-precision PID temperature controller of the drying device. Multiple embedded parts on the surface of the drying device... Platinum resistance thermometers collect the temperature of the drying unit in real time; a high-precision PID temperature controller dynamically adjusts the heating power; and an optimal stop-time drying endpoint determination algorithm reads the temperature slope sequence of the drying unit in each sampling period. Ratio sequence of heating power Calculate the expected cost gradient ,when The computer sends a shutdown command to the drying unit, and the drying of solid materials is completed. After the entire process is finished, the computer shuts down all equipment through integrated software.

[0053] The first key technical approach is online dynamic parameter identification under physical information constraints. Traditional pure data-driven models, when measurement data is sparse, suffer from a lack of physical priors, leading to identification results deviating from the physically reasonable range. This invention embeds the Arrhenius rate equation and material balance differential equation into the loss function with soft constraints, confining the parameter space to a subspace that satisfies physical laws. This maintains identification accuracy even under sparse observation conditions, providing continuously updated reliable prediction equation parameters for model predictive control. The second key technical approach is a confidence-adaptive learning rate adjustment mechanism. Existing online learning methods typically use a fixed learning rate, which can easily cause parameter oscillations when process variables fluctuate wildly, and suffer from slow updates when process variables are stable. This invention comprehensively evaluates the variance of the temperature slope, the drift rate of pH value in the reactor, and the fluctuation rate of feed flow rate through a dynamic identification confidence evaluation function, based on the confidence index value. The interval dynamic switching learning rate matches the parameter update rate with the current measurement confidence, thereby improving response flexibility while ensuring identification stability. The third key technical approach is adaptive drying endpoint determination based on optimal stopping time theory. Traditional drying relies on fixed time parameters, which cannot adapt to differences in material moisture content and particle size between batches. This invention models the drying process as a Markov decision process, replacing fixed time with dynamically calculated marginal costs to achieve adaptive drying endpoint determination. The synergistic effect of these three technical approaches is as follows: kinetic identification ensures the accuracy of temperature control during the reaction stage; the confidence-adaptive mechanism ensures the robustness of the identification process under abnormal conditions; and the optimal stopping time algorithm extends this adaptive decision-making idea to the post-processing stage. Together, they construct a full-process adaptive closed-loop control system from feeding and reaction to drying, enabling the entire solution to maintain stable process control quality in the face of batch differences and sensor anomalies.

[0054] It should be noted that this invention also solves the following technical problems: First, existing automated reactors commonly suffer from pH sensor reading oscillations during peristaltic pump feeding. This invention uses a pulse denoising median filter to perform sliding window median filtering on the original sampling sequence of the pH meter probe, eliminating the interference of local concentration fluctuations caused by the periodic pulse flow of the peristaltic pump on the pH reading from the signal processing level. Furthermore, a dead-zone nonlinear element is introduced at the control law output to block the path of frequent reverse adjustments triggered by small pH fluctuations, eliminating the limiting loop oscillation at its source and ensuring the stability of the feed rate regulation. Second, existing reactors face the problem of scaling and drifting of pH probes in high-concentration suspended solids reaction systems. Simply relying on probe cleaning cannot guarantee measurement continuity. This invention constructs a multivariate soft measurement model based on conductivity, the flow rate of the automatic discharge device flow meter, and the feed velocity. When the pH value deviation in the reactor exceeds the pH deviation threshold, it automatically switches to the soft measurement value and simultaneously triggers the pH meter probe cleaning cycle, achieving seamless switching of the control loop under abnormal measurement conditions and ensuring the continuity and reliability of pH feedback control during the reaction process. Third, existing reactors lack an integrated automatic control process from reaction to solid-liquid separation to drying. This invention, through the integrated design of a filtration device, liquid recovery tank, drying device and robotic arm slide rail transmission device, combined with the optimal stop-time drying endpoint determination algorithm, realizes fully unmanned operation from material discharge, solid-liquid separation, solid washing to drying endpoint determination, filling the gap in automation of the post-processing of small synthesis reactors.

[0055] Specifically, the principle of this invention is as follows: The technical solution of this invention can solve the aforementioned core technical problems because the physical information dynamics identification model embeds the reaction mechanism equation as a constraint into the training process of the neural network. This restricts the parameter space of the network to a subspace that satisfies physical laws. Even when the measurement data is sparse or noisy, the network cannot converge to a solution that violates the physical mechanism, thus ensuring that the identification results of the activation energy estimate and the frequency factor estimate are always within a physically reasonable range. Compared with a purely data-driven model, the introduction of physical constraints is equivalent to providing the network with additional prior information, significantly reducing the number of samples required for effective training, and suppressing the impact of noise on identification accuracy. The identification results are updated in real time to the prediction equation of the model predictive control through a sliding window method, enabling the controller to make multi-step advance predictions based on the real dynamic parameters of the current batch in each control cycle. This allows for early intervention in cooling regulation before the rate of temperature rise in the reactor accelerates, logically forming a complete closed loop from parameter identification to predictive control and then to actuator regulation, ensuring the timeliness and accuracy of control.

[0056] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0057] The specific implementation of step S01 is as follows: the operator loads the raw materials into raw material tank 1 and raw material tank 2 respectively, sends a start command to peristaltic pump 1 and peristaltic pump 2 through the computer, sets the feed rate and feed flow rate, starts the mechanical stirring motor synchronously and sets the stirring speed and stirring duration, and the peristaltic pump 1 and peristaltic pump 2 transport the raw materials into the vessel through the feed pipe, and the mechanical stirring motor drives the stirring paddle to mix the materials in the vessel.

[0058] The specific implementation of step S02 is as follows: the computer sends a programmed heating and cooling command to the temperature control device, sets the heating and cooling rate and the duration of constant temperature, and the temperature control device circulates the liquid heat transfer medium through the inlet and outlet of the reactor to adjust the temperature inside the reactor in real time. A PT100 temperature probe collects the temperature inside the reactor, and a pH meter probe collects the pH value inside the reactor. Both signals are uploaded to the computer via an RS485 communication line and recorded for archiving.

[0059] The specific implementation of step S03 is that the computer collects the internal temperature of the vessel from the PT100 temperature probe. pH value inside the vessel collected by pH meter probe Peristaltic pump 1 feed flow rate With peristaltic pump 2 feed flow rate As input, the data is fed into the physical information dynamics identification model, and the model outputs an estimated activation energy value. With frequency factor estimates The physical information dynamics identification model requires the network output to satisfy the reaction rate equation, which is expressed as follows:

[0060] ;

[0061] In the formula, The current reaction rate, in units of The output is derived from the online inference of the physical information dynamics identification model; For reference reaction rate, the unit is... The calibration experiment was conducted at the reference temperature. Compared with reference concentration Determined under the conditions; This is the estimated value of the frequency factor, in units of Output from the physical information dynamics identification model; Reference frequency factor, unit: It was obtained through offline identification of the calibration experiment; This is an estimated activation energy, in units of... Output from the physical information dynamics identification model; For reference activation energy, the unit is 1. It was obtained through offline identification in the calibration experiment; The temperature inside the vessel is measured in Kelvin (K) and is collected in real time by a PT100 temperature probe. The reference temperature, in Kelvin, is set by the calibration experimental conditions. The material balance constraint equation is expressed as follows:

[0062] ;

[0063] In the formula, The concentration of components inside the vessel, in units of The physical information dynamics identification model is used for online estimation. For reference concentration, the unit is... The conditions are set by the calibration experiment. For time, the unit is The computer bases its output on... and The predictive equation parameters of the model predictive control are updated in real time, dynamically adjusting the speeds of peristaltic pumps 1 and 2 and the power of the temperature control equipment. A pulse denoising median filter takes the median value of the original sampling sequence from the pH meter probe within a fixed-length sliding window as the filtered output, used to suppress periodic fluctuations in pH value within the reactor caused by the periodic pulse feeding of the peristaltic pumps. A dead-zone nonlinearity introduces a zero-centered insensitivity zone at the model predictive control output; when the absolute value of the control deviation is less than the dead-zone width, the control output remains unchanged, thus suppressing limiting loop oscillations. When the pH value deviation within the reactor exceeds the pH deviation threshold, the system switches to multivariable soft-sensing. As a pH feedback quantity From conductivity Automatic feeding device flow meter flow With feed flow rate , The estimation is based on a pre-established regression relationship, and the formula is expressed as follows:

[0064] ;

[0065] In the formula, The pH value inside the vessel is a multivariable soft measurement estimate, dimensionless; This is a measured value of the conductivity inside the vessel, in units of... The conductivity data is collected in real time by a conductivity sensor. Reference conductivity, in units of Obtained from calibration experiments; The measured flow rate of the automatic feeding device is given by the flow meter. The unit is [missing information]. The data is obtained from the RS485 communication feedback of the flow meter; For reference flow rate, the unit is... Obtained from calibration experiments; The reference feed velocity for peristaltic pump 1 is given in units of... The conditions are set by the calibration experiment. The reference feed flow rate for peristaltic pump 2 is given in units of... The conditions are set by the calibration experiment. The regression intercept is dimensionless and is obtained by fitting historical experimental data using the least squares method. , , , The regression coefficients for each input item are dimensionless and obtained from historical experimental data through least squares fitting. The regression residual term is dimensionless and reflects the model fitting error. The kinetic identification confidence assessment function calculates the confidence index value based on a weighted average of three normalized indices: the variance of the temperature slope within the vessel, the drift rate of the pH value within the vessel, and the fluctuation rate of the feed flow rate within the current sampling window. The formula is expressed as follows:

[0066] ;

[0067] In the formula, This represents the variance of the temperature slope within the vessel during the current sampling window, in units of... It is calculated from the temperature slope sequence at each sampling time within the current sliding window; This represents the maximum permissible variance of the temperature slope inside the vessel, in units of... The heat capacity of the vessel body and the rated heating power are pre-calibrated; This is the first derivative of the pH value in the reactor with respect to time, i.e., the pH drift rate, in units of... It is calculated by the ratio of the pH difference between adjacent sampling points to the sampling interval; for The absolute value, in units of ; This represents the maximum permissible absolute value of the pH drift rate, in units of... The settings are based on the sensor calibration results; The feed flow rate fluctuation rate is dimensionless, and its calculation formula is as follows:

[0068] ;

[0069] In the formula, This is the length of the feed flow rate statistics window, which defaults to 10. For the first Feed flow rate at each sampling time, in units of The data is obtained from the RS485 communication feedback of the peristaltic pump; This represents the average feed rate within the current window, in units of... ,Depend on Calculated; This is the maximum permissible value for feed flow rate fluctuation, dimensionless, and set by calibration experiments; , , Let be the weight coefficients of each term in the confidence evaluation function, satisfying Experience value , , ; This is a confidence level index, dimensionless. When... At that time, a high learning rate is used to accelerate the online updating of the physical information dynamics identification model parameters; when When, the online parameters of the model are updated normally using a moderate learning rate; when At that time, the online parameters of the frozen model were updated and switched to multivariate soft measurement values. Output, pending Updates will resume once the value drops to 0.6 or below.

[0070] The specific implementation of step S04 is that the computer continuously monitors the pH value and temperature slope inside the reactor. When the measured pH value inside the reactor... With multivariate soft measurement values When the deviation between them exceeds the pH deviation threshold, switch to As a pH feedback quantity, it triggers the automatic cleaning device valve to open, performing a short cleaning cycle on the pH meter probe. When the temperature slope inside the vessel exceeds the temperature rise rate threshold, the hardware-level emergency cut-off circuit intervenes independently of the computer control program, immediately cutting off the heating circuit and opening the cooling channel. The temperature rise rate threshold is in °C / s and is pre-calculated based on the vessel's heat capacity and the upper limit of the reaction exothermic rate, and is embedded in the hardware comparison circuit.

[0071] The specific implementation of step S05 is as follows: after the reaction is completed, the computer sequentially closes the solenoid valves of raw material tank 1 and raw material tank 2, opens the valve of the automatic feeding device, the flow meter of the automatic feeding device monitors the discharge flow rate in real time and feeds back to control the valve opening, the material flows into the filter device evenly, the liquid flows into the liquid recovery tank through the side pipe of the second layer of the filter device, and the solid is retained in the sintered metal filter plate of the first layer of the filter device.

[0072] The specific implementation of step S06 is as follows: The computer opens the valve of the automatic cleaning device, and the cleaning liquid, through the inlet pipe and multi-directional rotating nozzles, achieves 360° coverage cleaning of the inner wall of the vessel, the top cover, and the stirring paddle. The cleaning liquid flows out through the valve of the automatic discharge device and rinses the solids on the first layer filter plate of the filtration device. The computer controls the slide rail servo motor to drive the robotic arm, which uses a suction cup to move the first layer filter plate and solid material of the filtration device to the top of the drying device. The drying device determines the drying endpoint based on the optimal stop-time drying endpoint determination algorithm and automatically stops. The optimal stop-time drying endpoint determination algorithm is based on past... Temperature slope sequence of drying device at each sampling point Ratio sequence of heating power Calculate the expected cost gradient in the current state. The formula is expressed as follows:

[0073] ;

[0074] In the formula, For the first The temperature slope of the drying device at each sampling time, in °C / s, is calculated by the ratio of the temperature difference between adjacent sampling points to the sampling interval. This is the maximum reference value for the temperature slope of the drying device, in °C / s, obtained from the no-load calibration experiment. For the first The heating power ratio at each sampling moment, i.e. the ratio of the current heating power of the drying device to the rated power, is dimensionless and is obtained by RS485 communication feedback from the high-precision PID temperature controller of the drying device. For the first The ratio of heating power at each sampling time, dimensionless, with the meaning of... Same, is a sequence The Middle The value at time; This is the maximum reference value for the heating power ratio, dimensionless, and the default value is 1. and The weighting coefficients for the temperature slope term and the power ratio term in the optimal stop-drying endpoint determination algorithm satisfy the following conditions: Experience value , ; The sliding window length is the optimal stop-drying endpoint determination algorithm, which defaults to 10. Index of the current sampling time; The cost estimate at the stopping moment is dimensionless and obtained through offline regression of historical batch drying data, specifically based on the actual optimal stopping moment in the historical batch. sampling points mean as The fitting target is obtained through least squares regression. and The ratio is fixed in the computer program; The reference value is normalized at the cost of being dimensionless. Let be the gradient of the expected cost, which is dimensionless. When... When the drying endpoint is determined, the computer sends a shutdown command to the drying device. This allows the system to adaptively find the optimal drying endpoint, ensuring consistent product quality, even when there are differences in moisture content, ambient temperature, and particle size among different batches of material.

[0075] To better understand and implement this invention, the following is a specific application scenario of the invention, Example 2: To verify the effect of the invention, technicians set up a test environment and conducted a synthesis reaction experiment with sodium silicate aqueous solution and hydrochloric acid in a small-scale automated synthesis reaction control system to verify the entire process of the control method described in this invention.

[0076] The control system of the reactor used in the test consists of a mobile platform 16, a first raw material tank 5, a second raw material tank 6, a peristaltic pump 13, a peristaltic pump 24, a reactor 1, a temperature control device 2, an automatic cleaning device valve 7, an automatic discharging device valve 8, an automatic discharging device flow meter 9, a liquid recovery tank 11, a filter device 23, a robotic arm 12, a slide rail 13, a drying device 14, and a computer 15. All components are fixed on the mobile platform 16, and each actuator is connected to the computer 15 via an RS485 communication line. The reactor 1 consists of a vessel body, a vessel cover, a mechanical stirring motor 10, a stirring paddle 20, a pH meter probe 17, a PT100 temperature probe 18, a multi-directional rotating nozzle of the automatic cleaning device 19, a reactor inlet 21, and a reactor outlet 22. The overall structure is as follows. Figure 3 As shown, Figure 2 and Figure 3 As shown, the reactor 1 is located in the middle of the mobile platform 16. The temperature control device 2 is connected to the reactor inlet 21 and the reactor outlet 22 through two hoses respectively. The filter device 23 is located directly below the discharge port of the reactor 1. The drying device 14 is adjacent to the filter device 23. The slide rail 13 spans above the filter device 23 and the drying device 14. The robotic arm 12 is mounted on the slide rail 13. The computer 15 is placed on one side of the mobile platform 16.

[0077] Technicians injected 2000 mL of a 1.5 mol / L sodium silicate aqueous solution into the first raw material tank 5, and 2000 mL of a 1 mol / L hydrochloric acid solution into the second raw material tank 6. The feed flow rate of the peristaltic pump 13 was set to 20 mL / min, the feed flow rate of the peristaltic pump 24 was set to 15 mL / min, the speed of the mechanical stirring motor 10 was set to 200 r / min, and the stirring duration was set to 120 min using the integrated software on the computer 15.

[0078] Temperature control device 2 receives program heating instructions from computer 15, setting the first stage to raise the temperature from room temperature to 60℃ at a rate of 2℃ / min, and the second stage to maintain a constant temperature of 60℃ for 90min. PT100 temperature probe 18 collects the temperature inside the vessel in real time at a frequency of 1 time / s and uploads it to computer 15. pH meter probe 17 collects the pH value inside the vessel synchronously and uploads it to computer 15. As shown in Table 1, the typical process quantities of temperature and pH value inside the vessel in the first 30 minutes are recorded as follows.

[0079] Table 1. Process Quantity Recording Table for the First 30 Minutes

[0080]

[0081] As shown in Table 1, the physical information kinetic identification model completed its first online parameter update at the 10th minute of the reaction, outputting an estimated activation energy of 48.3 kJ / mol and an estimated frequency factor of [missing value]. Computer 15 updates the prediction equation of the model predictive control based on the above estimates. When the temperature slope inside the vessel reaches a maximum of about 2.3℃ / s, it starts the cooling power adjustment 3 control steps in advance to ensure that the temperature inside the vessel rises steadily to 60℃ and is maintained at a constant temperature.

[0082] At 15 minutes, a large amount of suspended solid particles began to form. pH meter probe 17 detected an increase in the absolute value of the pH drift rate within the vessel, with a confidence index value... The pH value rises to 0.67, exceeding the threshold of 0.6. Computer 15 automatically freezes the online parameter update of the physical information dynamics identification model and switches to a multivariate soft measurement value based on conductivity, flow rate of the automatic discharge device flowmeter 9, and feed velocity. Simultaneously, it triggers the automatic cleaning device valve 7 to open, continuously cleaning the pH meter probe 17 for 45 seconds. After cleaning is complete... The feed rate dropped back to 0.41 and resumed normal update mode. The feed rate was dynamically adjusted based on the pH feedback adjustment results of the multivariable soft measurement values, as shown in Table 1 from the 15th to the 30th minute.

[0083] At the 23-minute mark, the temperature slope inside the vessel reached 3.1℃ / s due to concentrated local heat release, exceeding the preset temperature rise rate threshold of 3℃ / s. The hardware-level emergency cut-off circuit immediately cut off the heating circuit and opened the cooling channel, with a response time of approximately 50ms. Independent of the computer control program, it intervened in advance, and the temperature inside the vessel dropped back to 59.2℃ within approximately 20 seconds before re-entering the constant temperature control range. No manual intervention was required throughout the entire process.

[0084] After the reaction proceeded for 120 minutes, computer 15 determined that the pH value inside the reactor had stabilized at around 6.8 and the feed flow rate had returned to zero. It then sequentially closed valves 5 of the first raw material tank and 6 of the second raw material tank, and opened valve 8 of the automatic discharge device. The flow meter 9 of the automatic discharge device monitored the discharge flow rate in real time and provided feedback to control the opening of valve 8. The material flowed into the filter device 23 at a uniform flow rate of approximately 30 mL / min. The solid silica gel remained on the first sintered metal filter plate of the filter device 23, while the liquid... The solution flows into the liquid recovery tank 11 through the second side pipe of the filter device 23. The discharge process takes about 65 minutes and recovers about 2350 mL of liquid.

[0085] After the material is discharged, the computer 15 opens the valve 7 of the automatic cleaning device. The cleaning liquid, under high pressure, is used to thoroughly rinse the inner wall of the vessel, the top cover, and the stirring paddle 20 with a 360° spray from the multi-directional rotating nozzle 19 of the automatic cleaning device. The rinsing waste liquid is discharged through the valve 8 of the automatic discharge device and simultaneously washes the solids on the first filter plate of the filter device 23. The cleaning process takes about 8 minutes. After the cleaning is completed, the servo motor of the slide rail 13 drives the robotic arm 12 to move horizontally above the filter device 23. The suction cup at the end of the robotic arm 12 engages with the first filter plate of the filter device 23. The servo motor drives the slide rail 13 in the opposite direction to move the first filter plate and the solid material of the filter device 23 horizontally above the drying device 14 and place them there.

[0086] Computer 15 sends a start command to the high-precision PID temperature controller of drying device 14, setting the initial drying temperature to 80℃. Platinum resistance thermometers collect the temperature of the drying device in real time, and the optimal stop-drying endpoint determination algorithm continuously reads the temperature slope sequence of the drying device. Ratio sequence of heating power As shown in Table 2, the typical drying process parameters are recorded below.

[0087] Table 2. Drying Process Parameter Recording Table

[0088]

[0089] As shown in Table 2, the expected cost gradient is as follows when drying reaches 30 minutes. When the value first turns positive and reaches 0.03, the optimal stop-time drying endpoint determination algorithm determines that the drying endpoint has been reached. The computer 15 sends a stop command to the drying device 14, and the drying device 14 automatically stops. The drying of the solid silica gel is completed, and a white powdery solid product is finally obtained.

[0090] This invention brings the following advancements compared to traditional methods: In parameter identification, the physical information dynamics identification model embeds physical equations as constraints into the loss function, ensuring that the identification results always fall within a physically reasonable range. This overcomes the inherent defects of pure data-driven models, which suffer from identification bias due to a lack of physical priors under sparse measurement conditions, and significantly reduces the sensitivity of identification accuracy to measurement noise. In temperature safety control, the hardware-level emergency cut-off circuit is independent of the software control program, guaranteeing the real-time over-temperature response at the circuit level and avoiding potential safety hazards caused by software scheduling delays. In drying control, the optimal stop-time drying endpoint decision algorithm replaces fixed time parameters with dynamic marginal cost decisions, enabling the drying device to adaptively determine the drying endpoint based on the actual water loss behavior of each batch of material. This eliminates the impact of batch-to-batch material differences on the consistency of drying results from the decision-making principle perspective.

[0091] It should be noted that the variables involved in this invention are explained in detail in Table 3.

[0092] Table 3. Variable Explanation Table

[0093]

[0094] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for intelligent automated control of small scale synthesis reactions, characterized by, Includes the following steps: The raw materials are loaded into raw material tank 1 and raw material tank 2 respectively. The peristaltic pump 1 and peristaltic pump 2 are started by computer, the feed rate and feed flow rate are set, and the mechanical stirring motor is started simultaneously and the stirring speed and stirring duration are set. Start the temperature control equipment, set the program heating and cooling rate and the duration of constant temperature. The PT100 temperature probe collects the temperature inside the vessel in real time and uploads it to the computer. The pH meter probe collects the pH value inside the vessel in real time and uploads it to the computer. All data is uploaded to the computer and recorded via RS485 communication line. The computer inputs the internal temperature of the vessel collected by the PT100 temperature probe, the internal pH value collected by the pH meter probe, and the feed flow rates of peristaltic pump 1 and peristaltic pump 2 into the physical information dynamics identification model. The physical information dynamics identification model outputs the current activation energy estimate and frequency factor estimate. The computer updates the model prediction control parameters based on the activation energy estimate and frequency factor estimate, dynamically adjusts the rotation speed of peristaltic pump 1 and peristaltic pump 2 and the power of the temperature control equipment, and at the same time, the pulse noise denoising median filter filters the internal pH value collected by the pH meter probe, and the dead zone nonlinear element suppresses the limit cycle oscillation. The computer continuously monitors the pH value and temperature slope inside the reactor. When the pH value deviation exceeds the pH deviation threshold, it switches to a multivariate soft measurement value based on conductivity, flow rate of the automatic discharge device flow meter, and feed flow rate, and triggers the automatic cleaning device valve to open for a short cleaning cycle of the pH meter probe. When the temperature slope inside the reactor exceeds the temperature rise rate threshold, the hardware-level emergency cut-off circuit intervenes independently of the computer control program and opens the cooling channel. After the reaction is completed, the computer sequentially closes the valves of raw material tank 1 and raw material tank 2, and opens the valve of the automatic feeding device. The flow meter of the automatic feeding device monitors the discharge flow rate in real time and feeds back to control the opening of the valve of the automatic feeding device. The material flows into the filter device evenly. The liquid flows into the liquid recovery tank through the second layer side pipe of the filter device, and the solid is retained in the first layer filter plate of the filter device. The computer opens the valve of the automatic cleaning device, and the cleaning liquid is used to clean the inner wall of the vessel, the top cover, and the stirring paddle 360° through the inlet pipe and the multi-directional rotating nozzle of the automatic cleaning device. The cleaning liquid flows out through the valve of the automatic discharge device and rinses the solids on the first layer of the filter plate of the filter device. The computer controls the slide rail servo motor to drive the robotic arm. The robotic arm moves the first layer of the filter plate and the solid material to the top of the drying device through the suction cup. The drying device determines the drying endpoint according to the optimal stop time drying endpoint judgment algorithm and automatically stops.

2. The intelligent automated small scale synthesis reaction control method of claim 1, wherein, The physical information dynamics identification model has the following structure: it uses a multi-layer feedforward neural network as the backbone. The input layer receives the temperature inside the vessel, the pH value inside the vessel, the feed flow rate and time. The output layer outputs the estimated activation energy and the estimated frequency factor. The loss function consists of two parts: the data fitting residual term and the physical equation residual term.

3. The intelligent and automated small-scale synthetic reaction control method according to claim 2, characterized in that, The residual terms of the physical equations embed the Arrhenius rate equation and the material balance differential equation in the form of soft constraints, requiring the network output to satisfy the reaction rate equation. and material balance constraint equations .

4. The intelligent and automated small-scale synthetic reaction control method according to claim 3, characterized in that, The training dataset for the physical information dynamics identification model is established by: collecting historical experimental data under different temperature programs and different feed flow rates, using the process quantity sequence at each sampling time as the input sample, and using the activation energy and frequency factor identified offline for the corresponding batch as labels. At the same time, virtual samples that satisfy the reaction rate equation and material balance constraint equation are generated through numerical simulation to expand the training set.

5. The intelligent and automated small-scale synthetic reaction control method according to claim 4, characterized in that, The training of the physical information dynamics identification model is specifically as follows: the Adam optimizer is used to minimize the total loss function, which is the weighted sum of the mean square error of the data fitting residuals and the physical equation residuals. The weight coefficients of the physical equation residuals are adaptively adjusted with each training round. After training, the input sequence is updated in real time using a sliding window method, and the current activation energy estimate and frequency factor estimate are output.

6. The intelligent and automated small-scale synthetic reaction control method according to claim 5, characterized in that, The physical information dynamic identification model also includes a dynamic identification confidence evaluation function. The dynamic identification confidence evaluation function is calculated based on a weighted average of three normalized indices: the variance of the temperature slope in the vessel, the drift rate of the pH value in the vessel, and the fluctuation rate of the feed flow rate. ,when When using a high learning rate, When using a moderate learning rate, The parameters are frozen and updated, and the output is switched to multivariate soft measurement values.

7. The intelligent and automated small-scale synthetic reaction control method according to claim 6, characterized in that, The confidence index value The calculation formula is ,in , , The weighting coefficients and , It is a dimensionless quantity.

8. The intelligent and automated small-scale synthetic reaction control method according to claim 7, characterized in that, The pulse denoising median filter specifically takes the median value of the original sampling sequence of the pH meter probe within a fixed-length sliding window as the filtered output. The dead zone nonlinear element specifically introduces an insensitive zone centered at zero at the model predictive control output. When the absolute value of the control deviation is less than the dead zone width, the control output remains unchanged.

9. The intelligent and automated small-scale synthetic reaction control method according to claim 8, characterized in that, The multivariate soft measurement value is specifically estimated as a substitute output value of the current pH value in the reactor based on conductivity, flow rate of the automatic discharge device flow meter and feed velocity through a pre-established regression relationship.

10. The intelligent and automated small-scale synthetic reaction control method according to claim 9, characterized in that, The hardware-level emergency shut-off circuit is a safety protection circuit that is independent of the computer control program, directly monitors the temperature inside the vessel by a hardware comparison circuit, and immediately shuts off the heating circuit and opens the cooling channel when the temperature rise rate threshold is exceeded. The unit of the temperature rise rate threshold is ℃ / s.

Images