Catalyst monitoring system based on sensing linkage and control method

Abstract

The invention belongs to the technical field of sewage treatment, and discloses a catalyst monitoring system based on sensing linkage and a control method.The system comprises a mixing pool, a settling pool and a backflow pool which are circularly arranged through pipelines, and further comprises a strong oxidant adding device, a powder catalyst adding device and a controller. According to the technical scheme, through combined use of the sensor, the flow meter and the valve, the problem of high loss rate of the powder type catalyst is solved, the service life of the catalytic reactor is prolonged, and the risk of environmental pollution is reduced. Meanwhile, the invention further discloses a powder type catalyst loss monitoring and controlling method based on the sensor and automatic control, in the method, hierarchical control is adopted on a controller control framework: a bottom-layer MPC is combined with a top-layer PINN prediction model to form a double-layer framework of network prediction and classical control, accurate monitoring and control are achieved, and the reliability of the controller is improved. Therefore, the purposes of reducing the loss rate of the catalyst, improving the stability of the catalytic reactor and prolonging the service life are achieved.

Description

Technical Field

[0001] The invention belongs to the technical field of sewage treatment and provides a catalyst monitoring system and a control method based on sensor linkage. Background Art

[0002] As one of the core equipment in modern sewage treatment technology, the catalytic reactor plays a vital role. It cleverly utilizes the catalytic effect of the catalyst to accelerate the chemical reaction of organic pollutants in sewage, thereby converting them into harmless or low-toxic substances and achieving effective sewage purification. Among the many types of catalysts, powder catalysts are highly favored due to their unique advantages. This type of catalyst has an extremely high specific surface area, which means that they can provide more active sites, allowing chemical reactions to proceed more quickly and thoroughly. At the same time, the activity of powder catalysts is relatively high, and they can exert significant catalytic effects at lower concentrations, which is of great significance for improving the efficiency of sewage treatment and reducing costs.

[0003] However, powdered catalysts have also exposed some thorny problems during their application, among which the loss problem is particularly prominent. During the operation of the catalytic reactor, due to factors such as water flow scouring and stirring, the powdered catalyst can easily flow out of the reactor with the sewage, resulting in the loss of the catalyst. This loss not only increases the consumption of the catalyst and increases the treatment cost, but may also have an adverse effect on the stable operation of the catalytic reactor. In order to meet this challenge, the prior art has proposed a series of solutions, among which precipitation and filtration are two more common methods. The precipitation method mainly uses gravity to gradually settle the lost catalyst particles in the sedimentation tank, thereby reducing the amount of them flowing out with the sewage. The filtration method intercepts the lost catalyst particles by setting a filter layer, such as a sand filter layer, a filter screen, etc. These methods can indeed reduce the loss of powdered catalysts to a certain extent, but they also expose some limitations. First, the physical capture and removal method may damage the activity of the catalyst during the reaction. For example, the filter layer may generate shear force on the catalyst particles, destroy their structure, and lead to a decrease in catalytic activity; during the precipitation process, the catalyst particles may agglomerate due to long-term standing, which will also affect their catalytic effect. Secondly, the operation and maintenance costs of these methods are not low. Sedimentation tanks need to be regularly desilted and filter layers need to be regularly replaced or cleaned, which increases the overall cost of sewage treatment. Finally, despite these measures, the loss of powdered catalysts has not been completely resolved, and the loss rate remains high, which not only shortens the service life of the catalytic reactor, but also may pollute the surrounding environment. For example, catalyst particles may be discharged into rivers, lakes and other water bodies with sewage, posing a threat to aquatic ecosystems.

[0004] Therefore, the existing technology still has a lot of room for improvement in dealing with the problem of powder catalyst loss. It is urgent to find a more effective and economical solution to reduce the catalyst loss rate, improve the stability and service life of the catalytic reactor, and reduce the overall cost of sewage treatment and protect the ecological environment. This is not only an urgent need in the field of sewage treatment technology, but also an important direction for promoting the sustainable development of the environmental protection industry. Summary of the invention

[0005] In view of this, the present invention proposes a catalyst monitoring system and control method based on sensor linkage, aiming to optimize the use efficiency of the catalyst, reduce the loss rate, and thus improve the overall catalytic effect.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] The present invention provides a catalyst monitoring system based on sensor linkage, including a mixing tank, a settling tank and a reflow tank arranged through a pipeline circulation, a sewage inlet pipe is provided at the bottom of the mixing tank, an inner cylinder extending downward from the top thereof and having a bottom opening is provided in the mixing tank, and a grid array filter plate is provided between the upper part of the mixing tank and the inner cylinder; it also includes a strong oxidant dosing device, a powder catalyst dosing device and a controller, the strong oxidant dosing device is connected to the inner cylinder through a strong oxidant metering pump, and the powder catalyst dosing device is connected to the inner cylinder through a powder catalyst metering pump; the strong oxidant metering pump and the powder catalyst metering pump are both electrically connected to the controller. In this way, the system design effectively reduces the loss rate of the catalyst through the coordinated circulation of the mixing tank, the settling tank and the reflow tank. The real-time adjustment of the controller ensures the optimization of the reaction conditions, improves the efficiency of sewage treatment, reduces the loss of the catalyst, reduces the replacement frequency and treatment cost of the catalyst, effectively removes organic pollutants in the sewage, reduces the impact on the environment, and meets the goal of sustainable development.

[0008] Optionally, the inner cylinder is provided with a side wall channel at the upper part higher than the grid array filter plate, and the side wall channel and the bottom opening of the inner cylinder are provided with a filter screen; the mixing tank is provided with a slurry pushing stirring device extending into the inner cylinder and electrically connected to the controller to enhance the circulation power; the mixing tank is also provided with a UV lamp and a microwave generator electrically connected to the controller. The UV lamp is usually used for disinfection and sterilization, and its wavelength can effectively destroy the DNA of bacteria and viruses, thereby achieving the effect of sterilization. The microwave generator can vibrate the substance to promote chemical reactions or physical changes; the grid array filter plate adopts a single-layer hexagonal honeycomb grid with a multi-level staggered stacking structure. The design of the hexagonal unit can make the filter plate have a larger filtration area per unit area, and the staggered arrangement of adjacent units also helps to enhance the stability and carrying capacity of the structure. This design is similar to a honeycomb in nature, realizing multiple stages from primary filtration to fine filtration. This graded filtration method can effectively remove impurities of different sizes and improve the overall efficiency of filtration.

[0009] Optionally, an agitator is provided on the reflux tank, which is mainly used to ensure uniform mixing of the liquid in the tank and sufficient dispersion of the catalyst; the powder catalyst dosing device is also connected to the reflux tank through a powder catalyst replenishing pump, which can replenish the catalyst in real time to ensure that the concentration of the catalyst is maintained at an ideal level during the reaction, and the agitator and the powder catalyst replenishing pump are both electrically connected to the controller, which can achieve precise control of the catalyst dosage; a diffuser connected to the reflux tank is provided in the mixing tank to promote uniform distribution and mixing of the fluid. In this way, it is helpful to realize the recycling of the catalyst and minimize material loss. At the same time, the precipitate can be effectively guided to the reflux tank for further processing, and it is also ensured that the reactants in the mixing tank are always maintained in the best reaction state.

[0010] Optionally, a barrel tube and a conical bucket are arranged in sequence from top to bottom in the sedimentation tank, and the bottom of the sedimentation tank is cone-shaped; a pulse microwave processor is provided on the sedimentation tank, which extends into the barrel tube and is electrically connected to the controller. Inside the barrel tube, the powder catalyst will settle, and the heavier particles will gradually settle to the bottom. The bottom is conical, which helps to concentrate the settled solid particles to the bottom of the tank, so as to facilitate subsequent treatment or discharge. The conical design can effectively improve the sedimentation efficiency and ensure that the solid particles can be quickly settled and collected. A pulse microwave processor is provided on the sedimentation tank, which extends into the barrel tube and is electrically connected to the controller. The main function of the pulse microwave processor is to use microwave energy to process the liquid. Microwave energy can effectively excite the molecular movement in the liquid, improve the sedimentation process, and improve the sedimentation efficiency.

[0011] Optionally, valve I and pump I are provided on the pipeline between the mixing tank and the settling tank, valve II and pump II are provided on the pipeline between the settling tank and the return tank, pump III is provided on the pipeline between the return tank and the mixing tank, the barrel pipe at the top of the settling tank is connected to a drain pipe with valve III, the cone at the bottom of the settling tank is connected to a discharge pipe with valve IV, and valve I, valve II, valve III, valve IV, pump I, pump II and pump III are all electrically connected to the controller. The water pump can optimize the working state of each treatment unit by adjusting the flow rate, and the valve is used to open or close the liquid flow for easy maintenance and operation.

[0012] Optionally, an ORP sensor and a COD sensor are provided on the mixing tank, which are used to monitor the oxidation-reduction potential (ORP) and the chemical oxygen demand (COD), respectively, and can provide real-time feedback on the water quality information in the mixing tank, thereby helping to optimize the reaction conditions and monitor water quality changes; a flow meter is provided on the water inlet pipe of the sedimentation tank, which is used to monitor the liquid flow in real time, ensure the flow matching between each unit, and improve the processing efficiency; a differential pressure sensor is provided between the water inlet pipe and the water outlet pipe of the sedimentation tank, which is used to monitor the pressure difference change of the sedimentation tank, help judge the sedimentation effect and the filtering status, and adjust the system operation in time; a temperature sensor, a pH value sensor and a first turbidity sensor are provided on the sedimentation tank and outside the barrel tube , respectively used to monitor the temperature, pH and turbidity of the liquid in the sedimentation tank, to ensure that the conditions of the sedimentation process are suitable, and to adjust the treatment strategy in time; a second turbidity sensor is provided on the sedimentation tank and located in the barrel tube, which can monitor the turbidity of the effluent from the sedimentation tank so as to evaluate the treatment effect, and the ORP sensor, COD sensor, flow meter, differential pressure sensor, temperature sensor, pH value sensor, first turbidity sensor and second turbidity sensor are all electrically connected to the controller to form a centralized control system, that is, the controller can adjust the operating status of the pump, the opening and closing of the valve, the reaction conditions, etc. in real time according to the information fed back by various sensors, so as to ensure the efficient operation of the entire system and the stability of water quality.

[0013] The present invention also provides a control method for a catalyst monitoring system based on sensor linkage. In the method, a hierarchical control is adopted in the controller control architecture: the bottom layer MPC is combined with the top layer PINN prediction model to form a two-layer architecture of "network prediction + classical control". The method includes the following steps:

[0014] S1, sewage enters the reactor, the sensor collects real-time data, and the controller pre-processes the collected data;

[0015] S2. Predict the temperature, COD, pH, and ORP water quality indicators in the reactor in the future through the PINN prediction model;

[0016] S3. The controller generates dynamic set values ​​or soft constraints based on the prediction results, including but not limited to: UV lamp power upper limit, target COD level, and valve optimal opening range;

[0017] S4. The controller performs rolling optimization in a fast cycle of seconds to minutes, and enables the controller to control valve opening, UV power, and catalyst addition variables in real time to ensure that the effluent quality meets the standard and energy consumption is minimized; during the operation of the system, a safety protection mechanism is retained: if the top-level prediction model or communication is abnormal, MPC or traditional PID is relied on to maintain basic operation;

[0018] S5. Conduct effluent testing and feedback, and ultimately achieve concentration monitoring and COD monitoring.

[0019] In step S2, the water quality index is predicted at intervals based on the physical information neural network PINN prediction model. When training the prediction model, the dynamic process of multiple parameters such as chemical oxygen demand COD, pH, temperature T and oxidation-reduction potential ORP in the sewage treatment system is targeted.

[0020] The PINN physical constraint equations specifically include the following:

[0021] 1) Construct a multi-parameter coupling mechanism model:

[0022] Chemical oxygen demand (COD) degradation kinetic model:

[0023]

[0024] parameter:

[0025] οQ in : Water inlet flow (L / min), determined by valve opening V valve Control: Q in =k valve ·V valve

[0026] ο: Reaction rate constant (hypothetical value)

[0027] οE a =5000 J / mol: activation energy (refer to photocatalysis literature)

[0028] ο: ideal gas constant

[0029] οT: reaction temperature (dynamic calculation)

[0030] pH dynamic model:

[0031]

[0032] parameter:

[0033] ο: ORP rise rate coefficient

[0034] οK ORP =50mg / L: half-saturation constant

[0035] ολ ORP =0.05min -1 : ORP natural attenuation coefficient

[0036] οORP 0 =200mV: basic redox potential

[0037] Temperature T dynamic model:

[0038]

[0039] parameter:

[0040] οη heat =0.2: UV lamp heat generation efficiency

[0041] οh=0.1W / (℃): heat dissipation coefficient

[0042] οmc p =1000J / ℃: Reactor heat capacity

[0043] οT env =25℃: Ambient temperature

[0044] οT in =20℃: water inlet temperature

[0045] Oxidation Reduction Potential (ORP) Dynamic Model

[0046]

[0047] parameter:

[0048] ο: ORP rise rate coefficient

[0049] οK ORP =50mg / L: half-saturation constant

[0050] ολ ORP =0.05min -1 : ORP natural attenuation coefficient

[0051] οORP 0 =200mV: basic redox potential

[0052] 2) Define control variables and random disturbances:

[0053] Manipulate variable scope:

[0054] Valve opening V valve∈[0, 1], controls the water flow rate Q in =10V valve

[0055] Dosing pump speed Q pump ∈[20,100]mL / min, adjust pH neutralization rate

[0056] UV power P UV ∈[200, 600]W, driving the photocatalytic reaction;

[0057] Random perturbation design:

[0058] Influent water quality disturbance:

[0059]

[0060] T in : Inlet water temperature (unit: ℃), example value: 20.

[0061] Q in : Water inlet flow (unit: L / min)

[0062] Dynamic Noise Injection:

[0063]

[0064]

[0065] Parameter definition:

[0066] N(0,σ): Gaussian noise with mean 0 and standard deviation σ.

[0067] σ COD : COD measurement noise intensity, dynamically adjusted with COD value.

[0068] σ ORP : ORP measurement noise intensity, including fixed error (5mV) and proportional error (0.02·|ORP|).

[0069] The training of the physical information neural network PINN in this method adopts a dynamic staged training framework, which realizes high-precision modeling and prediction of the sewage treatment system by integrating the multi-parameter coupling mechanism model and the adaptive weight adjustment mechanism. Specifically, it includes:

[0070] 1) Model architecture:

[0071] Network structure:

[0072] Input layer: historical time window data ([t-Tw:t]), including COD, pH, temperature, ORP and operating variables. The operating variables include valve opening, pump speed, and UV power.

[0073] Feature extraction layer: Use bidirectional LSTM to capture temporal dependencies and output hidden state ht;

[0074] Physical constraint layer: ht is input into multiple sub-networks to predict COD, pH, temperature, and ORP for the next Np steps respectively;

[0075] Output definition:

[0076]

[0077] 2) Construction of hybrid loss function:

[0078] Data-driven losses:

[0079]

[0080] Physics-driven loss: Calculates the physical residual of the predicted value:

[0081]

[0082] Dynamically weighted total loss:

[0083]

[0084] Adaptive Weight Regulator:

[0085]

[0086] Dynamically balance the weights of data fitting and physical constraints to avoid model divergence due to excessive physical residuals in the early stages of training;

[0087] 3) Time and space divide and conquer training strategy:

[0088] Phase 1: Pre-training of the rapid response process:

[0089] Objective: Prioritize learning the transient response of COD and ORP;

[0090] Data screening: select samples with COD change rate (|dCOD / dt|>5mg / L / min);

[0091] Loss weight: Increase the weight of the physical residual term of COD and ORP (λphy×2);

[0092] Phase 2: Slow process fine-tuning:

[0093] Objective: To optimize long-term predictions of temperature and pH;

[0094] Data screening: select samples with temperature change rate (dT / dt<0.5℃ / min);

[0095] Loss weight: Increase the weight of the temperature conservation equation and introduce energy constraint terms:

[0096]

[0097] Stage 3: Joint optimization of all parameters:

[0098] Objective: Global fine-tuning to improve the prediction accuracy of multi-parameter coupling;

[0099] 4) Virtual-Real Fusion Transfer Learning:

[0100] Synthetic data pre-training: Use the generated high-quality dataset to train the initial PINN and learn the physical laws in the mechanism model;

[0101] Real data fine-tuning: Collect a small amount of actual sewage plant data (such as 1% synthetic data volume) and align the feature distribution through domain adaptive algorithms (such as DANN):

[0102]

[0103] In this method, after the PINN model is deployed, an online learning update mechanism is used to ensure the long-term validity of the model. The online learning update mechanism includes:

[0104] Incremental data collection strategy:

[0105] 1) Real-time data caching:

[0106] Establish a real-time data buffer to store high-frequency data from the last 7-30 days;

[0107] Realize automatic labeling of data;

[0108] Set up an automatic data quality assessment mechanism;

[0109] 2) Key event trigger collection:

[0110] Increase the frequency of data collection when operating conditions change;

[0111] Save complete process data when abnormal loss events occur;

[0112] After equipment maintenance, focus on collecting initial operating data;

[0113] 3) Directed experimental data supplement:

[0114] Regularly conduct small-scale adjustments to supplement training data for specific areas;

[0115] Test system response under boundary conditions within safety limits;

[0116] Conduct targeted data collection in areas where model predictions are weak;

[0117] Model update strategy:

[0118] 1) Regular update mechanism:

[0119] Set a fixed period (every 2 weeks) to update model parameters;

[0120] Use the sliding window method to select the data of the last month for retraining;

[0121] Keep historical version models for comparison and rollback;

[0122] 2) Event-triggered updates:

[0123] Trigger model update when the prediction error exceeds the threshold;

[0124] Mandatory model update after equipment maintenance or catalyst replacement;

[0125] Targeted updates after changes in processes or operating methods;

[0126] 3) Incremental learning method:

[0127] Use transfer learning to preserve learned physical laws;

[0128] Adopt elastic weight update strategy to give different weights to new data and historical data;

[0129] Implement a partial parameter freezing mechanism to update only the parameters that are sensitive to new operating conditions.

[0130] In this method, dynamic gradient tracking and contribution quantification method are used

[0131] 1) Interpretability layer architecture design

[0132] Gradient separation module:

[0133] In the loss function, a residual term is defined separately for each physical equation (such as COD dynamics, temperature balance, ORP model) to construct a structured loss function:

[0134]

[0135] Symbol meaning:

[0136] λ data : Data-driven weight coefficient, dynamically adjusts the contribution of data fitting in the total loss.

[0137] : The weight coefficient of the ith physical equation, reflecting the importance of the physical constraint (such as COD dynamics > temperature balance).

[0138] L data: Data-driven loss, which calculates the mean squared error (MSE) between the model predictions and the true sensor data.

[0139] : The residual loss of the ith physical equation, based on the deviation between the predicted value and the actual value of the physical equation (such as the residual of the COD degradation kinetic equation).

[0140] N: The total number of physical equations, such as the 4 core constraints defined in the patent (COD, pH, temperature, ORP).

[0141] Gradient backpropagation separation

[0142]

[0143] Symbol meaning:

[0144] θ: All trainable parameters of the neural network, including the weights and biases of the LSTM layer and the physical constraint layer.

[0145] The gradient of the i-th physical equation with respect to the parameters is calculated independently via automatic differentiation (Autograd).

[0146] Application scenario: The patent uses a gradient hook to capture the contribution of each physical equation (such as the temperature model) to parameter updating.

[0147] The gradient tensor of each physical equation is extracted through automatic differentiation (autograd in PyTorch).

[0148] 2) Real-time contribution quantification indicators

[0149] Gradient Weight Ratio (GWR)

[0150] Calculate the contribution of each physical equation's gradient to the total gradient:

[0151]

[0152] Symbol meaning:

[0153] ||·|| 2 : L2 norm, calculates the Euclidean length of the gradient vector and measures the gradient magnitude.

[0154] GWR i : The gradient contribution ratio of the i-th physical equation, reflecting the relative importance of this equation to the update of model parameters.

[0155] Application scenario: If the GWR of COD dynamics is 70%, it means that the model relies more on the COD equation optimization parameters.

[0156] Reflects the relative importance of different physical equations to parameter updates during training.

[0157] Dynamic Residual Impact Factor (DRIF)

[0158] Dynamically adjust weights based on the proportion of physical residual in the total loss:

[0159]

[0160] Symbol meaning:

[0161] DRIF i : The residual proportion of the i-th physical equation, measuring the contribution of this equation to the total physical loss.

[0162] The difference from weight: is the unmultiplied weight The original residual reflects the difficulty of fitting the equation itself.

[0163] Application scenario: When the DRIF of the temperature equation is continuously higher than 50%, it indicates that the model has a large error in temperature prediction.

[0164] Used to evaluate the strength of constraints imposed by each physical equation on model predictions.

[0165] The innovation and advantages of this method are: transparency of physical constraints, separation of gradient-level physical contributions for the first time, breaking through the "black box" characteristics of traditional PINN. The actual impact of physical equations on model predictions is quantified through dynamic weight ratio (GWR) and residual impact factor (DRIF). Real-time interaction and debugging support: Engineers can manually adjust physical constraint weights (such as temporarily enhancing temperature equation constraints) through the dashboard and observe model responses. It supports backtracking of training history to locate physical law violations (such as sudden increases in energy conservation residuals).

[0166] Optionally, the algorithm flow of the MPC controller includes:

[0167] 1) Prediction model call:

[0168] Input current status: real-time sensor data (pH, ORP, temperature, turbidity, COD) and operating variables;

[0169] PINN prediction: call the trained PINN model to predict the multi-parameter prediction values ​​for the next (Np) steps; PINN input: current and historical data (pH, ORP, temperature, COD, valve opening, pump speed, UV power); PINN output: pH, ORP, T, COD; turbidity prediction values ​​are directly calculated through mathematical equations. Considering the limited computing power of the PINN model in actual operation:

[0170]

[0171] γ: flocculant sedimentation coefficient, related to the type of agent;

[0172] η: hydraulic retention time influencing factor;

[0173] κ: UV oxidation decomposition efficiency of suspended solids

[0174] 2) Optimization problem construction: Objective function (sorted by weight priority):

[0175] J=w 1 (COD-COD ref ) 2 +w 2 (pH-pH ref ) 2 +w 3 (Turbidity-Turbidity ref ) 2 +w 4 (Δu) 2

[0176] Constraints:

[0177] Valve opening: 20% ≤ opening ≤ 80% (to prevent blockage or excessive flow rate)

[0178] UV power change rate: ≤50W / minute (avoid thermal shock)

[0179] Turbidity threshold: turbidity ≤ 50NTU (to prevent catalyst loss)

[0180] 3) Numerical optimization solution:

[0181] Initialization: Generate an initial guess value (optimal solution in the previous cycle) based on the current operating variables and states;

[0182] Iterative solution: IPOPT based on CasADi is used to support fast nonlinear optimization, and the operating variables are iteratively adjusted in the prediction time domain to minimize the objective function; the objective function gradient needs to be calculated for each iteration (through automatic differentiation or finite difference);

[0183] Output result: Take the first step operation variable in the optimization sequence (rolling time domain characteristics of MPC).

[0184] Compared with the prior art, the present invention has the following beneficial effects:

[0185] 1. Solved the problem of catalyst loss: This technical solution uses turbidity sensors and valves to effectively reduce the loss rate of powdered catalysts by real-time monitoring and controlling the turbidity in the reactor. Compared with existing physical capture and removal methods, this technical solution is more accurate and effective and can more thoroughly solve the problem of catalyst loss.

[0186] 2. Improved catalyst activity: Since the technical solution avoids the influence of physical capture and removal methods in the prior art on catalyst activity, it can better maintain the activity of the catalyst, thereby improving the effect of the catalytic reaction.

[0187] 3. Reduced operation and maintenance costs: This technical solution reduces manual intervention and reduces operation and maintenance costs through automatic monitoring and control. Compared with existing technologies, this technical solution is more economical and efficient.

[0188] 4. Extending the service life of the catalytic reactor: Since the technical solution can effectively reduce the loss of catalyst, the service life of the catalytic reactor is extended, the frequency of equipment replacement and maintenance is reduced, and the utilization rate of the equipment is improved.

[0189] 5. Beneficial to environmental protection: Since this technical solution can effectively reduce the loss of catalyst, thereby reducing the impact of catalyst on the environment, it is beneficial to protect the environment and reduce environmental pollution problems.

[0190] In general, compared with the existing technology, this technical solution has higher efficiency, lower cost and better environmental benefits, and is a more advanced and practical sewage treatment technology solution.

[0191] Other advantages, objectives and features of the present invention will be described in the following description to some extent, and to some extent, will be obvious to those skilled in the art based on the following examination and study, or can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0192] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be described in detail below in conjunction with the accompanying drawings, wherein:

[0193] Figure 1 It is a schematic diagram of a powder catalyst loss monitoring and control system based on sensors and automatic control of the present invention;

[0194] Figure 2It is a schematic diagram of the process of the method of the present invention;

[0195] Figure 3 It is a control logic flow chart;

[0196] Figure numerals: 1-sewage inlet pipe, 2-mixing tank, 3-inner cylinder, 4-filter screen, 5-grid array filter plate, 6-side wall channel, 7-push-paddle stirring device, 8-ORP sensor, 9-COD sensor, 10-strong oxidant dosing device, 11-strong oxidant metering pump, 12-powder catalyst metering pump, 13-powder catalyst dosing device, 14-powder catalyst replenishing pump, 15-ultraviolet lamp, 16-microwave generator, 17-valve I, 18-pump I, 19-flow meter , 20- differential pressure sensor, 21- temperature sensor, 22- PH value sensor, 23- first turbidity sensor, 24- pulse microwave processor, 25- second turbidity sensor, 26- valve III, 27- barrel tube, 28- cone bucket, 29- sedimentation tank, 30- valve IV, 31- valve II, 32- pump II, 33- reflux tank, 34- agitator, 35- pump III, 36- diffuser, 37- controller, A- water inlet pipe, B- water outlet pipe, C- drain pipe, D- sewage pipe. DETAILED DESCRIPTION

[0197] The present invention is further described below in conjunction with specific implementation methods. The accompanying drawings are only used for exemplary descriptions and are only schematic diagrams, not actual drawings, and cannot be understood as limiting this patent; in order to better illustrate the embodiments of the present invention, some parts of the accompanying drawings may be omitted, enlarged or reduced, and do not represent the size of the actual product; for those skilled in the art, it is understandable that some well-known structures and their descriptions in the accompanying drawings may be omitted.

[0198] like Figure 1As shown, a catalyst monitoring system based on sensor linkage mentioned in the present invention comprises a mixing tank 2, a settling tank 29 and a reflow tank 33 arranged through a pipeline circulation, a sewage inlet pipe 1 is provided at the bottom of the mixing tank 2, which is responsible for introducing sewage into the mixing tank 2, an inner cylinder 3 extending downward from its top and having a bottom opening is provided in the mixing tank 2, a grid array filter plate 5 is provided in the upper part of the mixing tank 2 and between the inner cylinder 3, which is used to filter and capture the lost powder catalyst to reduce its loss in the subsequent process; and the inner cylinder 3 is provided with a side wall channel 6 at the upper part higher than the grid array filter plate 5, allowing the filtered liquid to flow out, thereby enhancing the fluidity of the mixing tank 2, that is, utilizing the bottom opening and the side wall channel 6 of the inner cylinder 3, and with the assistance of the push-paddle stirring device 7, a circulation passage is formed between the inner cylinder 3 and the mixing tank 2; the side wall channel 6 and the bottom opening of the inner cylinder 3 are both provided with The filter screen 4 has a reverse filtering effect; the mixing tank 2 is provided with a slurry pushing stirring device 7 which extends into the inner tube 3 and is electrically connected to the controller 37, which can make the catalyst more dispersed and ensure that the powdered catalyst is broken up as much as possible in the mixing tank and mixed evenly with the sewage, so that the sewage and catalyst in the mixing tank 2 can be effectively stirred; it also includes a strong oxidant dosing device 10, a powdered catalyst dosing device 13 and a controller 37. The strong oxidant dosing device 10 is connected to the inner tube 3 through a strong oxidant metering pump 11, and can accurately control the dosage of the strong oxidant as needed to improve the efficiency of sewage treatment. The powdered catalyst dosing device 13 is connected to the inner tube 3 through a powdered catalyst metering pump 12, ensuring that the powdered catalyst can be continuously and stably added during the reaction; the strong oxidant metering pump 11 and the powdered catalyst metering pump 12 are both electrically connected to the controller 37.

[0199] With the above scheme, sewage enters the mixing tank through the sewage inlet pipe and is fully mixed with the powder catalyst and the strong oxidant; the push-paddle stirring device stirs the sewage in the inner tube to promote full contact between the strong oxidant, the powder catalyst and the sewage, thereby enhancing the efficiency of the catalytic reaction; after mixing, the liquid in the mixing tank flows through the grid array filter plate to capture the lost catalyst, and the filtered liquid forms a circulation flow through the side wall channel and the bottom opening of the inner tube; the mixed sewage flows into the sedimentation tank, and the granular powder catalyst enters the reflux tank after settling under the action of gravity and then returns to the mixing tank, thereby ensuring the recycling of the catalyst and reducing the loss of the catalyst, and finally, the clean water is separated in the barrel tube of the sedimentation tank and discharged through the drain pipe C, or the deactivated powder catalyst is discharged through the discharge pipe D.

[0200] In this embodiment, the mixing tank 2 is also provided with a UV lamp 15 and a microwave generator 16 electrically connected to the controller 37; the UV lamp 15 installed on the mixing tank 2 can sterilize the substance in real time during the mixing process to ensure the safety and hygiene of the product. The use of the microwave generator 16 in the mixing tank 2 can accelerate the mixing process, increase the reaction rate, or in some cases, help the substance to be evenly dispersed.

[0201] In this embodiment, the grid array filter plate 5 adopts a single-layer hexagonal honeycomb grid with a multi-level staggered stacking structure. For example, the overall structure is a regular hexagonal honeycomb grid, but each hexagonal unit is provided with three levels of gradual channels to form a multi-level layered structure, the first level of large channels (100-200μm) as the primary filtration layer; the second level of medium channels (20-50μm) for secondary filtration; the third level of micro-channels (5-10μm) for fine filtration. In this way, adjacent hexagonal units are arranged in a staggered manner on the plane, similar to a honeycomb structure. The three-level channels in each hexagonal unit are not vertically connected, but are connected in a gradual transition manner to form zigzag or staggered channels. This design significantly increases the contact area between the airflow and the filter surface. It is recommended that the side length of each hexagonal unit be designed to be 500 microns and the unit wall thickness be 10 microns, which not only ensures the structural strength, but also does not excessively affect the filtration efficiency. The overall array can be made into a circular or rectangular plane according to actual needs. It is recommended that the size of a single array board be 100mm×100mm, which is convenient for modular installation and replacement.

[0202] In this embodiment, a stirrer 34 is provided on the reflux tank 33, through which the reaction rate can be increased and the contact efficiency between the reactants and the catalyst can be ensured; the powder catalyst dosing device 13 is also connected to the reflux tank 33 through the powder catalyst replenishing pump 14, and the stirrer 34 and the powder catalyst replenishing pump 14 are both electrically connected to the controller 37, which means that the entire system can be automatically controlled. The controller 37 can adjust the stirring speed and the amount of catalyst added according to the real-time monitoring data to optimize the reaction conditions.

[0203] In this embodiment, a diffuser 36 connected to the reflux tank 33 is provided in the mixing tank 2. The diffuser 36 can help the liquid form a more uniform flow state in the mixing tank, thereby increasing the uniformity and efficiency of the reaction.

[0204] In this embodiment, a barrel-shaped tube 27 and a conical bucket 28 are sequentially arranged in the settling tank 29 from top to bottom, and the bottom of the settling tank 29 is a cone; a pulse microwave processor 24 is arranged on the settling tank 29, which extends into the barrel-shaped tube 27 and is electrically connected to the controller 37. In this way, the liquid entering the settling tank 29 will flow through the passage between the barrel-shaped tube 27 and the conical bucket 28, so that the powder catalyst can flow along the conical bucket 28 to the bottom cone of the settling tank 29, and when the liquid rises and enters the barrel-shaped tube 27, the wall slope of the barrel-shaped tube 27 can be used to promote the powder catalyst to settle under gravity, and further, under the intermittent pulse microwave impact of the pulse microwave processor 24, it helps the granular powder catalyst to obtain a better sedimentation effect, further precipitate and separate the granular powder catalyst in the sewage, so as to facilitate the recovery of the catalyst.

[0205] In this embodiment, a valve I17 and a pump I18 are provided on the water inlet pipe A between the mixing tank 2 and the settling tank 29, a valve II31 and a pump II32 are provided on the water outlet pipe B between the settling tank 29 and the reflow tank 33, a pump III35 is provided on the pipeline between the reflow tank 33 and the mixing tank 2, a barrel pipe 27 at the upper part of the settling tank 29 is connected to a drain pipe C with a valve III26, a cone at the bottom of the settling tank 29 is connected to a discharge pipe D with a valve IV30, and valve I17, valve II31, valve III26, valve IV30, pump I18, pump II32 and pump III35 are all electrically connected to the controller 37. The function of these valves and water pumps is to control the flow and flow of liquids to ensure smooth liquid treatment processes at different stages of the system.

[0206] In this embodiment, an ORP sensor 8 and a COD sensor 9 are provided on the mixing tank 2; a flow meter 19 is provided on the water inlet pipe A of the settling tank 29; a differential pressure sensor 20 is provided between the water inlet pipe A and the water outlet pipe B of the settling tank 29; a temperature sensor 21, a pH value sensor 22 and a first turbidity sensor 23 are provided on the settling tank 29 and outside the barrel tube 27; a second turbidity sensor 25 is provided on the settling tank 29 and inside the barrel tube 27, and the ORP sensor 8, COD sensor 9, flow meter 19, differential pressure sensor 20, temperature sensor 21, pH value sensor 22, first turbidity sensor 23 and second turbidity sensor 25 are all electrically connected to the controller 37 to form a centralized control system. In this way, the controller can adjust the operating state of the pump, the opening and closing of the valve, the reaction conditions, etc. in real time according to the information fed back by various sensors to ensure the efficient operation of the entire system and the stability of water quality.

[0207] When the system is running, the PINN prediction model is used to capture the photocatalytic mechanism and predict the temperature, COD, pH, and ORP water quality indicators for a period of time in the future. The controller generates dynamic set values ​​or soft constraints based on the prediction results, including: UV lamp power upper limit, target COD level, and optimal valve opening range, so as to better cope with fluctuations in influent water quality. The controller performs rolling optimization within a fast cycle of seconds to minutes, and enables the controller to control valve opening, UV power, and catalyst addition variables in real time to ensure that the effluent water quality meets the standards and minimizes energy consumption. During the operation of the system, a safety protection mechanism is retained: if the top-level prediction model or communication is abnormal, MPC or traditional PID is relied on to maintain basic operation. In this embodiment, the PINN prediction model predicts the water quality indicators pH, ORP, and COD for the next 30 minutes every 15 minutes; and sends the prediction results or the optimized reference trajectory and constraint range to the MPC controller; the MPC controller uses 1 minute as the sampling period (Ts=1min) and rolls out predictions for 10 minutes (Np=10), but only optimizes the control input within the first 5 minutes (Nc=5), and controls the water inlet valve opening, UV lamp power, flow rate, and catalyst dosage. Under the premise of meeting the real-time process constraints, the key indicators of effluent pH, COD, ORP, temperature, and catalyst dosage are set or maintained in the optimal range, while taking energy saving into consideration.

[0208] Figure 2 The flowchart of the method of the present invention is shown in the figure, and the method includes the following steps: S1, sewage enters the reactor, the sensor collects data in real time, and the controller preprocesses the collected data; S2, the temperature, COD, pH, ORP water quality indicators in the reactor are predicted in the future through the PINN prediction model; S3, the controller generates dynamic set values ​​or soft constraints according to the prediction results, including but not limited to: UV lamp power upper limit, target COD level, valve optimal opening range; S4, the controller performs rolling optimization within a fast cycle of seconds to minutes, and realizes the controller to control the valve opening, UV power, and catalyst addition variables in real time to ensure that the effluent water quality meets the standard and the energy consumption is minimized; during the operation of the system, the safety protection mechanism is retained: if the top-level prediction model or communication is abnormal, the basic operation is maintained by MPC or traditional PID; S5, effluent detection and feedback are performed, and finally concentration monitoring and COD monitoring are realized.

[0209] In terms of the construction of training data sets, this paper proposes a synthetic data generation method based on a multi-parameter coupling mechanism model and dynamic noise injection. For the multi-parameter dynamic processes such as chemical oxygen demand (COD), pH, temperature (T) and oxidation-reduction potential (ORP) in sewage treatment systems, a high-fidelity, multi-scale, and highly generalized training data set is generated through physical and chemical mechanism modeling, random perturbation design, and dynamic verification mechanism. The core points of this training set construction method include:

[0210] 1. Dynamic model of multi-physical field coupling: Establish the joint differential equation of COD, pH, temperature and ORP to accurately describe the time-varying relationship between redox potential and pollutant degradation in the photocatalytic reaction.

[0211] 2. Adaptive noise injection mechanism: Dynamically adjust the noise intensity based on sensor characteristics to simulate measurement errors and sudden disturbances in real environments.

[0212] 3. Data enhancement driven by physical constraints: Introduce reaction rate boundary conditions and energy conservation constraints to ensure that the generated data conforms to the laws of thermodynamics and chemical kinetics.

[0213] In this embodiment, the training data set generation method steps include:

[0214] Step 1: Construct a multi-parameter coupling mechanism model

[0215] (1) Chemical oxygen demand (COD) degradation kinetic model

[0216]

[0217] -Introducing temperature-dependent Arrhenius correction terms to relate UV power to reaction efficiency and enhance model nonlinearity.

[0218] parameter:

[0219] οQ in : Water inlet flow (L / min), determined by valve opening V valve Control: Q in =k valve ·V valve

[0220] ο: Reaction rate constant (hypothetical value)

[0221] οE a =5000 J / mol: activation energy (refer to photocatalysis literature)

[0222] ο: ideal gas constant

[0223] οT: reaction temperature (dynamic calculation)

[0224] (2) Dynamic model of pH

[0225]

[0226] -The acid production effect term of COD degradation is proposed to reflect the inhibitory effect of organic matter decomposition on pH.

[0227] parameter:

[0228] ο: ORP rise rate coefficient

[0229] οN ORP =50mg / L: half-saturation constant

[0230] ολ ORP =0.05min -1 : ORP natural attenuation coefficient

[0231] οORP n =200mV: basic redox potential

[0232] (3) Temperature (T) dynamic model

[0233]

[0234] -A dual-path heat balance model that integrates UV heat generation and water inlet heat exchange can accurately predict the temperature accumulation effect.

[0235] parameter:

[0236] οη heat =0.2: UV lamp heat generation efficiency

[0237] οh=0.1W / (℃): heat dissipation coefficient

[0238] οmc p =1000J / ℃: Reactor heat capacity

[0239] οT env =25℃: Ambient temperature

[0240] οT in =20℃: water inlet temperature

[0241] (4) Oxidation-reduction potential (ORP) dynamic model

[0242]

[0243] Physical meaning:

[0244] UV photocatalysis generates oxidative free radicals (·OH), increasing ORP;

[0245] COD degradation consumes oxidants and reduces ORP.

[0246] parameter:

[0247] ο: ORP rise rate coefficient

[0248] οK ORP =50mg / L: half-saturation constant

[0249] ολ ORP =0.05min -1 : ORP natural attenuation coefficient

[0250] οORP 0 =200mV: basic redox potential

[0251] -The ORP kinetic model based on the Michaelis-Menten equation describes the competitive relationship between free radical concentration and COD consumption in the photocatalytic reaction.

[0252] Step 2: Define control variables and random disturbances

[0253] (1) Operation variable range

[0254] Valve opening V valve ∈[0, 1], controls the water flow rate Q in =10V valve .

[0255] Dosing pump speed Q pump ∈[20, 100]mL / min, adjust the pH neutralization rate.

[0256] UV power P UV ∈[200, 600]W, driving the photocatalytic reaction.

[0257] (2) Random perturbation design

[0258] - Influent water quality disturbance:

[0259]

[0260] -Dynamic Noise Injection:

[0261]

[0262] -Innovation: Noise intensity is dynamically adjusted with the measured value to simulate sensor nonlinear errors.

[0263] Step 3: Numerical solution and data generation

[0264] 1. Initial state sampling:

[0265] - Randomly generate initial COD, pH, temperature, and ORP, covering the typical operating range (COD: 50-150 mg / L, pH: 6.5-8.5, T: 25-35°C, ORP: 150-250 mV).

[0266] 2. Control sequence generation:

[0267] - Randomly adjust the operating variable (V every 10 minutes valve , Q pump , P UV ), ensuring coverage of the full parameter space.

[0268] 3. Fourth-order Runge-Kutta method solution:

[0269] -Use variable step size algorithm to solve differential equations and generate high-precision time series data.

[0270] Step 4: Physical Constraint Verification and Data Augmentation

[0271] 1. Energy conservation verification:

[0272] ο Check whether the temperature changes meet the Eliminate data that violates the first law of thermodynamics.

[0273] 2. Reaction rate boundary conditions:

[0274] οConstraining COD degradation rate Prevent the model from overestimating.

[0275] 3. Dynamic coupling verification:

[0276] οVerify whether the ORP rising trend is met when COD decreases

[0277] Step 5: Dataset construction and standardization

[0278] 1. Time window slicing:

[0279] - Input characteristics: historical 5-minute COD, pH, temperature, ORP and operating variables.

[0280] -Output label: Multi-parameter forecast value for the next 10 minutes.

[0281] 2. Dataset division:

[0282] -Training set (80%), validation set (10%), and test set (10%) to ensure time series continuity.

[0283] The training data set construction method provided by the present invention has the following advantages:

[0284] 1. High-fidelity data generation: By coupling the COD-ORP kinetic model, the problem of inaccurate redox potential prediction in traditional methods is solved.

[0285] 2. Strong generalization capability: random perturbation design and physical constraint verification mechanism ensure that data covers extreme working conditions and sensor failure scenarios.

[0286] 3. Computational efficiency: A lightweight model based on dynamic noise injection can generate millions of samples on an ordinary server.

[0287] 4. Cross-platform compatibility: The generated data can be directly used to train AI models such as PINN and LSTM, supporting virtual sensing, fault diagnosis and predictive control.

[0288] The application scenarios of the above method include: Digital twin of sewage treatment: building virtual reactors for plants that lack historical data. Sensor failure simulation: generating alternative data when ORP or pH is abnormal. Control strategy optimization: pre-training MPC controllers to shorten the on-site commissioning cycle.

[0289] Through mechanism modeling and dynamic noise injection, the present invention can generate high-quality, multi-parameter coupled training data sets without actual data, breaking through the limitations of traditional data-driven methods and providing core technical support for the rapid deployment of intelligent water systems. The generated data sets will contain rich dynamic information, enabling PINN to learn the physical nature of the system and achieve accurate prediction and control in practical applications.

[0290] The following are the steps for preprocessing sensor data and inputting it into the PINN model in this embodiment: A mapping relationship between turbidity (NTU) and catalyst concentration (mg / L) is established through laboratory calibration.

[0291] Real-time preprocessing process:

[0292] 1. Streaming data reception: Sensor data enters the processing queue in chronological order.

[0293] 2. Sliding window filtering: For each new data point, update the Kalman filter state.

[0294] 3. Dynamic standardization: Calculate the standardized value in real time based on the mean and standard deviation of historical data.

[0295] 4. Feature splicing: Splice the current moment data with the past \(N\) steps of historical data as the time window input.

[0296] Top-level PINN design and training

[0297] This embodiment adopts a PINN training method based on multi-physical field coupling and dynamic constraints, and proposes a dynamic phased training framework of physical information neural network (PINN). By integrating the multi-parameter coupling mechanism model and the adaptive weight adjustment mechanism, high-precision modeling and prediction of the sewage treatment system are achieved. The core innovations include:

[0298] 1. Hybrid loss function architecture: jointly optimize data-driven error and physical conservation residual, and embed joint differential equation constraints of COD, pH, temperature, and ORP.

[0299] 2. Time-space divide-and-conquer training strategy: Design a phased training process for fast reactions (COD degradation) and slow processes (temperature accumulation) to improve convergence efficiency.

[0300] 3. Virtual-reality fusion transfer learning: Use synthetic data for pre-training and fine-tune with a small amount of real data to break through the gap between simulation and reality.

[0301] PINN model architecture design:

[0302] (1) Network structure

[0303] -Input layer: historical time window data ([tT w :t]), including COD, pH, temperature, ORP and operating variables (valve opening, pump speed, UV power).

[0304] -Feature extraction layer:

[0305] -Use bidirectional LSTM to capture temporal dependencies and output hidden states (h t ).

[0306] -Innovation: Introducing the Temporal Convolutional Network (TCN) to enhance local feature extraction capabilities.

[0307] -Physical Constraint Layer:

[0308] -Write(h t ) is input into multiple sub-networks to predict the future (N p ) step COD, pH, temperature, and ORP.

[0309] - Innovation: Design of coupled residual connections to force sub-networks to share core features of photocatalytic reactions (such as free radical concentration).

[0310] (2) Output definition

[0311]

[0312] Hybrid loss function construction:

[0313] (1) Data-driven loss

[0314]

[0315] (2) Physical drive loss

[0316] Based on the differential equation in step 1, calculate the physical residual of the predicted value:

[0317]

[0318] A gradient enhancement mechanism is introduced to calculate high-order derivatives through automatic differentiation to strengthen the second law of thermodynamics (such as entropy increase constraint).

[0319] (3) Dynamically weighted total loss

[0320]

[0321] Propose an adaptive weight regulator:

[0322] λ data =1-λ phy

[0323] Dynamically balance the weights of data fitting and physical constraints to avoid model divergence due to excessive physical residuals in the early stages of training.

[0324] Time and space divide and conquer training strategy:

[0325] (1) Phase 1: Pre-training of the rapid response process

[0326] -Objective: Prioritize the study of transient response of COD and ORP.

[0327] -Data screening: Select samples with COD change rate (|dCOD / dt|>5mg / L / min).

[0328] -Loss weight: Increase the weight of the physical residual term of COD and ORP (λ phy×2 ).

[0329] (2) Phase 2: Slow process fine-tuning

[0330] -Objective: To optimize long-term predictions of temperature and pH.

[0331] -Data screening: Select samples with temperature change rate (dT / dt<0.5℃ / min).

[0332] -Loss weight: Increase the weight of the temperature conservation equation and introduce energy constraint terms:

[0333]

[0334] (3) Phase 3: Joint Optimization of All Parameters

[0335] -Objective: Global fine-tuning to improve the prediction accuracy of multi-parameter coupling.

[0336] -Innovation: Curriculum Learning is used to gradually increase the time window length (T w ) and prediction step size (N p ).

[0337] Virtual-Real Fusion Transfer Learning:

[0338] (1) Synthetic Data Pre-training

[0339] -Use the generated high-quality dataset to train the initial PINN and learn the physical laws in the mechanism model.

[0340] (2) Fine-tuning with real data

[0341] - Collect a small amount of actual sewage plant data (such as 1% synthetic data volume) and use domain adaptive algorithms (such as DANN)

[0342] Align feature distribution:

[0343]

[0344] Residual adversarial training is designed to force the synthesis to be consistent with the physical residual distribution of real data.

[0345] Model verification and patent advantages:

[0346] (1) Performance verification indicators

[0347] - Prediction accuracy: Mean squared error (MSE) and mean absolute percentage error (MAPE) on the test set.

[0348] - Physical consistency: Proportion of energy conservation and reaction rate constraints violated (must be less than 1%).

[0349] (2)Technological advantages

[0350] 1. High-precision prediction: Through the hybrid loss and time-space divide-and-conquer strategy, the COD prediction error is reduced by more than 30%.

[0351] 2. Strong generalization: Virtual-reality fusion transfer learning reduces the generalization error of the model in real scenarios by 50%.

[0352] 3. Real-time performance: The TCN-LSTM architecture supports millisecond-level multi-step prediction to meet online control requirements.

[0353] (3) Application scenarios

[0354] -Intelligent sewage treatment control system: integrated PINN to achieve real-time virtual sensing of COD and ORP.

[0355] -Fault diagnosis engine: Detects sensor anomalies and process deviations through physical residual analysis.

[0356] -Digital Twin Platform: Provides zero-sample modeling and optimization solutions for new sewage treatment plants.

[0357] The present invention solves the three major problems faced by traditional PINN in complex sewage treatment systems, namely, difficulty in multi-parameter coupling modeling, physical constraint conflicts, and weak migration from simulation to reality, through innovative hybrid loss functions, time-space divide-and-conquer training strategies, and virtual-reality fusion transfer learning, providing core technical support for autonomous decision-making of intelligent water systems.

[0358] Low-level MPC optimization design

[0359] In model predictive control (MPC), the core of solving the operating variables is to minimize the objective function under the premise of satisfying the dynamic model and constraints through an online optimization algorithm. The following is the specific solution process and key methods for the photocatalytic wastewater reactor:

[0360] 1. Specific solution steps of photocatalytic reactor

[0361] 1.1 Prediction model call

[0362] 1. Input current status: real-time sensor data (pH, ORP, temperature, turbidity, COD) and operating variables.

[0363] 2. PINN prediction: Call the trained PINN model to predict the future (N p )-step multi-parameter prediction values.

[0364] -PINN input: current and historical data (pH, ORP, temperature, COD, valve opening, pump speed, UV power).

[0365] -PINN output: pH, ORP, T, COD.

[0366] The turbidity prediction value is directly calculated by mathematical equations, considering the limited computing power of the PINN model in actual operation:

[0367]

[0368] ογ: flocculant sedimentation coefficient, related to the type of agent;

[0369] οη: hydraulic retention time influencing factor;

[0370] οκ: UV oxidation decomposition efficiency of suspended matter.

[0371] 1.2 Optimization Problem Construction

[0372] Objective function (sorted by weight priority):

[0373] J=w 1 (COD-COD ref ) 2 +w 2 (pH-pH ref ) 2 +w 3 (Turbidity-Turbidity ref ) 2 +w 4 (Δu) 2

[0374] -Constraints:

[0375] - Valve opening: 20% ≤ opening ≤ 80% (to prevent blockage or excessive flow rate)

[0376] -UV power change rate: ≤50W / min (avoid thermal shock)

[0377] - Turbidity threshold: turbidity ≤ 50NTU (to prevent catalyst loss)

[0378] 1.3 Numerical Optimization Solution

[0379] 1. Initialization: Generate an initial guess value (optimal solution in the previous cycle) based on the current operating variables and states.

[0380] 2. Iterative solution:

[0381] -Using IPOPT based on CasADi, it supports fast nonlinear optimization and iteratively adjusts the operating variables in the prediction time domain to minimize the objective function.

[0382] - The objective function gradient needs to be calculated at each iteration (via automatic differentiation or finite differences).

[0383] 3. Output results: Take the first step operation variable in the optimization sequence (rolling time domain characteristics of MPC).

[0384] 4. Example: Solution Process Decomposition

[0385] Assume the current status is:

[0386] - pH = 7.2, ORP = 450mV, temperature = 40°C, COD = 180mg / L (real-time sensor data or PINN prediction), turbidity = 35NTU (real-time sensor data or mathematical equation calculation)

[0387] - Target: COD reduced to 50mg / L, turbidity ≤50NTU

[0388] Operating variables: valve opening, pump speed, UV power

[0389] Step 1: Define candidate action variable sequences

[0390] Assume that the control domain N c =3, the optimization variables are:

[0391] U=[u k ,u k+1 ,u k+2 ]

[0392] _Each u i Includes valve opening, pump speed, and UV power.

[0393] Step 2: Predict future outputs via PINN

[0394] For each candidate U, call the PINN model to predict future COD and turbidity:

[0395] For example, if the candidate operating variables are U = [70%, 2500 rpm, 400 W], it is predicted that COD will drop to 60 mg / L and turbidity will rise to 48 NTU within 20 minutes.

[0396] -Minimize the objective function:

[0397] J=10(COD-50) 2 +5(pH-7.5) 2 +8(Turbidity -30) 2 +2(ΔUV power) 2 _

[0398] Step 4: Solve and select the optimal operating variables

[0399] The optimizer tries multiple candidate Us and finally selects the operation variable that minimizes J and satisfies the constraints:

[0400] ●

[0401] The essence of solving the operating variables in MPC is to find the optimal control action under the dynamic model and constraints through numerical optimization. For photocatalytic reactors, it is necessary to combine the nonlinear PINN model and turbidity constraints, select an efficient solver (such as IPOPT), and balance the calculation speed and accuracy. In actual deployment, it is recommended to first verify the solution stability through simulation, and then gradually migrate to industrial hardware.

[0402] Example:

[0403] The control system of the present invention adopts hierarchical control in the system control architecture: the bottom layer MPC, combined with the top layer PINN prediction model, forms a two-layer architecture of "network prediction + classical control" to improve industrial feasibility and dynamic tracking capabilities. In this embodiment, the control system is divided into two layers: the top layer (high layer): based on the physical information neural network (PINN) to predict future working conditions, combined with necessary optimization or strategy generation. The bottom layer (low layer): model predictive control (MPC), closed-loop regulation of process variables at a higher frequency (seconds to minutes). Figure 3 It is a control logic flow chart.

[0404] PINN prediction model input and output

[0405] Input (at time t):

[0406] Multidimensional time series such as pH (t-1..t-10), ORP (t-1..t-10), flow (t-1..t-10), COD (t-1..t-10), temperature (t-1..t-10) in the historical window;

[0407] Current operating status: valve opening, dosing pump speed, UV lamp power;

[0408] Mechanistic equation information (acid-base balance, redox kinetics) is written into the loss function when training PINN.

[0409] Output:

[0410] Predicted values ​​of pH(t+k), COD(t+k), ORP(t+k), flow(t+k), and temperature(t+k) at each time point in the next 30 minutes (t+1..t+30);

[0411] Additional product concentrations, intermediate reaction rates, etc. may also be given (optional).

[0412] The main mechanism equations include:

[0413] pH-Alkalinity-CO 2 Balance: pH = -log10([H + ]), and [H + ][HCO 3 - ]≈K 1 constant);

[0414] ORP - redox pair: E = E° - (RT / F) ln (Q_redox);

[0415] COD degradation kinetics: First-order approximation: dCOD / dt = -k·COD;

[0416] Flow conservation / steady-state control (small difference in inflow vs. outflow).

[0417] In the PINN training phase, the loss function is defined as:

[0418] L_total=w_data·L_data+w_phys·L_phys

[0419] L_data: mean square error of the difference with historical sensor data;

[0420] L_phys: The sum of squares of the residuals of the mechanism (e.g. pH-alkalinity balance, ORP balance, COD degradation) at the sampling point.

[0421] Output instructions to MPC

[0422] The top level runs every 15 minutes:

[0423] Predict pH (t+1..t+30), COD (t+1..t+30), etc. in the next 30 minutes;

[0424] Generate target trajectory / interval based on processing requirements or economic goals:

[0425] pH_ref(t+k), ORP upper and lower limits: ORP_min / max(t+k), COD upper limit COD_lim(t+k);

[0426] Add energy consumption constraints, such as: "The air pump opening in the next stage shall not exceed 85%."

[0427] This information is packaged and sent to the underlying MPC as control reference values ​​or soft constraint upper / lower limits.

[0428] In this embodiment, the specific MPC parameter values ​​are as follows:

[0429] Sampling period Ts = 1 minute

[0430] Prediction time domain Np = 15 steps (15 minutes)

[0431] Control time domain Nc = 5 steps (5 minutes), then keep the control input constant

[0432] State space model (linear approximation):

[0433] X(k+1)=AX(k)+Bu(k),y(k)=CX(k)

[0434] The input u(k) includes [valve opening, UV lamp power], and the output y(k) includes [COD outlet value, pH outlet value]

[0435] constraint:

[0436] 0≤Valve opening≤100%

[0437] 0≤UV lamp power≤5kW

[0438] Export pH ∈ [6.5, 8.5] (soft constraint)

[0439] Objective function:

[0440] J=Σ(COD_error 2 )+10×(pH_error 2 )+0.5×(Δu_valve 2 )+1.0×(ΔP_UV 2 )+0.2×(P_UV 2 )

[0441] Here, we increase the pH_error coefficient (10) to indicate that it is more sensitive to the pH constraint; we also increase the UV power by 0.2×(P_UV 2 )Punishment to save energy.

[0442] MPC Controller

[0443] Status and input and output definitions:

[0444] ·Input (Manipulated Variables, MV):

[0445] u 1 (k): Water inlet valve opening (determines inlet flow)

[0446] u 2 (k): Alkalinity dosing pump speed

[0447] u 3 (k): Aeration volume or UV lamp power (choose one, example)

[0448] Output (Controlled Variables, CV):

[0449] y 1 (k) = pH (k) (measurement of effluent pH)

[0450] y 2 (k) = COD (k) (outlet COD)

[0451] y 3 (k)=ORP(k)(outlet ORP)

[0452] Sampling period Ts = 1 minute.

[0453] Dynamic Model (Discrete State Space Example)

[0454] by represents the internal state vector (including pH related states, COD related states, ORP dynamic states reduced-order variables), then:

[0455] X(k+1)=A·X(k)+B·u(k)

[0456] y(k)=C·X(k)+D·u(k)

[0457] Wherein: k represents the kth sampling, time t = k·Ts; A, B, C, D are matrices (obtained by linearization or subspace identification), and the specific dimensions and parameters are as follows: A is a 6×6 matrix (n=6), B is a 6×3 matrix (3 control inputs), C is a 3×6 matrix (output 3 dimensions), and D is a 3×3 matrix.

[0458] MPC time domain configuration:

[0459] The prediction time domain Np=20, that is, the prediction is 20 steps, each step is 1 minute, a total of 20 minutes.

[0460] The control time domain Nc=5, each step has independent control input in the first 5 steps (5 minutes), and then the control amount of the 5th step is kept unchanged until the next rolling optimization.

[0461] constraint

[0462] ·Manipulated variables:0≤u 1 (k)≤100%(valve opening), 0≤u 2 (k)≤50(dosing pump speed, unit L / min), 0≤u 3 (k)≤1.0 (UV or aeration ratio, 0~1);

[0463] Controlled variables: 6.5≤pH(k)≤8.5, COD(k)≤50mg / L, ORP(k)≥100mV (example)

[0464] Operation increment: |Δu i (k)|≤0.1 (limit action smoothness), (if the top layer gives a stricter soft constraint, such as "UV≤0.8", u can be changed online at this time 3 (k)).

[0465] MPC cost function (objective function):

[0466] make is the control quantity of the jth step in the prediction time domain, y (k+j|k)is the predicted output. Definition:

[0467]

[0468] Where: y_ ref(k+j) is the target trajectory set by the top PINN (egpH_ ref =7.2, COD_ ref =30...); Indicates the change of control quantity; W 1 , W 2 , W 3 is the weight matrix (or diagonal).

[0469] Specific coefficients (example):

[0470] ·W 1 =diag([10,8,5]) # pH error weight = 10, COD error = 8, ORP error = 5

[0471] ·W 2 =diag([1, 1, 1]) # Penalty for operation increment

[0472] ·W 3 =diag([0.5, 0.3, 1]) # absolute value of operation (energy consumption, drug consumption and other penalties).

[0473] In this way, the controller will prioritize ensuring that the pH meets the target (because it has a greater weight), while also taking into account that COD and ORP do not deviate, and then taking into account smooth operating volume and minimized energy consumption.

[0474] MPC algorithm process:

[0475] At time k: Get the measured output y(k) and estimate the state X (k) (If Kalman filtering is needed to estimate the hidden state); read the last control value Form a rolling optimization problem: min J subject to dynamic equation + operation constraint + output constraint, solve Execute the first control input In the real process, the remaining input plan is kept until the next rolling update. k = k + 1, and the cycle repeats.

[0476] The controller adopts a hierarchical control strategy: the top layer generates a reference trajectory of valve opening based on a turbidity prediction model, and the bottom layer optimizes the opening range in real time through model predictive control (MPC).

[0477] (1) Calculating the catalyst loss rate of the sedimentation tank (29) according to the feedback value of the second turbidity sensor (25), and if the loss rate is greater than 5%, triggering the opening of the third regulating valve (V3) to increase by 10%-15%;

[0478] (2) When the COD value of the sedimentation tank (29) exceeds a set threshold, the controller (37) simultaneously increases the opening of the first regulating valve (V1) to increase the processing flow rate, and decreases the opening of the valve II (31) to reduce the backflow disturbance;

[0479] (3) The opening adjustment of each valve must meet the constraints: the opening range of V1 is 20%-80%, the opening range of V3 is 30%-90%, the opening of V2 / V4 is positively correlated with the turbidity value, and the change rate is ≤5% / minute.

[0480] Two-layer interaction mechanism:

[0481] References / restrictions issued by PINN to MPC:

[0482] PINN cycle: triggers a prediction every 5 minutes or 15 minutes;

[0483] Based on the prediction of pH, ORP, COD, flow rate, etc. in the next 20 to 60 minutes, if the inlet water quality is found to be deteriorating (COD rising) or pH is about to become acidic, PINN will: adjust pH_ ref ,COD_ lim , ORP_ min or to u 3 (Aeration / UV power) gives a new upper bound; pass this information to MPC.

[0484] MPC Response:

[0485] ·MPC will check the newly issued target (such as "pH_ ref =7.2","COD_ lim =40mg / L”) for rolling optimization;

[0486] If the top layer requires “UV power not to exceed 0.8”, MPC will use this as a hard constraint: 3 (k)≤0.8;

[0487] Generate final valve opening, dosing pump speed and other operating variables to avoid over-adjustment or loss of control.

[0488] Due to the advanced nature of the technical solution, it can be widely used in many fields such as sewage treatment, catalytic reactor design and powder catalyst application. First, in the field of sewage treatment, the technical solution can effectively reduce the loss rate of powder catalyst by adopting turbidity sensors and valves, thereby improving the efficiency and quality of sewage treatment. At the same time, since this method does not involve physical capture and removal processes, it will not affect the activity of the catalyst, ensuring the stability and sustainability of the catalytic effect. In addition, since this method has low operating and maintenance costs, it can greatly reduce the overall cost of sewage treatment and improve the competitiveness of the sewage treatment industry. Secondly, in the field of catalytic reactor design, the technical solution provides a new design idea, that is, to improve the performance and service life of the catalytic reactor by monitoring and controlling the loss rate of the catalyst. This method can not only be applied to powder catalysts, but also can be promoted to other types of catalysts, such as granular catalysts. This will promote the development of the field of catalytic reactor design and promote the development and application of new and efficient catalytic reactors. Finally, in the field of powder catalyst application, the technical solution provides a new solution for the application of powder catalysts. Since the loss problem of powder catalysts has always been a major challenge in this field, the application of this technical solution will greatly promote the promotion and application of powder catalysts. This will promote technological progress in related fields and meet the market demand for efficient, stable and economical sewage treatment solutions. In general, this technical solution has broad application prospects and large market demand, and is expected to play an important role in sewage treatment, catalytic reactor design and powder catalyst application.

[0489] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution, which should be included in the scope of the claims of the present invention.

Claims

1. A catalyst monitoring system based on sensor linkage, characterized in that: The system comprises a mixing tank (2), a settling tank (29) and a reflow tank (33) which are arranged in a pipeline circulation manner. A sewage inlet pipe (1) is provided at the bottom of the mixing tank (2). An inner cylinder (3) extending downward from the top of the mixing tank (2) and having a bottom opening is provided in the mixing tank (2). A grid array filter plate (5) is provided in the upper part of the mixing tank (2) and between the inner cylinder (3). It also includes a strong oxidant dosing device (10), a powder catalyst dosing device (13) and a controller (37), wherein the strong oxidant dosing device (10) is connected to the inner cylinder (3) via a strong oxidant metering pump (11), and the powder catalyst dosing device (13) is connected to the inner cylinder (3) via a powder catalyst metering pump (12); The strong oxidant metering pump (11) and the powder catalyst metering pump (12) are both electrically connected to the controller (37).

2. The catalyst monitoring system based on sensor linkage according to claim 1, characterized in that: The inner cylinder (3) is provided with a side wall channel (6) at an upper portion higher than the grid array filter plate (5), and the side wall channel (6) and the bottom opening of the inner cylinder (3) are both provided with a filter screen (4); the mixing tank (2) is provided with a slurry pushing and stirring device (7) extending into the inner cylinder (3) and electrically connected to the controller (37); The mixing tank (2) is also provided with a purple light lamp (15) and a microwave generator (16) electrically connected to the controller (37); The grid array filter plate (5) adopts a single-layer hexagonal honeycomb grid in a multi-level staggered stacking structure; The reflow tank (33) is provided with an agitator (34); the powder catalyst dosing device (13) is also connected to the reflow tank (33) via a powder catalyst replenishing pump (14), and the agitator (34) and the powder catalyst replenishing pump (14) are both electrically connected to the controller (37); a diffuser (36) connected to the reflow tank (33) is provided in the mixing tank (2); The sedimentation tank (29) is provided with a barrel-shaped tube (27) and a conical bucket (28) in sequence from top to bottom, and the bottom of the sedimentation tank (29) is in the shape of a cone. The sedimentation tank (29) is provided with a pulse microwave processor (24) extending into the barrel-shaped tube (27) and electrically connected to the controller (37).

3. The catalyst monitoring system based on sensor linkage according to claim 1, characterized in that: A valve I (17) and a pump I (18) are provided on the pipeline between the mixing tank (2) and the settling tank (29); a valve II (31) and a pump II (32) are provided on the pipeline between the settling tank (29) and the reflow tank (33); a pump III (35) is provided on the pipeline between the reflow tank (33) and the mixing tank (2); a barrel-shaped pipe (27) at the upper part of the settling tank (29) is connected to a drainage pipe (C) with a valve III (26); a cone at the bottom of the settling tank (29) is connected to a discharge pipe (D) with a valve IV (30); and the valve I (17), valve II (31), valve III (26), valve IV (30), pump I (18), pump II (32) and pump III (35) are all electrically connected to the controller (37); The mixing tank (2) is provided with an ORP sensor (8) and a COD sensor (9); the water inlet pipe (A) of the settling tank (29) is provided with a flow meter (19); a differential pressure sensor (20) is provided between the water inlet pipe (A) and the water outlet pipe (B) of the settling tank (29); a temperature sensor (21), a pH value sensor (22) and a first turbidity sensor (23) are provided on the settling tank (29) and outside the barrel-shaped pipe (27); a second turbidity sensor (25) is provided on the settling tank (29) and inside the barrel-shaped pipe (27), and the ORP sensor (8), COD sensor (9), flow meter (19), differential pressure sensor (20), temperature sensor (21), pH value sensor (22), first turbidity sensor (23) and second turbidity sensor (25) are all electrically connected to the controller (37).

4. A control method for a catalyst monitoring system according to any one of claims 1 to 3, characterized in that: The controller control architecture adopts hierarchical control: the bottom layer MPC is combined with the top layer PINN prediction model to form a two-layer architecture of "network prediction + classical control", which includes the following steps: S1, sewage enters the reactor, the sensor collects real-time data, and the controller pre-processes the collected data; S2. Predict the temperature, COD, pH, and ORP water quality indicators in the reactor in the future through the PINN prediction model; S3. The controller generates dynamic set values ​​or soft constraints based on the prediction results, including but not limited to: UV lamp power upper limit, target COD level, and valve optimal opening range; S4. The controller performs rolling optimization in a fast cycle of seconds to minutes, and enables the controller to control valve opening, UV power, and catalyst addition variables in real time to ensure that the effluent quality meets the standard and energy consumption is minimized; during the operation of the system, a safety protection mechanism is retained: if the top-level prediction model or communication is abnormal, MPC or traditional PID is relied on to maintain basic operation; S5. Conduct effluent testing and feedback, and ultimately achieve concentration monitoring and COD monitoring.

5. The control method according to claim 4, characterized in that: In step S2, the water quality index is predicted at intervals based on the physical information neural network PINN prediction model. When training the prediction model, the multi-parameter dynamic process of chemical oxygen demand COD, pH, temperature T and oxidation-reduction potential ORP in the sewage treatment system is modeled through physical and chemical mechanisms; Construct a multi-parameter coupling mechanism model; The training data collection methods include: laboratory simulation data: building a small catalyst system in the laboratory to simulate catalyst loss under different operating conditions; historical operation data: using historical operation data and catalyst loss records of existing factories; instrument measurement data: using a special sensor system to directly measure catalyst concentration, flow rate and other parameters in the actual production environment.

6. The control method according to claim 5, characterized in that: The physical constraint equations of the PINN prediction model specifically include the following: 1) Construct a multi-parameter coupling mechanism model: Chemical oxygen demand (COD) degradation kinetic model: pH dynamic model: Temperature T dynamic model: Oxidation Reduction Potential (ORP) Dynamic Model 2) Define control variables and random disturbances: Manipulate variable scope: Valve opening V valve ∈[0,1], controls the water flow rate Q in =10V valve Dosing pump speed Q pump ∈[20,100]mL / min, adjust pH neutralization rate UV power P UV ∈[200,600]W, driving the photocatalytic reaction; Random perturbation design: Influent water quality disturbance: Dynamic Noise Injection: CODE meas =COD+N(0,σ COD ),σ COD =a·COD ORP meas =ORP+N(0,σ ORp ),s ORP =b+c·|ORP|.

7. The control method according to claim 6, characterized in that: The training of the physical information neural network PINN in this method adopts a dynamic staged training framework, which specifically includes: 1) Model architecture: Network structure: Input layer: historical time window data ([t-Tw:t]), including COD, pH, temperature, ORP and operating variables. The operating variables include valve opening, pump speed, and UV power. Feature extraction layer: Use bidirectional LSTM to capture temporal dependencies and output hidden state ht; Physical constraint layer: ht is input into multiple sub-networks to predict COD, pH, temperature, and ORP for the next Np steps respectively; Output definition: 2) Construction of hybrid loss function: Data-driven losses: Physics-driven loss: Calculates the physical residual of the predicted value: Dynamically weighted total loss: Adaptive Weight Regulator: Dynamically balance the weights of data fitting and physical constraints to avoid model divergence due to excessive physical residuals in the early stages of training; 3) Time and space divide and conquer training strategy: Phase 1: Pre-training of the rapid response process: Objective: Prioritize learning the transient response of COD and ORP; Data screening: select samples with COD change rate (|dCOD / dt|>5mg / L / min); Loss weight: Increase the weight of the physical residual term of COD and ORP (λphy×2); Phase 2: Slow process fine-tuning: Objective: To optimize long-term predictions of temperature and pH; Data screening: select samples with temperature change rate (dT / dt<0.5℃ / min); Loss weight: Increase the weight of the temperature conservation equation and introduce energy constraint terms: Stage 3: Joint optimization of all parameters: Objective: Global fine-tuning to improve the prediction accuracy of multi-parameter coupling; 4) Virtual-Real Fusion Transfer Learning: Synthetic data pre-training: Use the generated high-quality dataset to train the initial PINN and learn the physical laws in the mechanism model; Fine-tuning with real data: Collect a small amount of actual operating condition data and align feature distribution through domain adaptive algorithms:

8. The control method according to claim 7, characterized in that: In this method, after the PINN model is deployed, an online learning update mechanism is used to ensure the long-term validity of the model. The online learning update mechanism includes: Incremental data collection strategy: 1) Real-time data caching: Establish a real-time data buffer to store high-frequency data from the last 7-30 days; Realize automatic labeling of data; Set up an automatic data quality assessment mechanism; 2) Key event trigger collection: Increase the frequency of data collection when operating conditions change; Save complete process data when abnormal loss events occur; After equipment maintenance, focus on collecting initial operating data; 3) Directed experimental data supplement: Regularly conduct small-scale adjustments to supplement training data for specific areas; Test system response under boundary conditions within safety limits; Conduct targeted data collection in areas where model predictions are weak; Model update strategy: 1) Regular update mechanism: Set a fixed period to update model parameters; Use the sliding window method to select the data of the last month for retraining; Keep historical version models for comparison and rollback; 2) Event-triggered updates: Trigger model update when the prediction error exceeds the threshold; Mandatory model update after equipment maintenance or catalyst replacement; Targeted updates after changes in processes or operating methods; 3) Incremental learning method: Use transfer learning to preserve learned physical laws; Adopt elastic weight update strategy to give different weights to new data and historical data; Implement a partial parameter freezing mechanism to update only the parameters that are sensitive to new operating conditions.

9. The control method according to claim 8, characterized in that: The algorithm flow of the MPC controller includes: 1) Prediction model call: Input current state: real-time sensor data and operating variables; PINN prediction: Call the trained PINN model to predict the multi-parameter prediction values ​​for the next (Np) steps; PINN input: Current and historical data; PINN output: pH, ORP, T, COD; 2) Optimization problem construction: objective function: J=w1(COD-COD ref ) 2 +w2(pH-pH ref ) 2 +w3(turbidity-turbidity ref ) 2 +w4(△u) 2 Constraints: Valve opening: 20% ≤ opening ≤ 80%, UV power change rate: ≤ 50W / min, turbidity threshold: turbidity ≤ 50NTU; 3) Numerical optimization solution: Initialization: Generate initial guess values ​​based on current operating variables and states; Iterative solution: IPOPT based on CasADi is used to support fast nonlinear optimization, and the operating variables are iteratively adjusted in the prediction time domain to minimize the objective function; the objective function gradient needs to be calculated for each iteration; Output result: Take the first step operation variable in the optimization sequence.

10. The control method according to claim 9, characterized in that: In this method, a dynamic gradient tracking and contribution quantification method is adopted, including: 1) Interpretability layer architecture design Gradient separation module: In the loss function, the residual term is defined separately for each physical equation to construct a structured loss function: Gradient backpropagation separation: Extract the gradient tensor of each physical equation through automatic differentiation; 2) Real-time contribution quantification indicators Calculate the contribution of each physical equation's gradient to the total gradient: Dynamically adjust weights based on the proportion of physical residual in the total loss: Used to evaluate the strength of constraints imposed by each physical equation on model predictions.

Images