Membrane process denitration automatic control system and control method based on AI prediction

Abstract

The invention provides a membrane process denitration automatic control system and method based on AI prediction. The membrane process denitration automatic control system comprises a sensor unit, an execution mechanism unit, a control and monitoring unit and an AI prediction hardware unit. The control and monitoring unit comprises a PLC (Programmable Logic Controller), a man-machine touch screen and an upper computer; the AI prediction hardware unit comprises an AI edge computing node, a data storage module and a communication module; the sensor unit is connected with the PLC; the execution mechanism unit is connected with the PLC; the PLC is respectively communicated with the man-machine touch screen and the upper computer; the AI edge computing node is bidirectionally connected with the PLC through the communication module; the AI edge computing node obtains real-time operation data from the PLC controller, stores the real-time operation data in the data storage module, generates a prediction result and a control instruction through model reasoning, and sends the prediction result and the control instruction to the PLC controller for execution. The problems that an existing membrane method denitration system is complex in operation, low in parameter control precision, lack of safety interlocking and lagging in membrane cleaning are solved.

Description

Technical Field

[0001] This invention belongs to the field of monitoring system technology, specifically relating to an AI-predictive membrane denitrification automatic control system and control method. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Membrane denitrification technology is a key process for industrial wastewater denitrification and brine purification in the chlor-alkali industry. However, it currently relies heavily on manual operation. System start-up and shutdown require manual step-by-step operations such as opening pretreatment valves, adjusting pump frequency, and starting the membrane system. Errors in this sequence can easily cause hydraulic shock and physical damage to the membrane elements. Simultaneously, traditional systems lack precision in controlling operating parameters, relying primarily on manual experience to adjust valve openings and pump power. This makes it difficult to achieve precise and stable control of key process indicators such as in-membrane pressure and recovery rate, leading to increased concentration polarization and accelerated membrane fouling. The resulting decrease in system permeate efficiency and shortened membrane element replacement cycles severely restrict the system's economic efficiency and treatment effectiveness.

[0004] The existing system lacks a comprehensive safety interlock mechanism for overpressure and water quality exceeding standards. When the pressure before the membrane exceeds the rated limit or the nitrate nitrogen concentration in the permeate is abnormal, it cannot automatically perform emergency operations such as pressure relief, backflow, or shutdown, creating a dual risk of equipment failure and excessive effluent discharge. In addition, the membrane cleaning process mostly adopts a passive strategy based on fixed time or fixed pressure difference, failing to dynamically adjust according to actual operating conditions such as fluctuations in influent water quality and changes in operating load. This leads to "over-cleaning" or "untimely cleaning," resulting in waste of chemical agents and irreversible decline in membrane flux. Summary of the Invention

[0005] To address the aforementioned problems, this invention provides an AI-predictive-based automatic control system and method for membrane denitrification, which solves the problems of cumbersome manual operation, low parameter control accuracy, lack of safety interlocks, and delayed membrane cleaning in existing membrane denitrification systems.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solution: In a first aspect, the present invention provides an AI-predictive-based automatic control system for membrane denitrification, comprising: a sensor unit, an actuator unit, a control and monitoring unit, and an AI prediction hardware unit; the sensor unit includes basic sensors and AI prediction supplementary sensors; the actuator unit includes basic actuators and actuators equipped with AI control interfaces; the control and monitoring unit includes a PLC controller, a human-machine interface screen, and a host computer; the AI ​​prediction hardware unit includes an AI edge computing node, a data storage module, and a communication module. The sensor unit is connected to the PLC controller; the actuator unit is connected to the PLC controller; the PLC controller communicates with the human-machine touch screen and the host computer; the AI ​​edge computing node is bidirectionally connected to the PLC controller through the communication module; the AI ​​edge computing node obtains real-time operating data from the PLC controller, stores it in the data storage module, and generates prediction results and control instructions through model inference, which are then sent to the PLC controller for execution.

[0007] As a further implementation, the basic sensors consist of a pressure sensor, a flow sensor, a water quality sensor, and a level sensor; the AI ​​predictive supplementation sensor consists of an influent turbidity sensor and an influent temperature sensor; the pressure sensor is installed at the inlet and outlet of the membrane module and the inlet and outlet of the security filter; the flow sensor is installed on the raw water inlet, membrane system permeate, and concentrate pipelines; the water quality sensors include a pH sensor installed on the pretreated effluent pipeline, a conductivity sensor installed on the permeate pipeline, and an ORP sensor installed at the inlet of the security filter; the level sensor is installed in the raw water tank, permeate tank, and chemical tank; the influent turbidity sensor is installed on the raw water inlet pipeline, and the influent temperature sensor is installed on the pretreated effluent pipeline.

[0008] As a further implementation, the basic actuator consists of a variable frequency pump set, an electric regulating valve, and a solenoid valve; both the cleaning pump in the variable frequency pump set and the drug dosage regulating valve in the electric regulating valve are equipped with AI control signal interfaces to receive cleaning flow rate and drug concentration commands.

[0009] As a further implementation, the PLC controller is equipped with an AI prediction result receiving module, which is used to parse and execute AI instructions in the loop interruption tissue block; the human-machine touch screen is equipped with an AI prediction interface, which is used to display the future transmembrane pressure difference prediction curve, cleaning suggestions and risk level; the host computer is equipped with an AI prediction model management module, which supports model training triggering, accuracy evaluation and data drift detection.

[0010] As a further implementation, the AI ​​edge computing node has a built-in ARM architecture processor and NPU, supporting TensorFlow Lite model inference; the data storage module is used to store historical running data; and the communication module supports Profinet IRT and 5G dual-link communication.

[0011] Secondly, the present invention also provides a control method for an AI-predictive membrane denitrification automatic control system, comprising the following steps: Step 1: With the command of one-button start, the system starts the pretreatment unit, performs low-pressure flushing of the membrane system, pressurizes and runs in sequence according to the preset order, and links with the posttreatment unit until it enters normal operating condition; Step 2: During normal operation, the safety interlock logic is executed in real time to monitor the membrane pressure, water quality parameters and liquid level, and trigger protection actions when the parameters exceed the standard. At the same time, the AI ​​prediction hardware unit performs model inference based on the real-time collected data to predict future changes in transmembrane pressure difference and overpressure risk, and provides cleaning strategies through the control and monitoring unit or issues adjustment instructions in advance when overpressure risk is predicted. Step 3: By using the one-click stop command or the safety interlock to trigger the shutdown, the system executes the membrane system depressurization, flushing, evacuation and shutdown steps in a preset sequence, and optimizes the flushing parameters based on the AI ​​prediction results; if the shutdown is caused by contamination, the cleaning mode is automatically triggered and the cleaning procedure is executed according to the cleaning strategy generated by AI.

[0012] As a further implementation, the cleaning strategy is generated differently based on the level of contamination: When the pollution level is low, delayed cleaning is performed during the system load off-peak period, using the first cleaning duration and the first reagent concentration. When the contamination level is medium, the standard cleaning process is triggered, including low-pressure rinsing, circulation of the first concentration of chemicals, and pure water rinsing, with the second cleaning duration. When the pollution level is high, an enhanced cleaning process is triggered, increasing the cleaning pump frequency and the second agent concentration, and extending the agent circulation time, thus adopting a third cleaning duration.

[0013] As a further implementation method, historical operating data covering different working conditions is collected, cleaned, labeled and feature-engineered to construct a training dataset containing time-series features and target labels; the training dataset is used to train the GRU time-series prediction model and the XGBoost risk classification model, and the accuracy is verified and the model is lightweighted to obtain the final combined AI prediction model.

[0014] As a further implementation, an AI-predicted overpressure interlock step is also included: when the AI ​​prediction model outputs a risk that the intermembrane pressure will exceed the rated value by 10% in the next 10 minutes, the PLC controller controls the pressure relief valve to open in advance and reduce the frequency of the high-pressure pump to prevent overpressure.

[0015] As a further implementation method, it also includes model operation and maintenance optimization steps: regularly using newly added real-time data to incrementally train and calibrate the AI ​​prediction model, and monitoring hardware status and data drift to ensure prediction accuracy and control stability.

[0016] Compared with the prior art, the advantages and positive effects of this invention are: This invention not only provides feedback control based on current measurements but also predicts future states using an AI-based predictive model, issuing control commands in advance for proactive feedback control, significantly improving the anticipation and accuracy of control. In membrane treatment processes, it precisely controls key parameters such as inlet pressure, recovery rate, and feed water pH with an error ≤5%, reducing membrane fouling and damage. The sensor unit, actuator unit, control and monitoring unit, and AI prediction hardware unit establish a safety interlock mechanism for overpressure, water quality exceeding standards, and abnormal liquid level, enabling automatic emergency handling of faults and reducing the risk of equipment failure and excessive emissions. The AI ​​prediction hardware unit can predict membrane fouling trends in advance, optimizing cleaning timing, reducing unnecessary cleaning frequency, and precisely controlling the cleaning mechanism to perform cleaning at the optimal time, rather than fixed-cycle cleaning, thereby reducing downtime, extending membrane life, and saving cleaning agents and energy consumption. Based on the predicted feed water load and water quality, pumps and valves are adjusted in advance to ensure the system always operates within its high-efficiency range, avoiding energy waste.

[0017] This invention transforms experience-based, step-by-step manual operations into precise, parallel execution controlled by a program, through one-button start-up, one-button shutdown, safety interlocks, and AI prediction and decision-making. This significantly improves operational efficiency and production flexibility, avoiding the risk of membrane damage due to human error. Safety interlocks and AI prediction and decision-making stabilize the control errors of parameters such as membrane inlet pressure, pH, and recovery rate within ≤5%. This effectively suppresses the acceleration of concentration polarization and membrane fouling, reducing the continuous operation failure rate of the system. Based on safety interlocks, it provides millisecond-level automatic emergency handling for faults such as overpressure, water quality exceeding standards, and abnormal liquid levels, and introduces an AI prediction hardware unit to provide early warning and intervention before faults occur. Through dynamic adjustment of the variable frequency pump unit and AI-predicted load adjustment, the system achieves fine-tuning, resulting in lower energy consumption of the variable frequency pump unit. The energy savings, combined with the aforementioned savings in operation and maintenance costs, significantly reduce the system's total lifecycle operating cost.

[0018] This invention triggers cleaning based on the level of contamination, enabling the matching of different energy consumption and material inputs to different contamination levels. For low-contamination conditions, cleaning is delayed until off-peak hours. By controlling the cleaning duration, pump frequency, and reagent concentration, energy is used efficiently and rationally, reducing operating costs. This tiered treatment avoids resource waste and membrane damage caused by heavy cleaning for lightly contaminated areas, reduces membrane replacement frequency, and prevents incomplete cleaning and continuous deterioration of membrane performance caused by light cleaning for heavily contaminated areas. Attached Figure Description

[0019] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0020] Figure 1This is a schematic diagram of the AI-predictive membrane denitrification automatic control system of the present invention; Figure 2 This is a schematic diagram of the safety interlocking and monitoring logic of the present invention.

[0021] In the diagram: 1. AI edge computing node; 2. Data storage module; 3. Communication module; 4. Sensor unit; 5. PLC controller; 6. Actuator unit; 7. Human-machine interface touch screen; 8. Host computer. Detailed Implementation

[0022] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0023] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, unless otherwise expressly indicated by the invention, the singular form is also intended to include the plural form. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof. Membrane denitrification: A process that uses membrane separation technologies such as nanofiltration (NF) and reverse osmosis (RO) to remove nitrate nitrogen and other impurities from water.

[0024] One-click start / stop: Through a single operation command, the membrane denitrification system can achieve fully automated control from start-up, operation to shutdown, without the need for manual intervention.

[0025] PLC / DCS control system: PLC (Programmable Logic Controller) or DCS (Distributed Control System) is used to integrate process logic, equipment linkage and safety interlocks, and is the core control unit to realize the automated operation of the system.

[0026] Transmembrane pressure difference: The pressure difference between the inlet and outlet of a membrane module is a key indicator for judging the degree of membrane fouling.

[0027] AI Prediction Hardware Unit: Based on deep learning algorithms, this functional module collects historical and real-time operational data to predict membrane fouling trends, changes in key parameters, and system failure risks, providing decision support for cleaning strategy optimization and safe operation.

[0028] GRU model: Gated Recurrent Unit, a deep learning model that excels at handling time-series data (such as transmembrane pressure difference changes). It controls information transmission through reset and update gates, avoiding the gradient vanishing problem in long-sequence data, and its prediction accuracy is better than that of the traditional LSTM model.

[0029] Data drift: The phenomenon in industrial scenarios where the distribution of sensor data changes over time (such as seasonal fluctuations in influent water quality or parameter shifts caused by membrane aging), leading to a decrease in the prediction accuracy of the original model.

[0030] Incremental training: A training method that updates the parameters of an existing model locally based on new data, without retraining the entire network, thus balancing model accuracy and training efficiency.

[0031] Example 1 This embodiment provides an AI-predictive-based automatic control system for membrane denitrification, such as... Figure 1As shown, it includes: a sensor unit 4, an actuator unit 6, a control and monitoring unit, and an AI prediction hardware unit; the sensor unit 4 includes basic sensors and AI prediction supplementary sensors for real-time acquisition of process parameters and water quality parameters; the actuator unit 6 includes basic actuators and actuators with AI control interfaces for executing flow, pressure, valve, and cleaning control; the control and monitoring unit includes a PLC controller 5, a human-machine interface 7, and a host computer 8 for realizing logic control, human-machine interaction, data storage, and monitoring; the AI ​​prediction hardware unit includes an AI edge computing node 1, a data storage module 2, and a communication module 3. The system includes a communication module for running AI prediction models and implementing AI prediction and decision-making; sensor unit 4 is connected to PLC controller 5 via wired or wireless means; actuator unit 6 is connected to PLC controller 5 via control signal lines; PLC controller 5 communicates with HMI touchscreen 7 and host computer 8 via industrial bus; AI edge computing node 1 is bidirectionally connected to PLC controller 5 via communication module 3; AI edge computing node 1 obtains real-time operating data from PLC controller 5 and stores it in data storage module 2, and generates prediction results and control commands through model inference, which are then sent to PLC controller 5 for execution. This system not only provides feedback control based on current measurements, but also predicts future states based on AI prediction models, issuing control commands in advance for early feedback control, greatly improving the anticipation and accuracy of control. In membrane treatment processes, precise control of key parameters such as inlet pressure, recovery rate, and feed water pH is maintained with an error of ≤5%, reducing membrane fouling and damage. Sensor unit 4, actuator unit 6, control and monitoring unit, and AI prediction hardware unit establish a safety interlock mechanism for overpressure, water quality exceeding standards, and abnormal liquid level, enabling automatic emergency handling of faults and reducing the risk of equipment failure and excessive emissions. The AI ​​prediction hardware unit can predict membrane fouling trends in advance, optimize cleaning timing, reduce unnecessary cleaning frequency, and precisely control the cleaning mechanism to perform cleaning at the optimal time, rather than fixed-cycle cleaning, thereby reducing downtime, extending membrane life, and saving cleaning agents and energy consumption. Based on predicted feed water load and water quality, pumps and valves are adjusted in advance to ensure the system always operates within its high-efficiency range, avoiding energy waste.

[0032] As a further implementation, the basic sensors consist of a pressure sensor, a flow sensor, a water quality sensor, and a level sensor; the AI ​​predictive supplementation sensor consists of an influent turbidity sensor and an influent temperature sensor; the pressure sensor is installed at the inlet and outlet of the membrane module and the inlet and outlet of the security filter; the flow sensor is installed on the raw water inlet, membrane system permeate, and concentrate pipelines; the water quality sensors include a pH sensor installed on the pretreated effluent pipeline, a conductivity sensor installed on the permeate pipeline, and an ORP sensor installed at the inlet of the security filter; the level sensor is installed in the raw water tank, permeate tank, and chemical tank; the influent turbidity sensor is installed on the raw water inlet pipeline, and the influent temperature sensor is installed on the pretreated effluent pipeline.

[0033] As a further implementation, the basic actuator consists of a variable frequency pump set, an electric regulating valve, and a solenoid valve. Both the cleaning pump in the variable frequency pump set and the dosage regulating valve in the electric regulating valve are equipped with AI control signal interfaces to receive cleaning flow rate and chemical concentration commands. The variable frequency pump set includes a raw water pump (30kW power, frequency range 20-50Hz), a high-pressure pump (75kW power, frequency range 30-60Hz), and a cleaning pump (15kW power), with speed controlled by a PLC output signal. The electric regulating valve includes a concentrate regulating valve (adjusting recovery rate, opening 0-100%), an inlet regulating valve (adjusting inlet water flow), and a dosage regulating valve (controlling cleaning chemical concentration), with a response time ≤1s. The solenoid valve includes a raw water inlet valve, a permeate discharge valve, a permeate return valve, a concentrate discharge valve, a cleaning liquid valve, a backwash inlet valve, a backwash drain valve, and a pressure relief valve, used for pipeline on / off control, with an action time ≤0.5s. A membrane system inlet valve is installed on the membrane system permeate pipeline.

[0034] As a further implementation, the PLC controller 5 is equipped with an AI prediction result receiving module, which is used to parse and execute AI instructions in the loop interruption tissue block; the human-machine touch screen 7 is equipped with an AI prediction interface, which is used to display the future transmembrane pressure difference prediction curve, cleaning suggestions and risk level; the host computer 8 is equipped with an AI prediction model management module, which supports model training triggering, accuracy evaluation and data drift detection.

[0035] As a further implementation, the AI ​​edge computing node 1 has a built-in ARM architecture processor and NPU, supporting TensorFlow Lite model inference; the data storage module 2 is used to store historical running data; the communication module is a module that supports Profinet IRT and 5G dual-link communication, realizing low-latency and high-reliability data exchange with the PLC controller 5.

[0036] Example 2 This embodiment provides a control method for an AI-predictive membrane denitrification automatic control system, including the following steps: Step 1: Using a one-button start command, the system sequentially starts the pretreatment unit, performs low-pressure flushing of the membrane system, pressurizes and operates in conjunction with the post-treatment unit, until it reaches normal operating conditions. One-button start / stop achieves full automation, reducing start / stop time from the traditional 30-40 minutes of manual operation to 10-15 minutes, minimizing manual steps, reducing operator workload, and avoiding human error. Specifically, the pretreatment unit must start with the raw water tank level ≥20%. During startup, the raw water inlet valve is opened, the raw water pump is started, and the inlet valves of the multi-media filter and activated carbon filter are opened sequentially. The vent valve is opened until no air bubbles are present and then closed. Simultaneously, the pH sensor monitors the pH of the pretreated effluent. 30% hydrochloric acid or 40% caustic soda is added via the dosage adjustment valve to stabilize the pH at 6-7, with an error within ±0.1. The membrane system... The trigger conditions for low-pressure flushing of the membrane system are: turbidity of pretreated effluent ≤1 NTU and residual chlorine ≤0.1 mg / L. During the low-pressure flushing process of the membrane system, the concentrate discharge valve and the permeate discharge valve are opened, the cleaning pump is started, and the membrane module is flushed for 5-8 minutes. When the turbidity is >5 NTU, the flushing time is extended to 10 minutes. The pressure before the membrane is controlled at 0.3-0.5 MPa by the inlet regulating valve. After flushing, the membrane system is pressurized and operated. The permeate discharge valve is closed, and the high-pressure pump is used to increase the pressure by frequency conversion. The pressure before the membrane is stabilized at 1.5-2.0 MPa (NF membrane) or 2.0-2.5 MPa (RO membrane) with an error within ±0.05 MPa. The opening of the concentrate regulating valve is adjusted to 40%-60%, and the recovery rate is controlled at 40%-50%. When the permeate tank level is ≥10%, the post-treatment linkage is activated, the permeate pump is started, the permeate pH sensor monitors the water quality, and caustic soda is added by the dosing pump to control the permeate pH at 6-9.

[0037] Step Two: During normal operation, the safety interlock logic is executed in real time, monitoring the membrane inlet pressure, water quality parameters, and liquid level, and triggering protective actions when parameters exceed limits. Simultaneously, the AI ​​prediction hardware unit performs model inference based on real-time collected data to predict future transmembrane pressure differential changes and overpressure risks, and provides cleaning strategies or issues adjustment commands in advance when overpressure risks are predicted through the control and monitoring unit. Figure 2As shown, the safety interlock logic specifically includes overpressure interlock, water quality interlock, and liquid level interlock. Through a dual mechanism of multi-parameter coupled safety interlock and AI predictive protection, it achieves automatic emergency handling of faults such as overpressure, water quality exceeding standards, and abnormal liquid level, reducing equipment damage rate by 30% and increasing water quality discharge compliance rate to over 99.5%, meeting the environmental protection requirements of the chlor-alkali industry. The overpressure interlock logic is as follows: if the pressure before the membrane exceeds the rated value by 10% (e.g., 2.2MPa for NF membrane and 2.75MPa for RO membrane), the PLC immediately stops the high-pressure pump, opens the pressure relief valve (100% opening), closes the membrane system inlet valve, and simultaneously triggers a human-machine audible and visual alarm (red alarm light + buzzer); if the pressure difference between the inlet and outlet of the security filter is >0.15MPa, the raw water pump is stopped, and the backwash command is triggered (opening the backwash inlet valve). Water valve and backwash drain valve, backwash time 10min); water quality interlock logic: if the permeate conductivity exceeds the design value (e.g., >500μS / cm), the permeate return valve will automatically open, returning the permeate to the raw water tank; if it does not meet the standard for 30 minutes, the membrane system will be shut down (the high-pressure pump will be stopped and the inlet valve will be closed); if ORP >200mV (free chlorine exceeds the standard), the raw water pump will be stopped and the reducing agent dosing pump will be started (adding 10% sodium bisulfite) until ORP ≤50mV; liquid level interlock logic: if the raw water tank level is <15%, the raw water pump will be stopped and the HMI will display a "low raw water level" alarm; if the permeate tank level is >90%, the membrane system will be shut down (the high-pressure pump will be stopped) and the permeate return valve will be opened; if the chemical tank (e.g., acid tank, alkali tank) level is <10%, the corresponding dosing pump will be stopped and a "chemical tank replenishment" prompt will be displayed; Step 3: By using the one-click stop command or the safety interlock to trigger the shutdown, the system executes the membrane system depressurization, flushing, evacuation and shutdown steps in a preset sequence, and optimizes the flushing parameters based on the AI ​​prediction results; if the shutdown is caused by contamination, the cleaning mode is automatically triggered and the cleaning procedure is executed according to the cleaning strategy generated by AI.

[0038] As a further implementation, the cleaning strategy is generated differently based on the level of contamination and sent to the PLC controller 5 for execution: When the pollution level is low, delayed cleaning is performed during the system load off-peak period, using the first cleaning duration and the first reagent concentration. When the contamination level is medium, the standard cleaning process is triggered, including low-pressure rinsing, circulation of the first concentration of chemicals, and pure water rinsing, with the second cleaning duration. When the contamination level is high, an enhanced cleaning process is triggered, increasing the cleaning pump frequency and the second chemical concentration, and extending the chemical circulation time, employing a third cleaning duration. The system triggers cleaning based on the contamination level, achieving different energy consumption and material inputs for different contamination levels. For low-contamination conditions, cleaning is delayed until off-peak hours. By controlling the cleaning duration, pump frequency, and chemical concentration, energy is used efficiently and rationally, reducing operating costs. This tiered treatment avoids resource waste and membrane damage caused by heavy cleaning for lightly contaminated areas, reduces membrane replacement frequency, and prevents incomplete cleaning and continuous deterioration of membrane performance caused by light cleaning for heavily contaminated areas.

[0039] As a further implementation method, historical operating data covering different working conditions are collected, cleaned, labeled and feature-engineered to construct a training dataset containing time-series features and target labels; the training dataset is used to train the GRU time-series prediction model and the XGBoost risk classification model, and the accuracy is verified and the model is lightweighted to obtain the final trained AI prediction model, which is then deployed on AI edge computing node 1.

[0040] As a further implementation method, it also includes AI prediction overpressure interlock logic: when the AI ​​prediction model outputs the risk that the in-mothermal pressure will exceed the rated value by 10% in the next 10 minutes, the PLC controller 5 controls the pressure relief valve to open in advance and reduce the frequency of the high-pressure pump, forming a double protection with the original overpressure interlock to prevent overpressure.

[0041] As a further implementation method, it also includes model operation and maintenance optimization steps: regularly using newly added real-time data to incrementally train and calibrate the AI ​​prediction model, and monitoring hardware status and data drift to ensure prediction accuracy and control stability.

[0042] The specific AI prediction and decision-making process includes data preparation, model development and training, and software integration; Data preparation: The data acquisition process includes second-level acquisition of process parameters (pre- / post-membrane pressure, flow rate), minute-level acquisition of water quality parameters (turbidity, pH, ORP), and hour-level acquisition of equipment status parameters (motor temperature, valve opening); and continuous acquisition for 6 months. It covers three different operating conditions: high turbidity during the rainy season, low temperature during winter, and normal operation, ensuring the dataset contains at least 5 million valid records. Data cleaning uses the 3σ criterion for screening; for example, if the pre-membrane pressure suddenly jumps to 5 MPa, it is judged as a sensor malfunction and directly discarded. For short-term missing data (<5 minutes), "linear interpolation of adjacent data" is used; for long-term missing data (>30 minutes), "data gaps" are marked, and this period is skipped during subsequent model training. For high-frequency fluctuating data such as pressure and flow rate, a 5-second sliding window average is used for smoothing; for example, pre-membrane pressures of 1.82, 1.78, and 1.80 MPa are smoothed to 1.80 MPa. A 32-dimensional feature vector is formed by calculating membrane pressure difference, recovery rate, and turbidity normalization value. The membrane pressure difference is the difference between the pressure before and after the membrane; the recovery rate is the percentage of permeate flow rate to feed flow rate; and the turbidity normalization value is the ratio of actual turbidity to the upper limit of the membrane feed water. Each data point is labeled with the predicted membrane pressure for the next 30 minutes and the overpressure risk level for the next 60 minutes, based on historical data from the same period (e.g., if a data point corresponds to an actual pressure of 2.1 MPa for the next 30 minutes, the predicted value is 2.1 MPa). The dataset is divided into training, validation, and test sets in a 7:2:1 ratio to ensure that all three sets contain an equal proportion of abnormal operating condition data. The cleaned CSV data is converted to TensorFlow-supported TFRecord format using Python's pandas library and stored in a "timestamp + 32-dimensional feature + 2-dimensional target value" structure to improve model retrieval efficiency.

[0043] Model development and training include core model selection and model training; core model selection: a combination of "GRU (Gated Recurrent Unit) and XGBoost" model is adopted (GRU processes time series data, XGBoost optimizes risk classification), the specific structure is as follows: The GRU temporal prediction layer consists of an input layer of 64 neurons, a hidden layer of 32 neurons, and an output layer of 16 neurons. The input is a 32-dimensional feature sequence of the past 120 seconds (transmembrane pressure difference, temperature, SDI), which is converted into a 64-dimensional vector through the embedding layer. The output layer uses a linear activation function to predict key parameters for the next 30, 60, and 120 minutes: transmembrane pressure difference (e.g., 0.75 MPa after 30 minutes, 0.9 MPa after 60 minutes); pre-membrane pressure (e.g., 2.1 MPa after 60 minutes); and product water conductivity (e.g., 480 μS / cm after 60 minutes). Training objective: The prediction error of transmembrane pressure difference per unit hour is ≤3% (predicted at 0.776-0.824MPa when the actual pressure is 0.8MPa). The XGBoost risk classification layer takes the "parameter prediction value" output by GRU and "current process characteristics" as input and outputs "probability of overpressure / water quality exceeding the standard risk in the next 60 minutes (0-100%)", and sets a threshold (e.g., risk > 85% is judged as high risk).

[0044] Model training environment setup: The server is configured with a GPU (NVIDIA Tesla V100) + 32GB of memory, the system is Ubuntu 20.04, TensorFlow 2.10 and XGBoost 1.7.5 are installed, and mixed precision training is used to accelerate convergence.

[0045] Training parameter settings: Adam optimizer (initial learning rate 0.001, decaying by 10% every 50 epochs), loss function (MSE mean squared error, used for GRU; cross-entropy loss, used for XGBoost), 200 training epochs, early stopping (patience=10) to prevent overfitting.

[0046] Model Training: Train the GRU model using the training set. Stop GRU training when the validation set MSE drops below 0.001 MPa² in each round (e.g., membrane pressure prediction error < ±0.03 MPa). Use the predicted values ​​output by the GRU as features and add them to the XGBoost model training until the validation set risk classification accuracy is >95% (e.g., model identification accuracy ≥95% in actual high-risk samples). Evaluate model performance using the test set, requiring the following: membrane pressure prediction R² ≥ 0.98, permeate conductivity prediction error ≤5%, and risk warning false alarm rate <3% (e.g., <15,000 false high-risk alarms in 500,000 test data points). If the accuracy requirements are not met, backtrack to the data cleaning stage or increase the number of neurons in the GRU hidden layer and retrain. Convert the trained model to TensorFlow Lite format and use quantization-aware training to compress the model size from 200MB to 15MB, increasing inference speed by 3 times and adapting to the computing power of edge computing nodes.

[0047] PLC Controller 5 Control Logic: Instruction Reception: A new OB35 cyclic interrupt organization block (10ms cycle) is added to the PLC to parse the instructions of the AI ​​predictive hardware unit (such as "high pressure pump frequency 40Hz") and drive the actuator; when the AI ​​instruction conflicts with the local safety interlock (such as frequency increase but the pressure before the membrane has exceeded 2.0MPa), the PLC safety logic (stop pump + pressure relief) is executed first, and the cause of the conflict is fed back to the AI ​​predictive hardware unit.

[0048] The HMI touchscreen interface 7 now features an "AI Prediction Monitoring Page," which displays real-time parameter prediction curves for the next 60 minutes, with black representing real-time values, red representing predicted values, and yellow indicating safety thresholds; risk level indicator lights showing green for low risk, yellow for medium risk, and red for high risk; and control command execution status. A manual intervention button has also been added: when operators determine that the AI ​​prediction hardware unit commands are unreasonable, they can click "Pause AI Control" to switch to PLC local control, while simultaneously recording the reason for the intervention for subsequent model optimization.

[0049] The host computer 8 supports manually triggering model training, viewing historical accuracy reports, and downloading model versions; it also automatically calculates AI control performance and generates monthly operation and maintenance reports.

[0050] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. An AI-predictive-based automatic control system for membrane denitrification, characterized in that, include: The system comprises a sensor unit, an actuator unit, a control and monitoring unit, and an AI prediction hardware unit. The sensor unit includes basic sensors and supplementary AI prediction sensors. The actuator unit includes basic actuators and actuators with AI control interfaces. The control and monitoring unit includes a PLC controller, a human-machine interface screen, and a host computer. The AI ​​prediction hardware unit includes AI edge computing nodes, a data storage module, and a communication module. The sensor unit is connected to the PLC controller; the actuator unit is connected to the PLC controller; the PLC controller communicates with the human-machine touch screen and the host computer; the AI ​​edge computing node is bidirectionally connected to the PLC controller through the communication module; the AI ​​edge computing node obtains real-time operating data from the PLC controller, stores it in the data storage module, and generates prediction results and control instructions through model inference, which are then sent to the PLC controller for execution.

2. The AI-predictive-based membrane denitrification automatic control system as described in claim 1, characterized in that, The basic sensors consist of a pressure sensor, a flow sensor, a water quality sensor, and a level sensor; the AI ​​predictive supplementation sensor consists of an influent turbidity sensor and an influent temperature sensor; the pressure sensor is installed at the inlet and outlet of the membrane module and the inlet and outlet of the security filter; the flow sensor is installed on the raw water inlet, membrane system permeate, and concentrate pipelines; the water quality sensors include a pH sensor installed on the pretreatment effluent pipeline, a conductivity sensor installed on the permeate pipeline, and an ORP sensor installed at the inlet of the security filter; the level sensor is installed in the raw water tank, permeate tank, and chemical tank; the influent turbidity sensor is installed on the raw water inlet pipeline, and the influent temperature sensor is installed on the pretreatment inlet pipeline.

3. The AI-predictive-based membrane denitrification automatic control system as described in claim 1, characterized in that, The basic actuator consists of a variable frequency pump set, an electric regulating valve, and a solenoid valve; both the cleaning pump in the variable frequency pump set and the drug dosage regulating valve in the electric regulating valve are equipped with AI control signal interfaces to receive cleaning flow rate and drug concentration commands.

4. The AI-predictive-based membrane denitrification automatic control system as described in claim 1, characterized in that, The PLC controller is equipped with an AI prediction result receiving module, which is used to parse and execute AI instructions in the loop interruption tissue block; the human-machine touch screen is equipped with an AI prediction interface, which is used to display the future transmembrane pressure difference prediction curve, cleaning suggestions and risk level; the host computer is equipped with an AI prediction model management module, which supports model training triggering, accuracy evaluation and data drift detection.

5. The AI-predictive-based membrane denitrification automatic control system as described in claim 1, characterized in that, The AI ​​edge computing node has a built-in ARM architecture processor and NPU, supporting TensorFlow Lite model inference; the data storage module is used to store historical running data; and the communication module supports Profinet IRT and 5G dual-link communication.

6. The control method for an AI-predictive membrane denitrification automatic control system as described in any one of claims 1-5, characterized in that, Includes the following steps: Step 1: With the command of one-button start, the system starts the pretreatment unit, performs low-pressure flushing of the membrane system, pressurizes and runs in sequence according to the preset order, and links with the posttreatment unit until it enters normal operating condition; Step 2: During normal operation, the safety interlock logic is executed in real time to monitor the membrane pressure, water quality parameters and liquid level, and trigger protection actions when the parameters exceed the standard. At the same time, the AI ​​prediction hardware unit performs model inference based on the real-time collected data to predict future changes in transmembrane pressure difference and overpressure risk, and provides cleaning strategies through the control and monitoring unit or issues adjustment instructions in advance when overpressure risk is predicted. Step 3: By using the one-click stop command or the safety interlock to trigger the shutdown, the system executes the membrane system depressurization, flushing, evacuation and shutdown steps in a preset sequence, and optimizes the flushing parameters based on the AI ​​prediction results; if the shutdown is caused by contamination, the cleaning mode is automatically triggered and the cleaning procedure is executed according to the cleaning strategy generated by AI.

7. The control method for an AI-predictive membrane denitrification automatic control system as described in claim 6, characterized in that, The cleaning strategy is generated differently based on the level of contamination: When the pollution level is low, delayed cleaning is performed during the system load off-peak period, using the first cleaning duration and the first reagent concentration. When the contamination level is medium, the standard cleaning process is triggered, including low-pressure rinsing, circulation of the first concentration of chemicals, and pure water rinsing, with the second cleaning duration. When the pollution level is high, an enhanced cleaning process is triggered, increasing the cleaning pump frequency and the second agent concentration, and extending the agent circulation time, thus adopting a third cleaning duration.

8. The control method for an AI-predictive membrane denitrification automatic control system as described in claim 6, characterized in that, Historical operational data covering different working conditions were collected, cleaned, labeled, and feature-engineered to construct a training dataset containing time-series features and target labels. The training dataset was used to train a GRU time-series prediction model and an XGBoost risk classification model, and accuracy verification and model lightweighting were performed to obtain the final combined AI prediction model.

9. The control method for an AI-predictive membrane denitrification automatic control system as described in claim 6, characterized in that, It also includes an AI prediction overpressure interlock step: when the AI ​​prediction model outputs a risk that the intermembrane pressure will exceed the rated value by 10% in the next 10 minutes, the PLC controller controls the pressure relief valve to open in advance and reduce the frequency of the high-pressure pump to prevent overpressure.

10. The control method for an AI-predictive membrane denitrification automatic control system as described in claim 6, characterized in that, It also includes model operation and maintenance optimization steps: regularly using newly added real-time data to incrementally train and calibrate the AI ​​prediction model, and monitoring hardware status and data drift to ensure prediction accuracy and control stability.

Images