Water plant dosing intelligent closed-loop control system and control method

Abstract

The invention discloses an intelligent closed-loop control system and method for dosing in a water plant. An integrated sensing unit of the system is used for synchronously acquiring an electrochemical impedance spectrum signal reflecting a floc coagulation state and a fluid entropy change signal reflecting mixed energy efficiency in situ; the control unit is used for extracting dielectric characteristic parameters and entropy yield parameters from the signals, inputting the dielectric characteristic parameters and the entropy yield parameters into a neural network model trained through reinforcement learning to generate a dosing control instruction, and can adaptively switch among a direct control mode, an online self-calibration mode and a safety keeping mode based on the entropy yield change rate or the signal quality; and the execution unit is used for executing the instruction to adjust the dosage. By directly sensing the micro state of the coagulation process and utilizing the intelligent model to make a decision in real time, the problems of lagging and strong model dependence of a traditional control mode are solved, second-level real-time accurate regulation and control are realized, the medicine consumption and the energy consumption are remarkably reduced, and the adaptability and the reliability of the system under complex working conditions are improved.

Description

Technical Field

[0001] This invention relates to the field of water treatment automatic control technology, specifically to an intelligent control system and method for the coagulation dosing process in water treatment facilities such as waterworks and sewage treatment plants. Background Technology

[0002] Coagulation is a key process unit for removing colloids, suspended solids, and other impurities from water. Its core control objective is to optimize the coagulant dosage while ensuring the stable compliance of the influent quality (usually measured by turbidity) of subsequent processes. This aims to minimize chemical consumption, associated energy consumption, and operating costs, while improving process stability. To achieve these goals, the technologies used and continuously evolving in this field mainly include the following: 1. Feedback control based on macroscopic parameters, which is the most common application. The system uses macroscopic water quality parameters such as turbidity, pH, and current flow rate (SCD) of the effluent from subsequent sedimentation tanks or filters as core feedback signals, adjusting the dosage through classic control algorithms such as PID or empirical rules. This approach is logically straightforward, but the feedback signals it relies on are the results after coagulation and sedimentation, rather than a direct reflection of the process itself, resulting in a significant time lag (usually several minutes to twenty minutes). When the raw water quality and quantity fluctuate rapidly, adjustments based on the lag signal inevitably lead to over- or under-dosing, resulting in wasted chemicals or water quality risks. 2. Soft measurement and decision support based on process models: To compensate for the shortcomings of direct online parameters, more advanced technologies attempt to construct process mathematical models (including mechanistic models based on chemical kinetics or machine learning models based on historical data) to simulate the coagulation process, predict effluent quality through simulation, and recommend dosage. However, this method has inherent defects. First, the model accuracy is heavily dependent on the breadth and quality of the training data. When faced with raw water characteristics not covered by the training set (such as sudden pollution or changes in algal populations), its predictive reliability drops sharply. Second, the complex model itself becomes a new black box, with opaque internal decision-making logic, making it difficult for operators to understand and trust, and requiring high computing power for deployment at the edge. 3. Flocculation state analysis based on image recognition: As an exploration of process state, some cutting-edge research attempts to capture floc images through cameras and use image processing algorithms to analyze floc particle size and morphology. However, this technology essentially obtains two-dimensional visual representation information of flocs and cannot perceive the key microscopic physicochemical properties that determine coagulation efficiency, such as internal charge distribution, interfacial potential, and coagulation strength of the flocs. In addition, optical sensors are easily interfered with in turbid water and their lenses are easily contaminated, resulting in poor long-term operational stability.

[0003] In summary, existing technologies face a fundamental technical contradiction in achieving precise, real-time, and reliable control of coagulation dosing: the effectiveness of the control system either relies on severely delayed macroscopic results or is constrained by complex and unreliable predictive models, lacking the ability to directly and stably perceive the immediate and essential state of the coagulation process itself. This leads to a series of chain reactions, including slow control response, weak adaptive capabilities, and low decision-making reliability in complex operating conditions, thus hindering further improvements in energy conservation, emission reduction, and stable operation of water plants.

[0004] Therefore, developing a novel closed-loop control paradigm that can bypass macroscopic hysteresis parameters and complex prediction models, directly, in situ, and online acquire microscopic physicochemical signals reflecting the core dynamic state of the coagulation process, and constructing a new closed-loop control paradigm with ultrafast response and inherent reliability based on these signals, has become a technical bottleneck that urgently needs to be overcome in this field, and has important practical necessity and urgency. Summary of the Invention

[0005] To address the shortcomings of existing technologies in water plant chemical dosing control, which rely on lagging macroscopic parameters, complex models, slow response, and poor adaptability, this invention provides an intelligent closed-loop control system and method for water plant chemical dosing based on direct perception of microscopic process states and intelligent decision-making. This system solves the problems of control lag, unreliable models, and poor robustness under complex operating conditions, ultimately achieving significant energy savings, reduced consumption, and improved water quality stability.

[0006] The solution to the technical problem of this invention is as follows: A smart closed-loop control system for water plant chemical dosing is adopted, comprising an integrated sensing unit configured to synchronously acquire, in situ, an electrochemical impedance spectroscopy signal reflecting the floc coagulation state during coagulation and a fluid entropy change signal reflecting the mixing energy efficiency; a control unit communicatively connected to the integrated sensing unit, configured to: extract dielectric characteristic parameters from the electrochemical impedance spectroscopy signal and entropy yield parameters from the fluid entropy change signal; input the dielectric characteristic parameters and the entropy yield parameters into a neural network model trained through reinforcement learning to calculate and generate dosing control commands; and switch between direct control, online self-calibration, and safety maintenance modes based on the rate of change or signal quality of the entropy yield parameters; and an execution unit communicatively connected to the control unit, configured to adjust the dosage of chemicals in response to the dosing control commands.

[0007] Preferably, the integrated sensing unit includes a multi-frequency electrochemical impedance spectroscopy measurement module and a micro-area thermal-current sensing module encapsulated in the same probe housing. The micro-area thermal-current sensing module is used to measure local temperature difference and flow velocity to calculate the fluid entropy change signal.

[0008] Preferably, the control unit is configured to switch operating modes according to the following rules: when the rate of change of the entropy yield parameter exceeds a first preset threshold, it enters the online self-calibration mode to optimize the parameters of the neural network model based on auxiliary water quality parameters; when the electrochemical impedance spectroscopy signal and / or the fluid entropy change signal are abnormal, it enters the safety maintenance mode to maintain the current state of the execution unit; otherwise, it is in the direct control mode to generate the dosing control command based on the dielectric characteristic parameter and the entropy yield parameter extracted in real time.

[0009] Preferably, the control unit is further configured to: when the process phase change rate calculated based on the electrochemical impedance spectroscopy signal exceeds a preset mutation threshold, switch from decision-making based on the neural network model to generating the dosing control command based on a stabilization control law based on the process phase deviation and entropy yield change rate.

[0010] Preferably, the stabilization control law is a proportional-feedforward composite control law, the output of which is a weighted sum of the proportional control term for the process phase deviation and the feedforward control term for the entropy yield change rate.

[0011] Another intelligent closed-loop control method for chemical dosing in water plants is adopted, including a synchronous sensing step: acquiring electrochemical impedance spectroscopy signals reflecting the floc aggregation state during coagulation and fluid entropy change signals reflecting mixing energy efficiency in situ through an integrated sensing unit; an intelligent decision-making and mode switching step: extracting dielectric characteristic parameters from the electrochemical impedance spectroscopy signals and entropy yield parameters from the fluid entropy change signals; inputting the dielectric characteristic parameters and the entropy yield parameters into a neural network model trained through reinforcement learning to generate dosing control commands; and adaptively switching between direct control, online self-calibration, and safety maintenance modes based on the rate of change or signal quality of the entropy yield parameters; and an execution step: adjusting the dosage of chemicals according to the dosing control commands.

[0012] Preferably, in the intelligent decision-making and mode switching steps, the mode switching rules are as follows: when the rate of change of the entropy production rate parameter exceeds the second preset threshold, the system enters the online self-calibration mode to optimize the parameters of the neural network model based on auxiliary water quality parameters; when the electrochemical impedance spectroscopy signal and / or the fluid entropy change signal are abnormal, the system enters the safety maintenance mode to maintain the current dosage of the reagent; otherwise, the system is in the direct control mode.

[0013] Preferably, the intelligent decision-making and mode switching steps further include: when the process phase change rate calculated based on the electrochemical impedance spectroscopy signal exceeds a preset mutation threshold, switching to generating the dosing control command based on the stabilization control law based on the process phase deviation and entropy yield change rate.

[0014] The beneficial effects of this invention are as follows: 1. This invention shifts the perception and decision-making basis of the control system from lagging process results to the immediate process state. By employing integrated sensing probes to synchronously acquire electrochemical impedance spectroscopy and local entropy yield signals that directly characterize the microscopic dynamic state of the coagulation process in situ, and by deploying a lightweight direct control model at the edge side for millisecond-level feature extraction and decision mapping, the response delay of the entire control closed loop undergoes a qualitative change. The total time from signal perception to actuator action is stably controlled within 150 milliseconds, which is a significant improvement over traditional schemes relying on effluent turbidity feedback (typically requiring 2-10 minutes), achieving true second / sub-second real-time control levels, thereby enabling near-instantaneous suppression of fluctuations in influent water quality and flow rate.

[0015] 2. This invention not only pursues control speed but also achieves systematic optimization of operating costs and resource efficiency through the collaborative design of architecture and algorithms. First, because the edge agent (ECU) can independently complete most real-time control tasks, interacting only with the cloud when uploading abnormal data packets, it greatly reduces reliance on continuously high-bandwidth networks and cloud computing power, simplifying system deployment and maintenance complexity. Second, because the control objective directly embeds the collaborative optimization of chemical and energy consumption, the system can automatically find the operating point with the lowest overall cost while ensuring water quality. Trial operation shows that compared to traditional PID control, this system can achieve a 10%-25% reduction in coagulant dosage, while optimizing mixing energy distribution and reducing the power consumption of associated stirring equipment by 5%-15%, achieving a synergistic reduction in chemical and power consumption.

[0016] 3. This invention significantly enhances the robustness of the system in complex and variable industrial environments by introducing a cloud-edge co-evolution mechanism and a hidden security loop. On the one hand, due to the establishment of a continuous learning closed loop (S104) between edge execution and cloud-based deep analysis and model optimization, the system possesses long-term adaptive capabilities, enabling it to self-update following the slow drift of water quality characteristics and avoiding the performance degradation of traditional models over time. On the other hand, and more importantly, the non-steady-state process phase synchronization analyzer features, as detailed in Example 2, provide a safety valve for the system to cope with unforeseen severe shocks. When a sudden change in the process phase is detected, the system can automatically and smoothly switch to a stabilizing control law based on physicochemical principles. This reduces the overshoot of the system's effluent water quality by approximately 40% and shortens the recovery time by approximately 35% when facing extreme conditions such as heavy rain and sudden pollution, significantly expanding the stable operating boundary of the intelligent control system in real and harsh scenarios, and laying a solid foundation for the large-scale and highly reliable application of the technology. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the system architecture provided in Embodiment 1 of the present invention; Figure 2 This is a flowchart of the intelligent closed-loop control method for chemical dosing in water plants provided in Embodiment 1 of the present invention; Figure 3 This is a schematic diagram of the working logic and switching of the unsteady-state process phase synchronization analyzer in Example 2; Figure 4 This is a schematic diagram comparing the technical effects of the present invention and traditional solutions. Detailed Implementation

[0018] To make the objectives, technical solutions, and beneficial effects of this invention clearer and more explicit, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of protection of this invention. Example

[0019] This embodiment addresses the core contradiction in coagulation dosing control in water treatment—namely, the lag in control basis and strong model dependence—by providing an intelligent closed-loop control system for water plant dosing. This embodiment aims to directly capture the microscopic physicochemical state of the coagulation process through a novel sensing and control paradigm, achieving second-level real-time control integrating feedforward and feedback. Figure 1 As shown, the system includes an integrated sensor probe 1, an edge computing and control unit (ECU) 2, a cloud / plant-level management platform 3, and a dosing execution unit 4. Figure 1 The display aims to illustrate the physical connections and core data flow between the integrated sensor probe, edge computing and control unit (ECU), cloud / plant-level management platform and dosing execution unit in a hierarchical block diagram, clarifying the system's edge-cloud collaborative architecture.

[0020] The integrated sensing probe 1 serves as both the first sensing unit (providing electrochemical impedance spectroscopy signals) and the second sensing unit (providing fluid entropy change signals). Specifically, a three-electrode electrochemical impedance spectroscopy measurement module and a micro-area heat-flow sensing module are encapsulated within a corrosion-resistant cylindrical probe housing. The working electrode of the three-electrode system is a platinum microarray electrode, the reference electrode is an Ag / AgCl electrode, and the counter electrode is a stainless steel electrode. The excitation signal frequency range is 1 Hz to 1 MHz. The micro-area heat-flow sensing module includes two pairs of precisely 5 mm-spaced miniature thermocouples (for measuring the axial temperature gradient ΔT) and a MEMS-based miniature eddy current velocity sensor (for measuring local flow velocity v). The probe is connected to a signal cable via a waterproof aviation connector and can be immersed and installed at the end of the rapid stirring zone of a coagulation reaction tank.

[0021] The edge computing and control unit (ECU) 2 is used to implement the functions of the control unit. Its hardware is an industrial-grade embedded computer. The signal processing module 21 running inside it specifically performs the following operations: receiving raw impedance spectrum data, performing least squares fitting using the Cole-Cole model, extracting the characteristic relaxation frequency (F_diel) and the reciprocal of the zero-frequency impedance (i.e., interface conductance G), and normalizing G by dividing it by the current influent conductance value to obtain the normalized interface conductance (G_norm); synchronously receiving ΔT and v signals, and calculating the local entropy production rate (σ_local) according to the formula σ_local=(η*v^2) / T+(k*ΔT^2) / (T^2*Δx), where η is the dynamic viscosity of water, k is the thermal conductivity, T is the absolute temperature, and Δx is the thermocouple spacing. The lightweight direct control model pre-installed in decision module 22 is a three-layer fully connected neural network. Its input layer consists of a feature vector composed of [F_diel, G_norm, σ_local], and its output layer is the frequency adjustment amount of the dosing pump (ΔF, unit: Hz). This model is trained by combining historical data and reinforcement learning, aiming to minimize the combined cost of drug consumption and disturbance suppression time in a digital twin simulation environment. The local cache and management module 23 is a circular data buffer that continuously stores the feature vector sequence, original signal segments, and control commands from the most recent 2 minutes.

[0022] The dosing execution unit 4 is a frequency-controlled diaphragm metering pump. Its frequency converter receives a 4-20mA analog signal (corresponding to a 0-50Hz frequency) from the ECU2 to realize continuous adjustment of the dosage.

[0023] like Figure 2 The system demonstrates the complete workflow (steps S101-S104) from in-situ signal acquisition, real-time feature extraction, edge intelligent decision-making, to command execution and data upload, reflecting the closed-loop control logic. The system workflow is as follows.

[0024] Step S101: Synchronous sensing. The integrated sensing probe 1 synchronously acquires and outputs a set of impedance spectrum data and a set of ΔT and v data at a frequency of once per second.

[0025] Step S102: Edge Computing and Decision Making. Signal processing module 21 completes feature extraction within 100 milliseconds, forming the feature vector for the current moment. The neural network of decision module 22 completes forward inference within 10 milliseconds, outputting ΔF. ECU2 immediately converts this ΔF into an analog signal and sends it to the dosing pump. Because the sensing signal originates directly from the core area of ​​the coagulation reaction, and the decision model is extremely lightweight, the total delay from signal acquisition to actuator action is controlled within 150 milliseconds, achieving second-level (sub-second) real-time closed-loop control, fundamentally overcoming the minute-level lag problem of traditional schemes that rely on effluent turbidity feedback.

[0026] Step S103: Data encapsulation and upload. Simultaneously, the local caching and management module 23 records this control event. If the absolute value of ΔF is greater than the threshold for five consecutive control cycles, or if σ_local changes by more than 30% within one minute, it is determined to be a "potential abnormal condition." At this time, module 23 packages all high-resolution data (original impedance spectrum, temperature, and flow rate sequences) and corresponding feature vectors and control commands within a 90-second time window before and after the trigger event into a single data file, marks it with a timestamp and device ID, and uploads it to the cloud platform 3.

[0027] Step S104: Cloud Optimization. The digital twin and deep analysis engine 31 of the cloud platform 3 uses a high-fidelity process model to perform in-depth review and attribution analysis on the data package. The model management and optimization center 32 collects such data packages uploaded by all edge nodes, performs incremental training and knowledge distillation on the lightweight direct control models deployed in each ECU2 once a month, generates optimized model parameters, and distributes them for updates in a unified manner. Example

[0028] Building upon Example 1, this example further discloses a technical feature that significantly improves the system's robustness under sudden water quality fluctuations. The inventors of this example discovered that, when the system of Example 1 encounters extreme unsteady conditions such as a sudden surge in raw water turbidity due to heavy rain or the instantaneous mixing of chemical wastewater, although σ_local may drastically change and trigger the uploading of potential abnormal conditions, the real-time control at the edge may still experience temporary inaccuracies because the model has not learned such extreme modes. Therefore, this example adds a parallel submodule called "Unsteady-State Process Phase Synchronization Analyzer" to the decision module 22 of ECU2. For example... Figure 3 This is a schematic diagram of the working logic and switching of the unsteady-state process phase synchronization analyzer. Figure 3 The aim is to demonstrate the core logic of this feature, describe the decision chain from phase calculation, mutation detection, and calming control, and the switching conditions between it and the conventional neural network control mode.

[0029] The submodule described in this embodiment treats the coagulation process as a dynamic oscillating system, analyzes its process phase through real-time impedance spectroscopy, and switches to an emergency stabilization control law based on first principles of physicochemical principles when a phase abruptly occurs. Its specific implementation includes the following steps.

[0030] 1. Phase Calculation: This submodule continuously monitors impedance spectrum data. For the impedance spectrum at each sampling time, it calculates the phase angle (θ) of its complex impedance at a specific characteristic frequency (e.g., 100Hz). This phase angle θ is defined as the "phase" of the current process state.

[0031] 2. Sudden Change Detection: This module maintains a phase angle sequence within a short time window (e.g., 10 seconds) and calculates its rate of change (dθ / dt) in real time. When the absolute value of dθ / dt exceeds a set high threshold (e.g., corresponding to a change of more than 50% in the main ionic components of the raw water within 30 seconds), the process is immediately determined to have entered a "non-steady-state phase sudden change period".

[0032] 3. Stabilization Control: Once this period begins, decision module 22 will temporarily bypass the lightweight neural network model. Instead, a pre-defined proportional-feedforward composite control law will be activated. The formula for this control law is: `ΔF_emergency=Kp(θ-θ_set)+Kff(dσ_local / dt)`. Where θ_set is the reference phase in steady state, Kp is the proportional coefficient, Kff is the feedforward coefficient, and `(dσ_local / dt)` is the rate of change of entropy production. The coefficients Kp and Kff are pre-calibrated based on the basic theory of coagulation reaction kinetics and numerous unsteady-state simulation experiments.

[0033] 4. Switching and Recovery: This stabilizing control will continue until the phase change rate dθ / dt falls back to the normal range and remains stable for 10 seconds, after which the system will automatically switch back to the conventional direct control mode based on the neural network. The entire switching process is smooth and uninterrupted.

[0034] This embodiment introduces a parallel safety control channel based on the detection of process intrinsic phase abrupt changes. This allows the system to abandon reliance on data-driven models and instead rely on more deterministic physicochemical relationships for rapid stabilization when faced with severe and sudden water quality shocks not covered by training data. This is equivalent to adding a conditioned reflex-like safety loop to the intelligent control system. Experiments show that under simulated sudden high turbidity shocks, the system using this technology can reduce effluent turbidity overshoot by approximately 40% and shorten the recovery time to a stable state by approximately 35%, significantly enhancing the system's resistance to shock loads and operational reliability in real-world complex environments. This feature provides an innovative and effective hybrid intelligent control approach to address the common industry problem of data-driven models failing to generalize under extreme conditions. Through the technical means disclosed in Embodiments 1 and 2, the present invention has achieved the following verifiable positive effects, such as... Figure 4As shown in the figure, by using impedance spectrum and entropy yield, which directly reflect the microscopic state of the coagulation process, as real-time feedback signals, and by deploying a lightweight neural network with millisecond-level response at the edge for direct mapping decision-making, the system's control closed-loop delay is reduced from 2-10 minutes in traditional schemes to less than 150 milliseconds, achieving near-instantaneous suppression of process disturbances. Because the control basis is shifted forward to the process itself, and the optimization objective directly includes chemical consumption and associated power consumption, the system can adjust the dosage earlier and more accurately, avoiding over-dosing. Long-term operational statistics show that compared to traditional systems based on effluent turbidity PID control, this system can achieve a 10%-25% reduction in coagulant dosage, while optimizing mixing energy distribution and reducing associated mixer energy consumption by 5%-15%. Due to the introduction of a cloud-edge collaborative model evolution framework and the non-steady-state phase synchronous stabilization control characteristics described in Example 2, the system can not only adapt to slow changes in water quality through continuous learning, but also maintain control stability when facing sudden and severe shocks, reducing the risk of water quality exceeding standards under extreme conditions by more than 50%.

[0035] It should be noted that the above embodiments and accompanying drawings are merely illustrative examples of the core principles and key structures of the present invention. The accompanying drawings are simplified schematic diagrams, intended to clearly illustrate the structural, process, or data flow relationships related to the innovative points of the technical solution, and are not intended to limit the complete form of the actual product. This specification focuses on the innovative technical means necessary to achieve the invention's objectives and solve the technical problems. While auxiliary or common-sense details such as probe packaging structures, cable protection, communication interface protocols, power management circuits, and industrial controller selection, which can be implemented by those skilled in the art without creative effort, are not described in detail, they should be understood as naturally included in the specific implementation of this invention and fall within the protection and implementation scope of this technical solution.

Claims

1. A water plant dosing intelligent closed-loop control system, characterized in that, The integrated sensing unit is configured to synchronously acquire in-situ an electrochemical impedance spectroscopy signal reflecting the flocculation state of flocs in the coagulation process and a fluid entropy change signal reflecting the mixing energy efficiency. The control unit is in communication connection with the integrated sensing unit, and is configured to extract a dielectric characteristic parameter from the electrochemical impedance spectroscopy signal and an entropy production rate parameter from the fluid entropy change signal, input the dielectric characteristic parameter and the entropy production rate parameter into a neural network model trained by reinforcement learning to calculate and generate a dosing control instruction, and switch among direct control, online self-calibration and safety maintenance modes based on the change rate or signal quality of the entropy production rate parameter. The execution unit is in communication connection with the control unit, and is configured to adjust the dosage of the reagent in response to the dosing control instruction. The integrated sensing unit includes a multi-frequency electrochemical impedance spectroscopy measurement module and a micro-zone heat flow sensing module packaged in the same probe shell, and the micro-zone heat flow sensing module is used to measure the local temperature difference and flow rate to calculate the fluid entropy change signal.

2. The system of claim 1, wherein, The control unit is configured to switch the working mode according to the following rules:

3. The system of claim 1 or 2, wherein, When the change rate of the entropy production rate parameter exceeds a first preset threshold, the online self-calibration mode is entered, and the parameters of the neural network model are optimized according to the auxiliary water quality parameters; When the electrochemical impedance spectroscopy signal and / or the fluid entropy change signal is abnormal, the safety maintenance mode is entered, and the current state of the execution unit is maintained; Otherwise, the direct control mode is adopted, and the dosing control instruction is generated according to the real-time extracted dielectric characteristic parameter and entropy production rate parameter. The control unit is further configured to switch from decision-making according to the neural network model to generating the dosing control instruction according to a stabilizing control law based on the process phase deviation and the entropy production rate change rate when the process phase change rate calculated based on the electrochemical impedance spectroscopy signal exceeds a preset mutation threshold.

4. The system of claim 1 or 2, wherein, The stabilizing control law is a proportional-lead composite control law, and the output is the weighted sum of the proportional control term of the process phase deviation and the lead control term of the entropy production rate change rate.

5. The system of claim 4, wherein, The integrated sensing unit is configured to synchronously acquire in-situ an electrochemical impedance spectroscopy signal reflecting the flocculation state of flocs in the coagulation process and a fluid entropy change signal reflecting the mixing energy efficiency.

6. A water plant dosing intelligent closed-loop control method, characterized in that, The control unit is in communication connection with the integrated sensing unit, and is configured to extract a dielectric characteristic parameter from the electrochemical impedance spectroscopy signal and an entropy production rate parameter from the fluid entropy change signal, input the dielectric characteristic parameter and the entropy production rate parameter into a neural network model trained by reinforcement learning to calculate and generate a dosing control instruction, and switch among direct control, online self-calibration and safety maintenance modes based on the change rate or signal quality of the entropy production rate parameter. The execution unit is in communication connection with the control unit, and is configured to adjust the dosage of the reagent in response to the dosing control instruction. In the intelligent decision-making and mode switching step, the switching rules are: When the change rate of the entropy production rate parameter exceeds a second preset threshold, the online self-calibration mode is entered, and the parameters of the neural network model are optimized according to the auxiliary water quality parameters; 7. The method of claim 6, wherein, ​ ​ When the electrochemical impedance spectroscopy signal and / or the fluid entropy change signal is abnormal, entering the safety maintenance mode, maintaining the current dosage; Otherwise, in the direct control mode.

8. The method according to claim 6 or 7, characterized in that, The intelligent decision and mode switching step further comprises: when the process phase change rate calculated based on the electrochemical impedance spectroscopy signal exceeds a preset mutation threshold, switching to generating the dosing control instruction according to a stabilization control law based on the process phase deviation and the entropy production rate change rate.

Images