Method for adjusting photodynamic inactivation parameters based on hybrid fuzzy-PID control

Abstract

The present invention discloses a method for adjusting photodynamic inactivation parameters based on hybrid fuzzy-PID control, which relates to the technical field of parameter fuzzy control. Aiming at the problem that it is difficult to balance the photodynamic inactivation efficiency and the safety of aquatic organisms in aquaculture, the dynamic water quality is grasped through multi-sensor fusion and environmental prediction; Subsequently, the fuzzy-PID adaptive dosing technology is coupled with the lighting strategy to achieve precise matching of the photosensitizer dose and light intensity; Then, real-time closed-loop and multi-scenario collaborative scheduling are used to cope with water body fluctuations and pathogen diffusion; Multidimensional indicators are used for evaluation and reinforcement learning is introduced for self-learning optimization to continuously improve the dosing and lighting parameters; Finally, it is extended to large-scale distributed control and cloud big data training, and high-efficiency inactivation and flexible deployment can be maintained in different aquaculture scenarios. It can effectively avoid the rapid spread of pathogens in adverse environments, reduce the damage to aquatic animals caused by excessive accumulation of reactive oxygen species (ROS), and at the same time reduce energy consumption and resource waste.

Description

Technical Field

[0001] The present invention relates to the technical field of parameter fuzzy control, and specifically to a method for adjusting the parameters of photodynamic inactivation based on hybrid fuzzy-PID control. Background Art

[0002] Currently, water disinfection plays a crucial role in municipal water supply, sewage treatment, industrial wastewater reuse, and large-scale aquaculture systems. With the acceleration of urbanization and the continuous development of industrial production, the content of suspended solids, organic matter, and pathogenic microorganisms in water shows highly dynamic and non-linear changes. Affected by multiple factors such as seasonal climate, rainfall, industrial emissions, aquaculture feeding, and biological metabolism, water quality parameters (such as pH, dissolved oxygen, temperature, and light intensity) often fluctuate violently. At the same time, modern water treatment systems are tending to adopt advanced sensor networks, distributed control, and cloud data analysis technologies to monitor and dynamically regulate the water state in real time, achieving more efficient and precise disinfection and purification. Various intelligent control devices are emerging continuously, providing new ideas and means for dealing with complex water quality and ensuring the safety of aquatic organisms, making water disinfection not limited to traditional chemical treatment, but gradually shifting towards environmentally friendly and energy-saving technologies such as photodynamic inactivation and photocatalysis.

[0003] In the Chinese invention patent with the application publication number CN101301538A, a fuzzy intelligent control system for adding water treatment photosensitizer is disclosed. The method adopted is: first, according to the current state and operation amount of the controlled object, an online prediction model of the controlled object is established, and the control result of maintaining the current operation amount is predicted by this model. In the control target evaluation, then, based on the current state and prediction result of the controlled object, the control effect of the system is evaluated and the result of adjusting the fuzzy control weight is generated, and the fuzzy control can thus determine the optimal operation increment.

[0004] However, in practical applications, the existing water disinfection control systems generally face the problems of response lag and inaccurate adjustment. Especially when using the photodynamic inactivation technology, the photosensitizer and light source irradiation parameters in water must be strictly matched to generate sufficient reactive oxygen species (ROS) to inactivate pathogens. However, traditional control strategies often use fixed set values or simple PID regulation, and cannot respond in real time to the sudden changes of key parameters such as pH, dissolved oxygen, and temperature caused by natural environmental changes, industrial load fluctuations, or aquaculture operation adjustments. In this case, too low photosensitizer dosing or insufficient light will lead to insufficient pathogen inactivation, while too high dosing or light may cause excessive ROS, thereby damaging the tissues of aquatic animals and disrupting the water ecological balance. More critically, the dimensional inconsistency and non-linear characteristics between multi-source sensor data make it difficult for traditional closed-loop control methods to simultaneously consider inactivation efficiency, energy consumption, and safety, and may ultimately lead to pathogen diffusion, unstable disinfection effect, and sharp increase in energy consumption.

[0005] To this end, the present invention provides a method for adjusting the parameters of photodynamic inactivation based on hybrid fuzzy-PID control. Summary of the Invention

[0006] (I) Technical problems to be solved

[0007] Aiming at the deficiencies of the prior art, the present invention provides a method for adjusting the parameters of photodynamic inactivation based on hybrid fuzzy-PID control. By fusing multi-sensors and predicting the environment, the dynamic water quality is grasped; then, the fuzzy-PID adaptive dosing technology is coupled with the lighting strategy to achieve precise matching of the photosensitizer dose and light intensity; furthermore, real-time closed-loop and multi-scenario collaborative scheduling are used to cope with water body fluctuations and pathogen diffusion; multi-dimensional indicators are used for evaluation and reinforcement learning is introduced for self-learning optimization to continuously improve the dosing and lighting parameters; finally, it is extended to large-scale distributed control and cloud big data training, which can maintain high-efficiency inactivation and flexible deployment in different aquaculture scenarios. It can effectively avoid the rapid spread of pathogens in adverse environments and reduce the damage to aquatic animals caused by excessive accumulation of ROS, thus solving the technical problems proposed in the background art.

[0008] (II) Technical solutions

[0009] To achieve the above objectives, the present invention is realized through the following technical solutions:

[0010] A method for adjusting the parameters of photodynamic inactivation based on hybrid fuzzy-PID control includes: when significant fluctuations simultaneously occur in key indicators such as the pH value, dissolved oxygen O2, and temperature of the water body, receiving the output of multi-source sensors and performing error correction and redundancy detection on the information of each channel through weighted filtering to generate an environmental fusion data vector. An adaptive prediction model is constructed based on this vector to identify extreme water quality changes and output the environmental trend in the future short time period for subsequent dosing and lighting strategy formulation, thereby enabling early warning in the stage of water quality mutation;

[0011] After generating and outputting the environmental fusion data vector, comparing the optimal activity range of the photosensitizer and the distribution of pathogens, calculating the dosing amount of the photosensitizer using the adaptive formulation dosing technology, and formulating a lighting plan according to the environmental cooperation coefficient and the water body dynamic prediction result to intensify the irradiation at local hot spots and reduce the energy consumption in the normal area. By coupling the dosing amount - dosing time - irradiation, the inactivation efficiency is maximized and the damage to the surface of aquatic animals is reduced, avoiding the harm caused by excessive accumulation of reactive oxygen species ROS due to excessive lighting or excessive photosensitizer;

[0012] After the photosensitizer is dispensed and light irradiation begins, if the sensor detects that the environmental condition data significantly deviates from the target again, real-time closed-loop regulation is performed through a hybrid fuzzy-PID. The scene difference vector is used to adapt to the water body regulation requirements. When the water body environment fluctuates violently or the risk of pathogen spread increases, local irradiation is strengthened in a timely manner or the isolation device is activated. After each control cycle ends, the deviation vector and compensation instructions are retained as data;

[0013] When the photodynamic inactivation process ends or reaches the expected duration, the treatment effect is comprehensively evaluated based on the aforementioned compensation instructions and control feedback indicators. The replay pool mechanism in deep reinforcement learning is used to train each state-action-reward sequence, update the adaptive dispensing and light irradiation strategy model. The data curve generated during this evaluation process is combined with the scene difference vector to summarize the key environmental impact factors and solidify the optimized parameters into the knowledge base;

[0014] In large-scale aquaculture scenarios with numerous partitions or across regions, local closed-loop scheduling is achieved by deploying edge control units in each sub-region. The environment integration data and control instruction logs are regularly uploaded to the cloud, and deep learning algorithms are used for multi-objective training to form a general decision-making model, which is then sent to each edge control unit for synchronous execution.

[0015] Preferably, the sensor array deployed in the aquaculture water body collects environmental condition parameters at a high frequency, performs redundant detection and error correction on multi-source sensor data, and then uses a custom fusion formula to form a fusion output The fusion formula is based on an exponential decay mechanism for the differences between sensors, and each environmental condition parameter is separately subjected to the above-mentioned fusion process once to obtain the fused data Rh.

[0016] Preferably, after obtaining the output fused data, an adaptive prediction model is used to construct a dynamic environment model Using a deep neural network or a fuzzy neural network, the trend of environmental condition parameters in the short term is predicted; the historical data and their fused values of each environmental condition parameter are input through multiple channels;

[0017] At time t, the fused data Rh and the fused values of several past moments are input into the trained or deployed neural network together to obtain the environmental prediction information of multiple parameters in the next period;

[0018] After each prediction is completed, the prediction result is stored in the environmental prediction database, and an environmental state model is generated If it is monitored that some parameters are about to break through the safety threshold, early intervention is carried out;

[0019] Preferably, based on the obtained environmental prediction information, the precise dispensing amount and dispensing timing of the photosensitizer are determined, and the basic dispensing amount of the photosensitizer is defined and environmental sensitivity parameters; when it is predicted that the environmental condition parameters will deviate significantly from the ideal range of the photosensitizer, the dosing amount is automatically reduced or the dosing is delayed;

[0020] Introduce an environmental cooperation coefficient to characterize the suitability of the water body for the dosing of the photosensitizer in the future time;

[0021] After obtaining the environmental cooperation coefficient , calculate the dosing amount of the photosensitizer , and implement the corresponding dosing according to the working mode of the dosing device. If the environmental cooperation coefficient is too low, the dosing time is appropriately delayed, or the auxiliary device is first called to correct the water quality, and then the dosing is carried out.

[0022] Preferably, after the determination and execution of the dosing amount of the photosensitizer , combined with the concentration distribution of the photosensitizer in the water body at the current moment and the predicted light environment and other auxiliary indicators, then formulate a light strategy; the light strategy needs to ensure the matching degree of the photosensitizer and light, and perform refined irradiation on the identified local hot spot areas, and perform enhanced irradiation on the areas with higher pathogen concentration through a mobile light source or an optical fiber array;

[0023] Define a light optimization function to determine the optimal cooperation relationship between the irradiation intensity and the irradiation duration . After the function is executed, the actual photosensitizer distribution state is used as the input variable, and combined with the environmental cooperation coefficient , it can be specifically described as:

[0024]

[0025] In the formula: is the benchmark coefficient of the light strategy, is the dosing amount of the photosensitizer; , are the sensitivity amplification coefficients respectively;

[0026] The light intensity and the light duration can be obtained according to the light optimization function under the following scheduling principles:

[0027] When the light optimization function is higher than the expected value, automatically select a higher light intensity and extend the light duration ; when the light optimization function When it is not higher than the expected value, it tends to reduce the light intensity and moderately shorten the light duration ; The light scheduling can be solved by offline or online optimization algorithms, or can be fine-tuned by a microprocessor in the subsequent real-time closed-loop stage.

[0028] During actual irradiation, spatial partitioning will be carried out according to the pathogen distribution:

[0029] The hot spot area is irradiated with relatively high intensity, and the non-hot spot area is irradiated with conventional intensity or not irradiated; The time interval between dosing and light is determined by the coordination degree of photosensitizer concentration reflected in the light optimization function ; When the expected photosensitizer is fully diffused in the water body and has not significantly decayed, the light can be started;

[0030] Preferably, relying on the achievements of photosensitizer dosing and light parameter setting, comparing with the real-time sensor data of the actual environment, if it is observed that the current state significantly deviates from the set target interval, then enter the closed-loop correction mode;

[0031] Obtain new environmental measurement values at time t to form an instant state vector , compare with the ideal interval vector to obtain an environmental deviation vector :

[0032]

[0033] Adopt a control method based on a non-linear potential function to map the environmental deviation vector to a control signal vector , including light intensity compensation and start / stop scheduling of auxiliary devices , which can be specifically written as:

[0034]

[0035] where, is a non-linear potential function;

[0036] is a mapping operation used to convert the output of the potential function into an instruction value that can directly drive the actuator;

[0037] Light compensation: If the light intensity compensation is positive, increase the light source power; if it is negative, reduce the light source power or shorten the effective irradiation duration ;

[0038] Auxiliary device scheduling: The start / stop scheduling of auxiliary devices can trigger oxygenation equipment, pH buffer dosing units, local water flow isolation devices, etc., and adaptively control according to the deviation source;

[0039] Preferably, according to the previously determined light mode and photosensitizer distribution plan, combined with the new round of sensor data and closed-loop control instructions, coordinated scheduling is carried out in the spatial and temporal dimensions to ensure ideal inactivation effects and aquatic animal safety in each scenario;

[0040] Define the environmental difference vector , and its components can include water flow velocity, aquaculture density, and light interference degree;

[0041] Construct a multi-scenario scheduling function , map the control signal vector jointly with the environmental difference vector to obtain a scenario-based instruction :

[0042] When the water body flow velocity is large, the multi-scenario scheduling function amplifies the light intensity compensation in a partitioned manner to make up for the decrease in photosensitizer concentration caused by water flow dilution;

[0043] When the aquaculture density is large, the multi-scenario scheduling function can call the auxiliary device to perform start-stop scheduling of the auxiliary device for priority enhancement;

[0044] If the light interference degree is too high, the multi-scenario scheduling function automatically suppresses the excessive increase in light source intensity and extends the light duration ;

[0045] The scenario-based instruction is finally sent by the scheduling module to each independent execution subsystem for execution:

[0046] Preferably, according to the sensor records, control instruction records, and final water body pathogen inactivation results generated during the actual execution process, and compared with the set target indicators, systematic multi-dimensional evaluation is carried out, mainly including the following content:

[0047] Record the pathogen clearance rate through a fluorescence imaging device or other pathogen detection modules , record the injury rate of individuals by observing and detecting the body surface injuries of aquatic animals or other health indicators ; record the data of oxygenation equipment, isolation devices, light source power, etc. to obtain the comprehensive energy consumption index ; and monitor after treatment to The recovery of pH, dissolved oxygen, temperature, etc. within a certain period of time, and form and output an effect evaluation report based on the above three types of evaluation results;

[0048] Preferably, using the output effect evaluation report, combined with the recorded execution actions, construct an adaptive optimization mechanism. The mechanism is based on the state-action-reward logic loop, providing comparison and improvement for each round of photodynamic inactivation process, where:

[0049] Define the state And define the action ;

[0050] To balance the pathogen inactivation efficiency, the safety of aquatic animals, and the energy consumption, define a comprehensive reward function ;

[0051] The larger the value, the better the balance obtained by the current action combination among the inactivation efficiency, the safety of aquatic animals, and the energy consumption; otherwise, corresponding strategy corrections must be made in the next round of training; Store the execution sequence of each cycle Into the replay pool, and form a mapping strategy after iterative training ;

[0052] Preferably, store the policy model parameters and key evaluation data obtained from this round of training in the knowledge base: Store the output of the reinforcement learning model separately according to different scenario labels, and retain the mapping relationship of each key variable for easy and quick matching of similar scenarios; In specific extreme situations, allow the parallel use of expert rules and the reinforcement learning model and store them in the same knowledge base. If an extreme state occurs, the expert rules are preferentially triggered to prevent serious accidents;

[0053] Before the start of a new batch of aquaculture, automatically load the corresponding set of policy parameters according to the scenario label, and preset the initial photosensitizer dosage 、environmental cooperation coefficient threshold, and light optimization plan, etc. If there are major environmental changes, start supplementary fine-tuning;

[0054] Preferably, deploy edge control units within each sub-region and quickly perform the following operations within a local range:

[0055] Each edge control unit receives the instant environmental condition data within the sub-region and uses the updated expert knowledge base or reinforcement learning strategy to perform edge-level rapid adaptive adjustment on the feedback control strategies such as the photosensitizer dosage 、irradiation intensity and irradiation duration ;

[0056] If the environmental cooperation coefficient Or when there are local extreme deviations in the pathogen distribution within the sub - partition, the edge control unit responds quickly under the condition of low latency; when multiple partitions are adjacent, each edge control unit exchanges key status information and coordinates the implementation of cross - partition isolation or in - partition water flow regulation;

[0057] Each edge control unit makes the first - layer quick decision in the local environment and at the same time transmits the periodically summarized data to the main control center, where the multi - partition information is fused.

[0058] Preferably, the environmental parameters, control actions and evaluation indicators generated by each edge control unit during local operation are regularly uploaded to the cloud big data platform and auxiliary information is attached;

[0059] Build a deep - learning framework for multi - objective optimization on the cloud platform, including a multi - task network structure or a multi - head output layer;

[0060] Perform large - scale batch training on the collected data based on the generalization function:

[0061] After training is completed, the cloud will send the new cloud decision model to each edge control unit;

[0062] If the breeding environment of a certain sub - partition or area highly matches the cloud training data, the cloud decision model can be directly adopted ; if there are still local differences, secondary fine - tuning can be performed locally at the edge control unit according to the reinforcement learning mechanism;

[0063] Preferably, based on the cloud decision model and the data of the partition edge control unit, cross - scenario data fusion and global policy iteration are carried out for different breeding scenarios, and the global optimal policy is regularly formed and fed back into the self - learning process;

[0064] In the proposed environmental difference vector to quantify elements such as flow velocity, breeding density, light interference, etc., and uniformly process the environmental difference vectors reported by different edge control units and compare with the cloud model to identify possible blind areas of scenario differences, supplement new feature dimensions or correct existing feature weights;

[0065] Combine the above - mentioned scenario difference information with the generated cloud decision model and adopt the distributed policy iteration algorithm:

[0066]

[0067] Where: is the distributed fusion function, which is used to adjust the cloud decision model after summarizing the feedback results of each edge control unit Fine-tuning is performed to generate a global decision-making model 。

[0068] (III) Beneficial effects

[0069] The present invention provides a method for adjusting the parameters of photodynamic inactivation based on hybrid fuzzy-PID control, which has the following beneficial effects:

[0070] The dosing amount of photosensitizer and the light parameters are dynamically optimized with high coupling. Through the hybrid fuzzy-PID algorithm, non-linear adjustment can be quickly made in response to fluctuations in water environment conditions, avoiding both the waste of resources caused by traditional fixed ratios and the damage to aquatic animals and pollution in extreme environments;

[0071] The environmental cooperation coefficient and the environmental difference vector are fully applied in real-time closed-loop control and multi-scenario scheduling: they seamlessly connect the obtained prediction model and dosing strategy parameters with online decision-making, enabling scenarios such as high-density indoor or net cage mobility to achieve differential disinfection and water pollution treatment through local water flow isolation or enhanced irradiation of hot spots, enhancing the system adaptability;

[0072] Effect evaluation and self-learning optimization introduce a reinforcement learning mechanism, balancing indicators such as pathogen inactivation rate, aquatic animal injury rate, and energy consumption. Through the comprehensive evaluation of the replay pool training and multi-objective reward function, continuous iterative improvement of the feedback control strategy is achieved;

[0073] The linkage between the edge control unit and the cloud platform gives the solution strong scalability: each edge control unit can execute the knowledge base and the cloud decision-making model locally, with fast edge closed-loop; the cloud integrates the current state data and control feedback data for in-depth training, extracts general strategies or models and then distributes them again to complete the global strategy iteration. This hierarchical collaborative method enables hybrid fuzzy-PID and reinforcement learning to cover diverse applications across regions or seasons rather than being limited to a single water body;

[0074] Based on the above, technical element combinations and synergistic effects are achieved in photosensitizer dosing and light control, environmental perception and real-time scheduling, as well as self-learning and distributed collaboration, significantly improving the accuracy, response speed, and ecological safety of aquaculture photodynamic inactivation. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] Figure 1 It is a schematic flow diagram of the method for adjusting the parameters of photodynamic inactivation based on hybrid fuzzy-PID control of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0076] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0077] Please refer to Figure 1 , the present invention provides a method for adjusting the parameters of photodynamic inactivation based on hybrid fuzzy-PID control, including,

[0078] Step 1: When significant fluctuations occur simultaneously in key indicators such as the pH, O2, and temperature of the water body, receive the output of multi-source sensors and perform error correction and redundancy detection on the information of each channel through weighted filtering to generate an environmental fusion data vector. Based on this vector, construct an adaptive prediction model to identify extreme water quality changes and output the environmental trend in the future short time period for subsequent dosing and light strategy formulation, so as to issue an early warning in advance during the water quality mutation stage;

[0079] The content of the above Step 1 is as follows:

[0080] Step 101: Multi-sensor data acquisition and fusion

[0081] The sensor array deployed in the aquaculture water body acquires the environmental condition parameters, including pH, dissolved oxygen O2, water body temperature T, and light intensity at a high frequency, and all measurement readings are uploaded to the information processing unit; to enhance the accurate identification of external mutations (such as sudden climate changes, inlet and outlet water fluctuations, etc.), first perform redundancy detection and error correction on the multi-source sensor data, and then use a custom fusion formula to form a fusion output , and the fusion formula is based on an exponential decay mechanism for the differences between sensors, used to introduce an inhibition factor between readings with large differences, thereby improving the reliability of the overall measurement result;

[0082] Specifically, assume that at time t there are effective observation values sent back by sensors, which are respectively: ,

[0083] where each can come from different types of sensors (such as pH sensors, dissolved oxygen sensors, etc.). First, perform dimensionless or normalization on the readings of different sensors, and define the fusion output as follows:

[0084]

[0085] In the formula: is the basic weight coefficient of the sensor, is the exponential amplification factor, is the difference suppression factor, and their values are all greater than 0; and : represents the distance between the maximum and minimum values of all sensor readings at time t, which is used to measure whether the degree of discreteness of the readings is too large; is the absolute value operator, which is used to calculate the amplitude of the numerical difference;

[0086] Perform the above-mentioned fusion process on the environmental condition parameters respectively to obtain the fused data Rh:

[0087]

[0088] The fusion result will be integrated into the unified dynamic environment model in the next step;

[0089] When in use, in a multi-sensor environment, the use of difference suppression and weighted processing can effectively filter out abnormal data points and improve the reliability of the final perception result; when there are drastic changes in the outside world, the exponential decay term avoids the excessive influence of abnormal readings on the overall result and ensures the stability of subsequent decisions. Introducing an exponential mechanism for difference suppression between multiple sensors avoids being too sensitive to outliers when only using simple weighted averaging.

[0090] Step 102, Dynamic environment modeling

[0091] After obtaining the output fused data, use an adaptive prediction model to construct a dynamic environment model , using a deep neural network or a fuzzy neural network, to predict the trends of environmental condition parameters in the short term, such as pH, dissolved oxygen O2, water temperature T, and light intensity In order to characterize the complex situation of the mutual coupling of multi-variables in the water environment, the historical data and their fused values of each environmental condition parameter are used for multi-channel input, and the model output can be subdivided into:

[0092]

[0093] Among them: is the prediction step (such as 30 minutes or 1 hour, etc.), and the obtained prediction results constitute the most core part of the dynamic environment model, and will be continuously compared with the real-time collected data, so as to be continuously updated and iterated;

[0094] At time t, input the fused data Rh and the fused values of several past moments into the trained or deployed neural network to obtain the environmental prediction information of multiple parameters in the next time period; after each prediction is completed, store the prediction results in the environmental prediction database and generate an environmental state model , which is used to describe the trend of key indicators in a short period of time in the future. If it is detected that certain parameters are about to break through the safety threshold, early intervention will be carried out;

[0095] Through the deep feature extraction ability of machine learning models or fuzzy neural networks, it is possible to more accurately identify potential abnormal extreme weather or water quality fluctuation trends, providing a time window for preparing photosensitizer dosing or starting auxiliary equipment in advance; with the arrival of new batches of sensor fusion data, the dynamic environment model continuously performs rolling predictions to ensure continuous tracking of water body changes; once the prediction results show that indicators such as pH or dissolved oxygen will tend to the critical range, more refined dosing amounts and light intensities can be formulated to reduce the potential damage risk to aquatic animals.

[0096] Step Two: Complete the generation and output of environmental fusion data vectors, compare the optimal activity range of the photosensitizer and the pathogen distribution, calculate the photosensitizer dosing amount using the adaptive formulation dosing technology, and formulate a light plan based on the environmental cooperation coefficient and the water body dynamic prediction results, so as to intensify irradiation at local hot spots and reduce energy consumption in normal areas, maximize the inactivation efficiency and reduce the surface damage of aquatic animals through the coupling of dosing amount - dosing time - irradiation, and avoid the accumulation hazard of reactive oxygen species ROS caused by excessive light or excessive photosensitizer;

[0097] The said Step Two includes the following contents:

[0098] Step 201: Adaptive photosensitizer dosing

[0099] Based on the obtained environmental prediction information, determine the precise dosing amount and dosing timing of the photosensitizer to enable the photosensitizer PS to produce an ideal inactivation effect in the subsequent light irradiation stage. To achieve adaptive dosing:

[0100] Define the basic dosing amount of the photosensitizer PS and the environmental sensitivity parameter , which is used to measure the degree to which the aquaculture water body deviates from the optimal activity range at present and in the short term;

[0101] When it is predicted that environmental condition parameters such as pH, dissolved oxygen or temperature will deviate significantly from the ideal range of the photosensitizer, appropriately reduce the dosing dose or delay the dosing time to avoid the generation of excessive reactive oxygen species in a harsh environment:

[0102] Introduce the environmental cooperation coefficient , which is used to characterize the suitability of the water body for photosensitizer dosing after time. This coefficient is related to the predicted deviations of the first three main indicators (pH, dissolved oxygen, temperature), and is defined as follows:

[0103] In the time interval , consider the three main predicted indicators of the water body: , , ;

[0104] wherein is any moment within the interval, is the prediction step. Additionally, the ideal reference values , , are preset respectively, and the readings of different sensors are made dimensionless or normalized; the environmental deviation vector and the sensitivity weight matrix W are defined as follows:

[0105]

[0106] This vector is used to simultaneously characterize the instantaneous deviations of the three indicators of pH, dissolved oxygen, and temperature from the ideal values;

[0107]

[0108] wherein the sensitivity parameters , , respectively measure the key importance of pH, dissolved oxygen, and temperature to the aquaculture environment and the amplification weights of the deviation degree, and usually take positive values;

[0109] The weighted environmental deviation is measured using the Euclidean norm ( norm), and it is integrated and accumulated within to obtain the environmental deviation cumulative function :

[0110]

[0111] where:

[0112]

[0113] Based on this deviation cumulative function, the defined environmental cooperation coefficient is:

[0114]

[0115] In the formula: is a three-dimensional deviation vector, which respectively reflects the instantaneous differences of pH, dissolved oxygen, and temperature from the ideal values; W is a diagonal matrix, and the main diagonal elements are respectively , , , and can be adjusted according to the aquaculture species and environmental characteristics; is the Euclidean norm; is the integral of the weighted deviation within the prediction interval;

[0116] After obtaining the environmental cooperation coefficient calculate the photosensitizer dosage , and its formula is:

[0117] In the formula: is the basic dosage, representing the dosage of photosensitizer PS defaultly put under ideal environment;

[0118] is used to map to the interval for non-linear correction of the basic dosage;

[0119] When the environmental cooperation coefficient is small (indicating that the future environment is poor), this item will significantly reduce the final dosage, or reduce it to a very low value in extreme cases;

[0120] After completing the calculation of the photosensitizer dosage , implement the corresponding dosing according to the working mode of the dosing device (such as an adjustable pump or a metering adder). If the environmental cooperation coefficient is too low, the dosing time should be appropriately delayed according to the actual aquaculture needs, or the auxiliary equipment (such as aeration or pH adjustment) should be called first for water quality correction, and then the dosing should be carried out.

[0121] When in use, by introducing the environmental cooperation coefficient , the dosage and dosing timing of the photosensitizer can be automatically fine-tuned according to the quality of the short-term predicted environment, reducing waste and environmental risks. When the environmental indicators are extremely abnormal, it will actively reduce or postpone the dosing to avoid generating excessive ROS in an unsuitable environment; the concentration of the photosensitizer after dosing is more likely to be in the optimal range, and the subsequent light parameters can be formulated more efficiently and the pathogen inactivation can be triggered, making the photosensitizer dosing have an adaptive attenuation characteristic in the face of large deviations, far superior to the traditional fixed-dose addition method. Combining the short-term prediction data with the non-linear exponential model avoids over-dosing or untimely dosing caused by relying solely on threshold judgment;

[0122] Step 202, Formulation of lighting strategy

[0123] After determining and implementing the photosensitizer dosage , formulate the lighting strategy in combination with the concentration distribution of the photosensitizer in the water body at the current moment, the predicted lighting environment and other auxiliary indicators; the lighting strategy should ensure the matching degree of the photosensitizer and light (wavelength, intensity, irradiation duration), avoid continuous over-strong irradiation causing heat stress or light stress of aquatic animals, and perform refined irradiation on the identified local hot spot areas, and perform enhanced irradiation on the areas with higher pathogen concentration through a mobile light source or an optical fiber array;

[0124] Define the light optimization function to determine the irradiation intensity and the irradiation duration between the optimal matching relationship. After the function is executed, the actual photosensitizer distribution state is used as the input variable, and combined with the environmental cooperation coefficient , which can be specifically described as:

[0125]

[0126] In the formula: is the benchmark coefficient of the light strategy; is the dosage of the photosensitizer; , are the sensitivity amplification coefficients of the photosensitizer and respectively, and their values are all positive;

[0127] The light intensity and the light duration can be obtained according to the light optimization function under the following scheduling principles:

[0128] When the light optimization function is higher than the expected value (indicating that both the dosage of the photosensitizer PS and the environmental cooperation degree are relatively good), automatically select a higher light intensity and extend the light duration to eliminate pathogens as soon as possible within the controllable range;

[0129] When the light optimization function is not higher than the expected value, it tends to reduce the light intensity and moderately shorten the light duration to avoid additional stress on aquatic animals caused by excessive irradiation, or wasteful irradiation in the case of insufficient photosensitizer concentration;

[0130] This light scheduling can be solved by offline or online optimization algorithms (such as greedy algorithms, genetic algorithms, etc.), and can also be fine-tuned by a microprocessor in the subsequent real-time closed-loop stage.

[0131] When implementing the actual irradiation, spatial partitioning will be carried out according to the pathogen distribution:

[0132] The hot spot area is irradiated with relatively high intensity, and the non-hot spot area is irradiated with conventional intensity or not irradiated. The time interval between dosing and irradiation is determined by the light optimization function Determined by the concentration matching degree of the photosensitizer reflected therein; when the expected photosensitizer is fully diffused in the water body and not significantly attenuated, the light irradiation can be started in a timely manner; among them, the pathogen concentration or related indicators in each area are obtained in real time through sensors (such as fluorescence imaging, microbial counting or other rapid detection means) and compared with the preset threshold. When the detected value in a certain area exceeds the threshold, this area is defined as the hot spot area.

[0133] During use, the coordinated delivery-irradiation is realized. and Key quantities such as etc. directly participate in the light irradiation decision-making, avoiding the waste or inefficiency caused by the separate control of the photosensitizer and irradiation in the traditional way; reducing environmental and biological stress, realizing the dynamic inhibition of over-irradiation by appropriately amplifying or shortening the light irradiation duration, and can be adjusted preventively according to the tolerance threshold of aquatic animals to improve the pathogen killing efficiency: the enhanced irradiation in the hot spot area and the conventional irradiation in the general area form a differential strategy, taking into account comprehensive killing and key strikes.

[0134] Through the linkage of adaptive photosensitizer delivery and light irradiation strategy formulation, the multi-dimensional environmental cooperation coefficient can not only accurately schedule the delivery amount, but also integrate with the photosensitizer distribution into the light irradiation strategy, generating a strong synergistic effect. This synergistic mechanism enables the photosensitizer and the light irradiation link to support and balance each other, maintaining both the efficient attack on pathogens and reducing the stress of aquatic animals and resource waste.

[0135] Step 3: After the photosensitizer has been delivered and the light irradiation has started, if the sensor detects that the environmental condition data deviates significantly from the target again, through the hybrid fuzzy-PID real-time closed-loop regulation, and use the scenario difference vector to adapt to the water body regulation requirements, strengthen the local irradiation in a timely manner or start the isolation device when the water body environment fluctuates violently or the risk of pathogen diffusion increases. After each control cycle ends, the deviation vector and compensation instruction are retained for data.

[0136] The said step 3 includes the following contents:

[0137] Step 301: Real-time closed-loop control

[0138] Relying on the achievements of photosensitizer delivery and light irradiation parameter setting, comparing with the real-time sensor data of the actual environment, if it is observed that the current state deviates significantly from the set target interval (taking or etc. as references), then immediately enter the closed-loop correction mode, which is executed by a microprocessor (such as STM32 or other high-performance embedded controllers), and the core logic is as follows:

[0139] Obtain a new environmental measurement value at time t to form an instant state vector :

[0140]

[0141] Compared with the previous stage according to or the ideal interval vector expected by the light irradiation strategy a comparison is made to obtain an environmental deviation vector :

[0142]

[0143] Adopt a control method based on a non - linear potential function to map the environmental deviation vector to a control signal vector , including light intensity compensation and start - stop scheduling of auxiliary devices , which can be specifically written as:

[0144]

[0145] where is a non - linear potential function used to amplify or suppress the magnitude of the deviation vector in segments;

[0146] is a mapping operation used to convert the output of the potential function into an instruction value that can directly drive the actuator. For example, when the absolute value of the deviation is small, the signal can be appropriately amplified to achieve agile adjustment; when the deviation is extremely large, it can limit overly drastic corrections within a safe limit to avoid causing secondary stress to aquatic animals;

[0147] Light compensation: If the light intensity compensation is positive, the power of the light source is increased; if it is negative, the power of the light source is reduced or the effective irradiation duration is shortened ;

[0148] Auxiliary device scheduling: The start - stop scheduling of auxiliary devices can trigger oxygenation equipment, pH buffer dosing units, local water flow isolation devices, etc., and perform adaptive control according to the source of the deviation. For example, when the pH is too low, buffer adjustment is triggered; when the dissolved oxygen is insufficient, oxygenation equipment is started, etc.;

[0149] During use, through the above - mentioned closed - loop control function, the environmental parameter errors are non - linearly and quickly corrected in each control cycle, which not only improves the immediate efficiency of photodynamic inactivation but also reduces the risk of rapid spread of pathogens due to local environmental deterioration; close connection with the light irradiation strategy: all correction actions are carried out around the ideal interval vector to achieve an adaptive linkage of dosing - irradiation - control; the introduction of a non - linear potential function To handle the deviation amount, it far exceeds the conventional linear PID or the method based only on variance detection, and has higher response flexibility.

[0150] Step 302, Multi-scenario collaborative scheduling

[0151] Furthermore, aiming at the characteristic differences in different scenarios such as indoor high-density aquaculture, outdoor pond aquaculture, and cage flow water aquaculture, a cross-scenario scheduling strategy is implemented. The core of this strategy is: according to the previously determined light mode and photosensitizer distribution plan, combined with the new round of sensor data and closed-loop control instructions, collaborative scheduling is carried out in the spatial and time dimensions to ensure that ideal inactivation effects and aquatic animal safety levels can be obtained in each scenario;

[0152] Define the environmental difference vector , and its components can include specific scenario factors such as water flow velocity, aquaculture density, and degree of light interference. For example:

[0153]

[0154] Among them: represents the water flow velocity (larger in cage aquaculture), represents the aquaculture density (larger in indoor high-density systems), represents the degree of external light interference (more obvious in outdoor ponds);

[0155] Construct a multi-scenario scheduling function , and jointly map the control signal vector and the environmental difference vector to obtain the scenario-based instruction :

[0156]

[0157] When the water flow velocity is large (i.e., in the cage water area), the multi-scenario scheduling function amplifies the light intensity compensation in a partitioned manner to make up for the decrease in photosensitizer concentration caused by water flow dilution;

[0158] When the aquaculture density is large (indoor high-density scenario), the multi-scenario scheduling function can prioritize the start-stop scheduling of the auxiliary device call to prevent local outbreaks in crowded areas (fish populations) where pathogens are prone to spread rapidly;

[0159] If the degree of light interference is too high (outdoor ponds are strongly affected by sunlight), the multi-scenario scheduling function Automatically suppress the excessive increase in light source intensity and extend the light duration , to reduce the risk of overheating or light burns in aquatic animals;

[0160] Scene-based instructions Finally, it is sent by the scheduling module to each independent execution subsystem and executed:

[0161] Mobile lighting device: Perform position tracking and irradiation power adjustment according to the local water flow velocity in the net cage or pond; Local enclosure / isolation device: Start a temporary surrounding network in high-flow scenarios to reduce the spread of pathogens; Aeration and water quality regulator: Prioritize the allocation of more power and resources in high-density scenarios; External light compensation / shielding equipment: If outdoor interference is too strong, add sunshade or reflection devices;

[0162] During use, through the introduction of the environmental difference vector , the same set of closed-loop control strategies can make differentiated execution plans according to the core characteristics of the actual aquaculture scenario. Combining the control signal vector with the environmental difference vector , dynamic allocation can be achieved for different regions and different time periods, enhancing resource utilization and inactivation accuracy; The same control signal vector only needs to add a multi-scenario scheduling function on the outer layer to adapt to different scenarios and reduce the need for system hardware modification; The multi-scenario scheduling function uniformly processes multiple factors such as flow velocity, density, and light interference in vector form, and no longer relies on simple empirical thresholds but realizes more delicate control allocation through a mapping mechanism.

[0163] Step 4: When the photodynamic inactivation process ends or reaches the expected duration, based on the aforementioned compensation instructions and control feedback indicators, summarize and evaluate the treatment effect, and use the replay pool mechanism in deep reinforcement learning to train each state-action-reward sequence, update the adaptive dosing and lighting strategy model, combine the data curve generated during this evaluation process with the scenario difference vector, to summarize the key environmental impact factors and solidify the optimized parameters into the knowledge base;

[0164] The said Step 4 includes the following contents:

[0165] Step 401: Multi-dimensional effect evaluation

[0166] Based on the sensor records, control instruction records, and the final water body pathogen inactivation results generated during the actual execution process, conduct a systematic multi-dimensional evaluation against the set target indicators, mainly including the following contents:

[0167] Record the pathogen clearance rate through a fluorescence imaging device or other pathogen detection modules ;

[0168] By observing and detecting the surface damage or other health indicators of aquatic animals, record the injury rate of individuals ;

[0169] Record data such as oxygenation equipment, isolation devices, and light source power, and obtain the comprehensive energy consumption index , representing the total energy consumption of photosensitizer dosing, light irradiation, and various auxiliary devices;

[0170] And monitor the recovery of pH, dissolved oxygen, temperature, etc. after treatment to within the time;

[0171] If the water quality can quickly return to the safe range, and at the same time the comprehensive energy consumption index remains at a relatively low level, it indicates that this round of treatment is more efficient in terms of energy conservation and ecological balance; form and output an effect evaluation report based on the above three types of evaluation results;

[0172] When in use, the inactivation rate of pathogens and the safety of aquatic animals can be considered simultaneously, and the energy consumption and water quality recovery status can be concerned to achieve a comprehensive evaluation of the entire treatment process.

[0173] Step 402, Deep Self-Learning Optimization

[0174] Using the output effect evaluation report, combined with the recorded execution actions, construct an adaptive optimization mechanism. The mechanism is based on the state-action-reward logic loop, providing comparison and improvement for each round of photodynamic inactivation process, where:

[0175] Define the state , which can include:

[0176]

[0177] At the same time, information such as multi-environment difference vectors can also be added to ensure that the complete environmental and operation characteristics are reflected in the reinforcement learning model;

[0178] Define the action :

[0179] The action refers to the adjustment of photosensitizer dosing and light irradiation parameters, such as changing the correction amount of photosensitizer dosing , fine-tuning the light intensity compensation value and the light duration compensation value , or controlling the start and stop of auxiliary devices, etc. All executed control instructions can be recorded as an action instance.

[0180] To balance the pathogen inactivation efficiency, the safety of aquatic animals, and the energy consumption, define a comprehensive reward function , extract key points from the evaluation metrics:

[0181]

[0182] In the formula: , , is the balance coefficient, all are positive numbers, and the value can be between 0.1 and 10, which is used to adjust the relative importance of the three indicators in the reward function;

[0183] , , is the non-linear amplification factor, all are positive numbers, and the value can be between 0.5 and 5;

[0184] The larger the value, the better the balance obtained by the current action combination between the inactivation efficiency, the safety of aquatic animals, and the energy consumption; otherwise, corresponding strategy corrections need to be made in the next round of training;

[0185] Store the execution sequence of each cycle into the replay pool. After iterative training through methods such as Q-learning, deep Q-network (DQN), and policy gradient, finally form a mapping strategy: ;

[0186] When in use, through the comprehensive reward function a better compromise can be found among pathogen elimination, animal safety, and energy consumption, avoiding local optima. The new data after each round of treatment will participate in the next round of training, and the algorithm will continuously evolve and correct unreasonable dosing or irradiation methods. The environmental difference vector is also incorporated into the reinforcement learning model to make the strategy adaptable to multiple scenarios such as indoor high density, outdoor water bodies, and the fluidity of cages.

[0187] Step 403, Knowledge base update and pre-configuration for the next batch of applications

[0188] Store the policy model parameters and key evaluation data obtained from this round of training in the knowledge base for direct invocation in the next batch or the next cycle of aquaculture practice. Among them: store the output of the reinforcement learning model (or neural network weights, Q-table, etc.) separately according to different scenario labels (such as indoor, outdoor, cage), and retain the mapping relationship of each key variable for quick matching of similar scenarios;

[0189] For situations where human or expert experience still needs to be involved in specific extreme situations (such as severe dissolved oxygen deficiency, extreme temperature deviation, etc.), it is allowed to use expert rules (such as limit thresholds or forced start / stop strategies) in parallel with the reinforcement learning model and store them in the same knowledge base. If a similar extreme state occurs, the expert rules can be triggered first to prevent serious accidents;

[0190] Before the start of a new batch of farming, the corresponding set of policy parameters is automatically loaded according to the scenario tags, and the initial dosage of photosensitizer is preset. and the environmental cooperation coefficient threshold, as well as the light optimization plan, etc. If there are major environmental changes, supplementary fine-tuning is initiated;

[0191] During use, new empirical data and the results of the reinforcement learning model are generated in each treatment process. The knowledge base is continuously updated in a rolling manner, and the overall system capabilities are enhanced over time. There is no need to rebuild the model from scratch before each start. The optimal strategy in a similar scenario can be directly imported, significantly shortening the preparation cycle. In cases of scarce data or extreme anomalies, expert rules can complement the self-learning model to ensure safety and controllability.

[0192] Step Five: In large-scale farming scenarios with numerous partitions or across regions, local closed-loop scheduling is achieved by deploying edge control units in each sub-region, and the environmental integration data and control instruction logs are regularly uploaded to the cloud. Deep learning algorithms are used for multi-objective training to form a general decision-making model, which is then sent to each edge control unit for synchronous execution;

[0193] The content of the above Step Five includes the following:

[0194] Step 501: Multi-node hierarchical control architecture

[0195] Regarding the problem that when the farming scale is large, there are many partitions, or the geographical locations are discrete, the real-time feedback speed and load capacity of a single main control center may be difficult to meet the high-frequency decision-making requirements, a multi-node hierarchical control architecture is proposed;

[0196] Specifically, edge control units are deployed within each sub-partition (or local farming unit) to quickly perform the following operations within a local range:

[0197] Each edge control unit receives the instant environmental condition data (pH, O2, T, etc.) within the partition, and uses the updated expert knowledge base or reinforcement learning strategy to quickly and adaptively adjust the feedback control strategies such as the dosage of photosensitizer , irradiation intensity and irradiation duration etc. at the edge level;

[0198] If there are local extreme deviations in the environmental cooperation coefficient or pathogen distribution within the sub-partition, the edge control unit quickly responds under low-latency conditions to avoid the time delay caused by remote instructions from the main control center;

[0199] When multiple partitions are adjacent, each edge control unit exchanges key status information (such as whether to start the aeration equipment, whether the pathogen breaks out and spreads in adjacent partitions, etc.), and coordinates the implementation of cross-partition isolation or in-partition water flow regulation: This coordination is usually achieved through a lightweight message exchange protocol (such as MQTT or a custom protocol);

[0200] Each edge control unit makes the first layer of rapid decisions in the local environment, and at the same time transmits the periodically summarized data (including the local inactivation rate , damage rate , energy consumption etc.) to the main control center. The main control center fuses the multi-partition information and makes more macroscopic full-field decisions, such as uniformly allocating the upper limit of the light source power, the dosage of photosensitizer, or comparing the water quality differences between different administrative regions.

[0201] During use, the edge control unit performs high-frequency closed-loop regulation locally to reduce the delay or network instability impact caused by long-distance communication. The failure of the edge control unit in any one partition will not immediately paralyze the entire system, and the main control center can dispatch other partitions to cooperate or implement an emergency mode.

[0202] The first layer of rapid decision-making refers to that after each edge control unit (ECU) locally collects sensor data in real time, it immediately analyzes and compares the local water quality parameters (such as pH, dissolved oxygen, temperature, etc.) according to the preset closed-loop control strategy and fuzzy-PID algorithm, and then quickly adjusts the light intensity, the dosage of photosensitizer, or activates the operation of auxiliary equipment (such as an aeration pump, isolation device). This decision-making process is completed within milliseconds to seconds, aiming to quickly respond to abnormal changes in the local environment, so as to ensure the overall stability and real-time control performance of the system without waiting for central control or cloud feedback.

[0203] Step 502, Cloud big data training platform

[0204] Centralize and process the data of different sub-partitions or different farms and conduct in-depth learning training to obtain a more general inactivation model and strategy, which can significantly reduce the local training cost of each edge control unit and improve the accuracy when issued later. The key steps are as follows:

[0205] Regularly upload the environmental parameters, control actions, and evaluation indicators (inactivation rate , aquatic animal damage rate , energy consumption etc.) generated by each edge control unit during local operation to the cloud big data platform, and attach auxiliary information such as geographical location, aquatic organism data, and seasonal temperature characteristics, so that the model training process can capture a wider range of environmental differences;

[0206] Build a deep learning framework for multi-objective optimization on the cloud platform, including a multi-task network structure or a multi-head output layer, corresponding to sub-objectives such as inactivation efficiency, aquatic biosafety, and energy consumption control;

[0207] Based on the following exemplary generalization function , perform large-scale batch training on the collected data:

[0208]

[0209] In the formula: , , respectively measure the differences between the true value and the model prediction value in terms of inactivation rate, damage rate, and energy consumption (which can be high-order loss metrics or more advanced non-linear metrics based on logarithmic mapping, etc.), and the values range from 0 to 1;

[0210] , , are adjustable target weights, supporting dynamic priority allocation in different application scenarios , , are the model prediction outputs, used to compare with the true value and backpropagate the gradient; the recommended value range is between 0.1 and 10;

[0211] After training, the cloud will send the new cloud decision model to each edge control unit. Among them, the cloud decision model refers to a general decision model trained through deep learning (such as reinforcement learning, deep Q-network, or policy gradient algorithm) on the cloud big data training platform after integrating the environmental monitoring data, control logs, and evaluation metrics from each edge control unit (ECU); if the aquaculture environment in a certain sub-region or area highly matches the cloud training data, the cloud decision model can be directly adopted to obtain high prediction accuracy; if there are still local differences, secondary fine-tuning can be performed locally at the edge control unit according to the reinforcement learning mechanism;

[0212] During use, by aggregating large-scale data from multiple regions, varieties, and seasons, the trained model has better generalization ability for new scenarios, shortening the local exploration and trial-and-error time; cloud training can utilize high-performance computing resources to process massive data in parallel, reducing the hardware requirements of each edge control unit; as the system deployment scope expands, the cloud platform can retrain and iterate the model regularly, quickly absorbing the latest aquaculture and pathogen prevention and control information globally or nationwide.

[0213] Step 503, Cross-scenario data fusion and global policy iteration

[0214] Cloud-based decision-making model Based on the data of the cloud decision model and the partition edge control unit, cross-scenario data fusion and global policy iteration are carried out for different aquaculture scenarios, further ensuring that in various differentiated scenarios such as indoor high density, outdoor ponds and cages, etc., it can flexibly respond and regularly form a global optimal policy and feedback it to the self-learning process;

[0215] In the proposed environmental difference vector To quantify elements such as flow velocity, aquaculture density, and light interference, and unify the environmental difference vectors reported by different edge control units Carry out unified processing, compare with the cloud model, identify possible blind areas of scenario differences, supplement new feature dimensions or correct existing feature weights;

[0216] Combine the above scenario difference information with the generated cloud decision model Adopt a distributed policy iteration algorithm:

[0217]

[0218] Where: Is a distributed fusion function, used to fine-tune the cloud decision model After summarizing the feedback results of each edge control unit, generate a global decision model ; In the above formula Represents the data provided by the th sub-partition or sub-scenario, corresponding to the indicators obtained by local execution or self-learning.

[0219] The global decision model after iteration Has the comprehensive advantages of taking into account different scenario differences and still maintaining local optimal characteristics on the basis of big data training.

[0220] Finally, the optimization results extracted from the global decision model Will flow back to the reinforcement learning or knowledge base update process, so that subsequent batches can also utilize distributed global experience in a single scenario, accelerate adaptation and reduce exploration costs.

[0221] When in use, fuse the fast feedback ability of the high-density indoor scenario, the rich water quality change data of the outdoor water body, and the fluidity characteristics of the cage scenario to obtain a more comprehensive global decision; every time a new aquaculture partition is connected, it can be systematically incorporated into this distributed architecture, and its data can provide more samples for cloud model training and partition self-learning. Use the distributed policy iteration distributed fusion function Realize the organic integration between the cloud decision model And the local reinforcement learning results of each partition edge control unit, break through the limitations of traditional single-center or single-node learning, and introduce the environmental difference vector The unified management mechanism enables the breeding data in different ecological environments to be cross-verified and enhanced with each other.

[0222] To ensure the consistency of physical dimensions during the calculation processes of the multi-sensor data fusion, environmental cooperation coefficient , photosensitizer dosage and light parameters etc. in the present invention, it is hereby declared that for different parameters involved in relevant formulas or functions, such as pH, dissolved oxygen in mg / L, temperature in °C, light intensity in W / m², etc., unit conversion or normalization processing shall be carried out before use, or they shall be unified by introducing a weight coefficient containing a unit adjustment factor. Specifically, if the pH difference, dissolved oxygen deviation, temperature deviation, etc. are superimposed, exponentiated or multiplied within the same expression, they shall be subtracted from their respective ideal values and then divided by their respective predetermined characteristic scales or normalization factors, so as to map variables in different dimensions to a comparable dimensionless interval; in addition, in fuzzy-PID control or environmental fusion calculation, corresponding dimension matching shall also be carried out for each sensing item, so that the operation model disclosed in the present invention has a clear and reproducible physical meaning during actual implementation, thereby avoiding the problem of insufficient disclosure caused by inconsistent dimensions.

[0223] Those of ordinary skill in the art can realize that the units and algorithm steps of each example described in combination with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are executed in a hardware or software manner depends on the specific application and design constraints of the technical solution. Professional technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of this application.

[0224] Those skilled in the art can clearly understand that for the convenience and brevity of description, the specific working processes of the systems, devices, and units described above can refer to the corresponding processes in the foregoing method embodiments, and will not be elaborated herein.

[0225] In several embodiments provided in this application, it should be understood that the disclosed systems, devices, and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of the units is only for some logical function divisions, and there can be other division methods during actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the displayed or discussed couplings or direct couplings or communication connections to each other can be through some interfaces, and the indirect couplings or communication connections of devices or units can be in electrical, mechanical, or other forms.

[0226] The unit described as a separation component may or may not be physically separated. The component shown as a unit may or may not be a physical unit, that is, it may be located in one place or distributed over multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0227] As described above, the foregoing is only a specific implementation manner of this application, but the protection scope of this application is not limited thereto. Any person skilled in the art within the technical scope disclosed by this application can easily think of changes or substitutions, which should all be covered within the protection scope of this application. Therefore, the protection scope of this application should be subject to the protection scope of the claims.

Claims

1. A method for adjusting the parameters of photodynamic inactivation based on hybrid fuzzy-PID control, characterized in that: Including, When the environmental condition parameters exceed the preset fluctuation range, fuse multi-sensor readings, correct errors, and construct an adaptive prediction model to convert the environmental state from multi-source discrete quantities into short-term trends to warn of extreme changes; When the environmental model shows that the water quality index deviates outside the positive threshold, calculate the photosensitizer dosage using the adaptive formulation dosing technology according to the environmental cooperation coefficient, configure the light parameters through the environmental cooperation coefficient and the water body dynamic prediction results, and through the linkage of the adaptive photosensitizer dosing and the light strategy formulation, the coupling of dosage - dosing time - irradiation is achieved, and the dosing - irradiation coordination is realized; The environmental cooperation coefficient is used to characterize the suitability of the water body for photosensitizer dosing after a period of time in the future, and the environmental cooperation coefficient is related to the prediction deviations of pH, dissolved oxygen, and temperature; If the real-time sensor readings deviate from the set target range, call the hybrid fuzzy-PID closed-loop to adjust the light parameters and auxiliary equipment instructions, and determine the oxygenation or isolation actions in combination with the environmental difference vector, so that the local water quality is restored from the unstable state to the safe range; among them, the components of the environmental difference vector include water flow velocity, aquaculture density, and light interference degree; After the end or the expiration of the photodynamic inactivation stage, summarize the control feedback data and the control signal sequence and input them into the reinforcement learning module together, conduct replay pool training and update the control strategy parameters, so that the single-operation data is transformed into a continuously iterative self-learning knowledge base; When the policy update is completed and it is necessary to face a large-scale cross-partition scenario, each edge control unit receives the cloud decision-making model and conducts multi-node hierarchical scheduling in combination with the environmental difference vector, so that the original single-point closed-loop is extended to cloud-edge collaboration, and the distributed environmental differences are integrated into a global strategy.

2. The photodynamic inactivation parameter adjustment method according to claim 1, characterized in that: The sensor array deployed in the water body collects environmental condition parameters at a high frequency, performs redundant detection and error correction, and then uses a custom fusion formula to form a fusion output quantity. The fusion formula is based on an exponential decay mechanism for the differences between sensors.

3. The photodynamic inactivation parameter adjustment method according to claim 2, characterized in that: Use an adaptive prediction model to construct a dynamic environmental model, and use a deep neural network or a fuzzy neural network to predict the trends of environmental condition parameters in the short term to obtain the environmental prediction information of multiple parameters in the next time period; Store the prediction results in the environmental prediction database, and generate an environmental state model. If it is monitored that some parameters are about to break through the safety threshold, early intervention will be carried out.

4. The photodynamic inactivation parameter adjustment method according to claim 3, characterized in that: Based on the obtained environmental prediction information, introduce an environmental cooperation coefficient to characterize the suitability of the water body for photosensitizer dosing, calculate and obtain the photosensitizer dosage, and implement corresponding dosing according to the working mode of the dosing device; If the environmental cooperation coefficient is lower than expected, adaptively delay the dosing time, or first call the auxiliary equipment to correct the water quality and then conduct the dosing.

5. The photodynamic inactivation parameter adjustment method according to claim 4, characterized in that: After determining and implementing the dosage of the photosensitizer, a light irradiation strategy is formulated based on the current concentration distribution of the photosensitizer in the water body and the prediction results, and the light irradiation strategy is solved by an offline or online optimization algorithm; Fine irradiation of local hot spots enables the optimal linkage of the photosensitizer concentration distribution and the light source output in the spatio-temporal domain.

6. The method for adjusting the photodynamic inactivation parameters according to claim 1, characterized in that: Relying on the photosensitizer dosage and light irradiation parameter settings, comparing with the real-time sensor data of the actual environment, if it is observed that the current state deviates significantly from the set target interval, then enter the closed-loop correction mode; Adopt a control method based on a non-linear potential function to map the environmental deviation vector to the control signal vector, including compensating for the light intensity and scheduling the start and stop of auxiliary devices, and adaptively controlling according to the deviation source.

7. The method for adjusting the photodynamic inactivation parameters according to claim 6, characterized in that: According to the previously determined light irradiation mode and photosensitizer distribution plan, combined with the new round of sensor data and closed-loop control instructions, perform coordinated scheduling in the spatial and temporal dimensions to ensure ideal inactivation effects and aquatic animal safety in each scenario; Construct a multi-scenario scheduling function to jointly map the control signal vector and the environmental difference vector, obtain scenario-based instructions, and send them to each independent execution subunit for execution.

8. The method for adjusting the photodynamic inactivation parameters according to claim 1, characterized in that: Based on the sensor records, control instruction records and execution feedback data generated during the actual execution process, conduct multi-dimensional evaluations against the set target indicators, form and output an effect evaluation report, and construct an adaptive optimization mechanism in combination with the recorded execution actions; The adaptive optimization mechanism is based on the state-action-reward logic loop, where: Define the state, define the action and the comprehensive reward function; make corresponding strategy corrections and optimizations during training according to the comprehensive reward function, store the execution sequence of each cycle in the reinforcement learning replay pool, and form a mapping strategy after iterative training.

9. The method for adjusting the photodynamic inactivation parameters according to claim 8, characterized in that: Store the parameters of the trained policy model and key evaluation data in the knowledge base, store the output of the reinforcement learning model separately according to different scenario labels, and retain the mapping relationship of each key variable; Use the expert rules and the reinforcement learning model in parallel. When the preset rules are met, the expert rules are triggered preferentially, and the corresponding policy parameter set is automatically loaded according to the scenario label. If there are major environmental changes, start supplementary fine-tuning.

10. The method for adjusting the photodynamic inactivation parameters according to claim 1, characterized in that: Deploy edge control units within each sub-zone to collect the water quality dynamics within the sub-zone and perform hybrid fuzzy-PID closed-loop adaptive regulation. When multiple zones are adjacent, each edge control unit exchanges key state information and coordinates the implementation of cross-zone isolation or in-zone water flow regulation; After making decisions in the local environment, the edge control unit transmits the periodically summarized data to the main control center, and the main control center fuses the multi-zone information.

11. The photodynamic inactivation parameter adjustment method according to claim 10, characterized in that: Regularly upload the environmental parameters, control actions, and evaluation indicators generated during the local operation of each edge control unit to the cloud big data platform and append auxiliary information; After constructing a multi-objective optimization deep learning framework on the cloud platform, perform large-scale batch training on the collected data based on the generalization function, and after the training is completed, send the new cloud decision model to each edge control unit.

12. The photodynamic inactivation parameter adjustment method according to claim 10, characterized in that: If the breeding environment of a certain sub-region or area highly matches the cloud training data, the cloud decision model can be directly adopted; if there are still local differences, secondary fine-tuning is performed locally at the edge control unit according to the reinforcement learning mechanism; Based on the cloud decision model and the data of the partition edge control unit, perform cross-scenario data fusion and global policy iteration for different breeding scenarios, and regularly form a global optimal policy and feedback it into the self-learning process.

13. The photodynamic inactivation parameter adjustment method according to claim 12, characterized in that: Uniformly process the environmental difference vectors reported by different edge control units, compare them with the cloud model, identify possible blind areas of scenario differences, supplement new feature dimensions or correct existing feature weights; Combine the scenario difference information with the generated cloud decision model, and use the distributed policy iteration algorithm to fine-tune the cloud decision model after summarizing the feedback results of each edge control unit to generate a global decision model.

Images