Intelligent dosing control method and system based on multi-source data fusion and time sequence prediction

Abstract

The invention relates to the field of water treatment intelligent dosing control, and provides an intelligent dosing control method and system based on multi-source data fusion and time sequence prediction.The method comprises the steps that a raw water working condition, colloid electrical property balance and floc form attribute set are extracted, and the basic dosing matching amount is determined; obtaining a dynamic delay constant according to the volume of the sedimentation tank and an instantaneous flow value, and generating a causal mapping sample set by combining the three types of attribute sets and effluent turbidity signals of the sedimentation tank; generating a dynamic weight factor sequence according to the causal mapping sample set, and performing point-by-point inner product on the causal mapping sample set to generate a process condition sensitive feature set; and generating a predicted effluent turbidity value based on the process working condition sensitive feature set, performing reverse game optimization through the effluent target turbidity, generating an agent supplementing correction increment, and performing amplitude limiting truncation in combination with the basic agent adding proportioning amount to generate an agent adding control instruction. According to the invention, multi-source heterogeneous working condition sensing features and space-time causal alignment logic are fused, and an intelligent dosing decision and feedback compensation closed-loop mechanism is constructed.

Description

Technical Field

[0001] This invention relates to the field of intelligent dosing control in water treatment, and particularly to an intelligent dosing control method and system based on multi-source data fusion and time-series prediction. Background Technology

[0002] With the deepening construction of intelligent water purification plants and the continuous improvement of water supply safety standards, the control of the coagulation dosing process has become a key technology to ensure the compliance of terminal effluent and the energy efficiency of water plant operation. How to achieve real-time dynamic optimization of reagent dosage in a multivariate environment with drastic fluctuations in the physicochemical properties of raw water by sensing microscopic reaction mechanisms, effectively compensating for the problems of insufficient identification accuracy and physical response lag caused by the reliance on single macroscopic indicators in traditional systems, has become an urgent technical challenge to be solved in the process of achieving intelligent control of water treatment dosing procedures.

[0003] Chinese patent application CN120736654B discloses a dosing control system for a water treatment process. The system includes: a programmable logic controller (PLC) for transmitting current feature values ​​of influent and effluent parameters collected by sensors to an edge gateway, and driving a camera to capture images of flocs; a camera for transmitting the captured floc images to the edge gateway; an edge gateway for obtaining a predicted dosing amount based on a preset dosing amount prediction model using the influent parameter features, the effluent parameter features, and the floc images, and feeding back the predicted dosing amount to the PLC; and the PLC for controlling the dosing pump to perform dosing operations during the water treatment process based on the predicted dosing amount.

[0004] However, current technology still faces many challenges. Under low-temperature and low-turbidity operating conditions, raw water turbidity indicators are prone to distortion. Even if the turbidity value is relatively stable, its colloidal electrical state and chemical requirements may still change significantly. Traditional control methods mainly rely on the linear relationship between raw water turbidity and flow rate, which makes it difficult to reflect the microscopic instability in the early stages of coagulation reaction, resulting in a lack of observability during the dosing process. Due to a physical lag of approximately 2 to 4 hours between chemical dosing and treatment effect, by the time effluent turbidity is detected as exceeding the standard, a large amount of water has often already failed to meet treatment standards, missing the effective control window. Under these conditions, compensating by increasing the dosage can easily lead to overdosing, resulting in increased aluminum and iron ion concentrations in the effluent, which not only increases operating costs but also poses a risk to water supply safety. Summary of the Invention

[0005] To achieve the above objectives, this invention provides an intelligent dosing control method based on multi-source data fusion and time-series prediction, the specific technical solution of which is as follows:

[0006] The sensor signals from the raw water intake, the electrical signals from the online flow current meter, and the real-time underwater image stream were collected and extracted into raw water operating condition attribute sets, colloidal electrical balance attribute sets, and floc morphology attribute sets, respectively, and the basic dosing ratio was determined simultaneously.

[0007] The dynamic delay constant is obtained based on the rated physical volume of the sedimentation tank and the instantaneous flow rate value in the raw water condition attribute set. Based on the dynamic delay constant, the comprehensive condition state vector, and the collected sedimentation tank effluent turbidity signal, a causal mapping sample set is generated. The comprehensive condition state vector includes the raw water condition attribute set, the colloidal electrical balance attribute set, and the floc morphology attribute set.

[0008] A dynamic weight factor sequence is generated based on the causal mapping sample set. The dynamic weight factor sequence is then used to perform a point-by-point inner product operation on the causal mapping sample set to generate a process condition sensitive feature set.

[0009] Based on the process condition sensitive feature set, the predicted effluent turbidity value is generated. By performing reverse game optimization on the water quality deviation between the predicted effluent turbidity value and the preset effluent target turbidity, the incremental correction of the reagent addition is generated. Combined with the basic dosing ratio, the amplitude is limited and truncated within the physical safety boundary to generate the dosing control command.

[0010] Furthermore, the method for extracting the raw water condition attribute set includes: extracting a raw water condition attribute set containing raw water turbidity value, water temperature value, and instantaneous flow rate value based on the collected raw water intake sensor signal;

[0011] The method for extracting the colloidal electrical balance property set includes: extracting a colloidal electrical balance property set containing the flow current value and the hydrogen ion concentration balance index based on the electrical signal of the online flow current meter at the back end of the mixing cell;

[0012] The method for extracting the floc morphology attribute set includes: generating a binarized feature map based on the collected real-time underwater image stream, and generating a floc morphology attribute set containing the fractal dimension of the floc, the equivalent average diameter of the floc, and the settling rate of the floc per unit volume by analyzing the geometric distribution of the pixel connected domains.

[0013] The method for determining the basic dosing ratio includes: determining the basic dosing ratio based on the mapping relationship between the raw water turbidity value and water temperature value in a pre-stored expert experience database.

[0014] Furthermore, the method for analyzing the geometric distribution state of the pixel connected regions includes:

[0015] The total number of pixels in each connected region of the binary feature map and the window step size are statistically analyzed to obtain the floc projection area and the observation scale feature length. The least squares method is used to determine the fitting residual constant based on the natural logarithm of the floc projection area and the natural logarithm of the observation scale feature length. The fractal dimension of the floc is generated analytically based on the natural logarithm of the floc projection area, the natural logarithm of the observation scale feature length, and the fitting residual constant.

[0016] The equivalent geometric diameter of all connected components in the binarized feature map is averaged to generate the floc equivalent average diameter.

[0017] Calculate the geometric center coordinates of the connected domain of the same pixel in the binarized feature maps of two adjacent frames, and define them as the centroid of a specific floc. Based on the ratio of the dynamic displacement vector of the specific floc centroid between the binarized feature maps of two adjacent frames and the inter-frame sampling period, the settling rate of the floc per unit volume is obtained.

[0018] Furthermore, the method for constructing the causal mapping sample set includes:

[0019] Based on the fluid dynamics equilibrium relationship, the dynamic delay constant is obtained by time delay calculation of the rated physical volume and instantaneous flow rate of the sedimentation tank; the dynamic delay constant is the product of the algebraic quotient of the rated physical volume and instantaneous flow rate and the preset hydraulic efficiency coefficient.

[0020] Based on the dynamic delay constant, and the timing of the drug administration action. The combined working condition state vector sum The turbidity signal of the sedimentation tank effluent collected at any time is used to obtain the minimum cumulative normalization cost in order to generate a causal mapping sample set.

[0021] Furthermore, the method for obtaining the minimum cumulative regularization cost includes:

[0022] The width of the time-domain search window is determined based on the turbulence intensity of the water flow in the sedimentation tank and the variance of the hydraulic residence time generated by the sludge discharge cycle, and the time-domain search window is locked with the dynamic delay constant as the center offset.

[0023] Using the local distance measurement operator within the time domain search window, the degree of deviation between the comprehensive operating condition state vector and the turbidity signal of the sedimentation tank effluent is calculated to obtain a sequence of deviation degrees.

[0024] The parameter optimization operator is used to perform cumulative summation on the deviation degree sequence, and extreme value retrieval is performed to lock the time point corresponding to the minimum value of the cumulative summation as the best matching site;

[0025] Extract the cumulative sum at the best matching point and analyze to generate the minimum cumulative normalization cost.

[0026] Furthermore, the method for generating the process condition sensitive feature set includes:

[0027] The contribution measure score of each feature component in the causal mapping sample set is calculated, and a dynamic weight factor sequence is generated based on the distribution ratio of the contribution measure score in the causal mapping sample set; the feature components include the raw water turbidity value, the flow current value, and the fractal dimension of the flocs.

[0028] Based on the dynamic weight factor sequence, the Adama product operator is used to perform pointwise inner product operations on the causal mapping sample set to generate a process condition sensitive feature set.

[0029] Furthermore, the method for calculating the contribution measure score includes:

[0030] A multi-dimensional working condition feature interaction space is constructed, and the raw water turbidity value, flow current value and floc fractal dimension of the causal mapping sample set are projected into the multi-dimensional working condition feature interaction space.

[0031] The dot product similarity is used to measure the geometric projection overlap between each feature dimension in the multi-dimensional working condition feature interaction space and the fluctuation deviation sequence of the sedimentation tank effluent turbidity signal relative to the preset ideal control target, and to generate a causal coupling strength score.

[0032] The contribution index of the causal coupling strength score is scalarized using the causal gain mapping function, and the contribution measure score of each feature dimension is obtained by parsing.

[0033] Furthermore, the method for performing reverse game optimization includes:

[0034] The policy gradient search optimization based on generalized advantage estimation is performed by iteratively adjusting the drug dosing action control component and solving the corresponding reward function until a convergent solution that maximizes the reward function is found.

[0035] The reward function is obtained by summing and taking the negative value of three constraint terms. The three constraint terms include: the first term is the product of the squared water quality deviation term and the preset water quality deviation penalty weight coefficient; the second term is the product of the chemical dosing cost function and the preset chemical consumption cost penalty weight coefficient; and the third term is the product of the absolute value of the chemical replenishment correction increment, the preset stability constraint coefficient, and the mean component of the dynamic weight factor.

[0036] The drug addition cost function uses the drug replenishment correction increment as the independent variable; the mean of the dynamic weighting factor is obtained by performing an arithmetic average operation on the dynamic weighting factor sequence.

[0037] Based on the convergence solution of the maximum value, the optimal replenishment amount is locked, and the reagent replenishment correction increment is generated analytically.

[0038] Furthermore, the method for generating the dosing control command includes:

[0039] The process condition sensitive feature set is input into the long short-term memory network, and phase calibration is performed in combination with the dynamic delay constant to generate the predicted effluent turbidity value.

[0040] The water quality deviation between the predicted effluent turbidity value and the preset effluent target turbidity is used as the control incentive. Combined with dynamic weighting factors, reverse game optimization is performed to analyze and generate the reagent replenishment correction increment.

[0041] The intelligent dosing control system based on multi-source data fusion and time-series prediction is used to implement the above-mentioned intelligent dosing control method based on multi-source data fusion and time-series prediction. The system includes a multi-source sensing module, a spatiotemporal causal alignment module, a sensitive feature sensing module, and an intelligent decision control module.

[0042] The multi-source sensing module is used to collect sensor signals from the raw water intake, electrical signals from the online flow current meter, and real-time underwater image streams, and extract them into raw water operating condition attribute sets, colloidal electrical balance attribute sets, and floc morphology attribute sets, and simultaneously determine the basic dosing ratio.

[0043] The spatiotemporal causal alignment module is used to obtain a dynamic delay constant based on the rated physical volume of the sedimentation tank and the instantaneous flow rate value in the raw water condition attribute set, and to generate a causal mapping sample set based on the dynamic delay constant, the comprehensive condition state vector and the collected sedimentation tank effluent turbidity signal; the comprehensive condition state vector includes the raw water condition attribute set, the colloidal electrical balance attribute set and the floc morphology attribute set.

[0044] The sensitive feature perception module is used to generate a dynamic weight factor sequence based on the causal mapping sample set, and to perform a point-by-point inner product operation on the causal mapping sample set using the dynamic weight factor sequence to generate a process condition sensitive feature set.

[0045] The intelligent decision control module generates a predicted effluent turbidity value based on the sensitive feature set of process conditions. It then performs reverse game optimization on the water quality deviation between the predicted effluent turbidity value and the preset target effluent turbidity value to generate a reagent replenishment correction increment. Combined with the basic dosing ratio, it limits the amplitude within the physical safety boundary and generates a dosing control command.

[0046] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0047] This invention performs multi-dimensional decoupling perception of the physicochemical properties of raw water, the colloidal electrical equilibrium state, and the microscopic geometric topological characteristics of flocs. It transforms the coagulation and destabilization reaction quality, which cannot be directly measured, into a visual morphological attribute with physical mechanism support, thereby improving the problem of missing perception dimensions and appearance deception caused by traditional control logic relying solely on macroscopic turbidity indicators.

[0048] This invention maps the physical migration delay of controlled water bodies within a water purification structure into a phase compensation operator that dynamically fluctuates with the flow rate. Within a time-domain search window, it uses the minimum cumulative regularization cost to lock the causal relationship between the front-end dosing action and the back-end water quality response. This improves the signal phase distortion problem caused by the large time delay effect in traditional dosing feedback logic and avoids excessive oscillation caused by feedback response deviation.

[0049] This invention calculates the contribution measure scores of features in each dimension of the causal mapping sample set and generates a dynamic weight factor sequence. It then performs a weight reprojection mapping based on mechanism compensation on multi-source heterogeneous sensing data, thereby achieving digital focus on the real evolution trend of the controlled process state. This improves the blindness and decision oscillation problems caused by the inability to identify the dominant criteria when facing complex working conditions such as low temperature and low turbidity in traditional control logic.

[0050] This invention couples time-series trend prediction with physical mechanism constraints with game-like optimization of control quantity based on deep reinforcement learning. Within the safety boundary of the actuator, it achieves nonlinear superposition of the dosing reference component and dynamic correction component, eliminating feedback lag deviation caused by large time delay effect in the dosing process. It also improves the risk of drug consumption fluctuation and excessive metal ions in effluent caused by blind operation in traditional dosing logic. Attached Figure Description

[0051] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0052] Figure 1 This is a flowchart illustrating the principle of the intelligent dosing control method based on multi-source data fusion and time-series prediction of the present invention.

[0053] Figure 2 This is a schematic diagram illustrating the principle of floc morphology analysis based on underwater visual imaging in this invention.

[0054] Figure 3 This is a schematic diagram illustrating the principle of the present invention for the sedimentation rate of flocs per unit volume based on multi-frame binarized feature maps;

[0055] Figure 4 This is a functional block diagram of the intelligent dosing control system based on multi-source data fusion and time-series prediction of the present invention. Detailed Implementation

[0056] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0057] Example 1:

[0058] Please see Figure 1 This embodiment provides an intelligent dosing control method based on multi-source data fusion and time-series prediction, including:

[0059] Step S1000: Acquire the sensor signal at the raw water intake. Online current meter electrical signal and real-time underwater image stream Extracted into raw water condition attribute sets respectively. Colloid electrical balance property set and floc morphology attribute set Simultaneously determine the basic dosage ratio. .

[0060] Specifically, this step aims to transmit the sensor signal from the raw water intake. Online current meter electrical signal and real-time underwater image streaming As the object of perception, the macroscopic physicochemical fluctuations during water flow are mapped into a set of raw water condition attributes containing information on turbidity, water temperature, and instantaneous flow rate. Furthermore, the instantaneous charge response after the agent undergoes an electron neutralization reaction with the colloid is mapped to a set of colloid electrical balance properties. Simultaneously, by utilizing morphological evolution principles, the microstructural characteristics of flocs at the end of their growth period are mapped into a set of floc morphological attributes including fractal dimension, equivalent diameter, and settling rate. The basic dosage ratio is calculated based on historical steady-state operating conditions. This enables the decoupling of the nonlinear and multivariable state space characteristics of the controlled concrete object at the data source, providing a data foundation with physical mechanism support for solving the problem of large hysteresis feedback.

[0061] Further, step S1000 includes:

[0062] Step S1100, based on the collected raw water intake sensor signal Extracting values ​​including raw water turbidity Water temperature value Instantaneous flow rate Raw water working condition attribute set Based on the turbidity value of the raw water and water temperature value The mapping relationship in the pre-stored expert experience database is used to determine the basic dosage ratio. .

[0063] Specifically, this step aims to use the dynamically flowing raw water at the water intake as the data sensing object, and to decouple and map the mass concentration characteristics of colloids and suspended solids in the water, the thermal motion intensity characteristics of water molecules, and the kinetic energy characteristics of the water flow into a raw water condition attribute set. Turbidity value of raw water Water temperature value and instantaneous flow rate The basic dosage ratio is determined by retrieving historical steady-state operating conditions at the data source. This achieves initial decoupling between external operating condition disturbance variables and basic control commands.

[0064] In practice, this step involves continuously scanning the raw water flowing through it in real time using a sensor array deployed at the water plant's intake, capturing the sensor signals from the raw water intake. The raw water intake sensor signal The analog current or voltage signal, which includes electrical components, is sampled and quantized by an analog-to-digital converter module, and linearized and calibrated according to the range linear slope term and intercept compensation term of each sensor, thereby mapping the amplitude of the electrical signal into parameters characterizing the physical characteristics of the controlled water body to be treated.

[0065] The parameters of the physical characteristics of the controlled water body to be treated are defined as the raw water condition attribute set. .in, The value represents the raw water turbidity, measured in NTU. It is obtained by linearly calibrating the electrical signal fed back by the raw water turbidity sensor. As the main dimension reflecting the concentration of impurities in the controlled water body, it is used to quantify the distribution of suspended particles that obstruct light transmission in a unit volume of water. The water temperature value, expressed in °C, is obtained by mapping the electrical signal fed back from the raw water temperature sensor through physical quantity transformation. It serves as a key variable characterizing the kinetics of the coagulation reaction and is used to calculate the dynamic viscosity correction coefficient. To compensate for the impact of changes in water viscosity caused by fluctuations in ambient temperature on particle collision frequency; This represents the instantaneous flow rate, in units of... Its value is obtained by linearizing and calibrating the electrical signal fed back from the flow monitoring instrument. It serves as the main feedforward control quantity for the dosing actuator, used to calculate the basic chemical dosing load in real time and determine the range baseline of the control system. This is achieved by restoring the electrical signal to a set of raw water operating condition attributes with clear physical meaning. This enables a smooth transition from front-end physical sensing to back-end digital control.

[0066] Secondly, a pre-stored expert experience database is invoked, which contains mapping relationships based on historical steady-state operating conditions. This step is based on the currently captured raw water turbidity value. and water temperature value Multidimensional linear interpolation calculations are performed on the mapping relationships in the pre-stored expert experience database to determine the basic dosage ratio. The aforementioned basic dosage ratio The value is equal to that of the multidimensional nonlinear mapping operator. Based on the raw water turbidity value and dynamic viscosity correction factor The result value after mapping the set of independent variables.

[0067] Among them, the multidimensional nonlinear mapping operator It represents the mapping logic from the operating condition attribute space to the dosing control space stored in the expert experience base. This logic is used to perform continuous processing on discrete historical optimal operating condition data points through a multidimensional linear interpolation algorithm, outputting deterministic command components adapted to the current operating condition; the dynamic viscosity correction coefficient... Based on water temperature value The constructed physical property operator serves as a key physical intermediary connecting environmental parameters and reaction kinetics, used to compensate for water viscosity changes caused by environmental temperature fluctuations. At the physical mechanism level, the dynamic viscosity correction coefficient... With water temperature The decrease in nonlinearity and the increase in nonlinearity thus guide the multidimensional nonlinear mapping operator. Turbidity value of raw water While keeping the dosage constant, the basic dosage ratio is adjusted by compensating for kinetic reaction resistance. . output intensity.

[0068] Step S1200: Based on the electrical signal from the online flow current meter at the back end of the mixing tank. Extracting values ​​including the flow current. and hydrogen ion concentration balance index colloidal electrical balance property set .

[0069] Specifically, this step aims to use the controlled water flow at the rear end of the mixing tank, which is undergoing dynamic destabilization, as the data sensing object. Utilizing the charge balance induction principle, the electrical signal from the online flow current meter is... The characteristics of the residual charge after neutralization of the drug contained in the colloidal substance are mapped to a set of colloidal electrical balance properties. The value of the flowing current in Furthermore, the chemical potential characteristics of hydrogen ions in water are mapped to a set of colloidal electrical equilibrium properties. Hydrogen ion concentration balance index This approach achieves characteristic coupling between the controlled electrical state of the water body and the hydrolysis acid-base environment at the data source, providing real-time operating condition data input on the destabilization performance of the colloid for subsequent steps, thereby eliminating the perception blind spot that exists in the traditional control framework during the reaction induction period.

[0070] In practice, this step involves capturing the electrical signal of an online flow current meter in real time, reflecting the destabilization state in the early stages of the reaction, through a sampling branch deployed at the rear end of the reagent addition and mixing tank. The online current meter's electrical signal It contains the induced electromotive force signal generated when the controlled water flows through the detection chamber, which is processed by a high-magnification circuit and nonlinear compensation logic to be analyzed and restored into physical parameters reflecting the charge properties of the colloidal substance.

[0071] The physical parameters that reflect the charge properties of colloids, as determined by analytical analysis, are defined as the set of colloid electrical equilibrium properties. This set of colloidal electrical balance properties Characterizes the completeness of charge neutralization of the controlled water body at the end of the mixing phase. Among them, The value represents the flow current, measured in SC. It is obtained by synchronously rectifying and converting the amplitude of the alternating current sensed by the online flow current meter sensor. As a core sensing indicator reflecting the degree of colloid destabilization, it is used to quantify the state of residual negative charge or excess positive charge in the mixed water. When the flow current value... When the charge approaches zero, it indicates that the dosage of the drug and the surface charge of the particles have reached a dynamic equilibrium, that is, the charge neutralization reaction has reached a steady-state saturation point. The hydrogen ion concentration balance index, expressed in pH, is obtained by mapping the electrical signal fed back from the online pH sensor through the Nernst equation. It is used to characterize the current acid-base environment of the controlled water body and reflect the charge release efficiency of the coagulant hydrolysis products.

[0072] Step S1300, based on the acquired real-time underwater image stream Generate a binarized feature map, and by analyzing the geometric distribution of pixel connected components, generate a fractal dimension containing flocculents. , equivalent average diameter of flocs and the settling rate of flocs per unit volume floc morphology attribute set .

[0073] Specifically, this step aims to use flocs in the dynamic growth completed state at the end of the flocculation process, such as alum flocs, as data sensing objects, and to map the potential geometric and topological features of particles in the controlled water body within the current field of view, which characterize the density of their internal structure, into a set of floc morphological attributes. fractal dimension of flocculents Mapping the spatial scale characteristics of particles to a set of floc morphological attributes The equivalent average diameter of the flocs in Furthermore, the dynamic displacement characteristics of particles under the action of gravity are mapped into a set of floc morphological attributes. The unit volume floc settling rate This enables the transformation of coagulation reaction quality, which cannot be directly measured, into visualized physical properties at the data source, thereby establishing timely prediction logic for subsequent sedimentation effects during the flocculation reaction stage.

[0074] In practice, this step utilizes an underwater visual imaging module deployed at the end of the flocculation tank to capture a real-time underwater image stream containing a large number of dynamic floc particles. To eliminate discrete optical noise caused by suspended impurities in the water, the real-time underwater image stream was first processed. Median filtering is performed to smooth background interference by sorting the grayscale values ​​of the pixel neighborhood and extracting the median. Subsequently, an adaptive threshold segmentation algorithm is employed. Based on the brightness gradient distribution within a local neighborhood sliding window centered on the current pixel, the segmentation threshold for each pixel is dynamically determined. This discretizes and separates the dark pixels representing the floc structure from the bright pixels representing the background water, generating a binarized feature map reflecting the microscopic geometric boundaries of the flocs. Based on this binarized feature map, the geometric distribution of the pixel connected components is calculated and defined as the floc morphology attribute set. .

[0075] Further, please refer to Figure 2 , Figure 2 This is a schematic diagram illustrating the principle of floc morphology analysis based on underwater visual imaging, as described in this invention. Figure 2 As shown, the underwater visual imaging module is deployed at the end of the flocculation tank to capture real-time underwater image streams of the controlled water body. The floc particles marked with red circles are the core sensing objects, while suspended impurities marked with purple circles are defined as background noise. By sequentially performing median filtering and an adaptive threshold segmentation algorithm, the real-time visual signal is transformed into a binary feature map reflecting the microscopic geometric boundaries of the flocs.

[0076] The fractal dimension of flocs, ranging from 1.0 to 2.0, is used to characterize the complexity of the microstructure and the density of space filling in the flocculation reaction products. Its calculation logic follows a linear regression mapping based on the relationship between pixel coverage area and observation scale, specifically as follows: the natural logarithm of the projected area of ​​the flocs is defined as the fractal dimension of the flocs. The product of the natural logarithm of the observation scale feature length and the algebraic sum of the fitting residual constant is used. Here, the floc projected area represents the numerical value of the floc projected area formed by connected pixel regions in the binarized feature map. Its value is determined by the total number of pixels contained within the connected regions, used to quantify the spatial physical extent occupied by floc particles in the two-dimensional field of view. The numerical value of the observation scale feature length corresponds to the coverage window step size for performing spatial measurements on the pixels, used to detect the self-similarity features of floc edges at different scales. The value of the fitting residual constant is obtained by performing least-squares fitting on multiple scale observation sample sets, i.e., the natural logarithm of the floc projected area and the natural logarithm of the observation scale feature length, used to compensate for observation biases caused by electronic thermal noise of the imaging module and controlled water flow disturbances, ensuring the anti-interference capability of fractal feature extraction.

[0077] Represents the equivalent average diameter of the flocs, in units of Its value is obtained by averaging the equivalent geometric diameters of all pixel connected regions in the binarized feature map, and serves as a macroscopic size indicator characterizing the growth and development status of flocs in the controlled water body.

[0078] This represents the settling rate of flocs per unit volume, in units of... The value is obtained by performing feature matching on a time-series image sequence composed of multiple frames of binarized feature maps. Specifically, the centroid algorithm is used to calculate the geometric center coordinates of the connected components of the same pixel in two adjacent frames of binarized feature maps, and these coordinates are defined as the centroid of a specific floc. The settling rate of floc per unit volume is calculated using the ratio of the dynamic displacement vector of the specific floc centroid in the time-series image sequence to the inter-frame sampling period. It is used to predict the sludge discharge efficiency after flocs enter the sedimentation tank.

[0079] Further, please refer to Figure 3 , Figure 3 This is a schematic diagram illustrating the principle of the present invention regarding the settling rate of flocs per unit volume based on multi-frame binarized feature maps. For example... Figure 3 As shown, the same connected region of a pixel, indicated by a green circle, is identified in the binarized feature maps of two adjacent frames. The geometric center coordinates of this connected region are locked and defined as the specific floc centroid indicated by the red dot. During the sampling period from time T1 to time T3, the motion trajectory of this specific floc centroid is tracked. The blue arrows represent the dynamic displacement vector from time T1 to time T2, and the purple arrows represent the dynamic displacement vector from time T2 to time T3. The floc settling rate per unit volume is also shown. The value of is obtained by calculating the ratio of the vector summation of the dynamic displacement vector to the inter-frame sampling period, and is used to realize the timely prediction of the sludge discharge efficiency after the flocs enter the sedimentation tank at the control level.

[0080] Specifically, this involves resolving the visual image stream into a set of floc morphological attributes with clearly defined physical properties. It enables real-time identification of the implicit reaction quality within the controlled coagulation object, prompting the control system to perceive the micro-level fluctuations in operating conditions. It captures the evolution trend of the floc microstructure in the early stage of the reaction, reduces the control decision lag caused by the hydraulic retention time in the sedimentation tank, and realizes real-time monitoring of the reaction quality.

[0081] Step S2000, based on the rated physical volume of the sedimentation tank. Raw water operating condition attribute set Instantaneous flow rate value Obtain the dynamic delay constant And based on dynamic delay constant Comprehensive operating condition state vector and collected sedimentation tank effluent turbidity signal Generate a causal mapping sample set The comprehensive operating condition state vector includes the raw water operating condition attribute set. Colloid electrical balance property set and floc morphology attribute set .

[0082] Specifically, this step aims to use the controlled water body in a dynamic migration state during the water purification process as the data synchronization object, and to synchronize the data with the raw water condition attribute set from step S1100. Instantaneous flow rate The changing physical migration delay is mapped to a dynamic delay constant. And will be at the moment when the drug administration action occurs. The raw water condition attribute set from step S1100 The set of colloidal electrical equilibrium properties from step S1200 , Floc morphology attribute set from step S1300 and in Turbidity signal of sedimentation tank effluent at any time By performing time-domain reorganization, the causal logic alignment between the front-end drug dosing action and the terminal process response is achieved at the data source, generating a causal mapping sample set. .

[0083] Further, step S2000 includes:

[0084] Step S2100: Based on the fluid dynamics equilibrium relationship, determine the rated physical volume of the sedimentation tank. and instantaneous flow rate The dynamic delay constant is obtained by performing time delay calculation. The dynamic delay constant Rated physical volume and instantaneous flow rate The product of the algebraic quotient and the preset hydraulic efficiency coefficient.

[0085] Specifically, this step aims to identify the controlled water body in a dynamic migration state within the water purification structure as a time-delay identification object, and utilize the principle of fluid continuity mapping to determine the rated physical volume of the sedimentation tank in the spatial dimension. The mapping is based on the raw water condition attribute set from step S1100. Instantaneous flow rate The reaction residence time window of controlled water body fluctuating in real time within the water purification structure realizes the logical transformation from spatial geometric boundary to time step signal at the data source, and analyzes and generates dynamic delay constants that characterize the physical delay boundary of the causal relationship of chemical dosing. .

[0086] In the specific implementation process, this step retrieves the geometric parameters of the water purification facilities, which reflect the physical structural boundaries of the water plant, stored in the production management database. These geometric parameters include the rated physical volume of the sedimentation tank. The rated physical volume of the sedimentation tank. The unit is Its value is determined by the engineering design geometry data of the sedimentation tank, serving as a fixed constant that defines the physical spatial boundary of the controlled water migration path.

[0087] Subsequently, this step acquires the instantaneous flow rate value in real time to quantify fluid load variations. Finally, based on the fluid dynamics equilibrium relationship, a time delay calculation is performed to analytically generate the dynamic delay constant. The dynamic delay constant The unit is It represents the physical migration time of the controlled water body from the front-end chemical dosing point to the back-end effluent monitoring point, and is used to determine the logical offset between the chemical dosing action and the water quality feedback in the chemical dosing adjustment decision logic.

[0088] The specific execution logic of the time delay calculation based on the fluid dynamics equilibrium relationship is as follows: the dynamic delay constant The value is equal to the rated physical volume of the sedimentation tank. With instantaneous flow rate The algebraic quotient is then multiplied arithmetically with the hydraulic efficiency coefficient used to correct for differences in flow field distribution. The hydraulic efficiency coefficient ranges from 0.85 to 0.95 and is used to compensate for deviations in effective retention volume caused by dead zones or short-circuiting in the sedimentation tank, ensuring that the physical time delay calculation results conform to the actual fluid flow pattern.

[0089] At the level of physical mechanisms, when the instantaneous flow value When the dosage is increased, the dosing regulation decision logic automatically compresses the dynamic delay constant. The numerical values ​​guide subsequent steps to perform causal logic alignment within a dynamic sampling time range that decreases with flow velocity. This identification method, which maps the geometric constraints of water purification structures to dynamic time operators, effectively characterizes and compensates for the hydraulic migration delay in traditional chemical dosing processes, ensuring a causal closed loop in the logical chain between multi-source operating conditions and terminal water quality feedback results.

[0090] Step S2200, based on the dynamic delay constant Based on the time of drug administration The combined working condition state vector sum Turbidity signal of sedimentation tank effluent collected at all times Obtain the minimum cumulative regularization cost To generate a causal mapping sample set .

[0091] Specifically, this step aims to use the comprehensive operating condition state vector and feedback indicators in the later stage of sedimentation, which are in asynchronous sampling cycles and at inconsistent locations in the water purification process, as data synchronization objects, and to synchronize the comprehensive operating condition state vector with the dynamic delay constant from step S2100. Turbidity signal of sedimentation tank effluent after offset The execution logic is reorganized, and the comprehensive operating condition state vector includes the raw water operating condition attribute set from step S1100. The set of colloidal electrical equilibrium properties from step S1200 , Floc morphology attribute set from step S1300 At the data source, the causal logic between the front-end drug dosing action and the terminal process response is reconstructed, and a causal mapping sample set is generated. .

[0092] In the specific implementation process, when obtaining the dynamic delay constant Subsequently, this step introduces variable flow velocity phase synchronization compensation logic to perform causal closed-loop reconstruction of the characteristic flow under continuous flow conditions. The specific execution logic of the variable flow velocity phase synchronization compensation logic is as follows:

[0093] This step first obtains the time corresponding to the drug administration action. Collected raw water operating condition attribute set characterizing physicochemical features Colloid electrical balance property set and floc morphology attribute set The system performs cascaded encapsulation processing according to a preset dimensional order, and parses to generate a comprehensive operating condition state vector. This comprehensive operating condition state vector is an ordered list used to characterize the initial energy state and physical properties of a specific reactive agglomerate before it enters the sedimentation tank.

[0094] Subsequently, The water quality response signal, namely the turbidity signal of the sedimentation tank effluent, is constantly monitored by back-end online analysis instruments. The turbidity signal of the sedimentation tank effluent. Used to characterize the dynamic delay constant The actual water quality improvement effect produced by the front-end chemical dosing action after a determined physical time delay.

[0095] Due to micro-circulation disturbances caused by uneven hydraulic gradients within the water purification structure, the actual migration path of the reagent in the water is not completely linear. To eliminate the resulting phase shift error, this step calculates a sequence containing the comprehensive operating condition state vector and a sequence containing the turbidity signal of the sedimentation tank effluent. The optimal association path between sequences is determined to obtain the minimum cumulative regularization cost. The minimum cumulative normalization cost The value is equal to that of the dynamic delay constant. Within a defined time-domain search window, the parameter optimization operator extracts the parameters that enable the local distance measure operator to work with the overall operating condition state vector and the turbidity signal of the sedimentation tank effluent. The path cost value that reaches its minimum after the calculation result is cumulatively summed. The width of the time-domain search window is determined by the hydraulic residence time variance driven by both the turbulence intensity and the sludge discharge cycle in the sedimentation tank, and is expressed as the dynamic delay constant. The center offset is locked; the parameter optimization operator is an operational logic used to perform extreme value retrieval in the multi-dimensional path space, and its value corresponds to a specific mapping time point that makes the cumulative deviation result reach a minimum, which is used to lock the best matching point within the time domain search window; the local distance measurement operator is a function operator that measures the geometric difference between data of different dimensions, and its value reflects the degree of deviation between the comprehensive working condition state vector and the back-end water quality signal under a single sampling section.

[0096] When the minimum cumulative regularization cost obtained by the solution is When the current operating condition is within a preset threshold range for determining causal relationships, it is determined that there is a strong causal relationship between the current operating condition characteristics and the water quality response. Then, the comprehensive operating condition state vector at the corresponding time point and the sedimentation tank effluent turbidity signal are structurally mapped and encapsulated to generate a causal mapping sample set. .

[0097] Step S3000, based on the causal mapping sample set Generate a dynamic weight factor sequence and use the dynamic weight factor sequence to map the causal sample set. Perform pointwise inner product operations to generate a set of process condition sensitive features. .

[0098] Specifically, this step aims to transform the causal mapping sample set from step S2200. As a sensing benchmark, using self-attention mapping logic, heterogeneous operating condition data from multiple sensor sources, including raw water physicochemical characteristics, colloidal electrical characteristics, and floc geometric and topological characteristics, are mapped into a dynamic weighting factor sequence characterizing the sensitivity to operating conditions. This restores key destabilization features originally submerged in a complex and multivariable environment into feature electrical signals with significant discriminative power, and reconstructs and generates a set of process operating condition sensitive features. .

[0099] Further, step S3000 includes:

[0100] Step S3100: Calculate the causal mapping sample set. The contribution measure scores of each feature component are used to evaluate the contribution measure scores in the causal mapping sample set. The distribution ratio in the model is analyzed to generate a dynamic weighting factor sequence; the feature components include the raw water turbidity value. Flow current value and fractal dimension of flocs .

[0101] Specifically, this step aims to transform the causal mapping sample set from step S2200. As a feature extraction object, the nonlinear correlation of potential nonlinear correlations in the coagulation and destabilization process, which characterize the differences in the contribution of different dimensional operating condition indicators to the coagulation effect improvement target, is mapped into a contribution measure score reflecting the dominant ability of feature terms to process response. The dynamic weight factor sequence is generated by parsing and generating the adaptive weight allocation for the execution of operating conditions, thereby decoupling the sensitivity of raw water physicochemical properties, colloidal electrical equilibrium state and floc geometric topology characteristics to effluent response at the data source.

[0102] In the specific implementation process, this step retrieves a causal mapping sample set with physical coherence. By constructing a multi-dimensional operating condition feature interaction space, the true contribution of different operating condition dimensions to the terminal water output effect is identified. First, the causal mapping sample set... The various feature dimensions in the data specifically include the raw water condition attribute set. Turbidity value of raw water Colloid electrical balance property set The value of the flowing current in Fiber morphology attribute set fractal dimension of flocculents The causal correlation retrieval is performed by projecting the data into the multidimensional working condition feature interaction space.

[0103] Subsequently, dot product similarity was used to measure each feature dimension in the multidimensional operating condition feature interaction space, and compared with the turbidity signal of the sedimentation tank effluent. The geometrical projection overlap between the fluctuation deviation sequences relative to the preset ideal control target is defined as the causal coupling strength score characterizing the synchronicity of characteristic fluctuations and process response. To achieve unified quantification of the influence weights of heterogeneous operating condition indicators, this step performs contribution index scalarization processing by projecting the causal coupling strength score onto a causal gain mapping function, parsing to obtain the contribution measure score for each feature dimension. The causal gain mapping function takes the causal coupling strength score of each feature dimension as input, and its mapping logic is as follows: multiply the causal coupling strength score of each feature dimension with a preset gain scaling factor, and superimpose a preset system deviation compensation constant to obtain the contribution measure score for each feature dimension. The preset gain scaling factor is a real number component greater than 0, used to adjust the sensitivity gradient in the process of converting causal correlation into contribution score; the preset system deviation compensation constant is a preset bias adjustment parameter used to correct the correlation benchmark deviation caused by sensor system errors or environmental noise. This contribution measure score is used to quantify specific feature components and the current sedimentation tank effluent turbidity signal. The strength of the guiding role of the evolutionary trend reflects the strength of the mutual information mapping between it and the current coagulation destabilization control target.

[0104] Finally, based on the contribution measure score in the causal mapping sample set... The distribution ratio in the model is subjected to exponential normalization mapping, and the dynamic weight factor components in the generated dynamic weight factor sequence are analyzed. Specifically, the contribution measure score of each feature dimension is nonlinearly enhanced using the natural exponential function operator, transforming it into an energy mapping value with positive monotonicity; subsequently, the causal mapping sample set is processed... The total summation of all energy mapping values ​​is used to construct the total amount of the full-dimensional contribution measurement benchmark under the current controlled operating condition. Finally, the energy mapping value of a single feature dimension is divided by the total amount of the full-dimensional contribution measurement benchmark to obtain the dynamic weight factor components.

[0105] Wherein, the dynamic weight factor component is in The normalized real number of the interval is used to characterize the influence of each feature dimension relative to other dimensions in the dosage decision under the current controlled operating conditions. A larger value indicates a more significant dominant position of that indicator in the dosage decision process. The natural exponential function operator is based on the natural logarithm. The exponential operation with a base is used to perform nonlinear enhancement mapping on the contribution measure score, which amplifies the weighted response with dominant characteristics through exponential difference, while weakening redundant indicators at the background noise level.

[0106] Step S3200: Based on the dynamic weight factor sequence, apply the Adama product operator to the causal mapping sample set. Perform pointwise inner product operations to generate a set of process condition sensitive features. .

[0107] Specifically, this step aims to transform the causal mapping sample set from step S2200, which represents the logical relationship between the application action and the response effect, into a single sample. As a data reconstruction object, this causal mapping sample set Mapped to a set of process condition sensitive features with significant sensitivity gradients At the data source, nonlinear gain amplification of key destability characteristics and effective suppression of background redundancy indicators are achieved, eliminating the risk of decision oscillation caused by the inability of traditional control logic to identify the dominant indicators of the operating condition.

[0108] In the specific implementation process, after completing the sensitivity identification of each dimension of the working condition indicators, this step reverse maps each dimension component in the dynamic weight factor sequence to the causal mapping sample set. The physical primitive quantity space to which it belongs. By performing pointwise inner product operations between dynamic weight factor components and various feature dimensions acquired through sensing, physical mechanism compensation and gain reconstruction of multi-source heterogeneous sensing data streams reflecting the coagulation reaction evolution process are achieved to generate a process condition sensitive feature set. .

[0109] The process condition sensitive feature set It is a set of digitally controlled object state descriptions that integrates dynamic weight distribution information and possesses physical determinism. The amplitude of each feature dimension component within it reflects its guiding strength on the current coagulation reaction effect, serving as the sole criterion for subsequent intelligent dosing decision generation. Its specific execution logic is as follows: using the dynamic weight factor sequence as a scaling operator, and using the Adama product operator to map the causal sample set... The raw water condition attribute set included Colloid electrical balance property set and floc morphology attribute set Perform the operation. The Adama product operator is an operation logic that performs term-by-term arithmetic multiplication on corresponding elements, which is used to assign nonlinear sensitivity weights to the original physical quantity, thereby achieving nonlinear stretching of the feature space.

[0110] Specifically, this step transforms the sensor values, which were originally in a state of equal weighting, into process characteristic quantities with distinguishable weighting. This enables key criteria that play a dominant role in the coagulation and destabilization reaction under specific operating conditions, such as the set of floc morphology attributes under low temperature and low turbidity seasons. fractal dimension of flocculents The energy contribution in the feature space is significantly enhanced, while the background redundancy index, which is weakly correlated with the current reaction process, is suppressed, thereby achieving digital focusing on the true evolution trend of the controlled process state.

[0111] Step S4000, based on the process condition sensitive feature set Generate predicted effluent turbidity values By analyzing the predicted turbidity value of the effluent and the preset target turbidity of the effluent The water quality deviations between the two sides are optimized through reverse game theory to generate the corrective increment for reagent replenishment. In conjunction with the basic dosage ratio Within the physical safety boundary, a limit cutoff is performed to generate dosing control commands. .

[0112] Specifically, this step aims to extract the process condition sensitive feature set from step S3200, which characterizes the true nature of the controlled process reaction. As a feedforward benchmark, the temporal evolution logic of the Long Short-Term Memory network is used to map the potential precursor fluctuations in the current field of view that characterize the coagulation destabilization trend to the predicted effluent turbidity values ​​of future sampling sections. It utilizes deep reinforcement learning agent logic to achieve the preset target turbidity of the effluent. Analytical generation of reagent supplementation correction increment under physical constraints This mechanism enables a shift from passive, delayed feedback to proactive trend prediction at the decision-making stage, thus generating defensive dosing control commands. Provide decision support.

[0113] Further, step S4000 includes:

[0114] Step S4100: Set the process condition sensitive features. Input the long short-term memory network and combine it with the dynamic delay constant. Phase calibration is performed, and the predicted effluent turbidity value is generated through analysis. .

[0115] Specifically, this step aims to integrate the process condition sensitive feature set from step S3200. As an evolutionary prediction source, the temporal cyclical characteristics of long short-term memory networks are utilized to map the potential destabilization features of chemical dosing in the current field of view, which are at the micro-reaction stage, to the predicted effluent turbidity value at the end of the sedimentation tank after the physical offset time. By simulating the migration and settling process of concrete agglomerates in digital space, transparent observation of the reaction state under physical time-delay conditions is achieved at the sensing layer.

[0116] In the specific implementation process, this step first involves creating a process condition sensitive feature set that includes weighted and calibrated information on the raw water's physicochemical, electrical, and floc morphology properties. As the input stream, a Long Short-Term Memory (LSTM) prediction model that incorporates physical latency characteristics is constructed. The specific execution logic is as follows:

[0117] First, the forgetting gate of the Long Short-Term Memory Network prediction model is used, along with the process condition-sensitive feature set. The embedded dynamic weighting factor serves as the activation criterion for the feature channel, and is sensitive to the feature set of the process conditions. The temporal sampling sequences of each component undergo dynamic noise masking based on weighted reliability. Digital filtering is performed on instantaneous electrical signal offsets caused by fluctuations in online instrument sampling, as well as random hydraulic noise errors caused by turbulence in the internal flow field of the water purification structure, to generate pure state components after removing environmental noise, ensuring the baseline stability for subsequent prediction evolution.

[0118] Subsequently, through the input gate of the Long Short-Term Memory Network prediction model, the process condition sensitive feature set within the current sampling period is used. The set of floc morphology attributes included fractal dimension of flocculents and colloidal electrical balance property set The value of the flowing current in The destabilization precursor signal, which is presented under nonlinear interaction and indicates the change in colloidal stability of controlled water, is selectively updated into the pure state component after being processed by the forget gate, to obtain the fused feature signal.

[0119] Next, the characteristic signal is updated into the latent state carrier reflecting the kinetic state of the coagulation reaction. The latent state carrier is a cell state in a long short-term memory network, serving as a long-term memory carrier for the controlled water mass migration and flow within the various levels of water purification structures, and is used to lock the abrupt change trend of raw water quality and the evolution trajectory of reactions within the structures.

[0120] Finally, based on the latent state carrier, the dynamic delay constant characterizing the spatial migration displacement of the controlled water mass is determined. The phase calibration step size, used as the prediction logic, is embedded in the long short-term memory network prediction model to generate the predicted effluent turbidity value. The predicted effluent turbidity value The value is predicted by the Long Short-Term Memory Network (LSTM) model based on the time of the current medication administration action. The moment the previous drug administration action occurred Until the end of history Sensitive feature sets of each group of process conditions The constructed ordered tensor, combined with the dynamic delay constant characterizing the physical migration offset of the water flow, The output value after performing the nonlinear mapping. This is the predicted effluent turbidity value. The specific formula is as follows:

[0121] ;

[0122] in, This represents the Long Short-Term Memory network evolution operator, which includes a forget gate, an input gate, and cell state update logic, and is used to capture the deep dynamic correlation between multi-source sensing indicators and process results; Indicates the time when the current drug administration action occurs. The process condition sensitive feature set is a vector tensor obtained by weight reprojection of the raw water physicochemical properties, colloidal electrical properties and floc morphological properties. It is used to characterize the initial destabilization of a specific reaction water mass before it enters the sedimentation tank unit. It is the starting state site of the ordered tensor. The hydraulic residence backtracking period is a time step constant that takes a positive integer value. Its value is determined by the quotient of the maximum residence time of the process structure and the sampling period. It is used to ensure that the prediction logic covers the complete residence period of the water mass. Indicates the time when the current drug administration action occurs. The process condition sensitive feature set of the previous sampling time is used to provide the prediction logic with the gradient information of feature evolution; Indicates the first in history The process condition sensitive feature set of each sampling period is used as the backtracking boundary value of the ordered tensor.

[0123] Step S4200: Predict the turbidity value of the effluent. Compared with the preset target turbidity of the effluent The water quality deviation between the two sides is used as the control incentive. Combined with dynamic weighting factors, an inverse game is used to find the optimal solution and generate the incremental reagent replenishment correction. .

[0124] Specifically, this step aims to incorporate the predicted effluent turbidity values ​​from step S4100, which characterize the future evolution of the coagulation process. Compared with the preset target turbidity of the effluent The discrete deviations between the parameters are used as control excitations. A reverse game-theoretic optimization logic based on dynamic weight factor sequence constraints from step S3100 is employed to map potential, nonlinear water quality deviations into incremental corrections for reagent replenishment. At the decision-making stage, precise closed-loop regulation of the coagulation and destabilization process is achieved, providing a basis for generating dosing control commands. It provides a dynamic compensation benchmark with adaptive operating conditions, thereby eliminating the blind dosing caused by physical time delay in traditional control logic.

[0125] In the specific implementation process, when predicting the turbidity value of the effluent Deviation from the preset target turbidity of effluent At this point, the Proximal Policy Optimization (PPO) controller is invoked to execute the inverse game optimization logic. To suppress the regulation lag and response oscillations caused by the physical inertia of the large time-delay water treatment system, this step introduces the dynamic weight factor sequence into the inverse game optimization logic to perform gain correction based on the sensitivity characteristics of the current operating conditions.

[0126] The specific execution logic of the reverse game optimization logic is as follows: by determining the timing of the current drug addition action... Using water quality deviation as input, a policy gradient search based on Generalized Advantage Estimation (GAE) is performed within the continuous action space to optimize the water quality. The water quality deviation predicts the water turbidity value. and the preset target turbidity of the effluent The difference between them. By continuously adjusting the control component of the agent addition action and solving the corresponding reward function, until a convergent solution that maximizes the reward function is found, the optimal addition amount with process stability is locked, and the agent addition correction increment is generated analytically. The reward function is numerically equal to the negative sum of the three constraint terms: the first term is the product of the squared water quality deviation term and the water quality deviation penalty weight coefficient; the second term is the product of the chemical dosage cost function and the chemical consumption cost penalty weight coefficient; and the third term is the chemical replenishment correction increment. The product of the absolute value, the stability constraint coefficient, and the mean component of the dynamic weighting factor.

[0127] Wherein, the water quality deviation square term is the square of the water quality deviation; the water quality deviation penalty weight coefficient is a preset proportional adjustment operator used to adjust the constraint strength of the feedback loop on the deviation of the predicted effluent turbidity from the target value; the reagent addition cost function is a mathematical expression describing the mapping relationship between reagent consumption and economic expenditure, used to correct the incremental reagent addition. The magnitude of the value is converted into a cost evaluation index, serving as a feedback basis for operational efficiency in the reverse game optimization logic; the chemical consumption cost penalty weight coefficient is an adjustment parameter used to constrain resource consumption, limiting the amplitude of the chemical dosing action control component during the maximization search process to prevent the control logic from sacrificing operational economy in pursuit of extreme water quality; the stability constraint coefficient is a damping parameter used to adjust the smoothness of the action, adjusting the adjustment range during the optimization process according to the sensitivity of the operating conditions to suppress the incremental correction of chemical replenishment. A step jump occurs; the mean component of the dynamic weighting factor is a real number between 0 and 1, and its value is obtained by performing an arithmetic average operation on the dynamic weighting factor sequence, used to reflect the current moment's change in the raw water condition attribute set. Colloid electrical balance property set and floc morphology attribute set The overall sensitivity of the set of operating conditions to chemical dosing disturbances.

[0128] Step S4300, based on the basic dosage ratio. and the increase in dosage of medicine Within the preset physical safety boundaries, a threshold truncation is performed, and the dosing control command is parsed, generated, and executed. .

[0129] Specifically, this step aims to convert the baseline dosage from step S1100, which characterizes the static process baseline, into a basic dosage ratio. The incremental drug addition correction from step S4200, which characterizes the dynamic trend prediction compensation. As a control source, it realizes the algebraic superposition of feedforward signals and feedback correction signals, and generates dosing control commands. The mapping is transformed into a physical signal flow that drives the variable frequency dosing terminal, enabling a shift from passive, lagging regulation to active intervention at the data execution level. This eliminates the dosing feedback blind spot caused by physical time delay and intercepts the risk of exceeding standards due to sudden changes in water quality.

[0130] In the specific implementation process, this step retrieves the basic dosing ratio generated based on the analysis of the raw water's physicochemical properties. It is used as a static feedforward component of the controlled object in the coagulation process to maintain the basic destabilization environment of the coagulation reaction; and the reagent addition correction increment is generated based on deep reinforcement learning game optimization. This is used as a dynamic adjustment component to actively compensate for predicted water quality deviations in future time periods.

[0131] To improve the reliability of the entire coagulation process control loop, this step introduces a physical boundary constraint check of the variable frequency dosing drive terminal, based on the preset physical safety boundaries in the variable frequency dosing drive terminal. The static feedforward component and the dynamic adjustment component are truncated and limited to generate the final dosing control command. The physical security boundary, in particular... Includes the rated minimum control quantity and rated maximum control quantity The rated minimum control quantity This is the lower limit threshold of the output that the variable frequency dosing drive terminal can maintain stable metering. Its value is determined by the minimum operating frequency of the dosing pump and is used to prevent physical interruption of the flow during the dosing process; the rated maximum control quantity It is the upper limit threshold of the output of the variable frequency dosing drive terminal under safe operating conditions. Its value is determined by the rated maximum load of the dosing pump and is used to limit the excessive waste of reagents caused by control overshoot.

[0132] The dosing control command Numerically equal to: the basic dosage ratio calculated using boundary constraint operators. and the increase in dosage of medicine The sum performs a nonlinear limiting mapping; the nonlinear limiting refers to the rated minimum control quantity at the variable frequency dosing drive terminal. With the rated maximum control quantity The constraint values ​​are executed within the defined physical safety boundaries, and the resulting output value is the dosing control command. The numerical value. The boundary constraint operator is a logic function used to perform nonlinear limiting on the coupled numerical value, ensuring that the control gain does not exceed the mechanical load boundary of the variable frequency dosing drive terminal.

[0133] Finally, this step will transmit the dosing control instructions. It is converted into a pulse width modulation signal, which drives the variable frequency dosing terminal to perform the dosing action, thus completing the intelligent closed-loop control of the entire coagulation process.

[0134] Example 2:

[0135] This embodiment, based on Embodiment 1, provides an intelligent dosing control system based on multi-source data fusion and time-series prediction, such as... Figure 4 As shown, the system includes a multi-source sensing module, a spatiotemporal causal alignment module, a sensitive feature sensing module, and an intelligent decision control module;

[0136] The multi-source sensing module is used to collect signals from the raw water intake sensor. Online current meter electrical signal and real-time underwater image stream Extracted into raw water condition attribute sets respectively. Colloid electrical balance property set and floc morphology attribute set Simultaneously determine the basic dosage ratio. .

[0137] The spatiotemporal causal alignment module is used to align the sedimentation tank according to its rated physical volume. Raw water operating condition attribute set Instantaneous flow rate value Obtain the dynamic delay constant And based on dynamic delay constant Comprehensive operating condition state vector and collected sedimentation tank effluent turbidity signal Generate a causal mapping sample set The comprehensive operating condition state vector includes the raw water operating condition attribute set. Colloid electrical balance property set and floc morphology attribute set .

[0138] The sensitive feature perception module is used to perceive the causal mapping sample set. Generate a dynamic weight factor sequence and use the dynamic weight factor sequence to map the causal sample set. Perform pointwise inner product operations to generate a set of process condition sensitive features. .

[0139] The intelligent decision control module is based on a set of process condition sensitive features. Generate predicted effluent turbidity values By analyzing the predicted turbidity value of the effluent and the preset target turbidity of the effluent The water quality deviations between the two sides are optimized through reverse game theory to generate the corrective increment for reagent replenishment. In conjunction with the basic dosage ratio Within the physical safety boundary, a limit cutoff is performed to generate dosing control commands. .

[0140] The parts of the technical solutions provided in the embodiments of this application that are consistent with the implementation principles of corresponding technical solutions in the prior art have not been described in detail to avoid excessive elaboration.

[0141] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A smart dosing control method based on multi-source data fusion and time-series prediction, characterized in that, include: The sensor signals from the raw water intake, the electrical signals from the online flow current meter, and the real-time underwater image stream were collected and extracted into raw water operating condition attribute sets, colloidal electrical balance attribute sets, and floc morphology attribute sets, respectively, and the basic dosing ratio was determined simultaneously. The dynamic delay constant is obtained based on the rated physical volume of the sedimentation tank and the instantaneous flow rate value in the raw water condition attribute set. Based on the dynamic delay constant, the comprehensive condition state vector, and the collected sedimentation tank effluent turbidity signal, a causal mapping sample set is generated. The comprehensive condition state vector includes the raw water condition attribute set, the colloidal electrical balance attribute set, and the floc morphology attribute set. A dynamic weight factor sequence is generated based on the causal mapping sample set. The dynamic weight factor sequence is then used to perform a point-by-point inner product operation on the causal mapping sample set to generate a process condition sensitive feature set. Based on the process condition sensitive feature set, the predicted effluent turbidity value is generated. By performing reverse game optimization on the water quality deviation between the predicted effluent turbidity value and the preset effluent target turbidity, the incremental correction of the reagent addition is generated. Combined with the basic dosing ratio, the amplitude is limited and truncated within the physical safety boundary to generate the dosing control command.

2. The intelligent dosing control method based on multi-source data fusion and time-series prediction according to claim 1, characterized in that, The method for extracting the raw water condition attribute set includes: extracting a raw water condition attribute set containing raw water turbidity value, water temperature value, and instantaneous flow rate value based on the collected raw water intake sensor signal; The method for extracting the colloidal electrical balance property set includes: extracting a colloidal electrical balance property set containing the flow current value and the hydrogen ion concentration balance index based on the electrical signal of the online flow current meter at the back end of the mixing cell; The method for extracting the floc morphology attribute set includes: generating a binarized feature map based on the collected real-time underwater image stream, and generating a floc morphology attribute set containing the fractal dimension of the floc, the equivalent average diameter of the floc, and the settling rate of the floc per unit volume by analyzing the geometric distribution of the pixel connected domains. The method for determining the basic dosing ratio includes: determining the basic dosing ratio based on the mapping relationship between the raw water turbidity value and water temperature value in a pre-stored expert experience database.

3. The intelligent dosing control method based on multi-source data fusion and time-series prediction according to claim 2, characterized in that, The analysis method for the geometric distribution state of the pixel connected regions includes: The total number of pixels in each connected region of the binary feature map and the window step size are statistically analyzed to obtain the floc projection area and the observation scale feature length. The least squares method is used to determine the fitting residual constant based on the natural logarithm of the floc projection area and the natural logarithm of the observation scale feature length. The fractal dimension of the floc is generated analytically based on the natural logarithm of the floc projection area, the natural logarithm of the observation scale feature length, and the fitting residual constant. The equivalent geometric diameter of all connected components in the binarized feature map is averaged to generate the floc equivalent average diameter. Calculate the geometric center coordinates of the connected domain of the same pixel in the binarized feature maps of two adjacent frames, and define them as the centroid of a specific floc. Based on the ratio of the dynamic displacement vector of the specific floc centroid between the binarized feature maps of two adjacent frames and the inter-frame sampling period, the settling rate of the floc per unit volume is obtained.

4. The intelligent dosing control method based on multi-source data fusion and time-series prediction according to claim 1, characterized in that, The method for constructing the causal mapping sample set includes: Based on the fluid dynamics equilibrium relationship, the dynamic delay constant is obtained by time delay calculation of the rated physical volume and instantaneous flow rate of the sedimentation tank; the dynamic delay constant is the product of the algebraic quotient of the rated physical volume and instantaneous flow rate and the preset hydraulic efficiency coefficient. Based on the dynamic delay constant, and the timing of the drug administration action. The combined working condition state vector sum The turbidity signal of the sedimentation tank effluent collected at any time is used to obtain the minimum cumulative normalization cost in order to generate a causal mapping sample set.

5. The intelligent dosing control method based on multi-source data fusion and time-series prediction according to claim 4, characterized in that, The method for obtaining the minimum cumulative normalization cost includes: The width of the time-domain search window is determined based on the turbulence intensity of the water flow in the sedimentation tank and the variance of the hydraulic residence time generated by the sludge discharge cycle, and the time-domain search window is locked with the dynamic delay constant as the center offset. Using the local distance measurement operator within the time domain search window, the degree of deviation between the comprehensive operating condition state vector and the turbidity signal of the sedimentation tank effluent is calculated to obtain a sequence of deviation degrees. The parameter optimization operator is used to perform cumulative summation on the deviation degree sequence, and extreme value retrieval is performed to lock the time point corresponding to the minimum value of the cumulative summation as the best matching site; Extract the cumulative sum at the best matching point and analyze to generate the minimum cumulative normalization cost.

6. The intelligent dosing control method based on multi-source data fusion and time-series prediction according to claim 1, characterized in that, The method for generating the process condition sensitive feature set includes: The contribution measure score of each feature component in the causal mapping sample set is calculated, and a dynamic weight factor sequence is generated based on the distribution ratio of the contribution measure score in the causal mapping sample set; the feature components include the raw water turbidity value, the flow current value, and the fractal dimension of the flocs. Based on the dynamic weight factor sequence, the Adama product operator is used to perform pointwise inner product operations on the causal mapping sample set to generate a process condition sensitive feature set.

7. The intelligent dosing control method based on multi-source data fusion and time-series prediction according to claim 6, characterized in that, The calculation method for the contribution measure score includes: A multi-dimensional working condition feature interaction space is constructed, and the raw water turbidity value, flow current value and floc fractal dimension of the causal mapping sample set are projected into the multi-dimensional working condition feature interaction space. The dot product similarity is used to measure the geometric projection overlap between each feature dimension in the multi-dimensional working condition feature interaction space and the fluctuation deviation sequence of the sedimentation tank effluent turbidity signal relative to the preset ideal control target, and to generate a causal coupling strength score. The contribution index of the causal coupling strength score is scalarized using the causal gain mapping function, and the contribution measure score of each feature dimension is obtained by parsing.

8. The intelligent dosing control method based on multi-source data fusion and time-series prediction according to claim 1, characterized in that, The method for performing reverse game optimization includes: The policy gradient search optimization based on generalized advantage estimation is performed by iteratively adjusting the drug dosing action control component and solving the corresponding reward function until a convergent solution that maximizes the reward function is found. The reward function is obtained by summing and taking the negative value of three constraint terms. The three constraint terms are: the first term is the product of the squared water quality deviation term and the preset water quality deviation penalty weight coefficient; the second term is the product of the chemical dosing cost function and the preset chemical consumption cost penalty weight coefficient; and the third term is the product of the absolute value of the chemical replenishment correction increment, the preset stability constraint coefficient, and the mean component of the dynamic weight factor. The drug addition cost function uses the drug replenishment correction increment as the independent variable; the mean of the dynamic weighting factor is obtained by performing an arithmetic average operation on the dynamic weighting factor sequence. Based on the convergence solution of the maximum value, the optimal replenishment amount is locked, and the reagent replenishment correction increment is generated analytically.

9. The intelligent dosing control method based on multi-source data fusion and time-series prediction according to claim 1, characterized in that, The method for generating the dosing control command includes: The process condition sensitive feature set is input into the long short-term memory network, and phase calibration is performed in combination with the dynamic delay constant to generate the predicted effluent turbidity value. The water quality deviation between the predicted effluent turbidity value and the preset effluent target turbidity is used as the control incentive. Combined with dynamic weighting factors, reverse game optimization is performed to analyze and generate the reagent replenishment correction increment.

10. An intelligent dosing control system based on multi-source data fusion and time-series prediction, used to implement the intelligent dosing control method based on multi-source data fusion and time-series prediction as described in any one of claims 1-9, characterized in that, The system includes a multi-source sensing module, a spatiotemporal causal alignment module, a sensitive feature sensing module, and an intelligent decision-making and control module. The multi-source sensing module is used to collect sensor signals from the raw water intake, electrical signals from the online flow current meter, and real-time underwater image streams, and extract them into raw water operating condition attribute sets, colloidal electrical balance attribute sets, and floc morphology attribute sets, and simultaneously determine the basic dosing ratio. The spatiotemporal causal alignment module is used to obtain a dynamic delay constant based on the rated physical volume of the sedimentation tank and the instantaneous flow rate value in the raw water condition attribute set, and to generate a causal mapping sample set based on the dynamic delay constant, the comprehensive condition state vector and the collected sedimentation tank effluent turbidity signal; the comprehensive condition state vector includes the raw water condition attribute set, the colloidal electrical balance attribute set and the floc morphology attribute set. The sensitive feature perception module is used to generate a dynamic weight factor sequence based on the causal mapping sample set, and to perform a point-by-point inner product operation on the causal mapping sample set using the dynamic weight factor sequence to generate a process condition sensitive feature set. The intelligent decision control module generates a predicted effluent turbidity value based on the sensitive feature set of process conditions. It then performs reverse game optimization on the water quality deviation between the predicted effluent turbidity value and the preset target effluent turbidity value to generate a reagent replenishment correction increment. Combined with the basic dosing ratio, it limits the amplitude within the physical safety boundary and generates a dosing control command.

Images