A multi-dimensional intelligent control system for defluorination

Abstract

The present invention discloses a multi-dimensional intelligent control system for defluorination, specifically related to the field of sewage defluorination, including a water quality fluctuation monitoring module, a normal measurement determination module, an abnormal stage judgment module, a deep learning prediction module, and a multi-task regression module; the water quality fluctuation monitoring module is used to obtain real-time water quality fluctuation information of the water source to be defluorinated, establish pre-measurement condition 1 and pre-measurement condition 2 based on the water quality fluctuation information, and determine whether there are water quality influencing factors in the current water source through pre-measurement condition 1 and pre-measurement condition 2. By introducing multi-dimensional water quality fluctuation information and a real-time monitoring mechanism, the present invention realizes the management of fluoride concentration control. Based on water quality parameters such as water body chromaticity, conductivity, reflectivity, and ammonia nitrogen concentration, combined with a deep neural network for real-time prediction, it is beneficial to identify water quality fluctuations and perform corresponding regulation, dynamically respond to water quality changes, and improve the convenience of water source control.

Description

Technical Field

[0001] The present invention relates to the technical field of sewage defluorination, and more specifically, the present invention relates to a multi-dimensional intelligent control system for defluorination. Background Art

[0002] In the traditional sewage defluorination process, it usually relies on fixed parameter settings and manual control, lacking a real-time dynamic adjustment mechanism. These methods are often lagging when facing water quality changes and cannot adjust the fluorine removal process in time to cope with sudden water quality fluctuations and environmental changes;

[0003] In addition, many existing water quality monitoring systems lack the analysis of the comprehensive water quality status, resulting in the inability to judge the status adaptation change trend of fluorine in the water source; these deficiencies make the traditional methods appear powerless in dealing with complex water quality fluctuations, especially abnormal water quality fluctuations, affecting the defluorination efficiency and the safety of the water source water quality. Summary of the Invention

[0004] In order to overcome the above-mentioned defects of the prior art, an embodiment of the present invention provides a multi-dimensional intelligent control system for defluorination, which includes a water quality fluctuation monitoring module, a normal measurement determination module, an abnormal stage judgment module, a deep learning prediction module, and a multi-task regression module to solve the problems raised in the above background art.

[0005] To achieve the above object, the present invention provides the following technical solution: A multi-dimensional intelligent control system for defluorination, including a water quality fluctuation monitoring module, a normal measurement determination module, an abnormal stage judgment module, a deep learning prediction module, and a multi-task regression module;

[0006] The water quality fluctuation monitoring module is used to obtain the water quality fluctuation information of the water source to be defluorinated in real time, establish a pre-measurement condition one and a pre-measurement condition two based on the water quality fluctuation information, and determine whether there are water quality influencing factors in the current water source through the pre-measurement condition one and the pre-measurement condition two;

[0007] The water quality influencing information in the water quality fluctuation monitoring module includes water body chromaticity, conductivity, water body reflectivity, and ammonia nitrogen concentration in water;

[0008] The normal measurement determination module is used to establish a pre-measurement condition one, and the pre-measurement condition one is the normal measurement state; when the pre-measurement condition one includes that the water body chromaticity is greater than the system preset water body chromaticity threshold and the conductivity is greater than the system preset conductivity threshold, then it is determined as a type of abnormal stage;

[0009] The abnormal stage judgment module is used to establish the second pre-measurement condition when the water source is in a first-class abnormal stage; the second pre-measurement condition includes that when the water body reflectivity is lower than the water body reflectivity threshold preset by the system, or the water body ammonia nitrogen concentration is greater than the water body ammonia nitrogen concentration threshold preset by the system, the water quality impact factors are marked for the current water source, otherwise it returns to the normal measurement state within the preset sliding time window;

[0010] The deep learning prediction module is used to predict three target variables for the water source marked with water quality impact factors through a deep neural network model; construct a feature tensor based on the input water quality parameter data, and perform end-to-end feature extraction through the backpropagation algorithm in the shared hidden layer;

[0011] The multi-task regression module parallelly regresses three target variables based on the task output layer of the multi-task learning framework, and optimizes the discrimination of the three target variables through independent loss functions; the three target variables include seasonal influence variables, treatment response lag variables, and abnormal water quality variables.

[0012] In a preferred embodiment, the water body chromaticity is calculated based on a weighted time decay model, and a weight coefficient w i is introduced to reflect the influence degree of historical data on the current chromaticity, and a time decay factor (1 + α·(t - t i )) is introduced to describe the decay effect of historical observation data over time; the weighted historical data is decayed to reflect the change trend of the water body chromaticity over time;

[0013]

[0014] where C color (t) is the water body chromaticity value at the current moment t; D i (t) is the observed value of the water body chromaticity at the historical moment t i ; w i represents the weight coefficient corresponding to the moment t i ; α is the time decay coefficient; t is the current moment; t i is the time stamp of the historical moment; n is the size of the time window.

[0015] In a preferred embodiment, by combining a weighting factor and a decay factor to reflect the change of conductivity over time and dynamic factors, through the weighting factor C j (t) and V j (t) respectively reflect the influences of factors such as ion concentration and flow rate in the water, and use β j to adjust its weight on the conductivity; it is assumed that E(t) represents the conductivity value at the current moment t;

[0016]

[0017] Among them, C j (t) represents the j-th factor related to conductivity; V j (t) represents the j-th factor affecting conductivity; β j represents the weight used to adjust the influence of the j-th factor on conductivity; δ j is the time decay coefficient; γ j represents the degree of adjustment for the effects of δ j and η; η is the decay exponent; t j is the timestamp of historical data; t is the current moment.

[0018] In a preferred embodiment, by integrating historical reflectance data and combining with a decay function to reflect the influence of historical data on the current reflectance; introducing non-linear time adjustment, through simulating the influence of the water surface state on reflectance; defining R(t) as the water reflectance at the current moment t

[0019]

[0020] where φ(t - t′) is the function of reflectance decay over time; S(t′) represents the parameter of the water surface state; τ(t′) is the coefficient used to adjust the influence of the water surface state on reflectance; ζ(t′) represents the adjustment coefficient used to control the change of reflectance over time; (t - t′) represents the time difference.

[0021] In a preferred embodiment, by weighting the historical ammonia nitrogen concentration-related factors to calculate the current ammonia nitrogen concentration, and its weight is controlled by the time decay factor λ(t - t′). As the time interval t - t′ increases, the influence of historical data on the current ammonia nitrogen concentration decreases. By integrating multiple historical data over a past period of time, the current ammonia nitrogen concentration is predicted; N(t) is used to represent the ammonia nitrogen concentration at the current moment t;

[0022]

[0023] where A(t′) is the ammonia nitrogen-related factor at the historical moment t′; λ(t - t′) is the time decay factor; μ is the adjustment coefficient; (t - t′) α is the time difference term; α is the time decay exponent.

[0024] In a preferred embodiment, inputting water quality parameter data to construct a feature tensor. The water quality parameter data includes C color (t), E(t), R(t), N(t), and constructing the water quality parameter data into a feature tensor X t at the time step t for training and feature extraction by a deep neural network;

[0025] X t= [C color (t), E(t), R(t), N(t)] where t ∈ [t 1 , t 2 , …, t n ;

[0026] Where X t represents the input feature tensor corresponding to the current time step t. The value of each water quality parameter in the water quality parameter data is indexed by time t, constituting an element of the input feature;

[0027] In the deep neural network, the input feature tensor X t undergoes feature extraction through multiple hidden layers, and finally outputs a predicted value. The training of the network is carried out through the backpropagation algorithm, and the loss function is used to optimize the weights;

[0028] The output of the hidden layer is passed through the ReLU activation function;

[0029] H t = f(W (1) ·X t + b (1) );

[0030] The output layer makes predictions and regresses three target variables in parallel;

[0031]

[0032] Where W (1) and b (1) are the weight matrix and bias term of the first layer respectively, acting on the input feature X t , and performing a non-linear mapping through the activation function f(·) to generate the hidden layer output H t ; In the multi-task learning framework, the task output layer takes the output H t of the hidden layer as the input and predicts three target variables respectively: the seasonal influence variable the treatment response lag variable the abnormal water quality variable f seasonal (H t ) represents the function for seasonal influence; f delay (H t ) represents the function for treatment response lag; f anomaly (H t ) represents the function for abnormal water quality detection; H t represents the historical water quality data at time t;

[0033] Under the multi-task learning framework, three target variables are predicted in the task output layer through parallel regression. The regression task for each target variable includes an independent loss function, and the learning of the model is improved by optimizing the loss function of each task.

[0034] including a loss function as the total loss; L total = L seasonal + L delay + L anomaly ;

[0035] where L total is the total loss function, which is the sum of the loss functions of multiple target tasks; L seasonal is the regression loss function for seasonal influence; L delay is the regression loss function for dealing with response lag; L anomaly is the regression loss function for abnormal water quality;

[0036] The loss function corresponding to each target variable is used to measure the error between the predicted value and the actual value. The weights of the network are optimized through backpropagation. Through parallel regression, each task is independently optimized on the basis of shared features to achieve the prediction of the target variable.

[0037] In a preferred embodiment, the backpropagation algorithm is introduced. The propagation algorithm calculates the gradient of the loss function of each task and backpropagates it to each layer of the network to update the weights and biases; perform gradient calculation;

[0038]

[0039] Calculate the gradient of the total loss function with respect to the network weight W through the backpropagation algorithm, and use the gradient to update the weights of each layer. The gradient contribution of each loss function to the weights is independent, but joint optimization is performed on the shared hidden layer parameters;

[0040] Enhance the learning ability of the model to distinguish three target variables through time series features and external environmental data;

[0041] Capture the delay effect of water quality changes through lag features, and construct lag feature X lag ;

[0042] X lag = [C color (t - 1), E(t - 1), R(t - 1), N(t - 1)];

[0043] External environmental feature X ext includes flow rate Q(t), temperature T(t), humidity H(t);

[0044] X ext= [Q(t), T(t), H(t)];

[0045] The trend analysis feature captures the long-term change trend of water quality based on the trend analysis of historical data.

[0046] The technical effects and advantages of the present invention:

[0047] 1. By introducing multi-dimensional water quality fluctuation information and a real-time monitoring mechanism, the present invention realizes the management of fluoride concentration control. Based on water quality parameters such as water chromaticity, conductivity, reflectivity, and ammonia nitrogen concentration, combined with a deep neural network for real-time prediction, it is beneficial to identify water quality fluctuations and perform corresponding regulation, dynamically respond to water quality changes, and improve the convenience of water source control.

[0048] 2. Through a dynamic determination mechanism for pre-set measurement conditions 1 and 2, it is possible to judge in real time whether the water source enters an abnormal stage, and accordingly adjust the fluoride removal strategy. The pre-set conditions are based on changes in water quality parameters, combined with time window and sliding window control, to ensure that the system can respond in a timely manner to any water quality changes, preventing the fluoride concentration from exceeding the standard due to abnormal water quality not being processed in time;

[0049] 3. A deep neural network model is used for feature extraction of water quality fluctuations. Through a multi-task learning framework, three target variables of seasonal influence, processing response lag, and abnormal water quality are regressed in parallel, improving the convenience of the model for predicting water quality fluctuations;

[0050] 4. By integrating external environmental data and time series analysis, and combining meteorological factors, the present invention enhances the model's prediction ability for water quality changes. Through the lag feature and the time dependence of historical data, it is beneficial for the model to capture the delay effect of water quality changes and make flexible adjustments based on historical data, improving the robustness of the prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 It is a schematic diagram of the system module of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0052] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0053] Referring to the attached Figure 1 , a multi-dimensional intelligent control system for defluorination according to an embodiment of the present invention includes a water quality fluctuation monitoring module, a normal measurement determination module, an abnormal stage judgment module, a deep learning prediction module, and a multi-task regression module;

[0054] The water quality fluctuation monitoring module is used to obtain water quality fluctuation information of the water source that needs to be defluorinated in real time, establish pre-measurement condition 1 and pre-measurement condition 2 based on the water quality fluctuation information, and determine whether there are water quality influencing factors in the current water source through pre-measurement condition 1 and pre-measurement condition 2;

[0055] The water quality impact information in the water quality fluctuation monitoring module includes water chromaticity, conductivity, water reflectivity, and ammonia nitrogen concentration in water;

[0056] The normal measurement judgment module is used to establish the pre-measurement condition 1, which is the normal measurement state; the pre-measurement condition 1 includes that when the water body chromaticity is greater than the water body chromaticity threshold preset by the system, and the conductivity is greater than the conductivity threshold preset by the system, then a type of abnormal stage is determined;

[0057] The abnormal stage judgment module is used to establish the second pre-measurement condition when the water source is in a type of abnormal stage; the second pre-measurement condition includes marking the water quality impact factor for the current water source when the water body reflectance is lower than the water body reflectance threshold preset by the system, or the water body ammonia nitrogen concentration is greater than the water body ammonia nitrogen concentration threshold preset by the system, otherwise it returns to the normal measurement state within the preset sliding time window;

[0058] The deep learning prediction module is used to predict three target variables for water sources marked with water quality influencing factors through a deep neural network model; construct feature tensors based on the input water quality parameter data, and perform end-to-end feature extraction in the shared hidden layer through the back propagation algorithm;

[0059] The multi-task regression module regresses three target variables in parallel based on the task output layer of the multi-task learning framework, and optimizes the distinction of the three target variables through independent loss functions; the three target variables include seasonal impact variables, treatment response lag variables, and abnormal water quality variables;

[0060] Regarding the above-mentioned solution, it should be further noted that in order to ensure effective water quality monitoring of the water source that needs to be defluorinated, this solution adopts the strategy of obtaining water quality fluctuation information in real time, and accurately judges whether there are water quality impact factors in the water source by establishing pre-measurement condition 1 and pre-measurement condition 2. First of all, by monitoring water quality parameters such as water body chromaticity, conductivity, water body reflectivity, and ammonia nitrogen concentration, we can build a complete water quality fluctuation monitoring system, in which pre-measurement condition 1 is used to judge whether the water source enters the abnormal stage. If the water body chromaticity is greater than the preset threshold and the conductivity exceeds the standard set by the system, the system determines that the water source is in the first type of abnormal stage. This stage is a preliminary screening for potential risks that cannot be identified under conventional monitoring conditions. When the water source enters this stage, the further pre-measurement condition 2 is used to further screen through water body reflectivity and ammonia nitrogen concentration to ensure more refined abnormal identification. If these parameters return to normal within the preset sliding time window, the system returns to the normal measurement state; otherwise, it continues to be marked as the abnormal state. These judgment conditions can be dynamically adjusted to reflect the changes in water quality fluctuations, ensuring the timeliness and accuracy of water source water quality control. Next, for the water source marked as a water quality impact factor, it is analyzed through a deep neural network model, a feature tensor composed of water quality parameter data is constructed, feature extraction is carried out through the shared hidden layer in deep learning, and end-to-end training is carried out using the backpropagation algorithm. In this process, the network can learn the implicit patterns of water quality fluctuations in the shared feature space, and then parallelly regress three target variables - seasonal impact variable, treatment response lag variable, and abnormal water quality variable through a multi-task learning framework.

[0061] The loss function of each target variable is independently optimized, which can improve the model's ability to distinguish different dimensions of water quality fluctuations and prediction accuracy. This design not only improves the identification ability of water quality impact factors, but also ensures that the system can conduct refined and dynamic monitoring and prediction of water quality from multiple dimensions, avoiding the over-simplification and limitations of traditional methods. Through such a deep learning framework, we can respond to the changes in the water source in real time and make more scientific and accurate treatment decisions. This design is based on a deep understanding of different types of water quality fluctuations. By combining multi-task learning and time series features, the system has stronger response capabilities in the complex water quality control process.

[0062] Calculate the water body chromaticity based on the weighted time decay model, and introduce the weight coefficient w i to reflect the influence degree of historical data on the current chromaticity, and introduce the time decay factor (1 + α·(t - t i )) to describe the decay effect of historical observation data over time. The weighted historical data is decayed to reflect the change trend of the water body chromaticity over time.

[0063]

[0064] Among which C color (t) is the water body chromaticity value at the current moment t. The water body chromaticity value is an index used to reflect the concentration of visible substances in water. In this solution, it is related to the concentrations of components such as organic matter and suspended solids in water; D i (t) is the observed value of the water body chromaticity at the historical moment t i of the water body chromaticity, and D i (t) is used to predict and analyze the future chromaticity value; w i represents the weight coefficient corresponding to the moment t i . In the above formula, w i is used to describe the influence degree of different historical observation data on the current water body chromaticity. The weight coefficient will be adjusted over time, and the more recent historical data will have a higher weight; α is the time decay coefficient, indicating the decay rate of the water body chromaticity changing with time; t is the current moment; t i is the time stamp of the historical moment. Each D i (t) corresponds to a specific historical moment t i ; n is the size of the time window, and n is used to define the retrospective range of data collection; i is the dataset index;

[0065] In the above formula, the closer D i (t) is to the current moment t, the greater its influence on the current chromaticity, and the more distant historical data has a smaller influence; in addition, w i and α are used to quantify the importance of historical data in time.

[0066] By combining the weighting factor and the decay factor to reflect the change of conductivity with time and dynamic factors, through the weighting factors C j (t) and V j (t) respectively reflect the influences of the ion concentration and flow rate in water, etc., and use β j to adjust its weight on the conductivity; it is proposed that E(t) represents the conductivity value at the current moment t;

[0067]

[0068] Among which C j (t) represents the j-th factor related to the conductivity; V j (t) represents the j-th factor affecting the conductivity; β j represents the weight used to adjust the influence of the j-th factor on the conductivity; δ j is the time decay coefficient; γ j represents the value used to adjust δ jThe degree of influence of and η, that is, the sensitivity to adjust the attenuation rate; η is the attenuation exponent, which determines the non - linear degree of the influence of historical data on the current conductivity. The larger η is, the more sharply the influence of historical data will decrease; t j is the timestamp of historical data, used to indicate the time point of historical conductivity observation; m is the total number of factors; t is the current moment;

[0069] In the above formula, as the attenuation factor part (1 + δ j ·(t - t j )) -η indicates that as time goes by, the influence of early observation data on the current conductivity gradually weakens; in addition, by introducing a non - linear attenuation function, the change of conductivity at different time points is simulated.

[0070] By integrating historical reflectivity data and combining with the attenuation function to reflect the influence of historical data on the current reflectivity; introducing non - linear time adjustment, through simulating the influence of the water surface state on the reflectivity; historical data at a relatively recent moment has a greater influence on the current reflectivity, and historical data far from the current moment has a smaller influence; it is assumed that R(t) is the water body reflectivity at the current moment t

[0071]

[0072] Among them, φ(t - t′) is the function of reflectivity decay with time, indicating that as time goes by, the influence of historical data on the current reflectivity gradually decreases; S(t′) represents the parameters of the water surface state, and S(t′) includes suspended solid concentration, temperature, etc., which affect the optical properties of the water surface and thus affect the reflectivity; τ(t′) is the coefficient used to adjust the influence of the water surface state on the reflectivity, which is related to the concentration of dissolved substances in the water body; ζ(t′) represents the adjustment coefficient used to control the change of reflectivity with time; (t - t′) represents the time difference, describing the time span from the historical time point t′ to the current moment t; t - n is the starting point of the time window.

[0073] The current ammonia - nitrogen concentration is calculated by weighting the factors related to historical ammonia - nitrogen concentration, and its weight is controlled by the time - decay factor λ(t - t′). As the time interval t - t′ increases, the influence of historical data on the current ammonia - nitrogen concentration decreases. By integrating multiple historical data within a past period of time, the current ammonia - nitrogen concentration is predicted; N(t) is used to represent the ammonia - nitrogen concentration at the current moment t;

[0074]

[0075] Among them, A(t′) is the ammonia nitrogen related factor at the historical moment t′, which represents the factors related to the ammonia nitrogen concentration in the water body at this time point. It includes the concentration of amino acids and nitrogen source substances in the water. A(t′) reflects the source or generation amount of the change in ammonia nitrogen concentration; λ(t - t′) is the time decay factor, which is used to describe the change of the influence of historical data on the current ammonia nitrogen concentration over time. As t′ increases, the influence of historical data gradually weakens; μ is the adjustment coefficient, and μ is used to control the sensitivity of time decay and affect the degree of influence of (t - t′) α on the calculation of ammonia nitrogen concentration. A larger μ value will accelerate the attenuation of the influence of historical data on the current ammonia nitrogen concentration; (t - t′) α is the time difference term, which represents the time interval between the current moment t and the historical moment t′; α is the time decay exponent, which is used to describe the rate of attenuation of the influence of historical data on the current ammonia nitrogen concentration over time. A larger α value indicates that the influence of historical data on the current value will weaken rapidly;

[0076] The (t - t′) term in the above formula α shows, through non-linear time decay, that the influence of historical data on the current ammonia nitrogen concentration has different weakening rates. By simulating the dynamic change law of pollutants in the water body through non-linear decay, the deficiencies of the simple linear decay model are avoided. And through the adjustment coefficient μ and the decay exponent α, the influence weight of historical data on the current ammonia nitrogen concentration is adjusted to meet the requirements of different water quality conditions. In addition, the weighted ammonia nitrogen factor A(t′) represents the contribution of the ammonia nitrogen concentration related factor of historical data to the current ammonia nitrogen concentration at each historical time point. In practical applications, reasonable attenuation can be selected to achieve a reasonable influence of historical data on the current ammonia nitrogen concentration;

[0077] Regarding the above solution, it should be noted that by establishing pre-measurement conditions one and two to detect water quality fluctuations in real time, and determining whether to enter the abnormal stage under different water quality thresholds. The establishment of the pre-conditions is based on water quality parameters such as chromaticity, conductivity, reflectivity, and ammonia nitrogen concentration. When the conditions are met, they are marked as water quality impact factors, and the conversion between the normal state and the abnormal state is dynamically adjusted by setting a sliding window. Next, the deep neural network model is trained with these water quality fluctuation data, constructs a feature tensor, and performs end-to-end feature extraction through a shared hidden layer. This network model adopts a multi-task learning framework and regresses three target variables in parallel through independent loss functions: seasonal impact, treatment response lag, and abnormal water quality, ensuring that each target variable can be independently learned and optimized. Through deep learning and the multi-task learning framework, water quality change patterns can be extracted from the data, combined with time series analysis and external environmental data to enhance the model's prediction ability for water quality changes, and achieve accurate differentiation and timely response to different water quality fluctuations. In this way, the defluorination scheme is quickly adjusted according to real-time data, thereby improving the intelligence and flexibility of water source management. The core of this method lies in the fusion analysis of multi-dimensional data, enabling accurate prediction and processing of the multi-level and dynamic changes of water quality fluctuations, avoiding the limitations and lag of traditional monitoring methods.

[0078] Construct a feature tensor by inputting water quality parameter data, and the water quality parameter data includes C color (t), E(t), R(t), N(t), and construct the water quality parameter data into a feature tensor X according to the time step t t , which is used for training and feature extraction by the deep neural network;

[0079] X t =[C color (t), E(t), R(t), N(t)] where t ∈ [t 1 , t 2 , …, t n ;

[0080] Among them, X t represents the input feature tensor corresponding to the current time step t, and the value of each water quality parameter in the water quality parameter data is indexed by time t, constituting an element of the input feature;

[0081] In the deep neural network, the input feature tensor X t undergoes feature extraction through multiple hidden layers, and finally outputs a predicted value. The training of the network is carried out through the backpropagation algorithm, minimizing the loss function to optimize the weights;

[0082] Output through the ReLU activation function for the hidden layer;

[0083] H t =f(W(1) ·X t +b (1) );

[0084] The output layer makes predictions and regresses three target variables in parallel;

[0085]

[0086] where W (1) and b (1) are the weight matrix and bias term of the first layer respectively, which act on the input feature X t , and perform a non-linear mapping through the activation function f(·) to generate the hidden layer output H t ; In the multi-task learning framework, the task output layer takes the output H t of the hidden layer as the input and predicts three target variables respectively: the seasonal impact variable the treatment response lag variable the abnormal water quality variable f seasonal (H t ) represents the function for seasonal impact; f delay (H t ) represents the function for treating response lag, which depends on the historical data and lag features of water quality, including the time dependence of water quality changes; f anomaly (H t ) represents the function for abnormal water quality detection, which identifies sudden abnormalities in water quality fluctuations through historical water quality data and an anomaly detection mechanism; H t represents the historical water quality data at time t;

[0087] In the multi-task learning framework, the three target variables are predicted in the task output layer through parallel regression. The regression task for each target variable includes an independent loss function, and the learning of the model is improved by optimizing the loss function of each task;

[0088] including the loss function as the total loss; L total =L seasonal +L delay +L anomaly ;

[0089] where L total is the total loss function, which is the sum of the loss functions of multiple target tasks; L seasonal is the regression loss function for seasonal impact; L delay is the regression loss function for treating response lag; L anomaly is the regression loss function for abnormal water quality;

[0090] The loss function corresponding to each target variable is used to measure the error between the predicted value and the actual value. By backpropagation, the weights of the network are optimized. Through parallel regression, each task is independently optimized based on shared features to achieve the prediction of the target variable.

[0091] The backpropagation algorithm is introduced. The propagation algorithm calculates the gradients of the loss functions of each task and backpropagates them to the layers of the network to update the weights and biases; gradient calculation is performed;

[0092]

[0093] Calculate the gradient of the total loss function with respect to the network weight W through the backpropagation algorithm. The gradient is used to update the weights of each layer. The gradient contribution of each loss function to the weights is independent, but joint optimization is performed on the shared hidden layer parameters;

[0094] Enhance the learning ability of the model to distinguish three target variables through time series features and external environment data;

[0095] Capture the delay effect of water quality changes through lag features, and construct lag feature X lag ;

[0096] X lag =(C color (t - 1), E(t - 1), R(t - 1), N(t - 1)];

[0097] External environment feature X ext includes flow rate Q(t), temperature T(t), humidity H(t);

[0098] X ext =[Q(t), T(t), H(t)];

[0099] The trend analysis feature captures the long-term change trend of water quality based on the trend analysis of historical data. The analysis modes include linear regression or wavelet transform;

[0100] During the learning process of the deep neural network, combined with the added features, distinguish three target variables:

[0101] Seasonal influence dimension: Driven by the periodic changes of water quality parameters, use seasonal features to extract seasonal fluctuations. Seasonal features include Fourier transform;

[0102] Processing response lag dimension: The lag effect of water quality fluctuations. Due to the reaction lag of water treatment facilities, the model identifies it through lag time features and time windows;

[0103] Abnormal water quality dimension: Sudden water quality anomalies, such as a sharp increase in pollutants, can be identified by the model through anomaly detection algorithms, including IsolationForest and One-ClassSVM.

[0104] To effectively distinguish the three target variables of seasonal influence, treatment response lag, and abnormal water quality, additional features are introduced. The lag effect needs to combine time series data and historical water quality trends, and capture the time dependence of water quality fluctuations through lag features and external environmental data such as temperature, precipitation, and flow. Seasonal influence involves periodic changes, so seasonal trend analysis and climate variables are introduced to help identify seasonal-related fluctuation patterns. Abnormal water quality fluctuations are usually sudden, so features such as instantaneous flow and anomaly detection models are used to distinguish such fluctuations. By combining multi-dimensional features with four water quality parameters, a deep neural network is used to distinguish these three target variables, improving the prediction accuracy and robustness of the model.

[0105] In addition, it should be noted that this solution realizes the prediction and control of fluoride concentration in the water source and the removal of fluoride during the sewage treatment process by real-time monitoring of water quality fluctuations and combining deep learning with a multi-task learning framework. The key to fluoride removal is to ensure that the fluoride concentration in the water source is maintained within a safe range to avoid the negative impact of excessive fluoride on the water ecosystem and human health. By introducing parameters such as chromaticity, conductivity, reflectivity, and ammonia nitrogen concentration in the water quality fluctuation analysis and using a deep neural network model for feature extraction of time series data, we can predict the change trend of fluoride concentration in real time and take corresponding treatment measures in advance. The pre-measurement conditions and sliding time window mentioned in the solution enable the model to flexibly judge whether the water quality enters the abnormal stage according to the influence of historical data and the external environment, and timely adjust the fluoride removal strategy during the treatment process. This method not only improves the automation and intelligence of the fluoride removal process, but also enhances the sensitivity to water quality changes, ensuring that the fluoride concentration is always within a controllable range in a complex water treatment environment, thus effectively protecting water resources and public health and achieving the ultimate goal of sewage defluorination.

[0106] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A multi-dimensional intelligent control system for fluorine removal, comprising a water quality fluctuation monitoring module, a normal measurement determination module, an abnormal stage determination module, a deep learning prediction module, and a multi-task regression module, characterized in that: The water quality fluctuation monitoring module is used to obtain water quality fluctuation information of the water source that needs to be defluorinated in real time, establish pre-measurement condition 1 and pre-measurement condition 2 based on the water quality fluctuation information, and determine whether there are water quality influencing factors in the current water source through pre-measurement condition 1 and pre-measurement condition 2; The water quality impact information in the water quality fluctuation monitoring module includes water chromaticity, conductivity, water reflectivity, and ammonia nitrogen concentration in water; The normal measurement judgment module is used to establish the pre-measurement condition 1, which is the normal measurement state; the pre-measurement condition 1 includes that when the water body chromaticity is greater than the water body chromaticity threshold preset by the system, and the conductivity is greater than the conductivity threshold preset by the system, then a type of abnormal stage is determined; The abnormal stage judgment module is used to establish the second pre-measurement condition when the water source is in a type of abnormal stage; the second pre-measurement condition includes marking the water quality impact factor for the current water source when the water body reflectance is lower than the water body reflectance threshold preset by the system, or the water body ammonia nitrogen concentration is greater than the water body ammonia nitrogen concentration threshold preset by the system, otherwise it returns to the normal measurement state within the preset sliding time window; The deep learning prediction module is used to predict three target variables for water sources marked with water quality influencing factors through a deep neural network model; construct feature tensors based on the input water quality parameter data, and perform end-to-end feature extraction in the shared hidden layer through the back propagation algorithm; The multi-task regression module regresses three target variables in parallel based on the task output layer of the multi-task learning framework, and optimizes the distinction of the three target variables through independent loss functions; the three target variables include seasonal influence variables, processing response lag variables, and abnormal water quality variables.

2. A multi-dimensional intelligent control system for fluorine removal according to claim 1, characterized in that: The water chromaticity is calculated based on the weighted time decay model, and the weight coefficient w is introduced i To reflect the impact of historical data on the current chromaticity, a time attenuation factor (1+α·(tt i )) is used to describe the attenuation effect of historical observation data over time; the weighted historical data is attenuated to reflect the changing trend of water body chromaticity over time; Among them C color (t) is the chromaticity value of the water body at the current time t; D i (t) is the historical moment t i Observed value of water chromaticity; w i Represents the time corresponding to t i The weight coefficient of ; α is the time attenuation coefficient; t is the current time; t i is the timestamp of the historical moment; n is the size of the time window.

3. A multi-dimensional intelligent control system for defluorination according to claim 2, characterized in that: By combining the weighting factor and the attenuation factor to reflect the change of conductivity with time and dynamic factors, the weighting factor C j (t) and V j (t) respectively reflects the influence of ion concentration and flow rate in water, and uses β j Adjust its weight on conductivity; propose E(t) to represent the conductivity value at the current time t; Among them C j (t) represents the jth factor related to conductivity; V j (t) represents the jth factor affecting conductivity; β j Indicates the weight used to adjust the influence of the jth factor on conductivity; δ j is the time attenuation coefficient; γ j Indicates the value used to adjust δ j and η are the degrees of action; η is the attenuation index; t j Timestamp for historical data; t is the current time; m is the total number of factors.

4. A multi-dimensional intelligent control system for defluorination according to claim 3, characterized in that: By integrating the historical reflectivity data and combining the attenuation function, the influence of historical data on the current reflectivity is reflected; nonlinear time adjustment is introduced, and the Simulate the influence of water surface state on reflectivity; assume R(t) is the reflectivity of water at the current time t; Where φ(tt′) is a function of the reflectivity decay over time; S(t′) represents the parameters of the water surface state, which include suspended matter concentration and temperature; τ(t′) is a coefficient used to adjust the influence of the water surface state on the reflectivity; ζ(t′) represents the adjustment coefficient used to control the reflectivity change over time; (tt′) represents the time difference; t′ is a historical time point; t is the current moment; tn represents the starting point of the time window, and n is the length of the time window.

5. A multi-dimensional intelligent control system for defluorination according to claim 4, characterized in that: The current ammonia nitrogen concentration is calculated by weighting the historical ammonia nitrogen concentration-related factors. The weight is controlled by the time decay factor λ(tt′). As the time difference (tt′) increases, the influence of historical data on the current ammonia nitrogen concentration decreases. The current ammonia nitrogen concentration is predicted by integrating multiple historical data over a period of time. N(t) is used to represent the ammonia nitrogen concentration at the current time t. Where A(t′) is the ammonia nitrogen related factor at the historical time t′; λ(tt′) is the time decay factor; μ is the adjustment coefficient; (tt′) α is the time difference term; α is the time decay exponent.

6. A multi-dimensional intelligent control system for fluorine removal according to claim 5, characterized in that: Input water quality parameter data to construct feature tensor. Water quality parameter data includes C color (t), E(t), R(t), N(t), and construct the water quality parameter data into a feature tensor X according to the time step t t , used for deep neural network training and feature extraction; X t =[C color (t),E(t),R(t),N(t)]wheret∈[t1,t2,…,t n ] Where X t Represents the input feature tensor corresponding to the current time step t. The value of each water quality parameter in the water quality parameter data is indexed by time step t and constitutes an element of the input feature; In a deep neural network, the input feature tensor X t After multiple hidden layers, feature extraction is performed and the predicted value is finally output. The network is trained through the back propagation algorithm, and the loss function is used to optimize the weights; Output the hidden layer through the ReLU activation function; H t =f(W (1) ·X t +b (1) ) The output layer predicts and regresses three target variables in parallel; Where W (1) and b (1) are the weight matrix and bias term of the first layer, acting on the input feature X t , and performs nonlinear mapping through the activation function f(·) to generate the hidden layer output H t ; In the multi-task learning framework, the task output layer converts the output H of the hidden layer t As input, three target variables are predicted separately: seasonal impact variable Handling lagged response variables Abnormal water quality variables f seasonal (H t ) represents the function for seasonal effects; f delay (H t ) represents the function used to deal with response lag; f anomaly (H t ) represents the function used for abnormal water quality detection; H t Represents the historical water quality data at time t; In the multi-task learning framework, the three target variables are predicted in the task output layer through parallel regression. The regression task of each target variable includes an independent loss function. The learning of the model is improved by optimizing the loss function of each task. Included as the loss function of the total loss; L total =L seasonal +L delay +L anomaly Where L total is the total loss function, which is the sum of the loss functions of multiple target tasks; L seasonal is the regression loss function of seasonal effects; L delay is the regression loss function for dealing with response lag; L anomaly is the regression loss function of abnormal water quality; The loss function corresponding to each target variable is used to measure the error between the predicted value and the actual value. The weight of the network is optimized through back propagation. Through parallel regression, each task is independently optimized based on the shared features to achieve the prediction of the target variable.

7. A multi-dimensional intelligent control system for fluorine removal according to claim 6, characterized in that: Introduce the back propagation algorithm, which calculates the gradient of the loss function of each task and propagates it back to each layer of the network to update the weights and biases; perform gradient calculation; The gradient of the total loss function with respect to the network weight W is calculated through the back-propagation algorithm, and the gradient is used to update the weights of each layer. The gradient contribution of each loss function to the weight is independent, but they are jointly optimized on the shared hidden layer parameters. The learning ability of the model to distinguish the three target variables is enhanced by time series features and external environment data; The delayed effect of water quality changes is captured by the hysteresis feature, and the hysteresis feature X is constructed. lag ; X lag =[C color (t-1),E(t-1),R(t-1),N(t-1)] External environment characteristicsX ext Including flow Q(t), temperature T(t), humidity H(t); X ext =[Q(t),T(t),H(t)] Trend Analysis Feature Trend analysis based on historical data captures long-term trends in water quality.

Images