An artificial intelligence automated dosing method for sewage treatment scenarios

Abstract

The present invention relates to an artificial intelligence automated dosing method in a sewage treatment scenario, comprising the following steps: performing predictive analysis based on the dosing effect in the formally put into production sewage treatment scenario, obtaining a biased data set D1 to form a forward dosing prediction model; under the same dosing production process environment, using an experimental environment to correlate with the forward dosing prediction model, obtaining an abnormal boundary sample data set D2 in the experimental environment under the same environment; merging the biased data set D1 and the abnormal boundary sample data set D2 as a training sample set for a reverse dosing prediction model, retraining the reverse dosing prediction model, and obtaining a final dosing pre-test dose. The present invention can be used for different sewage dosing scenarios, using only a few industry-wide common indicators, combined with previous biased actual production data and small-scale experimental data and trained through an MLP multi-layer perceptron network, to achieve an automated and precise dosing solution that can be implemented in mass production.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent sewage treatment, and mainly relates to an artificial intelligence automated drug dosing method in a sewage treatment scenario. Background Art

[0002] The process of setting up a sewage or wastewater treatment process involves complex considerations. Treatment processes must be tailored to the specific sewage source and water quality, and customized plant and wastewater treatment facilities must be constructed. Factors involved include physical pollution, inorganic pollution, organic pollution, nutrient contamination and eutrophication, biological contamination, and radioactive contamination. Chemical dosing, with its wide applicability, universality, high efficiency, and stable results, has become a crucial step in sewage and wastewater treatment. Due to the varying water quality and customized process flows, the entire sewage and wastewater treatment system, from initial design to ongoing operation and maintenance, requires the analysis and operation of experienced engineers. This is especially true for the various chemical dosing processes (especially those where the relationship between dosage and effect is not strictly linear, and where multiple chemical combinations are being administered simultaneously). Decisions regarding chemical dosage require continuous monitoring by engineers.

[0003] With the rise of artificial intelligence (AI) in recent years, modeling technology has been widely applied in various industries. The smart water sector, with its inherent characteristics of big data support, complex interactions, and clear indicator expectations, is well-suited to the application of AI technology to solve data prediction and reasoning problems in various links. Furthermore, combined with business evolution, it has developed application technologies such as automated data presentation, indicator prediction, indicator risk control, monitoring and early warning, and precise chemical dosing. However, judging from the current state of technological development in the industry, there are still relatively few cases of AI technology being implemented in the field of environmental wastewater treatment. The main problem can be attributed to the lack of valid data samples. In actual implementation, the influencing factors involved in wastewater treatment processes are complex, and the treatment systems of different water plants have their own particularities. In addition, due to the high cost of indicator collection sensors, only some key indicators are usually collected, resulting in only a rough understanding of the details of the system's operation. Unlike theoretical model derivation, on the one hand, since the water outlet indicators are required to be within the compliance range in the real production environment, the data are all optimistic samples, lacking the trial and perception of the values ​​exceeding the standard under different situations; on the other hand, in some cases, in order to obtain the dosage data when the standard is exceeded, engineers often use the data of the test environment as sample training, but the test system cannot ensure consistency with actual production. Therefore, it is difficult for engineers to obtain a large amount of real water plant data covering all scenarios. The models trained with such data are either too conservative, resulting in waste of dosage, or cannot accurately control the standard boundaries, resulting in risks such as water outlet indicators exceeding the standard.

[0004] Therefore, although some theories have emerged in the industry to solve the problem of intelligent drug dosing through modeling, they are often difficult to implement due to the high cost of data acquisition and large data deviations. The main reason is that when applying drug dosing technology based on artificial intelligence big data models, it is inevitable to face the problem of data loss. Even for some project scenarios with a certain amount of historical operation data, there is also the problem of insufficient data. The main reasons are:

[0005] (1) The cost of building big data that AI intelligent model training relies on is high, the benefits are unclear, and the equipment investment cost is high;

[0006] (2) Historical experimental data still differs from the actual production environment and cannot be directly reused;

[0007] (3) In the artificial stage form of supplementary indicator collection, due to the limitation of detection efficiency, the sampling interval is long and the data curve is not continuous and smooth.

[0008] Based on the above problems, the core difficulty lies in the fact that the small amount of historical data samples cannot represent the optimal dosing experience. The label itself deviates from the optimal true value. If the inverse model is trained with this data, it will not be able to learn the optimal dosing method. If the data samples need to be supplemented, the investment cost will increase significantly. For this reason, there will be: First, the dosage in the historical data, out of caution in actual production, various dosages may be over-dosed or the dosages may complement each other. Therefore, the dosage obtained through learning is also cautious. Second, in the historical dosing process, especially in scenarios such as the simultaneous dosing of multiple agents, different agents can complement each other through the increase and decrease of each other. However, due to the small amount of data, the model cannot capture the complementary relationship between the combined agents, and it is difficult to judge which one is more appropriate, which one is less appropriate, and which agents in the sample are actually over-dosed and wasted. Summary of the Invention

[0009] In response to the above problems, the present invention proposes a low-cost, efficient, and automated dosing method based on artificial intelligence in sewage treatment scenarios. The present invention applies an MLP multi-layer perceptron neural network and adopts a model construction scheme under biased data samples. For the drug dosing scenarios of different sewage treatment systems, only a few common industry indicators are used, and combined with previous biased actual production data and small-scale experimental data, a prediction model is constructed from multiple steps of the forward and reverse paths to achieve an automated and precise dosing scheme that can be implemented in mass production. In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0010] The present invention provides an artificial intelligence automated drug administration method for sewage treatment, comprising the following steps:

[0011] Based on the prediction analysis of the dosing effect in the scenario of the formally put into operation sewage treatment system, the biased data set D1 is obtained to form a forward dosing prediction model;

[0012] Under the same dosing production process environment, an experimental model of the experimental system scenario was created in a proportionally scaled-down manner based on the formally put into production sewage system scenario. The experimental model was correlated with the forward dosing prediction model to obtain the abnormal boundary sample data set D2 under the same experimental environment. The abnormal boundary sample data set D2 under the experimental environment was supplemented by the proportionally scaled dosing method to achieve sample enhancement.

[0013] The biased data set D1 and the abnormal boundary sample data set D2 are combined as the training sample set of the reverse drug administration prediction model, and the reverse drug administration prediction model is retrained to obtain the final drug administration pre-test dose.

[0014] The above scheme is further preferred, and the predictive analysis of the dosing effect based on the formal production sewage treatment scenario includes the following steps:

[0015] Step 10: Collect biased data from past drug administration as training samples in a formal production environment, and perform normalization and indicator true value recovery on the training samples;

[0016] Step 11: Select the inlet characteristic index and the outlet characteristic index, use the sewage inlet characteristic index as the input index, and extract the n-th power value for each input index as the basic feature;

[0017] Step 12: Input the basic features into the forward drug dosing prediction model for prediction analysis to obtain output indicators that support the estimation based on the input indicators;

[0018] The above solution is further preferred, wherein the normalized value satisfies:

[0019] a_u = (a - a_min) / (a_max - a_min), where a_min and a_max are the maximum and minimum values ​​of each indicator respectively;

[0020] The indicator true value recovery process satisfies: a=a_u*(a_max-a_min)+a_min.

[0021] The above scheme is further preferred, wherein the experimental environment is correlated with the forward drug dosing prediction model and comprises the following steps:

[0022] Step 20: Based on the formally put into production sewage system scenario, a set of experimental models of the experimental system scenario is prepared in a scaled-down manner. Using the experimental environment with the same process as the formally put into production environment, according to the dosage of the formally put into production environment, the inlet and outlet characteristic indicators of the experimental environment are collected in accordance with the proportional reduction dosage method.

[0023] Step 21: Select the same water inlet and outlet characteristic indicators as those in the formal production environment and use them as training samples in the experimental environment.

[0024] Step 22: Perform correlation training on the training samples in the experimental environment using the forward drug dosing prediction model to obtain prediction indicators for abnormal boundary sample data in the experimental environment;

[0025] Step 23: Extract the n-th power value as the basic feature for the prediction indicator of each abnormal boundary sample data;

[0026] Step 24: Normalize the extracted indicators to obtain the output indicators estimated based on the input indicators in the experimental environment.

[0027] The above solution is further preferred, wherein retraining the reverse dosing prediction model to obtain the final pre-test dose of the drug comprises the following steps:

[0028] Step 30: Combine the biased data of the formal production environment and the abnormal boundary sample data of the experimental environment as the training sample set for intelligent drug delivery;

[0029] Step 31: normalize the training sample set of the intelligent dosing, select the water inlet characteristic index and the water outlet characteristic index of the training sample set of the intelligent dosing, and extract the n-th power value of each input index as the basic feature;

[0030] Step: 32: Reversely input the basic features into the forward drug administration prediction model for reverse training to obtain the drug administration prediction index of the reverse training.

[0031] The above scheme is further preferred, and a three-layer serial structure MLP multi-layer perceptron neural network is used for predictive analysis based on the formal dosing effect, and automatic training is performed for different sewage treatment scenarios to analyze the relationship between multiple sewage inlet characteristic input indicators to obtain a pre-test dose of the dosing.

[0032] The above scheme is further preferred, wherein the three-layer series structure of the MLP multi-layer perceptron neural network is: the first layer node input layer, the second layer is a hidden layer, and the optimal weight is obtained by repeatedly training and comparing the relationship between multiple common indicators, and the output of the third layer is the prediction target, that is, the predicted dosage of the drug administration reagent.

[0033] In summary, the present invention adopts the above technical solution, and the benefits of the present invention are:

[0034] (1) The present invention is a low-cost and efficient automated dosing method based on artificial intelligence. It applies an MLP multi-layer perceptron neural network and adopts a model under biased data samples to construct this solution. For the dosing scenarios of different sewage treatment systems, it uses only a few common industry indicators and combines previous biased actual production data and small-scale experimental data. It uses a multi-layer series MLP perceptron neural network structure to make predictions, thereby realizing an automated and precise dosing solution that can be implemented in mass production.

[0035] (2) The present invention combines artificial intelligence technology with the use of multi-layer perceptron (MLP) technology, relying on only a small number of general indicators as core indicators. Detection capabilities are easy to obtain, and various environmental protection manufacturers have already built and possessed detection capabilities, thereby solving the problem of precise dosing in water treatment scenarios. Furthermore, the cost of promotion and use is low.

[0036] (3) Through the multi-layer serial MLP perceptron neural network, the relationship between multiple complex general indicators can be automatically trained and analyzed. Based on low-cost data construction, combined with production data and experimental sample data, it can provide better reverse dosing prediction training, thereby forming a feedback network, and combined with the supplement of boundary data, the accuracy of the prediction training can be adjusted. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 This is a schematic diagram of the node structure of the MLP multi-layer perceptron neural network model of the present invention;

[0038] Figure 2 This is a training model diagram of the MLP multi-layer perceptron neural network of the present invention;

[0039] Figure 3 This is the prediction result diagram of traditional MLP training;

[0040] Figure 4 This is the effect diagram of traditional MLP training;

[0041] Figure 5 This is the effect diagram of the MLP training of the present invention. DETAILED DESCRIPTION

[0042] The present invention is described in detail below with reference to the accompanying drawings and specific implementations. The scope of protection of the present invention is not limited by the specific embodiments, but is defined by the claims.

[0043] Combine Figure 1 The specific implementation of the method of the present invention is described as follows. This embodiment provides an artificial intelligence automated dosing method in a sewage treatment scenario. The present invention uses the Fenton reagent in the Fenton process section of the industrial wastewater treatment system (the core dosing agent is hydrogen peroxide <h2o2>and ferrous sulfate

[0044] <fe2so4>) is used as an example to explain in detail the dosing scenario. The specific implementation measures and steps are as follows: First, the purpose of the present invention is to overcome the insufficient technology and data reserves of existing sewage treatment plants, and propose a low-cost solution. With the help of the same process experimental environment, a method of sample supplementation is realized to improve the accuracy of the intelligent dosing effect. In terms of implementation strategy, it is divided into three stages: including the following steps: based on the dosing effect in the formal production sewage treatment scenario, a predictive analysis is performed to obtain a biased data set D1 to form a forward dosing prediction model; under the same dosing production process environment, the experimental environment and the forward dosing prediction model are correlated to obtain the abnormal boundary sample data set D2 in the experimental environment under the same environment; according to the proportional reduction dosing method, the abnormal boundary sample data set D2 in the experimental environment is supplemented to achieve sample enhancement; the biased data set D1 and the abnormal boundary sample data set D2 are merged as the training sample set of the reverse dosing prediction model, and the reverse dosing prediction model is retrained to obtain the final dosing pre-test dose.

[0045] In the present invention, combined with Figure 1 and Figure 2 As shown in the figure, the predictive analysis of the dosing effect based on the formal production sewage treatment scenario includes the following steps:

[0046] Step 10: In a formal production environment, biased data of past drug administration are collected as training samples, and the training samples are normalized and the true value of the indicators is restored; the normalized value a_u satisfies: a_u = (a-a_min) / (a_max-a_min), where a_min and a_max are the maximum and minimum values ​​of each indicator respectively. The purpose of normalizing each indicator is to converge each feature to a unified quantity and range, and to improve the iterative optimization efficiency of gradient descent during model training; the maximum and minimum values ​​a_min and a_max of each indicator are recorded so that the true value of the indicator can be restored later by the formula, and the true value recovery of the indicator satisfies: a = a_u*(a_max-a_min)+a_min;

[0047] Step 11: Select the inlet characteristic index and the outlet characteristic index (the common index of water characteristic index and outlet characteristic index), use the sewage inlet characteristic index as the input index, and extract the n-th power value as the basic feature for each input index, including: u itself, quadratic u_2, negative first power u_n1, negative second power u_n2, and linear offset u_bia, a total of 5;

[0048] Step 12: Input the basic features into the forward dosing prediction model for prediction analysis, obtain the output indicators that support the estimation based on the input indicators, and input the basic features into the forward dosing prediction model for prediction analysis; based on the biased data of previous dosing as training samples, select the inlet characteristic indicators and outlet characteristic indicators as the initial condition data, normalize the initial conditions, converge each feature to a unified number and range, and use sparse samples in the formal production environment for training. The total number of input features is 25, the number of output labels is 1, the number of samples is 89, the batch_size is 4, and the maximum training epochs is 100, thereby obtaining support for estimating the output indicators based on the input indicators, thereby forming the output model M1 (output indicator model) of the forward prediction indicator, completing the first phase of the implementation strategy work;

[0049] In the present invention, combined with Figure 1 and Figure 2 As shown, the steps of correlating the experimental environment with the forward dosing prediction model include the following:

[0050] Step 20: Based on the formally put into production sewage system scenario, a set of experimental models under the experimental system scenario is prepared in a scaled-down manner. An experimental environment with the same process as the formally put into production environment is used. According to the dosage of the formally put into production environment, the inlet and outlet characteristic indicators of the non-exceeding area in the experimental environment and the inlet and outlet characteristic indicators of the exceeding area are collected in a proportionally reduced dosage manner. According to the dosage in the production environment, the dosage is scaled proportionally, and data samples are collected for the non-exceeding area and the exceeding area respectively. In order to improve the accuracy of the training, some data can be proportionally added according to the dosage in the production sample to collect the inlet and outlet indicators in the experimental environment.

[0051] Step 21: Select the same water inlet and outlet characteristic indicators as those in the formal production environment as the training samples in the experimental environment; select the water inlet and outlet characteristic indicators in the experimental environment as input indicators, including: pH value of water inlet detection in the experimental environment<ph_in> , water volume<water_amount> , cod value<cod_in> , hydrogen peroxide dosage <h2o2>and ferrous sulfate dosage <fe2so4>, water output indicators under experimental environment<cod_out> Among them, the water volume indicator needs to be amplified proportionally according to the production environment; using the forward dosing prediction model formed by the first phase implementation strategy, the above indicators (except<cod_out> As input, the prediction results of the forward dosing prediction model<cod_out_prd> As output indicators;

[0052] Step 22: The training samples in the experimental environment are associated with the forward dosing prediction model to obtain the prediction indicators of the abnormal boundary sample data in the experimental environment and form the association model M2; the experimental environment is associated with the production environment. Its purpose is to capture the correlation between the experimental environment and the production environment in the non-exceeding area, and in the exceeding area, take advantage of the easy acquisition of a large amount of test data in the experimental environment to supplement the samples for the production environment.

[0053] Step 23: Normalize the extracted indicators to obtain the output indicators estimated based on the input indicators in the experimental environment; the normalization scheme adopts a_u=(a-a_min) / (a_max-a_min), and records the maximum and minimum values ​​a_min and a_max of each indicator so that the true value of the indicator can be restored later through the formula: a=a_u*(a_max-a_min)+a_min; converge each feature to a unified quantity and range, which can improve the iterative optimization efficiency of gradient descent during training.

[0054] Step 24: Extract the n-th power value as the basic feature for the prediction index of each abnormal boundary sample data; including: u itself, quadratic u_2, negative first power u_n1, negative quadratic u_n2, and linear offset u_bia, a total of 5. The present invention uses an MLP multi-layer perceptron neural network to perform association training on the training samples in the experimental environment with the help of a forward dosing prediction model, and with the help of the above feature selection method, performs nonlinear expansion, adopts a mean square error loss function, and uses a gradient descent algorithm for iterative optimization to obtain the best prediction index. The above samples are used for training, with a total input feature number of 30, an output label number of 1, a sample number of 80, a batch_size of 4, and a maximum training epochs of 100. In the experimental environment, 200 test samples are added, 80 of which are collected in the exceeding standard area. Sample enhancement is achieved by adding abnormal boundary samples, thereby obtaining an association relationship to support the estimation of output indicators based on input indicators, forming an output model M2 (association output indicator model) of the association prediction indicator, and completing the second phase implementation strategy work.

[0055] In the present invention, combined with Figure 1 and Figure 2 As shown, retraining the reverse dosing prediction model to obtain the final dosing pre-test dose includes the following steps:

[0056] Step 30: Combine the biased data of the formal production environment and the abnormal boundary sample data of the experimental environment as the training sample set for intelligent drug delivery;

[0057] Step: 31: Normalize the training sample set of intelligent dosing. The normalization scheme adopts a_u=(a-a_min) / (a_max-a_min), record the maximum and minimum values ​​a_min and a_max of each indicator, so that the true value of the indicator can be restored later by the formula: a=a_u*(a_max-a_min)+a_min, select the water entry characteristic index and water exit characteristic index of the training sample set of intelligent dosing, and extract the nth power value as the basic feature for each input indicator, including: u itself, quadratic u_2, negative first power u_n1, negative quadratic u_n2, linear offset u_bia, a total of 5; in the present invention, the pH value during water entry detection<ph_in> , water volume<water_amount> , cod value<cod_in> , water output indicators<cod_out> , cod elimination amount<cod_bias=cod_in-cod_out> ; Select the dosage of hydrogen peroxide <h2o2>and ferrous sulfate dosage <fe2so4>As an output indicator

[0058] Step: 32: Reversely input the basic features into the forward dosing prediction model for reverse training to obtain the dosing prediction index of the reverse training; form a reverse dosing prediction model M3 based on the dosing prediction index obtained through reverse training; in the present invention, the total number of input features is 25, the number of output labels is 2, the number of samples is 289, the batch_size is 4, the maximum training epochs is 100, and the dosing prediction model M3 is generated. Reverse training is performed based on the training index to obtain the final intelligent reverse dosing prediction model M3, which can reversely infer the Fenton dosing dosage based on the current water inlet detection index and the expected water outlet index of production, and perform predictive analysis based on the formal dosing effect. A three-layer series structured MLP multi-layer perceptron neural network is used, such as Figure 2 As shown, the relationship between multiple sewage inlet characteristic input indicators is automatically trained and analyzed for different sewage treatment scenarios to obtain the pre-test dosage of the drug; the three-layer series structure of the MLP multi-layer perceptron neural network is as follows: the first layer node layer (Input Layer layer, node R n =12), which has included the linear and nonlinear transformation based on the original input parameters into 12 features. The second layer is the hidden layer (Hidden Layer layer, node R n =10), the optimal weights are obtained by repeated training and comparison of the relationship between multiple common indicators. The optimal number of neurons obtained by repeated training and comparison is 10. The third layer is the output layer (Output Layer layer, node R n =2) is the prediction target, and the output layer is the predicted dosage of the two Fenton reagents, that is, the predicted dosage of the dosing reagent; the input parameters of the multi-layer perception neural network include: pH value at the time of water inlet detection, water volume in the treatment pool, and COD value of the water inlet. The label uses the COD value at the time of water outlet detection. The sample data is simple and universal; in the forward prediction of water outlet index, a single output node<cod_out> , while the reverse intelligent dosing model has two output nodes, namely the predicted dosage of the two Fenton reagents and the dosage of hydrogen peroxide. <h2o2>and ferrous sulfate dosage <fe2so4>, the present invention uses 3000 samplings as the experiment, each coordinate point is a data sample after thinning, that is, one dose, the vertical coordinate data is normalized to show the relative size of the predicted dose and the actual dose. At the same time, the traditional MLP method effect diagram (such as Figure 3 and Figure 4 As shown) and the effects of the present invention (as shown Figure 5 shown), Figure 3 The figure in the middle is a comparison between the automatic prediction results and the true value. Compared with directly using the MLP prediction model, the prediction accuracy of the present invention is significantly improved. Figure 5 As shown, the dosage is lower than manual dosage, and the operating sample exceedance rate is 0%, demonstrating practical application value and low cost of promotion and use. The proposed method relies on only a small number of common indicators, thus ensuring detection capabilities. This method utilizes multi-layer perceptron (MLP) technology to automatically train and analyze the complex relationships between multiple common indicators, thereby solving the problem of precise dosing in water treatment scenarios and providing excellent dosing prediction results.

[0059] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. An artificial intelligence automated drug delivery method for sewage treatment, characterized by: The steps include: Based on the prediction analysis of the dosing effect in the scenario of the formally put into operation sewage treatment system, the biased data set D1 is obtained to form a forward dosing prediction model; The predictive analysis of the dosing effect based on the formally put into production sewage treatment scenario includes the following steps: Step 10: Collect biased data from past drug administration as training samples in a formal production environment, and perform normalization and indicator true value recovery on the training samples; Step 11: Select the inlet characteristic index and the outlet characteristic index, use the sewage inlet characteristic index as the input index, and extract the n-th power value for each input index as the basic feature; Step 12: Input the basic features into the forward drug dosing prediction model for prediction analysis to obtain output indicators that support the estimation based on the input indicators; Under the same dosing production process environment, an experimental model of the experimental system scenario was created in a proportionally scaled-down manner based on the formally put into production sewage system scenario. The experimental model was correlated with the forward dosing prediction model to obtain the abnormal boundary sample data set D2 under the same experimental environment. The abnormal boundary sample data set D2 under the experimental environment was supplemented by the proportionally scaled dosing method to achieve sample enhancement. Correlating the experimental environment with the forward dosing prediction model includes the following steps: Step 20: Based on the formally put into production sewage system scenario, a set of experimental models for the experimental system scenario is prepared in a proportionally reduced manner. An experimental environment with the same process as the formally put into production environment is used. According to the dosage of the formally put into production environment, the inlet and outlet characteristic indicators of the experimental environment are collected in accordance with the proportionally reduced dosage method. The inlet and outlet characteristic indicators of the experimental environment are collected in accordance with the dosage of the formally put into production environment. The inlet and outlet characteristic indicators of the experimental environment are collected in accordance with the proportionally reduced dosage method. When using the forward dosage prediction model for association training, a three-layer serial structure MLP multi-layer perceptron neural network is used to automatically train and analyze the relationship between multiple sewage inlet characteristic input indicators for different sewage treatment scenarios to obtain the pre-test dosage of the dosage. Step 21: Select the same water inlet and outlet characteristic indicators as those in the formal production environment and use them as training samples in the experimental environment; Step 22: Perform correlation training on the training samples in the experimental environment using the forward drug dosing prediction model to obtain prediction indicators for abnormal boundary sample data in the experimental environment; Step 23: Extract the n-th power value as the basic feature for the prediction indicator of each abnormal boundary sample data; Step 24: Normalize the extracted indicators to obtain the output indicators estimated based on the input indicators in the experimental environment; Merge the biased data set D1 and the abnormal boundary sample data set D2 as the training sample set of the reverse drug dosing prediction model, retrain the reverse drug dosing prediction model, and obtain the final drug dosing pre-test dose; Retraining the reverse dosing prediction model to obtain the final pre-test dose includes the following steps: Step 30: Combine the biased data of the formal production environment and the abnormal boundary sample data of the experimental environment as the training sample set for intelligent drug delivery; Step 31: normalize the training sample set of the intelligent dosing, select the water inlet characteristic index and the water outlet characteristic index of the training sample set of the intelligent dosing, and extract the n-th power value of each input index as the basic feature; Step 32: Reversely input the basic features into the forward drug administration prediction model for reverse training to obtain the drug administration prediction index of the reverse training.

2. The artificial intelligence automated drug administration method for sewage treatment according to claim 1, characterized in that: The normalization process satisfies: a_u = (a - a_min) / (a_max - a_min), where a_min and a_max are the maximum and minimum values ​​of each indicator respectively; The indicator true value recovery process satisfies: a=a_u*(a_max-a_min)+a_min.

3. The artificial intelligence automated drug administration method for sewage treatment according to claim 1, characterized in that: The three-layer series structure of the MLP multi-layer perceptron neural network is as follows: the first layer is a node input layer, the second layer is a hidden layer, and the optimal weight is obtained by repeatedly training and comparing the relationship between multiple common indicators. The output of the third layer is the prediction target, that is, the predicted dosage of the drug administration reagent.

Images