Terminal ultrapure water quality intelligent pre-judgment device based on RBF neural network algorithm model and pre-judgment and control method

Abstract

The invention discloses a terminal ultrapure water quality intelligent pre-judgment device based on an RBF neural network algorithm model and a pre-judgment and control method.The pre-judgment device is arranged on a bypass in front of a raw water pool of an ultrapure water preparation system and comprises a processing system, a detection system and a water quality pre-judgment system, and the processing system is used for preprocessing raw water; the detection system is used for detecting the pretreated water in real time to obtain pretreated water parameter real-time data; the water quality pre-judging system is used for pre-judging the contents of TOC and B in the terminal ultrapure water by adopting an RBF neural network according to the pre-treated water parameter real-time data to obtain a prediction result, and regulating the dosage of chemicals in the ultrapure water preparation system in combination with the historical dosage of chemicals, so that the contents of TOC and B in the terminal ultrapure water stably reach the standard. According to the method, the TOC content and the B content in the ultrapure water production water can be pre-judged at the raw water end, historical data can be compared, the agent adding amount can be regulated and controlled, and it is ensured that the TOC content and the B content in terminal ultrapure water stably reach the standard.

Description

Technical Field

[0001] The present invention relates to electronic grade ultrapure water production technology, and mainly to a terminal ultrapure water quality intelligent prediction device and a prediction and control method based on an RBF neural network algorithm model. Background Art

[0002] With the improvement of manufacturing process, the integrated circuit industry has increasingly stringent requirements for ultrapure water quality. Trace pollutants such as total organic carbon (TOC) and boron (B) that are easily exceeded can easily affect the quality of water produced by electronic ultrapure water systems. Taking TOC as an example, at present, the TOC requirement for ultrapure water used in 12-inch integrated circuit production is less than 1μg / L or even 0.5μg / L, and Sievers M500e instruments are usually used for monitoring at the terminal water production point. B requirements are even lower. At present, B in ultrapure water usually needs to be less than 50ng / L, or even less than 10ng / L, and can only be detected online at the terminal using Sievers boron meters.

[0003] Although the detection data of the terminal can feedback the water quality of the water produced at this stage, there is a high hysteresis. For example, CN107758942A introduces an ultrapure water manufacturing system. The patent has an ultrapure water detection system. The detection system is located after the multi-stage resin column and the filter, that is, the terminal water production. When using this detection system, if the water quality of the terminal ultrapure water deteriorates, or even there is a risk of TOC or B exceeding the standard, the system operation and maintenance personnel will control the terminal TOC and B indicators by adding a fixed proportion of oxidants to the front-end raw water tank and other units or adjusting the pH value of the RO (reverse osmosis) section water. However, during on-site operation, the overall preparation process of ultrapure water has a time span. The raw water will stay in the raw water tank for nearly 20 hours. From the raw water tank to the terminal water outlet, it usually needs to go through the pretreatment system, process system and the final polishing system. The water will stay in the RO water tank and the pure water tank for nearly 10 hours. That is to say, after the TOC and B at the ultrapure water terminal do not meet the standards, adding advanced oxidants to the raw water tank to remove TOC or adjusting the pH of the RO section to remove B will still affect the water quality of the subsequent 10 hours or so, affecting the normal water use of the integrated circuit industry and causing greater economic losses.

[0004] The above method of judging the TOC and B content of the terminal water product and then feeding back to the front end for control has the problems of lack of water quality prediction and delayed control measures. Summary of the invention

[0005] Purpose of the invention: The purpose of the present invention is to provide a terminal ultrapure water quality intelligent prediction device and prediction and control method based on RBF (radial basis function) neural network algorithm model, which is used for prediction and control of TOC and B content in ultrapure water.

[0006] Technical solution: The prediction device described in the present invention is arranged on the bypass in front of the raw water pool of the ultrapure water preparation system, including a processing system, a detection system and a water quality prediction system. The processing system is used to pre-treat the raw water and output pre-treated water that meets the requirements of the detection system; the detection system is used to perform real-time detection on the pre-treated water and obtain real-time data of the pre-treated water parameters; the water quality prediction system uses the RBF neural network algorithm model to predict the total organic carbon TOC and boron B content in the terminal ultrapure water according to the real-time data of the pre-treated water parameters, obtains the prediction result, and adjusts the dosage of the reagent in the ultrapure water preparation system in combination with the historical dosage of the reagent, so that the TOC and B content in the terminal ultrapure water can be stably up to standard. The present invention can predict the TOC and B content in the ultrapure water at the raw water end, and can compare historical data, adjust the dosage of the reagent, and ensure that the TOC and B content in the terminal ultrapure water is stably up to standard.

[0007] Optionally, the treatment system includes an ultrafiltration UF unit, a reverse osmosis RO unit and an ion exchange resin tower connected in sequence.

[0008] Optionally, the real-time data of pre-treated water parameters include resistivity, TOC content, small molecule organic matter content and B content data.

[0009] Optionally, the RBF radial basis neural network algorithm model formula is:

[0010]

[0011] Among them, y j represents the jth result output after hidden layer processing, that is, the TOC content and B content of the terminal water produced according to the water quality of the pretreated water sample, ω ij represents the weight value from the i-th input value to the j-th output value; h is the number of hidden layer nodes; x c is the center value of the basis function, i.e., the cluster center, i represents the number of input values ​​of the neural network algorithm, j represents the number of output values ​​of the neural network algorithm, c i is the real-time water parameter data of the i-th sample in the basis function, d m"x Represents the i-th sample value x input in the basis function i The maximum distance from the cluster center.

[0012] Optionally, the flow rate of raw water introduced into the bypass is 20-50 L / h, the overall residence time in the water quality prediction device is 10-20 min, and the bypass outlet water is supplemented into the filter water pool of the ultrapure water preparation system.

[0013] The present invention also designs a terminal ultrapure water quality intelligent prediction and control method based on the prediction device and the RBF neural network algorithm model, which includes the following steps:

[0014] Pre-treat the raw water to obtain real-time data of pre-treated water parameters, predict the water quality based on the real-time data of pre-treated water parameters, and output the prediction results;

[0015] According to the prediction results and combined with the historical dosage of reagents, the dosage of each reagent in the dosing system is controlled to achieve water quality prediction and control.

[0016] Further, the raw water is pre-processed to obtain real-time data of pre-processed water parameters, the water quality is predicted based on the real-time data of pre-processed water parameters, and the prediction result is output; specifically, it includes:

[0017] (1) Pre-treating raw water to obtain pre-treated water that meets the testing requirements;

[0018] (2) Perform real-time detection on pretreated water to obtain real-time data of pretreated water parameters, including resistivity, TOC content, small molecule organic matter content, and B content data;

[0019] (3) The RBF neural network algorithm model is used to process the real-time data of pretreated water parameters to obtain the predicted values ​​of TOC content and B content in the terminal ultrapure water, that is, the prediction results.

[0020] Furthermore, the raw water is pretreated in step (1) as follows:

[0021] The UF membrane in the UF unit removes colloids and suspended particles in the raw water, the RO membrane in the RO unit removes ions, bacteria and viruses in the raw water, and the ion exchange resin further desalinates and is used for pretreatment of the raw water.

[0022] Furthermore, step (3) is specifically as follows:

[0023] The real-time data of the pre-processed water parameters are processed by a normalization function; the normalization function is:

[0024]

[0025] Among them, X i ' is the data obtained after normalization of the real-time data of pre-processed water parameters, x i is the real-time data of water parameters of the i-th sample in the basis function; x min is the minimum value of the sample data; x max is the maximum value of the sample data;

[0026] Calculate the cluster centers of the normalized preprocessed water parameter real-time data, and use the number of cluster centers as the number of hidden layer nodes in the RBF neural network algorithm model;

[0027] Solve for the variance σ of the basis function;

[0028] According to the cluster center point xC , the maximum distance d between the data and the cluster center m"x And the number of cluster centers, the RBF radial basis neural network algorithm model expression is obtained as follows:

[0029]

[0030] Among them, y j represents the jth result output after hidden layer processing, i.e. the TOC content and B content of the terminal water product predicted based on the water quality of the pretreated water sample, j represents the number of output values ​​of the neural network algorithm, and h is the number of hidden layer nodes;

[0031] make The above formula can be rewritten as:

[0032]

[0033] An electronic device for storing and executing the method, comprising:

[0034] A memory storing executable program code;

[0035] a processor coupled to the memory;

[0036] The processor calls the executable program code stored in the memory to execute the steps of the terminal ultrapure water quality intelligent prediction and control method based on the RBF neural network algorithm model.

[0037] Beneficial effects: Compared with the prior art, the significant technical effects of the present invention are: (1) the TOC and B content control process is improved from problem discovery-solving to risk prediction-risk avoidance, saving a lot of time for regulating TOC and B content and water quality recovery time; (2) a method for intelligent risk prediction based on RBF neural network algorithm model and big data training simulation is proposed, by connecting a water quality prediction device in front of the raw water tank, equipped with a processing system and a variety of water quality detection equipment with a process of only 10 to 20 minutes, and inputting real-time data through a large number of Sample data training, output results converged RBF neural network algorithm model, with the help of the prediction results of the RBF neural network algorithm model, ensure that when the raw water quality fluctuates, the changes in TOC and B content in the terminal ultrapure water can be accurately predicted, and fed back to the TOC and B removal units, and the dosage of the reagents is adjusted in real time to ensure the stability of the produced water quality; (3) In the early stage, a large amount of data was combined to train the RBF neural network algorithm model, which is suitable for ultrapure water preparation systems that provide training sample data. At the same time, it is also of reference significance for ultrapure water preparation systems with similar process flows. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1This is a schematic diagram of the traditional mode terminal ultrapure water quality control process;

[0039] Figure 2 This is a schematic diagram of the terminal ultrapure water quality control process in the intelligent prediction mode;

[0040] Figure 3 Model structure for RBF prediction of terminal ultrapure water quality DETAILED DESCRIPTION

[0041] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0042] Figure 1 This is a schematic diagram of a traditional semiconductor ultrapure water preparation system. In this process, an advanced oxidant can be added to the raw water tank to ensure that the TOC value in the produced water is stable; a pH adjuster can be added to the RO unit to ensure that the B value in the produced water is stable. In the traditional mode, there is a lag in the control process of TOC and B content in the terminal ultrapure water. It is a process from problem discovery to problem solving. There is a very long hydraulic retention time from the raw water tank to the ultrapure water use point, usually greater than 20 hours, and from the RO water production tank to the water use point is usually greater than 10 hours. When the signal of abnormal terminal TOC and B data is fed back to the front-end dosing system, the water quality in the filter water tank and the RO water production tank is no longer up to standard. From the start of the dosing system to the qualified TOC period of the produced water, there may be water quality problems in the produced water.

[0043] like Figure 2 As shown, the ultrapure water quality prediction device of the present invention is arranged on a bypass in front of the raw water pool of the traditional semiconductor ultrapure water preparation system, and a stream of raw water is introduced into the water quality prediction device. The water quality prediction device includes a treatment system, a detection system and a water quality prediction system. The treatment system includes a UF (ultrafiltration) unit, an RO unit and an ion exchange resin tower connected in sequence. The UF membrane in the UF unit removes colloids and suspended particles in the raw water, and the RO reverse osmosis membrane in the RO unit removes ions, bacteria and viruses from the raw water. The ion exchange resin further desalinates and is used to pre-treat the raw water, improve the quality of the raw water, and prevent the precision unit in the detection system from being contaminated and blocked, so that the outlet water meets the water inlet requirements of the detection system; the detection system is used to perform real-time detection of the pre-treated water, and obtain Get real-time data of pretreated water; the detection system imports the acquired real-time data of pretreated water into the water quality prediction system. The water quality prediction system predicts the TOC and B content in the terminal ultrapure water according to the real-time data of pretreated water, and outputs accurate prediction results. The water quality prediction system combines the historical dosage of reagents to feed back the dosage information of pH regulator and bromide salt / oxidant to the dosing system, and adjusts the dosage of the TOC removal unit (i.e. the dosing point of the raw water pool) and the B removal unit (i.e. the dosing point of the secondary reverse osmosis unit) in the system to ensure that the TOC and B content in the terminal ultrapure water are stable and meet the standards.

[0044] In this embodiment, before the raw water enters the raw water pool, the raw water flow rate introduced into the bypass is 20-50L / h, and the overall residence time in the water quality prediction device is 10-20min. The bypass effluent (i.e., the effluent of the detection system) will be added to the filter pool. In the treatment system composed of UF, RO and ion exchange resin towers, the raw water residence time is 4-7min, which serves to improve the quality of raw water so that the effluent meets the water inlet requirements of the detection system. The detection system includes a resistivity detection instrument, a TOC detection instrument, a small molecule organic matter detection instrument and a B meter, which can obtain the resistivity, TOC content, small molecule organic matter content and B content data of the raw water after pretreatment. The detection system will detect the real-time data of the four parameters (TOC, small molecule organic matter content, resistivity, and B content) in the pretreated water according to the preset frequency (detected once every 15min in this embodiment). In the water quality prediction system, an intelligent prediction model is used to predict the TOC and B content in the terminal ultrapure water according to the real-time data of the four input parameters, and compared with the actual terminal ultrapure water quality, error convergence is performed, and the intelligent prediction model is iteratively trained until the prediction model converges to the limit (error less than 0.05), outputs accurate prediction results, and feeds back to the dosing system; in the ultrapure water preparation process, the raw water tank and the secondary RO unit will adjust the dosage of the dosing according to the requirements of the qualified ultrapure water quality to ensure the stability of the water quality of the filtration tank, RO water production tank and ultrapure water use point. Compared with the traditional model, after adopting the intelligent prediction method, the TOC and B treatment feedback time of the ultrapure water use point can be shortened to less than 20 minutes.

[0045] The four parameters output by the detection system fluctuate irregularly, are highly nonlinear, and have multiple dimensions. The intelligent prediction model in the water quality prediction system is an RBF neural network algorithm model, which is a feedforward neural network with good performance such as unique best approximation (overcoming the local minimum problem), simple training, and fast learning convergence. It is suitable for rapid simulation of large amounts of data and fitting of multivariate data. Figure 3 As shown in FIG. 1 , the RBF neural network algorithm model includes an input layer, a hidden layer and an output layer. The input data are TOC content, small molecule organic matter content, B content and resistivity. After being processed by the hidden layer, TOC and B are output.

[0046] Furthermore, the formula of the initial RBF radial basis neural network algorithm model is as follows:

[0047]

[0048] Among them, ω ij represents the weight value from the i-th hidden layer node to the j-th output value; σ represents the variance of the basis function (also known as the basis function amplitude), where the basis function is φ(x i ,x c )=||x i -xc ||; h is the number of hidden layer nodes; x c is the center value of the basis function (cluster center), i represents the number of input values ​​of the neural network algorithm, j represents the number of output values ​​of the neural network algorithm, and x i Represents the input value of the i-th sample in the basis function, y j Represents the jth result output after hidden layer processing, that is, the TOC and B content of the terminal water product predicted based on the water quality of the pretreated water sample.

[0049] Before obtaining accurate prediction results, the RBF neural network algorithm model needs to be trained with a large amount of sample data. The prediction results are obtained by inputting the four types of parameters obtained by the detection system into the RBF neural network algorithm model. According to the difference between the prediction results and the water quality information results of the actual ultrapure water use point, the neural network is self-iterated. When the water quality error value is set to be less than 0.05, the iteration ends and the prediction model is considered to have converged to the limit. At this time, the prediction results can be considered to have high accuracy, and the expected water quality of the terminal ultrapure water can be fed back based on the prediction results of the water quality prediction system.

[0050] The water quality prediction system feeds back the signal to the back-end dosing program based on the predicted values ​​of TOC and B content in pure water at the ultrapure water use point and the historical reagent dosage, and adjusts the reagent dosage of the TOC and B removal units to ensure that the TOC and B content in the terminal ultrapure water is stable and meets the standards.

[0051] The water quality prediction system stores the dosage of the bromide / oxidant dosing unit in the raw water pool for TOC removal in the historical ultrapure water preparation system and the pH regulator dosing unit in the secondary reverse osmosis unit for boron removal. The prediction results will be compared with the dosage of the reagents corresponding to the TOC and B content in the historical ultrapure water, and the dosage will be adjusted at any time in the raw water pool and the primary RO unit.

[0052] The method for intelligently predicting and controlling the water quality of terminal ultrapure water by using the ultrapure water quality prediction device comprises the following steps:

[0053] S1. Pre-treat the raw water through the ultrapure water quality prediction device, obtain the real-time data of the pre-treated water, predict the water quality according to the real-time data of the pre-treated water, and output the prediction result; specifically:

[0054] (1) Pre-treat the raw water through a treatment system consisting of UF, RO, and ion exchange resin to improve the quality of the raw water and obtain pre-treated water that meets the water inlet requirements of the detection system.

[0055] (2) The pre-treated water is tested in real time by the detection system to obtain real-time data of the pre-treated water parameters.

[0056] First, it is necessary to clarify which inlet water parameters will affect the TOC content and B content in the produced water. Usually, the inlet water parameters include: flow rate, conductivity, pH, small molecule organic matter content (urea, etc.), TOC, silicon, B and other information. Through the screening of a large amount of sample data, in this embodiment, the four inlet water parameters of TOC content, small molecule organic matter content (LW), B content, and resistivity have a great influence on the TOC and B content of the produced water. After the raw water passes through the treatment system in the water quality prediction device, the effluent obtains the above four parameters through the detection system as key influencing factors, and then as input factors of the water quality prediction system.

[0057] In this embodiment, a resistivity detection meter is used to detect the resistivity of the pretreated water in real time, a TOC detection meter is used to detect the TOC content of the pretreated water in real time, a small molecule organic matter detection meter is used to detect the small molecule organic matter content in the pretreated water in real time, and a B meter is used to detect the B content of the pretreated water in real time to obtain the resistivity, TOC content, small molecule organic matter content and B content data of the raw water after pretreatment.

[0058] (3) The real-time data of pre-treated water parameters are processed by the water quality prediction system to obtain the predicted values ​​of TOC and B content in the produced water at the point where the produced water, i.e., the ultrapure water, is used, i.e., the prediction results. Specifically:

[0059] First, the four key influencing factors (i.e., real-time data of pre-processed water parameters) obtained by the detection system are normalized using the following function:

[0060]

[0061] Among them, X i ' is the data obtained after normalization of the real-time data of pre-processed water parameters, x i is the input value of the i-th sample in the basis function, i.e., the real-time data of water parameters; x min is the minimum value of the sample data; x m"x is the maximum value of the sample data. Normalization limits the data to [-1,1], making it easier for the algorithm to converge to the optimal solution.

[0062] Furthermore, to clarify the number of basis function center points in the RBF radial basis neural network algorithm model, the K-means mean clustering method or subtraction clustering algorithm can be used to solve and find the cluster center x c , and the obtained number of cluster centers is used as the number of RBF hidden layer nodes h.

[0063] Furthermore, when the hidden layer activation basis function is selected as a Gaussian function, the cluster center x c After obtaining the number of cluster centers h, the following formula can be used to solve the variance σ of the basis function.

[0064]

[0065] Among them, d max Represents the i-th sample value x input in the basis function i The maximum distance to the cluster center, calculated using the Euclidean distance function.

[0066] Further, we get the cluster center point, d max After substituting the number of cluster centers into formula (1), we can get the following expression:

[0067]

[0068] in, is the activation function expressed by a Gaussian function, Expression. Command Then formula (4) can be rewritten as:

[0069]

[0070] Furthermore, the pseudo-inverse method is used to solve the function weight value ω ij . Let J = J(x i ) is the expected output value of TOC and B content at the function fitting terminal, assuming that J ij is the i-th input vector x i At the jth output node y j The expected output value, ω kj is the weight from the kth hidden layer node to the jth output node, i∈(1,n), k∈(1,n), k∈(1,h). Then the weight matrix {ω kj} can be solved using the following formula:

[0071] {ω kj =G + J (6)

[0072]

[0073] Among them G @ Denote G = {g iA The pseudo-inverse matrix of} can be solved by singular value decomposition (SVD). iA Represents the i-th input vector x i The output value of the kth hidden layer node has a total of k input vectors.

[0074] Assume that G is a matrix of M rows × N columns, and perform singular value decomposition on G, then [U, S, V] = SVD (G), and the matrices U, S, and V can be obtained respectively.

[0075] Where U is an M×M matrix, the column vectors in U are left singular vectors, V is an N×N matrix, and the column vectors in the V matrix are called right singular vectors; S is an M×N diagonal matrix; the main diagonal elements of the S matrix are the singular values ​​of the matrix G, S1=1. / S, which is an N×M matrix determined by the singular values. The above matrices have the following relationship:

[0076] G=U×S×V T (8)

[0077] G + =V×S1×U T (9)

[0078] Furthermore, based on the RBF function weight ω solved above, it is substituted into equation (5) to obtain the expected output value y corresponding to the input parameter.

[0079] S2. According to the prediction results and combined with the historical dosage of reagents, the dosage of each reagent in the dosing system is controlled to achieve water quality prediction and control;

[0080] In this embodiment, based on the prediction results and the historical reagent dosage, the signal is fed back to the back-end dosing program to adjust the reagent dosage in the bromide / oxidant dosing unit of the raw water pool for removing TOC and the reagent dosage in the secondary reverse osmosis unit dosing unit for removing boron, so as to ensure that the TOC and B contents in the terminal ultrapure water are stable and meet the standards.

[0081] The following is a further explanation using an electronics company as an example. The TOC content of the ultrapure water of this electronics company needs to be less than 1μg / L, and the boron content of the effluent water needs to be less than 50ng / L. Its inlet water source is a mixture of tap water and recycled water. Compared with municipal tap water, the TOC and small molecule organic matter content in recycled water fluctuates more significantly. It is necessary to always pay attention to the impact load on the ultrapure water preparation system that may be caused by water quality fluctuations.

[0082] In this implementation example, by inputting 4 dimensions (TOC content, small molecule organic matter content, B content, resistivity) parameters, a total of 200 input quantities, when the number of iterations is 105 times, the RBF prediction model error value is less than 0.05, the model converges to the limit, the prediction result obtains the optimal solution, and the error with the measured value is close to the limit. The model structure is as follows Figure 3 shown.

[0083] When the RBF prediction model converges to the limit and the prediction result obtains the optimal solution, the TOC and B content of the effluent can be predicted through the water quality data of the prediction device. When the raw water enters the prediction device, the detection system result is displayed as x (TOC content, small molecule organic matter content, B content, resistivity) = (0.5, 0.9, 0.8, 0.3). The BRF model outputs the prediction result y1 value of 1.2, which means that the terminal ultrapure water TOC content is predicted to be 1.2μg / L, and the y2 value is 0.9, which means that the terminal B content is predicted to be 45ng / L, that is, TOC has a risk of exceeding the standard, and the B content index does not have a risk of exceeding the standard.

[0084] The prediction signal will be transmitted to the advanced oxidation system of the raw water pool, and based on the comparison of historical data, the dosage of sodium bromide and oxidant added to the raw water pool will be increased to remove small molecules effectively. If the B content index exceeds the standard, the pH regulator can be added in the secondary RO unit stage to increase the pH value and thus improve the B removal rate.

[0085] After 15 minutes, the raw water passes through the pretreatment system and enters the detection system again to obtain the latest detection data. The detection data is then imported into the intelligent prediction model, and the expected y value is obtained according to the RBF model. The dosage of the reagent for the TOC and B content removal process is adjusted by comparing the historical data. Repeat this step to ensure that the water in the raw water pool can still meet the terminal ultrapure water TOC and B content requirements after subsequent process treatment.

[0086] The present invention proposes a smart prediction method for the organic matter and B content in the terminal ultrapure water. The first of the methods requires a water quality prediction device to be connected in front of the raw water tank. The device has a processing system, a detection system and a water quality prediction system. By detecting the pre-treated water quality, the smart prediction model in the water quality prediction system can predict the TOC and B content of the terminal water production, which has the advantages of being timely and accurate. The second of the method is that an RBF neural network algorithm model is constructed in the prediction device. By inputting the water quality of the prediction device into the RBF neural network algorithm model, comparing the water quality of the prediction device with the actual terminal ultrapure water quality, the neural network can be self-iterated. After a large amount of sample data simulation training, when the difference between the prediction result and the actual result is less than 0.05, it is considered that the function converges to the limit. At this time, the prediction result can accurately represent the terminal ultrapure water quality. The method can predict the TOC and B content in the ultrapure water production at the raw water end, and can compare historical data, regulate the dosage of the reagent in the ultrapure water preparation system, and ensure that the TOC and B content in the terminal ultrapure water are stable and up to standard.

Claims

1. A terminal ultrapure water quality intelligent prediction device based on RBF neural network algorithm model, characterized in that: The device is arranged on the bypass in front of the raw water pool of the ultrapure water preparation system, and includes a processing system, a detection system and a water quality prediction system. The processing system is used to pre-treat the raw water and output pre-treated water that meets the requirements of the detection system; The detection system is used to perform real-time detection on pretreated water and obtain real-time data on pretreated water parameters. The water quality prediction system uses the RBF neural network algorithm model to predict the total organic carbon TOC and boron B content in the terminal ultrapure water based on the real-time data of pretreated water parameters, obtains the prediction results, and adjusts the dosage of reagents in the ultrapure water preparation system in combination with the historical dosage of reagents, so that the TOC and B contents in the terminal ultrapure water are stable and meet the standards.

2. The terminal ultrapure water quality intelligent prediction device based on the RBF neural network algorithm model according to claim 1 is characterized in that: The treatment system includes an ultrafiltration UF unit, a reverse osmosis RO unit and an ion exchange resin tower which are connected in sequence.

3. The terminal ultrapure water quality intelligent prediction device based on the RBF neural network algorithm model according to claim 1 is characterized in that: The real-time data of pre-treated water parameters include resistivity, TOC content, small molecule organic matter content and B content data.

4. The terminal ultrapure water quality intelligent prediction device based on the RBF neural network algorithm model according to claim 1 is characterized in that: The RBF radial basis neural network algorithm model formula is: Among them, y j represents the jth result output after hidden layer processing, that is, the TOC content and B content of the terminal water produced according to the water quality of the pretreated water sample, ω ij represents the weight value from the i-th input value to the j-th output value; h is the number of hidden layer nodes; x c is the center value of the basis function, i.e., the cluster center, i represents the number of input values ​​of the neural network algorithm, j represents the number of output values ​​of the neural network algorithm, x i is the real-time water parameter data of the i-th sample in the basis function, d m"x Represents the i-th sample value x input in the basis function i The maximum distance from the cluster center.

5. The terminal ultrapure water quality intelligent prediction device based on the RBF neural network algorithm model according to claim 1 is characterized in that: The flow rate of raw water introduced into the bypass is 20-50L / h, and the overall residence time in the water quality prediction device is 10-20min. The bypass water is replenished into the filter pool of the ultrapure water preparation system.

6. A method for intelligent prediction and control of terminal ultrapure water quality based on RBF neural network algorithm model, characterized in that: The following steps are involved: Pre-treat the raw water to obtain real-time data of pre-treated water parameters, predict the water quality based on the real-time data of pre-treated water parameters, and output the prediction results; According to the prediction results and combined with the historical dosage of reagents, the dosage of each reagent in the dosing system is controlled to achieve water quality prediction and control.

7. The method for intelligent prediction and control of terminal ultrapure water quality based on RBF neural network algorithm model according to claim 6 is characterized in that: Pre-treat the raw water to obtain real-time data of pre-treated water parameters, predict the water quality based on the real-time data of pre-treated water parameters, and output the prediction results; specifically including: (1) Pre-treating raw water to obtain pre-treated water that meets the testing requirements; (2) Perform real-time detection on pretreated water to obtain real-time data of pretreated water parameters, including resistivity, TOC content, small molecule organic matter content, and B content data; (3) The RBF neural network algorithm model is used to process the real-time data of pretreated water parameters to obtain the predicted values ​​of TOC content and B content in the terminal ultrapure water, that is, the prediction results.

8. The method for intelligent prediction and control of terminal ultrapure water quality based on RBF neural network algorithm model according to claim 7 is characterized in that: The pretreatment of raw water in step (1) is specifically as follows: The UF membrane in the UF unit removes colloids and suspended particles in the raw water, the RO membrane in the RO unit removes ions, bacteria and viruses in the raw water, and the ion exchange resin further desalinates and is used for pretreatment of the raw water.

9. The method for intelligent prediction and control of terminal ultrapure water quality based on RBF neural network algorithm model according to claim 7 is characterized in that: Step (3) is specifically as follows: The real-time data of the pre-processed water parameters are processed by a normalization function; the normalization function is: Among them, X i ' is the data obtained after normalization of the real-time data of pre-processed water parameters, x i is the real-time data of water parameters of the i-th sample in the basis function; x min is the minimum value of the sample data; x max is the maximum value of the sample data; Calculate the cluster centers of the normalized preprocessed water parameter real-time data, and use the number of cluster centers as the number of hidden layer nodes in the RBF neural network algorithm model; Solve for the variance σ of the basis function; According to the cluster center point x C , the maximum distance d between the data and the cluster center m"x And the number of cluster centers, the RBF radial basis neural network algorithm model expression is obtained as follows: Among them, y j represents the jth result output after hidden layer processing, i.e. the TOC content and B content of the terminal water product predicted based on the water quality of the pretreated water sample, j represents the number of output values ​​of the neural network algorithm, and h is the number of hidden layer nodes; make The above formula can be rewritten as:

10. An electronic device, characterized in that: The device comprises: A memory storing executable program code; a processor coupled to the memory; The processor calls the executable program code stored in the memory to execute the steps of the terminal ultrapure water quality intelligent prediction and control method based on the RBF neural network algorithm model as described in any one of claims 6 to 9.

Images