JPH0484799A - Water quality control method for plant - Google Patents

Abstract

PURPOSE:To contrive water quality control by allowing the neural circuit model of a hierarchic structure to learn the typical pattern of an input parameter value of various time points as a teaching signal to input a water quality pattern which is not learned by the neural model so as to operate. CONSTITUTION:An input value is input into an input layer and this signal is output into the neurone element model of an intermediate layer 2. The multiplication and addition of these values and weighting factor are calculated by the use of specified equations on the neurone element model of the intermediate layer 2, and an output value for an output layer 3 is determined according to this compensation. Also the output value of the intermediate layer successively calculates the multiplication and addition of a next intermediate layer and a next output layer to obtain the calculated value of the output layer 3. In order to execute the learning of a neural network, after the output layer 3, a comparator layer 4 and a teacher signal layer 5 are provided further, the signal of the output layer 3 and a teaching signal are input into the comparator layer 4 and the output signal is compared with the teaching signal. The size of the weighting factor is modified so as to lessen the errors. With the use of the modified value, recalculation and the teaching signal are compared with each other to successively repeat modification so as to lessen the errors.

Description

[Detailed Description of the Invention] [Industrial Application Field] The present invention provides optimal water quality control by processing information obtained by sensor systems installed in each part of a plant using New Nil Network '7. Regarding how to.

[Conventional technology]

An example of the application of New Nil Network to conventional analysis techniques is shown in 4. (1988), pp. 280-285, p. 284.

[Problem to be solved by the invention]

The above-mentioned conventional technology uses a neural network itself to obtain the weighting function itself necessary for quantitative analysis. The present invention applies this neural network to plant water quality diagnostic control. That is, the purpose is to provide optimal water quality control means for the input information by inputting water quality information obtained by a sensor system into a neural network, thereby controlling the water quality of a plant.

[Means to solve the problem]

In order to achieve the above object, the present invention uses a neural circuit model with a hierarchical structure in which at least one intermediate layer is provided between the input layer of water quality information from the sensor and the output layer. A neural circuit model is caused to learn by using a representative pattern of a plurality of input variable values at a time as an input signal and a control variable value corresponding to the representative pattern as a teacher signal. By inputting an unlearned water quality pattern as an input variable value to this learned neural circuit model, variable values for the target water quality control were obtained and the control system was operated.

[Effect]

The neural network uses past water quality information obtained by the sensor system or laboratory data to determine what kind of control signal should be generated, and generates output signals for control system operation in response to several input signal patterns. It can be learned in advance as Therefore, by inputting an unlearned water quality information pattern, the optimal output signal corresponding to that input pattern will be output according to the learning level. Since it is possible to obtain an output control signal that enables optimal water quality control operation even for a water quality pattern input signal of a seemingly low level, highly reliable water quality diagnostic control of the plant can be performed.

〔Example〕

<Example 1> In this example, dissolved oxygen (D
o), dissolved water quality (DH), dissolved hydrogen peroxide (DHzOz
) The reactor shown in Figure 1 shows an example of learning the monitoring pattern of structural material corrosion potential (ECP) using a new network, estimating the cause of the data pattern, and activating the water quality control system. In the surrounding block diagram, Do, DH, DHzOz, EC are in the instrumentation pipe 4.
P Various electrochemical sensors 1 are set. Details of the electrochemical sensor are shown in Figure 3. Information from this electrochemical sensor is sent through the potentiostat (5), data processing (6), and electrochemical measurement controller (8) to the neural net (8), and depending on the information pattern processed by the neural network, gas and chemicals (9) are sent. The injection system receives commands to control and manage water quality. Information such as the state of the water quality that gives the pattern of the obtained water quality information, i.e., whether it is in a normal state or whether the cause of the fluctuation in the pattern is a seawater leak, is sent to the remote control command device and the analysis result display and output device in 7 at any time. Is displayed. 10 is the reactor water supply pipe, 1
1 is a reactor recirculation system, 12 is a reactor pressure vessel, 13 is a turbine, 2 is a dryer, and 3 is a reactor core.

Figure 2 shows the hierarchical structure of the neural network. In FIG. 4, among the water quality information obtained by the electrochemical sensor 1 of FIG. 1, the concentrations of three chemical species, D O, D H, and D H2O2, are shown on the display system 7.

As a result of processing using the neural network shown in Figure 2 from the pre-learned input pressure cover pattern, it was thought that the cause of the increase in DO and DH (measurement times) was an increase in reactor output, and countermeasures were taken. When hydrogen gas was injected from the gas injection system No. 9, the DO was successfully returned to the level at the start of the measurement display, and the sensor information pattern output display through the neural network became a display of normal water quality.

In the enlarged view of the electrochemical sensor in Figure 3, 14 is a platinum microwire working electrode with a diameter of 50 μm, and 15 is a reference electrode, or EC.
P meter, 16 is pH sensor, 17 is platinum counter electrode, 18 is hole for liquid junction at a level that allows reactor water to pass through, 19 is tungsten metallized alumina, 20 is sensor case, 5US316
Made of L steel, 21 is MI cable (Mineral In5
22 is an MI cable integration pipe, and 23 is a welded part.

Next, the analysis method for obtaining the electrochemical sensor output results shown in FIG. 4 will be described in detail. Neural networks are also applied to this analysis method.

Using the electrochemical sensor shown in FIG. 3, installing this sensor in the instrumentation tube 4 shown in FIG. 1, and applying random pulse portammetry and the neural network shown in FIG. 2, dissolved oxygen can be detected.

Hydrogen peroxide, hydrogen concentration (C0 (021v C0 (H2O
21C01H2)) was simultaneously quantitatively analyzed.

Random pulse portammetry is a portammetry that introduces a batan recognition method. By shaking the quantitative analysis system in various ways using a complex disturbance signal (random pulse potential signal), a large amount of electrolytic current data (bang information) including information on the electrochemical reaction of each system to be measured can be obtained. I tried. In other words, even when the reaction potentials of the above chemical species are close to each other and separation and quantification are difficult, it is possible to analyze the electrochemical reaction process of each analyte based on the differences in the rate of electron transfer reaction, parameters, etc. , we focused on the characteristic patterns (feature vectors) in the data matrix obtained by the data compression process shown below, and conducted quantitative analysis based on the size of the feature vectors and the calibration curve. Perform qualitative analysis.

Random pulse portammetry consists of: 1. Obtaining baton information (electrolytic current data), 2. Compressing baton information, and 2. Purpose information (C' + oz+ ) from the compressed baton information.

This is executed in the following steps: extraction of COu+zoz++C'u+zt value). The outline of each step of ■, ■, and ■ is as follows. That is, in (2), first, the electrolytic currents of oxygen, hydrogen peroxide, and hydrogen are measured as response signals to the random pulse potential input to the measurement system. In the case of this embodiment, the random pulse potential signal is described using two potential waveform parameters (m, n). That is, the potential corresponding to the pulse potential of the first random pulse is represented by a parameter m, and the potential corresponding to the pulse potential of the second random pulse is represented by a parameter n. Therefore, 5m is set on the X axis, n is set on the y axis, and the electrolytic current, which is a response signal corresponding to the input potential on each (xt y) coordinate, is arranged in the form of a matrix according to its parameter value. ■
Now, the data matrix of the current component, which is the response signal of the input potential signal of 64 (8x8) points obtained in step (2), is converted and compressed into a data vector having several components. This data compression method includes Fourier transform, Hadamard transform, Karhunen-Löve transform, Barr transform, etc. which are used in the field of image processing.In this example, among these, the partial Hadamard transform was used. . In ■, extract the feature vector from the data vector obtained in ■, and
oz+*C0+uzoz+, C0utz+ are determined. That is, input each component value of the obtained data pertle, and calculate C' (021, C' (8202) +
Find and set a function whose output is each value of C'+o2+. In this example, this function was determined using a neural network. Below, ■, ■ in this example
.

③ Describe the details of each step.

First, details of ``obtaining slam information'' will be described.

FIG. 15 is an explanatory diagram of a method of applying a random pulse potential signal. Although it is called a random pulse potential signal, if it is applied completely arbitrarily, C0+oz++C0tHxox+,
In this example, portammetry using both a normal pulse and a reverse pulse detects C0tOZl+C.
Noting that Ou+zoz+ and C'+Hz+ can be determined separately, we changed the random pulse potential signal using this pulse application method as the basic type. The upper part of FIG. 5 is a schematic diagram of a current (i)-potential (E) curve in a ternary mixed system of oxygen, hydrogen peroxide, and hydrogen. In the positive potential region, the oxidation currents of hydrogen peroxide and hydrogen are Ii!, and in the negative potential region, the reduction currents of oxygen and hydrogen peroxide are Ii! be measured. The potential regions where oxidation current and reduction current are observed are respectively E^EC, and 2' (Q=1.2.3...
・Divide into ) levels. The reason why the number 21 is adopted is to facilitate processing by a computer. In this embodiment, each level is divided into eight levels (CQ=3), and each level is divided into eight levels: E^(n), EC(n) (n=1.
2.3...8). Furthermore, the potential at which the electrode reaction of oxygen, hydrogen peroxide, and hydrogen does not occur is Ec, and the potential at which the diffusion limit current for the oxidation reaction of hydrogen peroxide and hydrogen is obtained is El.

B1-B3 at the bottom of FIG. 5 represent three types of random pulse voltage signals applied to the working electrode of the electrochemical sensor.

The first type of random pulse is Ec-+EC(m)→EC
(n) are applied in this order. In this example, Ec is 1 second,
Eo(m) and Eo(n) are each 0.05 seconds, and E
The current value was sampled at the end of C(n). m, n
each represents a divided level (1 to 8) of EC. There are 8×8=64 combination patterns of the potentials of Eo(m) and EC(n), and current values corresponding to the pairs of m and n are stored in the 8×8 data matrix XCNP. Since the XCNP obtained in this way is based on the current data of the reduction reaction of oxygen and hydrogen peroxide in the test water, CO+o
z+.

It has almost the same information as the information regarding CON+zo2+.

Second kind of random pulse Ec-)E^(m)→E^(n
) are applied in this order. In this example, EC is set to 1 second E^(
m) and E^(n) are each 0.05 seconds, and E^(
Current values were sampled at the end of step n).

In this case, as in the case of the first type, an 8×8 data matrix XANP corresponding to the pair m and n is obtained.
Since ANP is determined from the current data of the oxidation reaction of hydrogen peroxide and hydrogen in the test water, C0 (ozr.

Contains information regarding c0utz+.

The third type of random pulse is Ec→E1→EC(m)→
In this example, Ec,
EC(m) + EC(n) is the same time as in the case of the first type random pulse, El is 0.1 second,
Current values were sampled at the end of EC(n). In this case, as in the case of the first type and the second type, an 8×8 data matrix X CRP corresponding to the set of m and n is obtained. The third type of random pulse differs from the first type in that EC
(m) and EC(n) are sequentially applied with El. In other words, X CRP includes oxygen in the test water,
In addition to information on the reduction reaction of hydrogen peroxide, information on the reduction reaction of oxygen generated by oxidation of hydrogen peroxide with El is included.

Next, the details of ``compression of button information'' will be described.

Before transforming and compressing into a data matrix vector, the matrix XCRP NP is defined by equation (1).

Xo(RP-NPI =X0Rp-X"pip
-(1) Through this operation, ``oxygen in the test water,''
Since the information about the reduction reaction of hydrogen peroxide is almost canceled out, the information about the reduction reaction of oxygen produced by the oxidation of hydrogen peroxide is concentrated in XC(RP-11PI).

Data Matrix XCRP NP t XCNP,
XANP is determined by matrix Y according to equation (2), respectively.
'(RP-NPI, Y'NPI Converted to YANP.

Y=H−X−H...(2) However,...(3). For each component of the matrix Y, information possessed by X appears in a specific component among the 64 components of Y. In other words, the information that was dispersed over 64 components of X has been compressed into several specific components of Y. In this example, y11+ y11a+ 3'
Information was compressed into each component of fit+y17+y71. Here Yo(RP-NPI, YONP, YA
NP's y11+ y151'/61+ y17+
ZC (RP-NPI, ZC
Create a vector NPI ZANP.

Z” (y11+ yi13+ y51t y
17t 3'71) ''(5) According to the above explanation, vector Z C+pp-NPI.

2 ONP I Z ANP has C' (Hz
oz+ + C0+oz+ten C0 (H2O21, C'o
This means that the information regarding +zo2+ +C01H21 is compressed with high density.

Next, details of ``extraction of target information from compressed button information'' will be described.

Any C01H21C0tHzz+, C0(82
1, the vector ZC (RP-NP
For I, C' n+zoz+ = f (Z 0ut
p-NPI ) - Prepare a scalar field f that satisfies (6). The scalar field f is defined in advance as C0(H2O
2) is known and C0toz+ + C0(H2) is determined for several arbitrary types of water to be tested.

In this example, this scalar field f was determined using a neural network that inputs each component of the vector ZC+RP-NPI and outputs CO+ozoz ).

C0(Hz+) Even for unknown test water, C0(Hz+) can be found by calculating equation (6).Similarly, any C'+oz+, C0tHz+, C
For Z ANP obtained in the test water of 01H21,
C0+oz+ = g (Z'NP) -C0(Hzo
z+ −(7)C0(Hz+ = h (ZANP
) -C' +82021 - Prepare scalar fields g and h that satisfy (8). The scalar field g+h is previously written as C
0ntzoz+ and C0 (021 is known, or C' +
Hz+ and CO+oz+ are determined in advance for the known test water. In this example, vector Z ONp
Input each component of C0toz+ + C0<ozoz
A neural network that outputs
This scalar field gvh was determined using each neural network with 202+ outputs. C0
(ov, COH4zoz+ * C'+nz+ Also for unknown test water, C0+oz+ e C0+oz+z+ , C0+ by equations (6), (7), and (8).
oz+ is determined respectively.

Next, we will show the basic calculation method using neural networks. As an example, an example in which the above-mentioned random pulse is applied will be shown.

First, a weighting coefficient W is assigned to each of the set signal values Zl−Zn.
An operation that multiplies IJ and then adds these (product-sum operation)
is calculated.

CJ (2) = ΣW J + (2←1)・zt(i)
...(1) n=1 Here, Zi(1): value of input layer (-th layer), Wo(2
←l) From the first variable of the two-manpower layer (layer -) to the middle layer (
Weighting coefficient for the j-th neuron element model of the second layer (second layer), CJ(2): is the total input value to the j-th neuron element model of the intermediate layer (second layer).

In the neuron element model of the input layer, the output value here is calculated according to the compensation of CJ(2) using the following equation.

ZJ(2)=1/(1e-"j(2))...
12) The calculation contents of equation (2) are as shown in FIG. 20. That is, as shown in the graph of the same figure, CJ(2)
The value between 0'' and Pa1'' is ZJ(2) depending on the value of
obtained as. The calculated value ZJ(2) is further sent to the output layer, where similar calculations are performed.

Next, an overview of the calculation method using a neural network will be explained.

The input value Zl(1) described above is input to the input layer of FIG. 20, and this signal (value) is output to the neuron element model of the intermediate layer. In the neuron element model of the intermediate layer, the sum of products Cj (2) of these output values ztD) and the weighting coefficient w+, (2←1) is calculated using equation (1), and the output to the output layer is adjusted according to this compensation. The value Zl(2) is determined using equation (2). Similarly, the output value z, (2) of the intermediate layer is further calculated by the weighting coefficient w lJ (s
The sum of products CJ(3) with ←2) is calculated using the following formula.

Here, Zl(2): value of the middle layer (second layer), WJI(
3←2): Weighting coefficient, C, from the i-th variable of the intermediate layer (second layer) to the j-th neuron element model of the output layer (third layer), (3): Output layer (third layer) is the total input value to the j-th neuron element model.

Furthermore, the output value ZJ is sent to the output layer depending on the size of CJ(3).
(3) is calculated using the following formula.

ZJ(3)=1/(le-cj(3))
-(4) In this way, the calculated value ZJ(
3) is obtained.

In order to perform learning in a neural network, a comparison layer and a teacher signal layer are further provided after the output layer as shown in FIG. 20, and the output layer signal and teacher signal are input to the comparison layer, and the The output signal and the teacher signal are compared. In order to reduce this error, the weighting coefficient W(3←2) and W
Correct the size of J I (2←1).

If the calculation of equations (1) to 4) and comparison with the teacher signal are performed again using this corrected value, errors will similarly occur. In order to reduce this error, the weighting coefficient W
The magnitudes of J i (3←2) and “tLi (2←1) are corrected.

In this way, the weighting coefficient WJ i is repeatedly corrected until the error becomes sufficiently small.

Initially, the weighting coefficients are given randomly (generated by random numbers), so
Although the error is large, the output signal gradually approaches the teacher signal value. Here, symbols c and z correspond to density and feature vector, respectively.

The method of correcting errors in this way is called the error backpropagation method.
) and others.

〔Effect of the invention〕

According to the present invention, in a water quality control method that corresponds to the obtained water quality information pattern, it is possible to make a neural network learn the optimal output signal pattern corresponding to the past water quality information pattern, and even with complex water quality information patterns. Since it is possible to obtain the output signal pattern necessary for optimal water quality control according to the learning level, it has the effect of obtaining the optimal water quality control method corresponding to the current water quality information pattern based on past water quality information. . Furthermore, by increasing the learning level of the neural network as needed, water quality control can be made even more precise.

[Brief explanation of the drawing]

Fig. 1 is a system diagram showing a plant water quality diagnosis control system adopted in an atomic couplant according to an embodiment of the present invention, Fig. 2 is a block diagram showing the hierarchical structure of a neural network, and Fig. 3
Figure 4 is a cross-sectional view of the electrochemical sensor, Figure 4 is a characteristic diagram showing the concentration of dissolved chemical species and water quality control connections using a neural network, and Figure 5 is the current (i )-potential (E) curve diagram and a graph representing a random pulse signal applied to a working electrode of an electrochemical sensor. 1... Electrochemical sensor, 2... Dryer, 3...
Reactor core, 4... Instrumentation tube, 5... Potentiostat,
6... Electrochemical measurement controller, 7... Remote control command device and analysis result display/output device, 8... Neural network, 9... Gas and chemical injection system, 1o...
Reactor water supply piping, 11...Reactor reli ring system, 12.
...Reactor pressure vessel, 13...Turbine, 14...
Working electrode, 15... Reference electrode or ECP meter, 16...
・pH meter, 17...Counter electrode, 18...Liquid junction hole, 19
...Tungsten metallized alumina, 20...Sensor case. 21... MI cable, 22... M■ cable accumulation pipe, 23... welded part. Takubetsu Map 8 --- 21-U1 Shiki, 7 Toto 3 Diagrams High 50 Hours (Support)

Claims (1)

[Scope of Claims] 1. In a water quality control method for a plant based on various types of water quality information obtained by water quality sensors installed in monitoring target parts of a plant, an input layer for inputting information from various water quality sensors. have,
Using a neural circuit model with a hierarchical structure having at least one intermediate layer between the input layer and the output layer, representative patterns of multiple input variable values at different points in time among past plant water quality information are input signals. Then, the neural circuit model is trained using variable values for water quality control corresponding to the representative pattern as a teacher signal, and the unlearned pattern is input as the input variable value to the learned neural circuit model to achieve the objective. A plant water quality diagnosis control system, characterized in that the variable value for water quality control is determined. 2. In claim 1, water quality control variable values of the controlled objects are determined so that at least one water quality controlled object reaches a target state according to the water quality information of the plurality of plants that changes over time, and the water quality of the plant is controlled. The neural circuit model has an input layer that inputs information from various water quality sensors, and has a hierarchical neural circuit model with at least one intermediate layer between the input layer and the output layer. Among the water quality information, a representative pattern of a plurality of input variable values at different times is used as an input signal, and a variable value for water quality control corresponding to the representative pattern is used as the teacher signal to the neural circuit model. A plant water quality diagnosis and control method that involves learning and inputting an unlearned pattern as the input variable value to the learned neural circuit model to obtain the target water quality control variable value. 3. According to claim 1 or 2, when learning the neural circuit model, the input variable value pattern at a certain point in time and the input variable value pattern at a certain point in time before the certain point in time are simultaneously used as input signals, and After learning about a plurality of input signals using the control variable value at a certain time as the teacher signal, the neural circuit model is given the input variable value pattern at the current time and the input variable value pattern at a certain time before the current time. A plant water quality diagnosis control method that calculates the current control variable values by inputting the following simultaneously. 4. A plant water quality diagnosis control method according to claim 1, 2 or 3, for determining a control variable value of a controlled object such that the controlled object reaches a target state in accordance with the input variable values of the plurality of water quality information that change with time. . 5. In claim 1, 2, 3 or 4, a causal relationship between each of the input variables and each control variable is extracted based on the learned result of the neural circuit model, and based on the causal relationship, the A plant water quality diagnosis control method that supports process operation.