KR0165017B1 - Temperature control method and device using neuro and fuzzy theory of a refrigerator - Google Patents

Abstract

The present invention relates to a temperature control method and a temperature control apparatus of a refrigerator using a neural network and fuzzy inference, and to calculate a fuzzy model according to the measured value of the rate of change of temperature at a predetermined place in the refrigerating chamber according to the stopping angle of the rotary blades Calculate the optimum rotor blade inference position angle to maintain the temperature equilibrium in the refrigerating chamber, take the temperature change value indicated by the temperature sensor in the refrigerating compartment at the input node, and determine the optimum rotor blade equilibrium angle to maintain the temperature equilibrium at the output node. By constructing the neural network to be calculated and using the temperature change value and the inference position angle, and to control the rotary blades according to the equilibrium angle output from the learned neural network. Accordingly, the fuzzy theory accurately infers the temperature in the refrigerating chamber and reduces the error according to the difference of each refrigerator by the neural network, and thus calculates the optimum rotor blade equilibrium angle so that the temperature distribution in the refrigerating chamber is always maintained evenly. .

Description

Temperature control method and temperature control device of refrigerator using neural network and fuzzy inference

1 is a partial front view of a refrigerator.

2 is an enlarged perspective view of a rotary blade.

3 is a block diagram illustrating a temperature control method according to the present invention.

4 is a graph showing a region partition state according to a three-part fuzzy structure.

5 is a graph showing a region partition state according to a 4-split fuzzy structure.

6 is a schematic structural diagram of a neural network according to the present invention.

7 is the final neural network of FIG.

8 is a graph showing a decrease in error according to an increase in the number of learning of a neural network.

9 is a control block diagram of a temperature control apparatus for performing a temperature control method according to the present invention.

10 is a temperature drop rate table according to the stop position of the rotary blade.

FIG. 11 is a table calculating weights applied to each node of a neural network. Table 2 is a weight value applied to an intermediate layer, Table 3 is a weight value applied to an output layer, and Table 4 is a weight value of a bias applied to an intermediate layer. , Table 5 shows the weight values of the bias applied to the output layer.

The present invention relates to a temperature control method and a temperature control apparatus of a refrigerator using a neural network and fuzzy inference, and to infer the temperature distribution in the refrigerator using a fuzzy inference algorithm to increase the accuracy of inference and to add a learning function by the neural network. By minimizing the error caused by the difference between each refrigerator, it is possible to precise temperature control.

In refrigerators, particularly large refrigerators, as loads of food and the like accommodated in the refrigerator compartment are different according to each part in the refrigerator compartment, it is difficult to maintain the temperature distribution in the refrigerator compartment evenly, and many methods have been studied to maintain the internal temperature evenly. As the temperature control method in the refrigerating chamber, a method of controlling the temperature by introducing a fuzzy theory to infer the temperature in the refrigerating chamber and discharging the cold air to a high temperature part according to the inferred value is used. There is only a reasoning function based on the disadvantage that the accuracy is limited.

In other words, the refrigerator has a slight error due to the temperature sensor or other parts that occur during the mass production of the product, and the difference in size and capacity of each refrigerator, and all the data obtained from several refrigerators that are a certain standard. Uniform application to the refrigerator results in a failure to sufficiently reflect such errors, which limits the accuracy of the temperature control.

Accordingly, an object of the present invention is to apply accurate inference by applying fuzzy theory with inference function and neural network theory with learning function, and to reflect the distribution of each set through learning, so that the temperature distribution in the refrigerating chamber is maintained evenly. It is to provide a temperature control method and a temperature control device for a refrigerator.

According to the present invention, in the temperature control method of a refrigerator for controlling the direction of cold air discharging in the refrigerating chamber according to the angular position of the rotary blade, the object of the present invention is to measure the rate of change of the temperature at a predetermined point in the refrigerating chamber according to the stopping angle of the rotary blade. Calculating a fuzzy model according to the present invention; Calculating optimal rotation wing inference position angles to maintain equilibrium temperature in the refrigerating chamber by performing fuzzy inference according to the fuzzy model; Constructing a neural network which takes the temperature change value indicated by the temperature sensor in the refrigerating compartment at the input node according to the stop angle of the rotary blade, and calculates an optimum rotational blade equilibrium angle at the output node to maintain the equilibrium of the temperature in the refrigerating compartment; ; Learning the neural network using the temperature change value and the inference position angle; It is achieved by the temperature control method of the refrigerator comprising the step of controlling the rotary blades in accordance with the equilibrium angle output from the learned neural network.

The calculating of the speculative position angle may include: performing fuzzy division on the basis of the measurement value for each of the predetermined points; Obtaining each individual discomfort norm for each partition structure of the fuzzy partitioned region and constructing a linear equation corresponding to the conclusion of the fuzzy inference based on the partition structure having the smallest discomfort norm value According to the general principle of, the accuracy of inference can be increased, and at this time, the fuzzy division for the partition having the smallest discomfort norm among the current partitioning structures is repeated until the discomfort norm reaches the partition structure having the smallest value. By further including the step, it is desirable to repeat the purge split until the structure is the most accurate inference.

The calculating of the discomfort norm value may include calculating a value of a parameter representing a fuzzy region of the divided structure; Preferably, the method includes calculating a discomfort standard value based on the parameter value, wherein calculating the parameter value comprises: determining a number of parameters for the fuzzy region forming the division structure; Providing a string subdividing the possible temperature range that the refrigerator compartment can have into a predetermined number of bits; Constructing a plurality of random strings in which the number of bits corresponding to the number of parameters among the bits of the string are different from the remaining bits; Calculating a neutral phase correlation coefficient between the random strings and the measured value; By including the step of taking the information of the random string having the highest correlation coefficient as the parameter value, a method of selecting data representing the current fuzzy region best among hundreds of random strings can be selected.

In this case, replicating an upper group composed of random strings having a high correlation coefficient and culling a lower group composed of random strings having a low correlation coefficient; Crossing the upper group and the middle group not corresponding to the lower group with the upper group; The cross-correlation coefficient is obtained only by the modified upper group including the random string having the high correlation coefficient in the upper group by the cross, and the GA purge principle can be introduced to more accurately select the random string having the highest correlation coefficient. have.

In the configuring of the linear equations, the linear equations are formed by reflecting weights that contribute to maintaining the temperature equilibrium in the refrigerating compartment of each purge region of the divided structure, such that each region of the current divided structure is an equilibrium of the entire refrigerator temperature. Considering the degree of contribution to maintain, the most optimal inference angle position calculation is possible.

Further, a value taken by the input node may include: a current stop angle at which the rotary blade is located; Present temperature values measured by temperature sensors in the refrigerator compartment; By configuring the temperature sensor in the refrigerating chamber to include the temperature value before the predetermined time measured, it is possible to configure an input node that directly reflects the change according to the time course of the measured value to the input side, wherein the learning of the neural network is based on the back propagation method. It can increase the learning effect of neural network.

The learning of the neural network may include obtaining an error by comparing the inference position angle and the equilibrium angle with respect to the same input value; By reflecting the error to the neural network, the inference position angle is used as the learning data of the neural network, and by continuing to perform this process, the error due to the dispersion of each product of the refrigerator neural network It is possible to obtain a continuous improvement effect.

In addition, according to the present invention, in the temperature control apparatus of the refrigerator for adjusting the direction of the cold air discharge in the refrigerating chamber according to the position of the rotary blade, the calculation by the measured value of the temperature change rate of a predetermined place in the refrigerating chamber according to the stop angle of the rotary blade. A fuzzy computation unit configured to perform fuzzy inference according to the estimated fuzzy model, and to calculate an optimal position of the rotor blade inference for maintaining the equilibrium of temperatures in the refrigerating chamber; A neural network calculation unit having a neural network calculating the optimum angle of rotation of the blades at the output node to take the temperature change value indicated by the temperature sensor in the refrigerating chamber according to the stop angle of the rotary blade and to maintain the equilibrium of the temperature in the refrigerating chamber; ; A learning unit learning the neural network by reflecting an error between the temperature change value and the inference position angle to the neural network computing unit; It is provided with a temperature control device of a refrigerator comprising a control unit for controlling the rotary wing according to the balance angle.

Hereinafter, the present invention will be described in detail with reference to the drawings.

1 is a partial front view of a refrigerator. The refrigerating compartment 10 of the refrigerator is usually installed at the bottom of the refrigerator. The refrigerating chamber 10 is partitioned in the vertical direction, and the lowermost part is used as the vegetable chamber 1. Typically, the top cell is ¾H (5), the second one is ½H (6), and the third one is ⅓H (7), except for the topmost compartment (2) (commonly referred to as freshroom, etc.). This name is used in the following description. In the refrigerating compartment 10, two temperature sensors 11 and 22 are installed. On the left side of ¾H (5), the S1 sensor 11 is installed to measure the temperature of the upper left part of the refrigerating compartment, and On the right side of the) S2 sensor 12 is installed to measure the temperature of the lower right part of the refrigerator compartment. There is a cold air discharge portion 15 in the inner rear surface of the refrigerating chamber 10, the cold air from the cold air discharge portion 15 is controlled by the rotary blades.

2 is an enlarged perspective view of the rotor blade. Rotating blade 20 is divided into three stages in the up and down direction consisting of the upper blade 21, the middle blade 22 and the lower blade (23), each wing ¾h (5), ½H (6), ⅓H ( 7) is located at the height of. Each blade 21, 22, 23 is made to rotate integrally with respect to the rotating shaft 25. As shown in FIG. Directions indicated by the respective blades 21, 22, and 23 are different from each other, and the distorted angles between the wings are 60 degrees apart. In the refrigerating chamber 10, a rear side of the rotary blade 20 is provided with a blowing fan (not shown) for blowing cold air transferred from a cold air duct (not shown). The cold air from the evaporator (not shown) of the refrigerator is discharged into the refrigerating compartment 10 by the blowing fan through the cold air duct, and according to the vane stop angle of the rotary vane 20, the respective compartments 5, 6, The discharge direction of the cold air discharged to 7) is adjusted.

The rotary blade 20 serves to disperse the cold air discharge direction of the refrigerating compartment 10 evenly by continuously rotating during the refrigerating operation of the refrigerator, or to discharge cold air in a stopped state in a specific direction.

In the present invention, by inferring the temperature of each part in the refrigerating compartment, by learning based on the data according to the inferred temperature to ensure that the cold air is discharged toward the portion determined to be the highest temperature, so that the inside of the refrigerating compartment always has an even distribution of cold air. The implementation of the method is largely divided into two phases. This step is shown in FIG. 3, which is a block diagram illustrating a temperature control method according to the present invention.

The first step is to infer the stopping angle of the rotor blade 20 to maintain the temperature balance in the refrigerating chamber 10 by fuzzy inference. First, the temperature change at six places (a position indicated by t1 to t2) in the refrigerating chamber is measured according to the angle of the rotor blade 20. This measurement is repeated in multiple refrigerators. The data thus obtained is tabulated and used as the basis for fuzzy inference. At this time, fuzzy inference uses TSK purge (Takagi-Sugeno-river fuzzy), and GA (Genetic Algorithm) purge algorithm is used in the inference process to increase the accuracy of inference. The GA purge is performed by the GA purge calculation unit 31 and the TSK purge is performed by the TSK purge supervisor 33.

The second step is to calculate the optimal equilibrium angle by correcting the error caused by each product by correcting the inference value deduced in the first step by the neural network. First, it constructs a neural network with the temperature variable input value according to the angle of the rotary blade 20 and outputs the angle of the rotary blade to be controlled, and learns the neural network using the measured value and the fuzzy inference value. It is possible to correct the distribution of each set according to each refrigerator by learning, thereby allowing the inferred temperature to be more accurate, so that the cold air discharge direction by the rotary blades is most appropriate. The configuration of the neural network and the calculation thereof are performed by the neural network controller 33, and the optimum equilibrium angle calculated accordingly is inputted to the rotary wing control unit 37 to control the rotary wing 20.

At this time, the comparison of the output value for learning is made by feeding back the difference between the output value of the neural network and the output value of the TSK purge supervisor 33 to the neural network controller 35. The change in temperature in the refrigerating chamber by the rotary blade 20 controlled according to the output value of the neural network is detected by the temperature sensor S1, S2 again and used as the neural network controller 35 input.

Each step will be described below.

First, as a first step described above, the process of calculating the inference position angle for maintaining the temperature equilibrium in the refrigerating chamber by fuzzy inference is permanently established.

The site | part to be deduced in the refrigerating chamber 10 is divided into 6 parts. That is, the left side of ¾H is called t1, the right side is called t2, the left side of ½H is called t3, the right side is t4, the left side of tH is called t5 and the right side is t6. In order to prepare reference data for applying fuzzy inference, a temperature sensor is first installed at the six sites (t1 to t6) to measure the temperature change. That is, after making the inside of the refrigerating chamber (10) at the proper temperature for refrigeration, the angle of the rotary blade (each wing is configured differently, so the angle is selected based on a specific wing, in this case the upper wing (21) based on In addition, the reference plane for measuring the angle can be handled according to the selection, and in the measurement of the present invention, the case toward the leftmost side of the rear wall surface of the refrigerating chamber 10 was selected as 0 °. ) Becomes 0 ° when discharging cold air toward the leftmost side of the refrigerating chamber 10, and becomes 180 ° when discharging toward the rightmost side.) Measure at intervals. At this time, since the direction of cold air discharge by the rotary blades is different for each wing and the structure inside the refrigerating chamber 10 is also different for each position, each part in the refrigerating chamber is different in the rate of falling of the temperature. The rate of temperature drop at each of these sites is recorded and used as data for 0 °. In this way, change the blade's angle by 10 ° to 180 ° and record the temperature drop rate of each part. The data thus constructed are shown in Table 1 of FIG.

Hundreds of data as shown in Table 1 are used to create a virtual temperature distribution from it, and find the optimal angle of rotor blade accordingly.

The inputs and outputs for inferring the angle of the rotor blade 20 (hereinafter, referred to as an inferred position angle) to achieve an optimum internal temperature balance are as follows.

input ; t1: ¾H left temperature t2: ¾H right temperature

t3: ½H left temperature t4: ½H right temperature

t5: ⅓H left temperature t6: ⅓H right temperature

Print ; ang: inference position angle

Hereinafter, a process for calculating the inference position angle will be described by dividing each stage STAGE.

[Stage 1]

Repeat this procedure as shown in Table 1 to construct 500 sets of these tables and use them as data to obtain the TSK purge (Takagi-Sugeno-steel FUZZY) model. First, use the least-square method commonly used for numerical analysis from all the data. The linear equation of the following equation (1) corresponding to the conclusion of TSK purge inference is generated. At this time, the variable reduction method based on the error rate is performed to at least reduce the variable.

The evaluation norms are applied to the equations obtained in this way. U. C (Unbiasedness Criterion) is used in GMDH (Group Method of Data Handling).

Divide the input data into two groups, A and B. At this time, care should be taken not to cause data scattering to be equal to A and B, for example, a large amount of data having a small value of t1 in A and a large amount of large data in reverse in B. Then, calculate the discomfort standard by substituting the following equation (2) defined to find U.C.

Where yi ^AA is an estimate of the output of group A by the fuzzy model obtained according to group A, y _i ^AB is an estimate of the output of group A by the fuzzy model obtained according to group B, and y _i ^BB is obtained according to group B An estimate of the output of group B by the fuzzy model, y _i ^BA is an estimate of the output of group B by the fuzzy model obtained according to group A, n _A is the number of data in group A, n _B is the number of data in group B, The first term is the difference between the estimated values of the output by the group A and the group B with respect to the input data of the group A, and the second term is the difference between the estimated values of the output by the group A and the group B with respect to the input data of the group B.

The value of U.C according to the above data is less than U.C1 and calculated by U.C1 = (2.16). After that, the process of selecting the structure that minimizes the value of U.C.

[Stage 2]

Set up a fuzzy model with two plant laws. In the construction of the premise, the selection of variables and the fuzzy division are considered at the same time.

First, a structure having only one of t1, t2, t3, t4, t5, and t6 as a predicate variable is considered, and the space is divided into two. Therefore, the following six structures are considered as the structure of the premise part. That is, the purge state that each variable t1 to t2 of each premise can have is divided into low temperature (hereinafter referred to as SMALL) and high temperature (hereinafter referred to as BIG), and a fuzzy indicating the degree of SMALL and BIG. The function of obtaining the parameters necessary to show the fuzzy function and the process of obtaining the inferred position angle using this parameter will be described later, and the following results are described together with the structures.

Get the U.C values for each of the above six. In order to obtain the UC value, it is necessary to find the FUZZY partition (preliminary parameter setting) for each. In the present invention, a genetic algorithm (GA) is used instead of the complex method, which is a general method of determining preamble parameters. Applied.

For example, the premise parameter of the structure corresponding to said (1) (henceforth "2-1) structure") is as follows.

Here, P1 and P2 represent the lower limit and the upper limit of the range corresponding to SMALL, and P3 and P4 represent the lower limit and the upper limit of the range corresponding to BIG. Therefore, this fuzzy function structure is sufficiently determined by four parameters P1, P3, P2 and P4.

It is assumed that the temperature in the refrigerator compartment is controlled between -10 ° C and 20 ° C. This assumes a sufficient temperature range in which the refrigerator compartment is used. This temperature range is subdivided into 0.1 ° C units to form a string with 300 bits, and any four bits of 300 bits of each string are filled with a value of 1 and the rest are all filled with 0 to form a random string. Make. Hundreds of random strings are constructed.

Then, the GA purge algorithm is applied using the random strings and the measurements configured as shown in Table 1. First, we calculate the neutral correlation between each random string and the measured value.The upper 10% of the random strings with good (high) defamatory correlations are the upper group, and the lower 10% of the random strings with lower (malignant) correlations are subgroups. , And the rest are separated into infix groups. The upper group is REPRODUCTION and the lower group is removed. The median group crosses with the parent group to create a new random string. Repeat the steps of cloning, culling, and mating using the reconstructed random strings. Continuously compare the neutral correlation coefficients of randomly generated random strings, and if the value is no longer high, the data of the random strings at that time is the final P1, P2, P3, and P4 values. .

When the predicate parameter is determined, the U.C value is determined accordingly, and the U.C value at this time becomes the U.C value of the 2-10 structure described above.

The above method is applied to all of the structures shown in (2) to (6) (hereinafter, referred to as '2-2) structure' to '2-6) structure') to obtain respective U.C values. Comparing the obtained U.C values are as follows.

Here, U.C1 is U.C1 as described above, and the number represented by x indicates the number of divisions in the part where (xy) is in the subscript next to UC. The number represented by y means the yth structure among the structures 2-10 to 2-6), and the numbers in parentheses mean the calculated UC values.

As shown in the above comparison, it can be seen that the U.C value of the '2-2) structure' is the smallest in the two-divided structure. Thus, we create a new three-part structure based on the structure that divides the variable t2.

[Stage 3]

To create a three-part structure, we need to add a new variable to make room for t2-ti. At this time, the possible variables that can be taken as ti are t1, t3, t4, t5, and t6, so many structures are generated.In order to eliminate unnecessary structures, a variable having a UC value larger than the value of U.C1 in stage 2 is omitted. The structure is simplified. Therefore, in the current system, the fuzzy division is performed by dividing the space of t2 and t3 by three. The structure at this time is shown in FIG.

4 is a graph showing a region division state according to a three-division purge structure, in which t2 is represented by a horizontal axis and t3 is represented by a vertical axis. Since we divide about t2, there are three ways to divide.

In (a) of FIG. 4, it is divided into three spaces: space L1 with t2 = SMALL, space L2 with t2 = BIG and t3 = SMALL, space L3 with t2 = BIG and t3 = BIG, The following figure shows the fuzzy function and calculates the output ang, as shown in the following (1) (hereinafter referred to as '3-1) structure'). As in the stage 2 described above, the parameter values according to this structure, the fuzzy function based on the parameter values, and the inferred position angle are shown with each fuzzy structure. The same method is used for the description in the following stage.

In (b) of FIG. 4, the space is divided into three spaces: space L1 with t2 = SMALL and t3 = SMALL, space L2 with t2 = SMALL and t3 = BIG, space L2 with t2 = BIG, The following figure shows the fuzzy function and calculates the output ang, as shown in the following (2) (hereinafter referred to as '3-2) structure').

In c) of FIG. 4, the space is divided into three spaces, a space L1 with t2 = SMALL, a space L2 with t2 = MEDIUM, and a space L3 with t2 = BIG. Is the same structure as (3) (hereinafter referred to as '3-3)').

Accordingly, the divided region according to c) of FIG. 4 of the fuzzy division region, that is, the '3-3) structure' is as follows.

3-3) The preliminary parameter of the structure is calculated by applying GA purge as in STAGE 2.

Thus, as in stage 2, it is assumed that the temperature in the refrigerator compartment is controlled between -10 ° C and 20 ° C. This assumes a sufficient temperature range in which the refrigerator compartment is used. This temperature range is subdivided into 0.1 ° C units to form a string with 300 bits, and any 8 bits of 300 bits of each string are filled with a value of 1, and the rest are all filled with 0 to form a random string. Make. Hundreds of random strings are constructed.

Then, the GA purge algorithm is applied using the random strings and the measurements configured as shown in Table 1. First, we calculate the neutral correlation between each random string and the measured value.The upper 10% of the random strings with good (high) defamatory correlations are the upper group, and the lower 10% of the random strings with lower (malignant) correlations are subgroups. , And the rest are separated into infix groups. The upper group is REPRODUCTION and the lower group is removed. The median group crosses with the parent group to create a new random string. Repeat the steps of cloning, culling, and mating using the reconstructed random strings. Continuously compare the neutral correlation coefficient of randomly generated random strings, and if the value is no longer high, the data of the random strings at that time correspond to the final P1, P2, P3, P4, P5, P6, P7, and P8. It is a parameter of the premise as a value to make.

When the predicate parameter is determined, the value of U.C becomes the U.C value of the structure 3-3) shown in FIG. 4C.

This method is applied to both the structures (1) and (2) (ie, '3-1)' and '3-2)' to obtain respective UC values, and the smallest UC value is obtained. The structure is selected and fuzzy partitioned to create four spaces based on this structure. At this time, the obtained values of UC _(3-1), UC _(3-2) , and UC _(3-3) should be better than ₍₂₎ smaller than UC _(2-2) and divided into four. If worse than UC _(2-2) (i.e., larger), the fourth division is stopped and a fuzzy rule having a UC _(2-2) value is used as the final purge rule. The following is a comparison of the magnitudes of the UC values found in the current system.

That is, the 3-3) structure has the smallest U.C value.

Therefore, a new quadrant structure is created based on 3-3) structure.

[Stage 4]

In this stage, the premise structure of the fuzzy model of stage 3 is further refined, and a model consisting of four plant laws is set. If the value of UC is smaller than UC _{(2-2) in} addition to stage 3, it should be considered as a three-division structure, which is the starting point for four-partition, but with the minimum UC-value to reduce the search process. The structure, in this case 3-3), is considered based on the structure. FIG. 5 shows a four-divided structure structured accordingly.

FIG. 5 is a graph showing a state partitioning state according to a 4-division purge structure, in which t2 is represented by a horizontal axis and t3 is represented by a vertical axis. 3-3) There are four ways to divide the structure into four.

For each of these structures (hereinafter referred to as "4-1) structure" to "4-4) structure", the U.C value is calculated in the same manner as in stage 3. The magnitude of U.C value obtained in this system is as follows.

Therefore, purge 5 division of stage 5 is performed based on the structure 4-1) corresponding to UC _(4-1) . However, in the present experiment, when each UC value was obtained by performing 5 divisions, all of the results were larger than the UC _(4-1) value.

As a result, the inferred position angle of the rotor blade for optimum high temperature equilibrium has the first precondition structure of the fuzzy quadrant.

Finally, the premise structure, parameter values, and conclusion structure finally applied according to 4-1) structure are as follows.

Thus obtained parameters of the premise obtained as follows, and the method for obtaining them using the GA purge algorithm as in stage 2, stage 3.

According to this value, the value of W1, W2, W3, W4, W5, W6 is calculated using the following equation (6). W1 to W6 are weights for applying the reflectivity of each region in the finally determined structure 4-1), and are obtained as follows according to the general theory in the TSK purge.

Using W1, W2, W3, W4, W5, W6, and ang1, ang2, ang3, and ang4, the final angle of inference in the refrigerating chamber is as follows.

This result is reflected in the neural network and used to learn the neural network. Marked as.

Next, as a second step, a step of learning by neural network and obtaining an equilibrium angle will be described.

From the past and present values of the S1 and S2 sensors, information of the actual product, and the stop positions of the past rotary blades, the stop positions of the rotary blades for the next sample period are obtained. As follows.

There is one output mode, as follows.

Y: ang (k + 1) (Rotating blade stop angle for the next sample time)

6 shows a neural network constructed according to such a configuration. In this figure, W1 is a weight value between the input layer and the middle layer, W2 is a weight value between the middle layer and the output layer, B1 is a bias applied to the middle layer, and B2 is a bias applied to the output layer. . B1 and B2 are constant input values externally applied to increase the accuracy of learning, and a value of '1' is typically applied.

At this time, the hidden layer is composed of one layer and the number of nodes is 20. The neural network thus constructed is shown in FIG.

The neural network is trained by the back propagation method using the output value of the TSK purge model for the supervisor obtained in the first step described above in the neural network having the input node and the output node. As described above, the data used as the reference for learning are the actual reference values of a1 to a5 as the reference data on the input side, and the reference data on the output side is the ang (k + 1) value. As the difference from the value, the degree of learning is reflected by comparing with the output position angle by fuzzy inference.

The relationship between the input node and the intermediate node in FIG. 7 is as follows according to the general equation in the neural network.

The relationship between the intermediate node and the output node is as follows.

Tan-sigmoid function was used as the transfer function of the middle layer, and log-sigmoid function was used as the transfer function of the output layer. We also used an improved back propagation method that uses momentum to avoid falling into the local minimum when learning.

The number of data used for learning is 316. 3000 times were learned, and the error was as shown in FIG. As the number of learning increases, the error gradually decreases.

The weights and biases according to the configuration of the neural network finally obtained are shown in Tables 2 to 5 of FIG. Table 2 shows the weights from the input layer toward the middle layer, where the superscript of W is the number of input layer nodes and the subscript is the number of middle layer nodes. Table 3 shows the weight values from the middle layer to the output layer. The superscript of W is the number of the middle layer node and there is one output node, so there is no subscript. Table 4 shows the weight values of the bias applied to the middle layer, and the superscript of b is the number of the middle layer node. Table 5 shows the weight values of the bias applied to the output node.

9 is a control block diagram of a temperature control device for performing the temperature control method according to the present invention.

Overall control is achieved by the microprocessor 50. Inference Position Angle by Fuzzy Inference Is input to the memory of the microprocessor 50 and is used as reference data for learning neural networks. The input value of the microprocessor 50 is the current rotation angle ang (k) and S1 (S1, S2), the current temperature measurement value detected by the temperature sensor and the temperature measurement value S1 before one sampling cycle. (k-1) and S2 (k-1). The microprocessor 50 performs an arithmetic routine by the neural network calculation routine section 55 according to this input value, and performs the load control routine 56 as a result of the balance angle through the rotary blade angle control section 53. Control the temperature of the refrigerator by controlling towards. The operation of the microprocessor 50 is not limited to operations and the like, and is actually used to control the compressor 51, the damper 52, and the like to control whether the refrigerator is refrigerated, and the amount of cold air discharged.

According to the temperature control method and the temperature control apparatus of the refrigerator using the neural network and the fuzzy inference as described above, the temperature in the refrigerating chamber is accurately inferred by the fuzzy theory and the error according to the difference of each refrigerator product by the neural network is reduced, Accordingly, the optimum rotor blade equilibrium angle is calculated, so that the temperature distribution in the refrigerating chamber is always maintained evenly.

Claims (14)

A temperature control method of a refrigerator for controlling a cooling air discharge direction in a refrigerating compartment according to each position of a rotary blade, the method comprising: calculating a fuzzy model based on a measured value of a rate of change of a predetermined temperature in a refrigerating compartment according to a stop angle of the rotary blade; Calculating optimal rotational wing inference position angles to maintain equilibrium temperature in the refrigerating chamber by performing fuzzy inference according to the fuzzy model; Constructing a neural network which takes the temperature change value indicated by the temperature sensor in the refrigerating compartment at the input node according to the stop angle of the rotary blade, and calculates an optimum rotational blade equilibrium angle at the output node to maintain the equilibrium of the temperature in the refrigerating compartment; ; Learning the neural network using the temperature change value and the inference position angle; And controlling the rotary blades according to the equilibrium angles output from the learned neural networks. The method of claim 1, wherein the calculating of the inferred position angle comprises: performing fuzzy division based on a measurement value for each of the predetermined points; Obtaining a respective individual discomfort norm for each divided structure of the fuzzy divided region and constructing a linear equation corresponding to the conclusion of the fuzzy inference based on the divided structure having the smallest discomfort norm. How to control the temperature of the refrigerator. The method of claim 2, further comprising repeating the fuzzy division for the partition structure having the smallest discomfort norm among the current partition structures until the discomfort norm reaches the partition structure having the smallest value. How to control the temperature of the refrigerator. The method of claim 2 or 3, wherein the calculating of the discomfort norm value comprises: calculating a value of a parameter representing a fuzzy region of the divided structure; Refrigerating temperature control method comprising the step of obtaining a discomfort norm value based on the parameter value. The method of claim 4, wherein the calculating of the parameter value comprises: determining a number of parameters for the purge region forming the division structure; Providing a string subdividing the possible temperature range that the refrigerator compartment can have into a predetermined number of bits; Constructing a plurality of random strings in which the number of bits corresponding to the number of parameters among the bits of the string is composed of mouths differently from the remaining bits; Calculating a neutral phase correlation coefficient between the random strings and the measured value; And taking the information of the random string having the highest correlation coefficient as the parameter value. 6. The method of claim 5, further comprising: duplicating an upper group composed of random strings having a high correlation coefficient and culling a lower group composed of random strings having a low correlation coefficient; Crossing the upper group and the middle group not corresponding to the lower group with the upper group; The method of controlling the temperature of the refrigerator further comprising the step of obtaining a correlation coefficient with only the modified upper group including the random string having a high correlation coefficient due to the crossing. 5. The method of claim 4, wherein the configuring of the linear equations comprises configuring the linear equations by reflecting weights that contribute to maintaining the temperature equilibrium in the refrigerating compartment of each purge region of the divided structure. . The method of claim 1, wherein the input node takes a value including: a current stop angle at which the rotary blade is located; Present temperature values measured by temperature sensors in the refrigerator compartment; And a temperature value before a predetermined time measured by temperature sensors in the refrigerator compartment. The method of claim 1 or 8, wherein the neural network learning is performed by a back propagation method. The method of claim 1 or 8, wherein the learning of the neural network comprises: obtaining an error by comparing the inference position angle and the equilibrium angle with respect to the same input value; And reflecting the error to the neural network. In the temperature control apparatus of the refrigerator for adjusting the direction of the cold air discharge in the refrigerating chamber according to the angular position of the rotary blade, purge according to the purge model calculated by the measurement value of the temperature change rate of a predetermined place in the refrigerating chamber according to the stop angle of the rotary blade. A fuzzy calculation unit for inferring to calculate an optimal position of the rotor blade inference position for maintaining the equilibrium of temperatures in the refrigerating chamber; A neural network computing unit having a neural network that takes the temperature change value indicated by the temperature sensor in the refrigerating compartment as an input node and calculates an optimum rotational wing equilibrium angle at the output node to maintain the equilibrium of the temperature in the refrigerating compartment; ; A learning unit learning the neural network by reflecting an error between the temperature change value and the inference position angle to the neural network computing unit; And a control unit for controlling the rotating blades according to the equilibrium angle. 12. The method of claim 11, wherein the fuzzy calculation unit obtains individual discomfort norms for each divided structure that is fuzzy-split based on the measured values for each of the predetermined points, and based on the divided structure having the smallest discomfort norm, Comprising a linear formula corresponding to the conclusion portion, the temperature control apparatus of the refrigerator, characterized in that to perform fuzzy inference according to the linear formula. The method according to claim 11 or 12, wherein the value that the input node takes includes: a current stop angle at which the rotary blade is located; Present temperature values measured by temperature sensors in the refrigerator compartment; Refrigerator temperature control device comprising a temperature value of a predetermined time measured by the temperature sensors in the refrigerator compartment. The temperature control apparatus of the refrigerator according to claim 11 or 12, wherein the learning unit learns by a reverse propagation method.