JPH06110863A - Discriminating method using neural network - Google Patents

Abstract

PURPOSE:To surely produce a product by displaying new setting, which accompanies the change of an initializing production condition, as a graph on a display device in contradistinction to a control curve and enabling a user to visually recognize new setting. CONSTITUTION:A central control part 7 has a neuro-simulator 14 learnt based on actually measured instance data by a device maker, and further, it is repeatedly learnt by plural instance data. As the result, the central control part 7 is so adjusted that a neuro-network can produce uniform fried bean curds in various conditions. When an input value is set with respect to a set item, a CPU 8 calculates an optimum output value in accordance with the learnt neuro-simulator 14 and obtains coordinate values by the new set input value and the output value and displays these coordinate values on the display device in contradistinction to a continuous line (control curve) generated by connecting coordinate values of input values and output values of instance data.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a discrimination method using a neural network. More specifically, when setting a new value as an input value for a neural network previously learned from a plurality of case data. The present invention relates to a discrimination method using a neural network for visually confirming and confirming the adequacy of the input value on a display means, and for discriminating whether or not the neuro simulator takes in the input value as new learning data.

[0002]

2. Description of the Related Art With the spread of computers, new roles are beginning to be demanded, and there is an expectation that they will be transformed into machines that can handle knowledge or information. Artificial intelligence (AI: Artifi) has been studied as an important technology to support this transformation in recent years with the aim of handling information other than numerical values.
cial Intelligence) is a big close-up.

As this artificial intelligence, there are neuro and fuzzy. In the case of neuro, in computerizing the knowledge of an expert who can be shown only by a case, on the computer side, a learning mechanism is used. ,
It is an attempt to create an expert judgment theory in a computer system in a black box. This learning function is different from the one in which the expert's judgment theory is expressed in the form of rules and the complicated input / output values are calculated by mathematical formulas, as in the fuzzy theory described later, and the judgment mechanism can be easily stored in a computer. It is extremely good in that it can be created.

On the other hand, in the fuzzy theory, since a membership function is prepared and inference is carried out using this, it is possible to control by a specific numerical value, but it is difficult to make a membership function with elements intricately entangled. Is.

[0005]

That is, in the neuro technology, since the learning function is used to build up the judgment theory to determine the answer, complicated calculations such as membership functions are unnecessary, and many element functions can be easily performed. Although I could handle it, I could not actually read the control contents, so
It has been considered unsuitable for direct control of machines.
In other words, in the case of neuro, the problem remains that the judgment theory relating to control cannot be confirmed with concrete numerical values. For this reason, some neuro technologies are used in combination with other artificial intelligence as a part of control, but they are not used alone.

The present invention has been made based on the above situation. For example, a theory of judgment, which has been difficult to quantify in the past, such as a craftsman's intuition, is constructed by "learning" using a plurality of case data, and the correct answer is inferred. In the neuro technology, the neurosimulator learned in advance from the case data is made to visually confirm the validity of the input / output values set for new control on the display screen, and the input / output values are updated. It is an object of the present invention to provide a discrimination method using a neural network that can be used for direct control of a machine by determining whether or not to take in as learning data.

[0007]

The above object of the present invention is to use a neural network which is pre-learned with a plurality of case data consisting of an input value and an output value corresponding to the input value, and which can further be learned a required number of times. In the discriminating method described above, a coordinate value is obtained from the input value and the output value corresponding to the input value to form a continuous line connecting the coordinate values of each case data, and a new input value excluding each of the input values is formed. When the coordinate value determined by the output value corresponding to the new input value is near the continuous line, it is determined whether or not the new input value is taken in as learning data. It is achieved by the discrimination method using.

[0008]

The learned neurosimulator compares the control curve created based on the case data with the coordinate values obtained from the newly set input / output values and displays them on the screen, so that these can be visually recognized. Then, the validity of this new set value can be easily confirmed and judged.

Further, by further inputting the input / output value by this new setting as learning data, the neurosimulator is additionally learned, and the inference accuracy thereof can be further improved.

[0010]

Embodiments of the present invention will be described below with reference to FIGS. FIG. 1 shows a schematic configuration of a control device to which a discrimination method in a neural network according to an embodiment of the present invention is applied, and FIG. 2 shows a flowchart for explaining the operation of the control device.

The control device in this embodiment is used as a frying machine for a coagulating liquid (a mixture of soymilk and a coagulant mixed with soybean milk) that is a source of frying, and the quality is made uniform by a neuro technique. To achieve. Incidentally, for reference, in the frying process, the coagulating liquid is poured onto a stainless steel belt covered with a polypropylene filter cloth, and a constant pressure is applied with a stainless steel belt covered with a filter cloth. It is made to a thickness.

In the control device shown in FIG. 1, various input values which can be arbitrarily set and changed by a user's operation from the control panel 1 with respect to preset manufacturing conditions described later are I / O. It is adapted to be input to the central control unit 7 via the 0 interface 6. The flow sensor 2, the coagulating liquid concentration, the coagulating liquid concentration, the ambient temperature, the belt pressure of the pressurizing belt, etc., are respectively supplied from a coagulating tank (not shown).
It is designed to be detected by the densitometer 3, the temperature sensor 4, the pressure sensor 5, and the like.
0 CPU of the central control unit 7 through the interface 6
8 is input. Each sensor output is a factor that causes a change in the thickness of the frying product that has been commercialized due to various conditions such as the installation state of the machine and the coagulating liquid that changes daily.

The central control unit 7 is provided outside the CPU 8.
It is configured to include a RAM 9 and a ROM 10. The CPU 8 stores the output from each sensor and the input value set by the user in the RAM 9, and reads the manufacturing conditions preset by the device maker from the ROM 9 to output these, each sensor output, and the input value. Are compared and calculated, and the control signal is transmitted via the I / O interface 11 to the valve 12 and the molding machine 13 in the subsequent stage.
To each drive system of. The valve 12
Is provided in a pipeline for supplying the coagulating liquid to the belt and controls the supply amount of the coagulating liquid.

The CPU 8 has a neurosimulator 14 for constructing a later-described hierarchical neural network.
Is built in. This neuro simulator 14
Is inferred from learning based on case data conducted in advance by the equipment manufacturer, for example, by inferring the supply amount of coagulating liquid to the belt and the degree of pressing of the belt according to the factors that change day by day, etc. The central control unit 7
To control.

Next, the operation of this control device will be described with reference to FIG. In this control device, the neurosimulator 14 is first learned by the device manufacturer based on the actually measured case data (step ST1), and is repeatedly learned by a plurality of case data (step S).
T2). As a result, the control device is adjusted so that the neuronetwork can always produce a uniform frying under various circumstances (step ST3). That is, the coagulation molding machine is shipped after the control device is adjusted so as to change how the coagulation liquid is taken out and how much the coagulation liquid is pressed according to the situation. The solidification molding machine adjusted and shipped in this way can set input values for a plurality of specific setting items by the user's own judgment. Then, an input value is set for this setting item (step ST4). When the setting is completed (step ST5), the CPU 8 then calculates the optimum output value according to the learned neurosimulator 14 (step ST6), and at the same time, obtains the coordinate value from the new set input value and the new output value. The coordinate values are compared with a continuous line (hereinafter referred to as a control curve) formed by connecting the coordinate values of each input value and output value of the case data, and are displayed on a display not shown (step). S
T7). As a result, the user can visually recognize and confirm the correspondence between the control curve and the coordinate value obtained by the new input setting on the display. Then, in this correspondence relationship, when the neuro simulator determines that the new setting is within the predetermined allowable range (step ST8), the control device transfers the new input value and output value to the RAM 10 ( Simultaneously with step ST9), the neuro simulator 14 is additionally learned based on this new data (step ST10). Further, when the neuro simulator 14 determines that the new data is out of the allowable range (for example, due to an input error), the control device clears the input value and the output value (step ST11), and the user is informed. On the other hand, after inquiring about the necessity of re-input (step ST12), when the re-input is necessary, the device is returned to the initial state, and when it is unnecessary, it is ended.

The characteristic requirements of the present invention are shown in steps ST6 to ST10 of the above flow chart,
This requirement is further detailed by FIGS. 3 and 4. The control device obtains each coordinate value as shown in FIG. 3 from the input / output values based on a plurality of case data (11 data in this embodiment) for learning the neuro simulator 14, and stores the coordinate values in the ROM 10. I am letting you. The neuro simulator 14 predicts a value between coordinate values from each coordinate value in the ROM 10 so as to make up for the shortage of the case data,
A control curve L connecting these coordinate values is displayed on the display by the control device as shown in FIG.

Under such circumstances, when a new input value is set by the user, the control device determines, for example, that the input value is a value between coordinate values excluding the value indicated by the case data. And when set as an appropriate value, the coordinate value by this new setting is plotted on the said control curve and the vicinity of this control curve according to the prediction of a neuro. As a result, the user can easily confirm the validity of the new setting made by himself / herself, and can carry out the fried food production under the new set value with confidence. After that, the control device takes in the newly set input / output values as new learning data and causes the neurosimulator 14 to additionally learn. At this time, this additional learning acts to improve the accuracy of the control curve when the coordinate value by the new setting is plotted on the control curve, and when it is plotted in the vicinity of the control curve. It acts to correct the control curve. When the control curve is corrected, the neurosimulator 14 similarly corrects the front and rear portions F of the coordinate value T including the newly set coordinate value T as shown in FIG. The control curve L is automatically corrected so that it becomes a continuous line.

On the other hand, if the setting of a new input value by the user is abnormal due to, for example, an input error, the control device determines that the neurosimulator 14 is out of the allowable range and does not converge. Display the situation on the display. Therefore, the user can immediately recognize a setting error or the like and confirm it. The permissible range is set as a permissible error in the neurosimulator to be described later, and the setting of such permissible error excludes a state in which the simulator cannot give a reply.

Incidentally, FIG. 8 shows a state of control by a control device which does not include a conventional neurosimulator shown for comparison. In the control by such a conventional device, when a new setting is made by the user so as to correct the preset control curve L, the correction of the coordinate value T before and after the setting can be performed. Not done That is, it indicates that the control device corrects only the changed set value and does not participate in the correction of the vicinity thereof. However, in general, once the set value is changed, the user usually misunderstands that the control unit also corrects the vicinity of the set value, and as a result, in this vicinity, control is performed with a value different from the predicted value. Done and did not produce the desired product.

Finally, the neurosimulator applied to this embodiment will be outlined. As this neuro simulator, a commercially available general-purpose product can be used, and for example, “NEUROSIM / L” manufactured by Fujitsu can be used. Figure 6 shows the structure of this "NEUROSIM / L".
The OSIM / L 50 is composed of a portion 51 that simulates a hierarchical neural network and an MMI (man-machine interface) portion 52 that displays / inputs to the user 100 via the MS-WINDOWS 60. The NEUROSIM / L 50 can be executed independently as an application of MS-WINDOWS 60 or incorporated into an application of MS-DOS 70 and executed. Further, by using dedicated hardware (neuro board) 80 for high-speed calculation by incorporating it into the simulator part, it is possible to perform learning and recognition of the neural network at high speed. And the network file 90 is a ROM
Stored in 10.

Such a neural network is to finally obtain the correct answer by repeating the learning process to adjust the weight between the neurons. That is, the learning process is performed as shown in FIG. 7, and the weights between the neurons are initially set to random values, and the following operations are repeated until the error becomes sufficiently small. Presentation of input Calculation of output of neural network Calculation of error between presentation of correct answer (also called teacher signal) and correct answer Adjustment of weight between each neuron to reduce error By this operation, neural network learns , Gradually output correct answers. The neural network learned in this way stores the correct relationship between input and output due to the weight between neurons.

[0022]

As described above, according to the discrimination method using the neural network according to the present invention, the new setting accompanied by the change of the initially set manufacturing condition is compared with the control curve created by the initial setting. Since it is displayed as a graph on the display, the user can visually recognize this new setting and reliably manufacture the product.

That is, even in the conventional system using the neural network, the coordinate value obtained from the input value and the output value obtained corresponding to the input value is not displayed in the form of a graph on the display, and is never changed. Although it was only shown as a result value, it is possible to visually confirm and grasp the state of the entire control by displaying a graph in contrast with the control curve created by the initial setting as in the present invention. Can drive the machine with confidence.
Also, in the control that does not use the neuro, it is difficult to obtain the function formula of the complicated curve when it is attempted to display the graph by fuzzy control, and even if it is obtained, it is possible to find it in a narrow range. You can only draw, not the whole thing. On the other hand, the neuron does not need to be established as a functional formula by relying on the learning function to create the control curve, so that the display can be simplified. or,
According to the method of the present invention, a new setting is made by the user, and the setting is used as new learning data for a previously learned neurosimulator, so that the neurosimulator further enhances its inference accuracy. be able to. In other words, the more conditions are used, the more various conditions are input, and optimal control is performed even under various conditions, so that the quality of the fried food to be produced can be made more uniform.

[Brief description of drawings]

FIG. 1 is a block diagram of a control device to which a discrimination method in a neural network according to the present invention is applied.

FIG. 2 is a flowchart illustrating an operation of the control device shown in FIG.

FIG. 3 is a diagram showing a state in which coordinate values obtained from case data are plotted.

FIG. 4 is a diagram showing a state in which the coordinate values in FIG. 3 are connected and displayed as a continuous line on a display.

FIG. 5 is a diagram illustrating a state when a setting item is changed with respect to a control curve by initial setting.

FIG. 6 is a system configuration diagram of a neurocomputer applied to a control device.

FIG. 7 is a diagram illustrating a learning process of a neurosimulator.

FIG. 8 is a diagram illustrating a state when a setting item is changed with respect to a control curve in a conventional example shown for comparison.

[Explanation of symbols]

 1 Control Panel 2 Flow Rate Sensor 3 Density Meter 4 Temperature Sensor 5 Pressure Sensor 7 Central Control Section 8 CPU 12 Valve 13 Molding Machine 14 Neuro Simulator

Claims (1)

[Claims] 1. A discrimination method using a neural network, which is preliminarily learned so as to obtain predetermined output values based on a plurality of input values, and which can be further learned a required number of times. A coordinate value determined from the corresponding output value to form a continuous line connecting the coordinate values, and the coordinate value determined by the new input value excluding the respective input values and the output value corresponding to the new input value Is near the continuous line, it is determined whether or not the new input value is taken in as learning data. A discrimination method using a neural network.