JP2599985B2 - Pattern recognition system using neural network - Google Patents

Description

The present invention relates to a pattern recognition system for recognizing various patterns using a neural network connected to a sensor group.

[Prior Art] Conventionally, as a system for recognizing various types of gases, for example, a system using a group of various types of sensors and a neural network connected to the group is known (pattern recognition using a neural network). Regarding systems, Nikkei Electronics, 1987.8.10, P.115-, Nature Vol.3
23, Oct 9, 1986).

In this system, for example, back propagation is applied to a neural network as shown in FIG. 4 to recognize the type of gas (the figure is simplified for simplicity of explanation).

In the figure, 1 is a multi-type sensor, 21 is a plurality of first-stage synapse elements connected to each sensor 1, 22
Are connected to the first-stage synapse elements 21, 21,... Respectively.
A third-stage synapse element is connected to each of the stage synapse elements 22, 22,....

The synapse element 21 weights a signal having a magnitude corresponding to the magnitude of the sensor signal input thereto and outputs the weighted signal to each of the synapse elements 22, 22,. Further, the synapse elements 22, 22,... Weight the respective signals having the magnitudes corresponding to the magnitudes of the respective signals thus inputted, and the respective synapse elements 23,
23, ... are output.

Using such a system, pattern recognition is performed as follows.

That is, at the time of the initial supervised learning, a gas type is assigned to each of the synapse elements 23 at the highest stage. For example, alcohol is assigned to the left synapse element 23, gasoline is assigned to the central synapse element 23, and LPG gas is assigned to the right synapse element 23. Then, the sensor group 1 is
For example, put in alcohol gas. The weights are appropriately changed and adjusted so that a sufficiently large signal (0.999) is output from the left synapse element 23 and a sufficiently small signal (0.001) is output from the center and right synapse elements 23. . Similarly, for gasoline and LPG gas,
The weight is changed and adjusted so that in the case of gasoline, a sufficiently large signal (0.999) is output from the central synapse element 23, and a sufficiently small signal (0.001) is output from the left and right synapse elements 23. In the case of LPG gas, a sufficiently large signal (0.99
9) is output, and a sufficiently small signal (0.001) is output from the left and synapse elements 23.

In this way, with the system in which each weight of the neural network is appropriately adjusted by the initial supervised learning,
The sensor group is placed in an unknown gas, and the signal magnitude of the left synapse element 23, the signal magnitude of the central synapse element 23, and the signal magnitude of the right synapse element 23 are relatively compared. However, for example, when the synapse element 23 to which 0.999 is output is the left synapse element 23, the unknown gas is recognized as alcohol.

By the way, such a sensor changes its characteristics due to so-called aging. In other words, even if the output signal value of the sensor is the same alcohol, it will be different. Nevertheless, if the weighting of the neural network set at the beginning is left as it is, the recognition rate may be reduced and erroneous recognition may be performed.

Therefore, conventionally, for example, after a lapse of one year, the adjustment is performed again by supervised learning using a predetermined gas in the same manner as in the initial case.

[Problem to be Solved by the Invention] However, such a method of performing supervised learning again with respect to the secular change of the characteristic of the sensor is too labor-intensive because the operator needs to perform the learning periodically. There is a problem that.

An object of the present invention is to provide a pattern recognition system using a neural net that can respond to a change in sensor characteristics without requiring any human intervention in view of the problems of such a conventional neural network pattern recognition system.

[Means for Solving the Problems] The present invention relates to a gas sensor group for detecting a gas to be recognized, and a neural network recognizing means for inputting data output from the gas sensor group and performing pattern recognition using a neural network. A recognition result output means for outputting a result recognized by the neural network recognition means; and a timing instructing means for instructing the timing of unsupervised learning at predetermined time intervals predetermined according to the properties of the gas sensor group. After having the operator supervisedly learn the neural network by the operator at the beginning, the control means outputs the teaching data at the time of the recognition based on the recognition result of the pattern recognition at a predetermined time according to the instruction from the timing instruction means. The neural network so as to approach the corresponding teaching data value at the time of the initial learning. This is a system for changing the weight value of a set.

[Operation] The present invention performs supervised learning on the neural network at an initial stage, and then, for example, every one week from the initial setting, performs pattern recognition based on the teaching data of the neural network at that time, and based on the recognition result. The weight value of the neural network is changed so that the teaching data of the neural network approaches the corresponding teaching data value at the time of the initial learning.

EXAMPLES Hereinafter, the present invention will be described with reference to the drawings showing the examples.

FIG. 1 is a block diagram showing one embodiment of a pattern recognition system using a neural network according to the present invention.

The various sensors 10 are sensors that detect a recognition target such as gas. For example, one of the sensors 10 outputs a large signal for alcohol gas, and outputs a small signal for other gas such as gasoline.
Numeral 10 outputs a large signal to gasoline and outputs only a small signal to other gases.

Numeral 100 is a neural network recognizing means that receives data output from the sensors 10, 10,... And performs pattern recognition using a neural network.

This neural network recognition means 100
A plurality of first-stage synapse elements 210, 210,
, And those first-stage synapse elements 210, 210, ...
Are connected to the second-stage synapse elements 220, 220, respectively.
... and their second-stage synapse elements 220, 220, ...
Are connected to the third-stage synapse elements 230, 230, respectively.
.. Are provided. The synapse element 210 weights a signal having a magnitude corresponding to the magnitude of the sensor signal input thereto and outputs the weighted signal to each of the synapse elements 220, 220,. Further, the synapse element 220 weights a signal having a magnitude corresponding to the magnitude of each signal thus input, and the next-stage synapse elements 230, 230,...
Output to Each synapse element 230, 2 at the last stage
.. Output a teaching data signal corresponding to the magnitude of the signal input thereto.

Further, the neural network recognizing means 100 inputs the teaching data of the synapse elements 230, 230,... Of the last stage, and based on those signals, recognizes alcohol, gasoline or A recognition means 30 for recognizing LPG gas is provided.

The recognition result output means 40 is a recognition result output means for outputting a result recognized by the neural network recognition means 100. For example, display means such as a liquid crystal, an LED element, and a lamp.

The timing instructing means 50 is means for instructing the timing of unsupervised learning. For example, the timing instructing means 50 instructs the timing of the unsupervised learning at predetermined intervals such as one week, one month, or the like.

The control means 60, in accordance with the instruction from the timing instruction means 50, based on the recognition result of the recognizing means 30 when the instruction is given, based on the teaching data of the synapse element 230 at that time corresponding to the teaching data at the time of the initial learning. This is a means for changing the weight value of the neural network so as to approach the value.

 Next, the operation of the embodiment of the present invention will be described.

First, the operator performs supervised learning at an early stage. This initial supervised learning is the same as in the past. That is, a gas type is assigned to each of the synapse elements 230 in the uppermost stage. For example, alcohol is assigned to the leftmost synapse element 230, gasoline is assigned to the central synapse element 230, and LPG gas is assigned to the rightmost synapse element 230.
Then, the sensor groups 10, 10,... Are put in, for example, alcohol gas. And the left synaptic element 230
Are appropriately changed and adjusted so that a sufficiently large signal (0.999) is output from the, and a sufficiently small signal (0.001) is output from the center and right synapse elements 230.
Similarly, for gasoline and LPG gas, the weights are changed and adjusted. In the case of gasoline, a sufficiently large element (0.999) is output from the central synapse element 23, and a sufficiently small signal is output from the left and right synapse elements 23. (0.001) is output, and in the case of LPG gas, a sufficiently large signal (0.999) is output from the right synapse element 23,
A sufficiently small signal from the left and center synapse elements 23 (0.
001) is output.

In this way, a system in which each weight of the neural network is appropriately adjusted by the initial supervised learning is created.

Thereafter, the unknown gas is recognized by this system. That is, in the unknown gas, the sensors 10, 10,
... and the synapse element 23 on the left by the recognition means 30
The magnitude of the signal of 0, the magnitude of the signal of the central synapse element 230, and the magnitude of the signal of the right synapse element 230 are relatively compared, for example, a synapse element 230 from which a signal of magnitude 0.999 was output. Is the left synaptic element 230, the unknown gas is recognized as alcohol. When a signal having a magnitude of 0.999 is output from the central synapse element 230, the unknown gas is recognized as gasoline. Further, when a signal having a magnitude of 0.999 is output from the left synapse element 230, the unknown gas is recognized as an LPG gas. The same applies to other gases.

Next, from the time of the initial setting, when a predetermined period determined in advance, for example, one week, the timing instruction means 50,
Instructs the timing of unsupervised learning.

The control unit 60 receives the instruction from the timing instruction unit 50, and based on the recognition result of the recognizing unit 30 when the instruction is given, the teaching data of the synapse element 230 at that time corresponds to the initial learning. The weight value of the neural network is changed so as to approach the teaching data value. That is, the recognition result at the time of the instruction is alcohol gas, the teaching data value of the left synapse element 230 is 0.911, the teaching data value of the central synapse element 230 is 0.051, and the left synapse element 23 is 23.
Assume that the teaching data value of 0 is 0.051. These numerical values are different from the initial values because the characteristics of the sensor 10 have changed over time. Therefore, the teaching data of the left synapse element 230 is the initial value of 0.999, the teaching data of the central synapse element 230 is the initial value of 0.001, and the left synapse element 2 is
30 teaching data is 0.001 of the initial value,
The weight is changed.

One week after the change, the timing instructing means 50 again instructs re-learning. Control means receiving the instruction
60 changes the weight again as described above.

In this way, at appropriate intervals, the system of the present invention automatically performs unsupervised learning.

At the time of re-learning, the teaching data does not approach the initial corresponding value itself as in the above embodiment, but at least approaches the teaching data and a predetermined value determined from the corresponding initial teaching data. Is also good. For example, a certainty factor CF may be introduced so as to approach a predetermined value determined by the teaching data at the time of re-learning, the corresponding initial teaching data, and the certainty factor CF. The contents will be described next with reference to FIG.

That is, in FIG. 2, the initial teaching data (20
1) is set to X, teaching data (202) one week later is set to D,
Assuming that the recognition result is R and the confidence is CF, the new target teaching data (203) is expressed by N = [N _i , N _O ] N _i = (1−CF) X + CF · D N _O = R Can be represented.

Note that the confidence CF when the recognition result is i is It is.

n: the number of patterns (the number of types of recognition targets) O _i , O _j : teaching data values Further, based on the recognition result of the recognition means 30, the timing instructing means 50 determines the teaching data at the time of the recognition and the initial learning time. May be always compared with the corresponding teaching data, and when the difference exceeds a predetermined value, the timing of unsupervised learning may be instructed. That is, assuming that the current recognition result is alcohol gas, the initial teaching data values of 0.999, 0.001, and 0 in the case of alcohol gas of each synapse element 230.
001 current teaching data value, e.g. 0.987, 0.006, 0.0
08, respectively, and when the difference exceeds a predetermined value determined in advance, for example, for the leftmost synapse element 230, the difference is 0.012, since it exceeds the predetermined value 0.008,
This is for giving an instruction for unsupervised learning. This is because re-learning is considered necessary if there is a certain difference. This means that if you have to re-learn,
Re-learning will be performed.

 Next, the results of a simulation experiment of the present invention will be described.

FIG. 3 is a graph showing a characteristic change of the sensor. The output values of the sensor groups S1, S2,... Relating to gasoline change over time as indicated by the solid line A 'five months after the solid line A.
It is assumed that the output value of the sensor group S1, S2,... Regarding alcohol changes as shown by a solid line B 'five months after the broken line B.

As described above, the sensor groups S1, S2,.
In a system that learns and recognizes at an early stage using, there is a risk that after five months, gasoline and alcohol may not be recognized.

In this simulation, it is assumed that the characteristic changes by 1/5 of the characteristic change for 5 months every month.

 The results are shown in Tables 1 and 2.

Table 1 shows the recognition result when the supervised learning is performed only at the initial stage in the related art.

As is evident from Table 1, after the third month, gasoline is mistaken for alcohol. The correct answer rate for 5 months
70%.

Table 2 shows recognition results when unsupervised learning is performed again every month according to the present invention.

As is clear from Table 2, since the characteristic change is learned again, the recognition rate does not decrease. 100% accuracy for 5 months
%.

The neural network recognizing means 100, the control means 60, and the timing instructing means 50 described above can be realized not by discrete hardware but by software using a computer.

Further, it goes without saying that the present invention can be applied to other pattern recognition systems using a plurality of sensors, such as temperature and character pattern recognition, in addition to the above gas type recognition.

[Effects of the Invention] As is clear from the above description, the pattern recognition system using the neural network according to the present invention includes:
Since the unsupervised learning is automatically repeated according to the instruction of the timing instruction means, there is an advantage that the recognition can be correctly performed for a long time in response to the secular change of the characteristic of the sensor group without manual intervention.

[Brief description of the drawings]

FIG. 1 is a block diagram showing an embodiment of a pattern recognition system using a neural network according to the present invention.
The figure is a configuration diagram showing a method of creating training data of the embodiment,
FIG. 3 is a graph showing a characteristic change of a sensor in the simulation of the embodiment, and FIG. 4 is a schematic diagram showing a main part of a pattern recognition system using a neural network. 10: sensor, 30: recognition means, 40: recognition result output means, 50: timing instruction means, 60: control means, 21
0, 220, 230 ... synapse element, 100 ... neural network recognition means

 ──────────────────────────────────────────────────続 き Continued on the front page (56) References IEICE Technical Report Vol. 88 No. 177 PP. 79-94 PR U88-58 "Character Recognition Using Neural Networks" Yamada, Kami, Mizoguchi, Tenma IEICE Technical Report Vol. 88 No. 177 PP. 71-77 PR U88-57 "Handwritten Alphanumeric Recognition Using Neural Networks-Evaluation of Model Size and Number of Learning Data-" Kada et al.

Claims (6)

(57) [Claims] 1. A group of gas sensors for detecting a gas to be recognized, a neural network recognizing means for inputting data output from the group of gas sensors and performing pattern recognition using a neural network, and the neural network recognizing means. A recognition result output unit that outputs a recognized result; and a timing instruction unit that instructs the timing of unsupervised learning for each predetermined period that is predetermined according to the property of the gas sensor group. After the network has been supervised by an operator, the control means, based on the recognition result of the pattern recognition at a predetermined time according to the instruction from the timing instruction means, converts the teaching data at the time of the recognition into the corresponding data at the time of the initial learning. Changing the weight value of the neural network so that it approaches the teaching data value Pattern recognition system using a neural network characterized. 2. A gas sensor group for detecting a gas to be recognized, a neural network recognizing means for inputting data output from the gas sensor group and recognizing a pattern using a neural network, and the neural network recognizing means. A recognition result output unit that outputs a recognized result; and a timing instruction unit that instructs the timing of unsupervised learning for each predetermined period that is predetermined according to the property of the gas sensor group. After the network has been supervised by an operator, the control means, based on the recognition result of the pattern recognition at a predetermined time in accordance with the instruction from the timing instructing means, transmits the teaching data at the time of the recognition at least during the initial learning. To the value determined by the teaching data value to be recognized and the teaching data at the time of recognition. , The pattern recognition system using neural network and changes the weight value of the neural network. 3. The method according to claim 1, wherein the neural network is supervised and learned by an operator at an initial stage, and then, based on a recognition result of the pattern recognition at a predetermined time in accordance with an instruction from the timing instructing means, the teaching data is used by the control means. The weight value of the neural network is changed so as to approach at least a value determined from the corresponding teaching data value at the time of the initial learning, the teaching data at the time of its recognition, and the certainty factor CF. 2. A pattern recognition system using the neural network described in 2. 4. A gas sensor group for detecting a gas to be recognized, a neural network recognizing means for inputting data output from the gas sensor group and performing pattern recognition using a neural network, and the neural network recognizing means. A recognition result output unit that outputs a recognized result, and a timing instruction unit that instructs the timing of unsupervised learning, and after the supervised learning is initially performed by the operator on the neural network by the operator, by the control unit,
Based on the recognition result of the pattern recognition at a predetermined time according to the instruction from the timing instructing means, the weight value of the neural network is set so that the teaching data at the time of the recognition approaches the corresponding teaching data value at the time of the initial learning. Further, the timing instruction means constantly compares the teaching data at that time with the corresponding teaching data at the time of the initial learning based on the recognition result, and the difference is determined in advance in accordance with the property of the gas sensor. A pattern recognition system using a neural network, wherein when a predetermined value is exceeded, the timing of unsupervised learning is instructed. 5. A gas sensor group for detecting a gas to be recognized, a neural network recognizing means for inputting data output from the gas sensor group and performing pattern recognition using a neural net, and the neural network recognizing means. A recognition result output unit that outputs a recognized result, and a timing instruction unit that instructs the timing of unsupervised learning, and after the supervised learning is initially performed by the operator on the neural network by the operator, by the control unit,
Based on the recognition result of the pattern recognition at a predetermined time according to the instruction from the timing instructing means, the teaching data at the time of the recognition is determined at least from the corresponding teaching data value at the time of the initial learning and the teaching data at the time of the recognition. The timing instruction means constantly compares the teaching data at that time with the corresponding teaching data at the time of the initial learning based on the recognition result, so as to approach A pattern recognition system using a neural network, wherein when the difference exceeds a predetermined value which is predetermined according to the properties of the gas sensor, the timing of unsupervised learning is instructed. 6. After the neural network has been supervised by an operator at an initial stage, the control means, based on the recognition result of pattern recognition at a predetermined time in accordance with an instruction from the timing instruction means, teach data for the recognition. The weight value of the neural network is changed so as to approach at least a value determined from the corresponding teaching data value at the time of the initial learning, the teaching data at the time of its recognition, and the certainty factor CF. 5. A pattern recognition system using the neural network according to 5.