JPH0744517A - Signal discrimination method under noise background - Google Patents

Abstract

PURPOSE:To provide a signal discrimination method for a soner or the like having a high discrimination capability under a noise background. CONSTITUTION:In the signal discrimination method using a neural network extracting a signal from a noisy background to discriminate a characteristic such as a soner, a weight coefficient calculation means 120 calculates a square error between teacher data output and learning data output from a learning data generating means 110 after a lapse of a cycle in the learning process, and succeeding learning is conducted by eliminating learning data with the large square error of a predetermined rate in the learning data to obtain the weight coefficient 120a. Based on the weight coefficient 120a, a signal discrimination means 140 discriminates the signal with respect to an input signal 101a to provide the output of a signal discrimination result 140a.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a signal discrimination method under a noise background for extracting a signal from background noise and discriminating a feature in a sonar or the like.

[0002]

2. Description of the Related Art Conventionally, as a technique in such a field,
For example, some documents were described in the following documents. Reference 1: Information Processing Society of Japan, 30 [10] (1989)
-10) Tateishi and Yamazaki, "A Study on the Middle Layer of Multilayer Neural Networks in Handwritten Digit Recognition," p. 12
81-1288 Reference 2; Hideki Aso, "Neural Network Information Processing" (Sho 63) Sangyo Tosho, P. 39-54 Conventionally, in the extraction of a target narrowband signal and the determination of characteristics in a sonar, the received signal is subjected to frequency analysis to obtain a spectral distribution, and its temporal change is displayed as a grayscale image on a plane of (frequency × time). It was done by humans monitoring and confirming the temporal sequence of the maximum points of the spectrum that appears on the screen (called the "loafergram"). FIG. 2 is a flowchart showing an example of a conventional loagram generation process. The received signal (channel or beam output) of the sonar is output as power within the analysis width and integration time for each frequency by frequency analysis processing 1 and integration processing 2. The output of each frequency is x (f _i ), i = 0, 1, ..., N
Put it as -1. When the f _i will put analysis width and _{Δf, f i = f 0 +} i · Δf, i = 0,1, ..., is given by the N-1 ··· (1). The local average m _i is calculated in the local average calculation process 4 through the frequency loop process 3. The local average m _i is, for example, centered on the i-th bin output, and takes L / 2 bin outputs on both sides of the i-th bin output.
Can be obtained by

[0003]

[Equation 1] By the normalization process 5, a high level can be obtained for the frequency bins forming the peak with respect to the nearby frequencies, without being influenced by the signal levels of the frequency components that are far away. When it is determined by the determination processing 6 that the processing up to i = N−1 is completed, the result of the normalization processing 5 is subjected to luminance conversion and CRT.
Or as a display on recording paper. By this, the person confirms the presence of the signal from the time series of the displayed local peaks. However, the ability of a person to detect a desired signal from a loafergram varies greatly depending on the degree of fatigue. In addition, with the increase in size of sonars, the increase in multi-sensors, and the diversification of information, the amount of information to be monitored is rapidly increasing, making it difficult for a person to monitor all the information. In order to solve this problem, an automated system for signal extraction and feature discrimination is being developed, and methods such as pattern matching have been attempted, but sufficient signal extraction and feature discrimination performance has been obtained. Not not.

Therefore, a method using a neural network can be considered. The neural network simulates the structure of a human neural network to give ambiguous information the same judgment ability as a human.
Conventionally, in other fields, it has been applied to image / graphic information having a good signal-to-noise ratio (S / N ratio), such as recognition of handwritten characters, and good results have been obtained. Reference 1 is an example of applying a neural network to recognition of handwritten numbers. FIG. 3 shows a configuration example in the case of applying the method of character / figure recognition by this neural network to the extraction of a target narrowband signal and the discrimination of the feature in a sonar. FIG. 3 is a diagram showing a signal discrimination method using a conventional neural network. In this method, the hierarchical neural network 20 is used for classifying narrow band signals on the loafergram 10. Neural network 2
On the input layer 21 of 0, the screen 10 of the loafergram 10 is displayed.
a is cut out and input, and is output through the intermediate layer 22 to the output layer 23.
Outputs the degree of similarity with the registered catalog target as the class identification result.

As described in Reference 2, the neural network 20 is designed in advance to store all patterns of catalog targets by the entire weighting coefficient in the network by learning. Learning is performed using a method called an error back propagation algorithm. Data of each catalog target is input from the input layer 21, and the pattern that appears in the output layer 23 and the correct pattern (for example, +1 for the unit corresponding to the correct target, -1 for the other output units).
The weight in the network is adjusted by the difference between the error and the error (this is called "error"), and the sum of squares of the error is suppressed to a certain value or less. Usually, as the learning pattern, a pattern artificially created by simulation or a pattern cut out by human from actual data is used.

[0006]

However, the conventional signal discrimination method has a problem that a pattern unsuitable for learning is introduced, which prevents convergence of the weighting coefficient to the optimum value. The reason why patterns that are not suitable for learning are introduced is that accidental occurrence of inappropriate patterns based on random numbers used when generating learning patterns by simulation, or patterns that are difficult for machines to distinguish when humans cut out patterns. There are cutouts, etc. It is difficult to remove these patterns in advance before learning. The present invention solves the problem that the conventional technique has such that a pattern unsuitable for learning is introduced and the discrimination ability is deteriorated in a noise background, and an unsuitable pattern is determined by evaluating an error in the learning process. However, by excluding it from the subsequent learning, a signal discriminating method such as a sonar having a high discriminating ability under a noisy background is provided.

[0007]

In order to solve the above-mentioned problems, a first invention uses a neural network trained on the basis of teacher data and learning data to extract a signal from a noise background and The following measures are taken in the method of discriminating a signal such as a sonar under a noise background for discriminating a feature. That is, in the learning process, a squared error between the output of the teacher data and the output of the learned data after a certain cycle has passed is calculated, and a certain proportion of the learned data having a large squared error is removed, The following learning is done.

In the second invention, as in the first invention, in the signal discriminating method of the sonar or the like, the square error of the outputs of the teacher data and the learning data after a certain cycle has elapsed in the learning process is calculated, The learning data exceeding the set threshold value is removed from the average value of the square error of the previous cycle, and the subsequent learning is performed.

[0009]

According to the first and second aspects of the present invention, the signal discrimination method under the noise background is configured as described above. Therefore, when there is an unsuitable pattern in which an error in the learning process exceeds a certain threshold, This pattern is removed from the learning of, and the discriminating ability in a noisy background can be improved. Therefore, the above problem can be solved.

[0010]

First, the principle of learning for the neural network 20 shown in FIG. 3, which is the principle of this embodiment, will be described. Now, when the learning pattern (x _s1 , x _s2 , ..., X _sN ) is given as an input to the neural network 20 of FIG. 3 at time t, the correct answer (“( (Teaching data)) is y _sj . At this time, the output of the unit i of the intermediate layer 22 is m _i (t), and the output of the unit j of the output layer 23 is o.
_{If j} (t), then the squared error R is Calculated by Then, the coupling w _ij between the unit i of the intermediate layer 22 and the unit j of the output layer 23 is corrected as in the following Expression (4). w _ij (t + 1) = w _ij (t) + ε · (y _sj −o _j (t)) · m _i (t) (4) Here, the function f is a function that gives the input / output relationship of each unit. Is. As is clear from the expressions (3) and (4), 2
When the multiplication error R is 0, w _ij (t + 1) = w _ij (t), and no correction is necessary. In this way, the learning is repeated until the sum of the squared errors R converges to a certain value T or less.

Learning patterns (x _s1 , x _s2 , ...,
x _sN ) is generally given, but when there is a learning pattern in which the squared error R becomes very large, it will hinder the convergence of the weighting factors. Therefore, it is desirable to exclude these data. Therefore, learning progresses to some extent, and 2
When the total sum of the squared errors R reaches n times (for example, 8 times) T or less, x% (for example, 2%) of learning data of all the learning data is removed in order from the largest squared error R. To As a result, the calculation time can be shortened and the accuracy can be improved.

Next, an embodiment of the present invention will be described with reference to FIGS. FIG. 1 is a functional block diagram of a signal discriminating apparatus used in a signal discriminating method according to an embodiment of the present invention, and FIG. 4 is a functional block diagram of weighting factor calculating means in FIG. The signal discriminating apparatus shown in FIG. 1 generates learning data 110a of a neural network based on a target signal pattern 100a and teacher data 100b and outputs learning data generating means 110 for learning data 100b.
And a weighting factor calculation means 120 is connected to the output side thereof. The weighting factor calculation means 120 uses the learning data 1
Based on 10a and the teacher data 100b, the neural network is trained and the weighting factor of the network is set to 1
It has a function of calculating 20a. Further, this signal discriminating apparatus is provided with an input signal converting means 130 for normalizing the input signal 101a and outputting the normalized data 130a, and the signal discriminating means 140 is connected to the output side thereof. The signal determining means 140 is the weighting factor calculating means 12
It has a function of performing signal discrimination based on the weighting coefficient 120a from 0 and the normalized data 130a from the input signal conversion means 130 and outputting the signal discrimination result 140a. The weighting factor calculation means 120 shown in FIG. 4 has a learning data management means 121 for managing the learning data 110a input from the learning data generation means 110 and removing learning data having a large error, and the output calculation is performed on the output side thereof. Means 12
The weighting factor correction means 123 is connected via 2.
The output calculation unit 122 has the same structure as the signal determination unit 140 and has a function of performing a calculation process on the learning data output from the learning data management unit 121. The weighting coefficient correction means 123 calculates the sum of squared errors of the teacher data 100b and the output data of the output calculation means 122, corrects the weighting coefficient from the squared error, and sends the control signal to the learning data management means 121. It has a function of sending the corrected weighting coefficient 120a as well as sending it.

Next, a signal discriminating method using the apparatus shown in FIGS. 1 and 4 will be described. First, in FIG. 1, the target signal pattern 100a and the teacher data 100b are input to the learning data generating means 110. In the learning data generation means 110, Gaussian noise with an average of 0 and a variance of 1 is added to the input target signal pattern 100a to obtain learning data 110a,
It is sent to the weighting factor calculation means 120 together with the input teacher data 100b. In the weighting factor calculation means 120, the learning data 110 sent from the learning data generation means 110.
Learning is performed based on a and the teacher data 100b, the weighting coefficient 120a of the neural network is calculated, and the weighting coefficient 120a is sent to the signal discriminating means 140. For example, a method called an error backpropagation method is used to calculate the weighting coefficient 120a. Repeatedly adjust. That is, in the weighting factor calculation means 120 of FIG. 4, the learning data 110a from the learning data generation means 110 is sent to the learning data management means 121 and sequentially input to the output calculation means 122. The output calculation means 122 has the same structure as the signal determination means 140, generates an output when the learning data 110a is input, and sends it to the weighting coefficient correction means 123. The weighting factor correction means 123 calculates the sum of squared errors between the input teacher data 100b and the output data of the output calculation means 122,
When it becomes n times the threshold value T or less, the learning data management means 121 is notified so that the learning data management means 121 removes x% of the learning data having a large error. Further, the weighting factor correction means 123 considers that the error has converged when the error becomes smaller than the threshold value T, and the weighting factor 120a.
Is output. If neither is the case, the weighting coefficient correction means 123 corrects the weighting coefficient in accordance with the error backpropagation algorithm and sends it to the output calculation means 122 to perform the calculation again.
Weighting factor 120 output from weighting factor modifying means 123
a is sent to the signal discriminating means 140 in FIG. Upon receiving the sent weighting coefficient 120a, the signal determining means 140 sets a value in the weighting coefficient 120a. A received signal from a sensor or the like is subjected to frequency analysis and integration processing, and then input to the input signal conversion means 130 in the form of an input signal 101a at regular intervals. When the input signal conversion means 130 receives the input signal 101a, it performs a normalization process to generate normalized data 130a, and sends it to the signal determination means 140. The signal discrimination means 140 extracts a desired signal from the normalized data 130a and outputs a signal discrimination result 140a which is the extraction result.

As described above, in the signal discriminating method of this embodiment, the weighting factor calculating means 120 calculates the squared error between the outputs of the learning data 110a and the teacher data 100b after a certain cycle has passed in the learning process. Since a certain proportion of the learning data having a large square error in the learning data 110a is removed and the subsequent learning is performed, the contribution of the pattern having a large error to the weight update can be eliminated. . This promotes convergence in the correct direction, shortens the convergence time, and improves the performance after convergence.

The present invention is not limited to the above embodiment,
Various modifications are possible. The following are examples of such modifications. (A) In the above embodiment, the weighting factor calculation means 120 determines the inappropriate data only once when the sum of the squared errors after a certain cycle has converged to a certain value or less. It may be performed every cycle after the cycle elapses, or every several cycles. (B) As a determination method in the weighting factor calculation means 120, a threshold value is set from the average value of the squared error of the previous cycle,
Even if learning data that exceeds the threshold value is removed and subsequent learning is performed, the same effect as that of the above-described embodiment can be obtained.

[0016]

As described in detail above, according to the first and second inventions, an inappropriate pattern is determined by the evaluation of the error in the learning process and is excluded from the subsequent learning. It is possible to eliminate the contribution of the pattern having a large value to the weight update. This promotes convergence in the correct direction, shortens the convergence time, and improves the performance after convergence. Therefore, a signal discriminating method such as a sonar having a high discriminating ability under a noise background can be obtained.

[Brief description of drawings]

FIG. 1 is a functional block diagram of a signal discriminating apparatus showing an embodiment of the present invention.

FIG. 2 is a flowchart showing a conventional loafergram generation process.

FIG. 3 is a diagram showing a signal discrimination method using a conventional neural network.

FIG. 4 is a functional block diagram showing a weighting factor calculation means in FIG.

[Explanation of symbols]

 10 Lofergram 20 Neural Network 21 Input Layer 22 Intermediate Layer 23 Output Layer 100a Target Signal Pattern 100b Teacher Data 101a Input Signal 110 Learning Data Generating Means 110a Learning Data 120 Weighting Factor Calculating Means 120a Weighting Coefficient 121 Learning Data Managing Means 122 Output Computing Means 123 Weighting coefficient correcting means 130 Input signal converting means 130a Normalized data 140 Signal discriminating means 140a Signal discriminating result

Front page continuation (72) Inventor Tsuneo Ishiwata 1-7-12 Toranomon, Minato-ku, Tokyo Oki Electric Industry Co., Ltd. (72) Inventor Shunji Ozaki 1-7-12 Toranomon, Minato-ku, Tokyo Oki Electric Industry Co., Ltd. In the company

Claims (2)

[Claims] 1. A signal discrimination method under a noise background for extracting a signal from a noise background and discriminating a feature by using a neural network trained based on teacher data and learning data. Calculate a squared error between the outputs of the teacher data and the learning data after a certain cycle has passed, remove learning data having a large square error of a certain ratio in the learning data, and perform subsequent learning. A method for discriminating a signal under a noisy background. 2. A signal discrimination method under a noise background for extracting a signal from a noise background and discriminating a characteristic by using a neural network trained based on teacher data and learning data. The square error between the outputs of the teacher data and the learning data after a certain cycle is calculated, the learning data exceeding the set threshold value is removed from the average value of the square error of the previous cycle, and the subsequent learning is performed. A method for discriminating a signal under a noisy background.