JPH10227820A - Method and apparatus for correction of time response of sensor - Google Patents

Abstract

PROBLEM TO BE SOLVED: To obtain a method and an apparatus in which an electromagnetic environment can be measured so as to remove a measuring strain error generated in a sensor by a method wherein a feature amount which features an output waveform from the sensor is received on the side of an input, a feature amount which features an input waveform is received on the side of an output and a neuronetwork which expresses the relationship between the input and the output is constituted. SOLUTION: In a measuring system, the time waveform of a sensor input signal generated by a pulse generator 4 is input directly to a channel ch1 at an oscilloscope 2 via a conductor 6, a current which flows in the conductor 6 is detected by a current probe 3, and a detection time waveform (a sensor response output waveform) is measured simultaneously with the original waveform (the time waveform of the sensor input signal) by a channel ch2 at the oscilloscope 2. By the measuring system, the discrete value and the feature parameter of an input waveform and a response waveform can be grasped. As the feature parameter, the peak value of a time waveform, a rise rate, a pulse half-width, the integrated value of the waveform of the amplitude value of a preliminarily prescribed time width and the like are enumerated.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a technique for correcting a waveform measured by an EMI (Electromagnet Interference) evaluation sensor (hereinafter, simply referred to as a "sensor") such as an antenna and a current probe. The present invention relates to a technique for estimating an input waveform corresponding to a measured waveform from a sensor measured waveform, and enabling measurement of an electromagnetic environment in which a measurement distortion error generated by the sensor is removed.

[0002]

2. Description of the Related Art In recent years, electromagnetic interference caused by an electromagnetic pulse generated by a lightning surge, electrostatic discharge, or the like affects electronic devices, and electromagnetic interference that causes a failure in electronic devices has become a problem. This problem has been exacerbated by a decrease in the durability of electronic devices against electromagnetic interference due to an increase in the speed and power saving of the semiconductor devices that constitute the devices.

However, electromagnetic interference causing electromagnetic interference in electronic equipment is generated by various mechanisms, and furthermore, the characteristics of the electromagnetic interference are uncertain and the reproducibility of the electromagnetic interference is not good. It is often the case that it is difficult to elucidate the cause of the generated electromagnetic interference. Therefore, in order to take various measures against electromagnetic interference, it is important to measure electromagnetic interference and clarify its characteristics.
Various sensors have been developed for measuring the electromagnetic environment. For example, various types of antennas for detecting radiated electromagnetic interference waves propagating in space and current probes for detecting conductive electromagnetic interference waves transmitted through conductor lines such as power lines have been proposed.

[0004]

However, in order to take measures against electromagnetic interference, it is necessary to accurately measure electromagnetic interference waves and clarify the characteristics thereof, but conventionally proposed sensors have been proposed. The pulse waveform (electromagnetic interference) detected by the sensor is distorted on the time axis due to the limited measurement frequency band and structural reasons, etc., so that the electromagnetic interference can be measured accurately. There was a problem that there was no. Regarding this, "Masagi Masao, et al .:" Measurement and Analysis of Electromagnetic Pulses Associated with Indirect ESD "", EMI (Electromagnet Interfer
ence) Evaluation Sensor Society Transactions B-II, Vol. J75
-B-II, no. 9, pp. 647-654,199
2. And the like.

[0005] For example, FIG. 1 shows an example of a pulse waveform of an electromagnetic interference wave (peak value is normalized by 1), and FIG. 2 shows the pulse waveform shown in FIG. 3 shows a measurement waveform (peak value is normalized by 1) received by a biconical antenna installed at a position 3 (m) away from. As can be seen with reference to FIGS. 1 and 2, when a single pulse waveform is input to a biconical antenna which is a sensor, the response output becomes a vibration waveform having two peaks, and the electromagnetic interference wave is accurately detected. It turns out that it has not been detected.

As a technique for correcting response distortion (distortion of a response output signal with respect to an input signal) due to a sensor, a response waveform is converted from a time domain to a frequency domain, and a frequency characteristic of the sensor itself is considered in the conversion result. Again, a method of transforming the waveform from the frequency domain to the time domain has been proposed.However, when deriving a phase term, which is a conversion coefficient required when transforming between the time domain and the frequency domain, it is difficult to measure the phase term. It has been difficult to correct response distortion, that is, to accurately reproduce an input pulse waveform that is an electromagnetic interference wave due to an analysis error. Further, in order to obtain this phase term, sufficient measuring equipment is required, and in many cases, the phase term cannot be easily obtained and conversion processing to the frequency domain cannot be performed in most cases.

In order to solve such a problem that the measurement accuracy of the electromagnetic interference is not good, a sensor using an optical modulator (“Nobuo Kuwahara, et al .:“ Frequency characteristics of a printed antenna using a Mach-Zehender interferometer ”), IEICE Technical Report, EMCJ92-11, pp.33-7,
1992. "Kihiro Tajima, et al .:" Method of Measuring Electric Field Distribution Near a Form Wireless Device Using an Optical Electric Field Sensor ", 1995 IEICE Communications Society Conference, B-17
1, 1995. ) Are proposed, but there are problems in terms of manufacturing cost, reception sensitivity, and the like, and the fact is that it is not widely used.

Further, in order to effectively use a sensor that has already become widespread, it has been desired that a simple technique for correcting a measured waveform using an existing sensor appears. The present invention has been made in order to solve such a conventional unsolved problem, and an object thereof is to correct a waveform measured by a sensor, that is, to convert an input waveform corresponding to the measured waveform from a waveform measured by a sensor. Is to provide a simple means for performing electromagnetic environment measurement by removing the measurement distortion error caused by the sensor by estimating the measurement result.

[0009]

In order to achieve the above object, according to the first aspect of the present invention, there is provided a method for characterizing a known time response output waveform from an EMI evaluation sensor ("sensor"). A method for estimating one or more extraordinary traces characterizing an unknown input waveform to the sensor from a plurality of extraordinary traces, comprising inputting one or more extraordinary traces characterizing an output waveform from the sensor. Side, and the output side receives one or more characteristic amounts characterizing the input waveform to the sensor, and configures a neural network so as to express the relationship between input and output. One or more features that characterize one or more known input waveforms, as teacher data to the output side of the neural network, one corresponding to the one or more known input waveforms Also One or more traces characterizing a plurality of known time response output waveforms, as teacher data to the input side of the neural network, perform optimization processing of the neural network, and perform the optimization processing. For neural networks,
Input one or more features that characterize a known time response output waveform corresponding to an unknown input waveform to the sensor, and output the neural network corresponding to the input to an unknown input to the sensor. There is provided a sensor time response correction method characterized by estimating an input waveform as one or a plurality of characteristic amounts.

According to a second aspect of the present invention, in the first aspect, the one or more characteristic amounts that characterize the input waveform to the sensor are discrete values of the amplitude level of the time waveform of the input waveform. A sensor time response correction method, characterized in that the output waveform is a discrete value of the amplitude level of a time waveform of the output waveform. You.

According to a third aspect of the present invention, in the first aspect, the one or more characteristic amounts characterizing the input waveform to the sensor include a peak value of the input waveform,
Rise rate, pulse half width, and at least one or more of the integral amount of the waveform amplitude value of the time width defined in advance, one or more features that characterize the output waveform from the sensor, A sensor time response correction method is provided which is at least one of a peak value of an output waveform, a rise rate, a pulse half width, and an integral amount of a waveform amplitude value of a predetermined time width. .

According to a fourth aspect of the present invention, in the first aspect, the one or more characteristic amounts that characterize the input waveform to the sensor include a peak value, a rise rate, and a pulse of the input waveform. Half-width, and at least one or more of the integration amount of the waveform amplitude value of the predetermined time width, one or more features that characterize the output waveform from the sensor, of the output waveform, A sensor time response correction method is provided, which is a discrete value of an amplitude level of a time waveform.

According to another aspect of the present invention, that is, an apparatus aspect, the following means is provided. That is, according to the fifth aspect of the present invention, an unknown input waveform to the EMI evaluation sensor (“sensor”) is obtained from one or a plurality of extraordinary amounts that characterize a known time response output waveform from the sensor. An apparatus for estimating one or a plurality of characteristic amounts, which characterizes an output waveform from the sensor at an input side, and characterizing an input waveform to the sensor. A neural network configured to accept one or more features at the output and to represent the relationship between input and output, and one or more features characterizing one or more known input waveforms to the sensor And one or more second features that characterize one or more known time response output waveforms corresponding to the one or more known input waveforms. Storage means, and control means for correcting the response output from the sensor, the control means acquiring the first and second extraordinary amounts stored in the storage means, and Is given as teacher data to each of the output side and the input side of the neural network, and the neural network performs an optimization process on the neural network. One or more features that characterize a known time response output waveform corresponding to an unknown input waveform to the input, and output the neural network corresponding to the input to an unknown input waveform to the sensor The sensor time response correction device is characterized in that it is estimated and output as one or a plurality of extraordinary amounts characterizing

According to a sixth aspect of the present invention, in the fifth aspect, the neural network is characterized in that the extraordinary amount received on the input side is a discrete value of the amplitude level of the time waveform of the output waveform from the sensor. And a feature amount received on the output side is configured to be a discrete value of an amplitude level of a time waveform of an input waveform to the sensor, and a sensor time response correction device is provided. You.

Further, according to the invention according to claim 7, in claim 5, in the neural network, the extraordinary amount received on the input side includes a peak value, a rise rate, and a pulse half width of an output waveform from the sensor. And at least one of the integral amounts of the waveform amplitude values of the predetermined time width, and the characteristic amount received on the output side is a peak value, a rise rate, and a pulse half of the input waveform to the sensor. A sensor time response correction device is provided which is configured to be at least one of a value width and an integral amount of a waveform amplitude value of a predetermined time width.

Further, according to the invention according to claim 8, in claim 5, in the neural network, the extraordinary amount received on the input side is a discrete form of the amplitude level of the time waveform of the output waveform from the sensor. Value, and the characteristic amount received on the input side is the peak value of the input waveform to the sensor,
A sensor time response correction device is provided which is configured to be at least one of a rise rate, a pulse half width, and an integral amount of a waveform amplitude value of a predetermined time width. .

The integral amount of the waveform amplitude value having the predetermined time width is, for example, an amount obtained by integrating the time waveform with the time interval at which the amplitude level is reduced to 10% relative to the peak value as an integration range. Is mentioned.

Also, a flexible disk, a CDRO
M, a program stored in a storage medium such as a memory card is installed on a hard disk or the like, and the computer executes the program to input one or more characteristic amounts characterizing an output waveform from the sensor. Receiving at the output side, receiving at the output side one or more features that characterize the input waveform to the sensor, configuring a neural network to represent the relationship between input and output; One or more features that characterize one or more known input waveforms, as teacher data to the output side of the neural network, one corresponding to the one or more known input waveforms Alternatively, one or more features that characterize a plurality of known time-response output waveforms may be used as teacher data to the input side of the neural network. Performing a network optimization process; and, for the neural network after the optimization process, one or more features that characterize a known time response output waveform corresponding to an unknown input waveform to the sensor. Inputting a trace amount and estimating an output of the neural network corresponding to the input as one or more trace amounts characterizing an unknown input waveform to the sensor. It is possible.

[0019]

Embodiments of the present invention will be described below with reference to the drawings. First, a description will be given of two types of measurement systems for measuring the characteristic amount of a response output waveform when an input waveform is applied to a sensor, and then using a supervised learning type neural network to characterize a known response output waveform. A method of estimating one or a plurality of extraordinary amounts that characterize an unknown input waveform to the sensor from one or more extraordinary amounts to be added will be described.

FIGS. 3 to 5 are explanatory diagrams showing a measuring method for measuring a characteristic amount of a response output waveform when an input waveform is given to a current probe 3 as a sensor. FIG. 3 (a)
FIG. 3B shows a block diagram of the entire measurement system, and FIG.

As shown in FIG. 3A, this measuring system
A current probe 3 serving as a sensor, a pulse generator 4 for supplying a current to a conducting wire to give an input waveform to the sensor, and an original input waveform and a response output waveform of the current probe 3 for each of two channels (ch1, ch2). ) And an oscilloscope 2 for measurement. Then, FIG.
As shown in the figure, a connector 5a connected to the oscilloscope 2 and a connector 5a connected to the pulse generator 4
And a conductor 6 for electrically connecting the two connectors and for supplying an input waveform from the pulse generator 4 to the current probe 3 is connected thereto. The conductor 7 is also connected.

In this measuring system, the time waveform of the sensor input signal generated by the pulse generator 4 is directly input to the channel ch1 of the oscilloscope 2 via the conductor 6, and the current flowing through the conductor 6 is detected by the current probe 3. Then, the detection time waveform (sensor response output waveform) is measured on the channel ch2 of the oscilloscope 2, and the original waveform (sensor input signal time waveform) generated by the pulse generator 4 and the detection time waveform (sensor response) detected by the current probe 3 are measured. Output waveform) at the same time.

FIG. 4 shows time waveforms (channel ch1) of three types (input a, input b, input c) of input waveforms given to the current probe 3, and FIG. 5 shows each input waveform, input a, The response waveform for each of the input b and the input c, and the time waveform of the response a, the response b, and the response c are shown. Each input waveform shown in FIG. 4 is standardized by the maximum value of each amplitude level, and each response waveform shown in FIG. 5 is standardized by setting the peak value of the amplitude level of the response b to 1. .

As can be seen by comparing FIGS. 4 and 5, the waveform detected by the current probe 3 does not accurately detect the input waveform, and the electric field pulse is accurately detected using the current probe 3. In order to perform measurement, it is necessary to perform a correction process on the response waveform.

This measurement system makes it possible to grasp the discrete values of the input waveform and the response waveform (sampled amplitude levels of the time waveform at a predetermined time) and characteristic parameters. Here, as the characteristic parameters, the peak value of the time waveform, the rise rate (from 10 (%) of the peak value to 9)
0 (%), the half width of the pulse, and the integral amount of the waveform amplitude value of the predetermined time width (for example, the time interval at which the amplitude level is reduced to 10 (%) with respect to the peak value is the integration range. Is the amount obtained by integrating the time waveform).

Next, FIG. 6 shows a block diagram of another measuring system. This measurement system uses a shield room 10 in which a radio wave absorber 10 is attached to a wall or a ceiling, and an inverted cone antenna 8 for generating an electric field pulse, and a distance of 3 (m) from an installation position of the antenna 8. The receiving antenna 9 (biconical antenna), which is a sensor, set apart
And a pulse generator 4 for supplying a current to the inverted cone antenna 8 to give an input waveform to the sensor. Each of the original input waveform and the response output waveform of the receiving antenna 9 is divided into two channels (ch1, ch2). And an oscilloscope 2 for measuring.

It is known that the inverted cone antenna 8 having an inverted conical shape has a characteristic capable of radiating an electric field pulse waveform ("Masao Masugi, et al .:
“Measurement and Analysis of Electromagnetic Pulse Associated with Indirect ESD”, IEICE Transactions B-II, No. 9, pp. 647-
654, 1992. In the example shown in FIG. 6, the current pulse wave generated by the pulse generator 4 is radiated into the space as an electric field pulse wave via the inverted cone antenna 8, and the receiving antenna 9 The emitted electric field pulse wave is detected.

In this measuring system, the time waveform of the sensor input signal generated by the pulse generator 4 is directly input to the channel ch1 of the oscilloscope 2, and the electric field pulse wave is detected by the receiving antenna 9, and the channel of the oscilloscope 2 is detected. The detection time waveform (sensor response output waveform) is measured at ch2, and the original waveform (time waveform of the sensor input signal) generated by the pulse generator 4 and the detection time waveform (sensor response output waveform) by the receiving antenna 9 can be measured simultaneously. .

Although the measurement system shown in FIG. 8 shows the case where the radio wave absorber 10 is attached to the wall or ceiling of the shield room, a measurement system where the radio wave absorber 10 is not attached may be considered. In order to generate an electric field pulse wave, it is conceivable to use another type of antenna other than the inverted cone antenna 8.

FIG. 1 described above shows an original time waveform (channel ch) generated by the pulse generator 4 in order to apply an electric field pulse wave as an input signal to the receiving antenna 9.
FIG. 2 described above shows a time waveform of a response waveform corresponding to this waveform.

As can be seen by comparing FIGS. 1 and 2, the waveform detected by the receiving antenna 9 does not accurately detect the input waveform, and the electric field pulse is accurately detected using the receiving antenna 9. In order to perform measurement, it is necessary to perform a correction process on the response waveform.

Note that this measurement system also makes it possible to grasp the discrete values of the input waveform and the response waveform (sampled amplitude levels of the time waveform at a predetermined time) and characteristic parameters. Here, as the characteristic parameters, the peak value of the time waveform, the rise rate (time from 10 (%) to 90 (%) of the peak value), the half-width of the pulse, and the waveform amplitude value of the predetermined time width (For example, the amount obtained by integrating the time waveform with the time interval at which the amplitude level is reduced to 10% with respect to the peak value as an integration range).

Using the measurement system described above, it is possible to grasp the characteristic values of the waveform such as the discrete values of the input waveform and the response waveform and the characteristic parameters. Now, next, using a supervised learning type neural network, one or more characteristic amounts (discrete values, characteristic parameters) characterizing the known response output waveform are calculated.
A method for estimating one or more characteristic amounts (discrete values, characteristic parameters) characterizing an unknown input waveform to a sensor will be described. In order to facilitate understanding, first, a supervised learning type neural circuit is used. An outline of the network will be described.

FIG. 7 is an explanatory diagram for explaining an outline of operation of the neural network 1 of the supervised learning type. Neural network 1
Has an input side and an output side on its left and right sides, respectively, and has a function of expressing an input / output relationship between input and output. In general, neural networks can be roughly classified into supervised models that use teacher data for optimization, unsupervised models that use self-organization methods that do not use teacher data, and further, mixed models of both. it can.

In the present invention, it is assumed that a neural network of a supervised model is used, and the operation of this type of neural network can be considered in two steps as shown in FIG. . As shown in FIG. 7 (a), in the first stage, a plurality of pairs of teacher data prepared in advance are given to each of the input side and the output side of the neural network 1 so that the units in the neural network 1 Optimization processing (learning) is performed to determine the optimal coupling coefficient (synapse load) of.

Next, as shown in FIG. 7B, in the second stage, known input data corresponding to the estimated data is provided to the input side, and unknown estimated data is output from the output side.
In this way, by giving a plurality of pairs of teacher data to the neural network 1 and performing an optimization process for determining an optimum coupling coefficient, unknown data can be estimated, and various nonlinear system identification problems can be obtained. It can be applied to

FIG. 8 shows a specific example of the neural network 1, and the neural network 1 has an input layer, an intermediate layer, and an output layer including a plurality of units.
Now, the suffixes representing the respective representative units of the input layer, the intermediate layer, and the output layer are represented by “i”, “j”, “k”.
And then, the input layer, "z _i" outputs from each of the intermediate layer and output layer, "y _i", and "o _k", further, "v _ji" the coupling coefficient between the input layer and the intermediate layer, The coupling coefficient between the input layer and the intermediate layer is “w _kj ”.

Here, the outline of the procedure for optimizing the neural network 1 shown in FIG. 8 will be described as follows (1) to (5). (1) Initialize the model. Specifically, the initial value of the coupling coefficient between the layers is set by a random number. The number of intermediate layers and the number of units of the input layer and the output layer can be arbitrarily determined.

(2) Input an input teacher vector to the input layer. The input teacher vector (x) is input to the input layer, and an output value is derived from the sum of the coupling coefficients. The output of each layer is the sigmoid function f (x) = 1 / ｛1 + exp (−α
x)｝ (0 ≦ f ≦ 1, α is a sigmoid function weighting coefficient). In this case, the intermediate output y _j and the final output o _k is determined by the following equation.

The intermediate output y _j = f (Σv _ji z _i ) (Σ is
All j, indicating the sum of i), output = o _k = f
(Σw _kj y _j ) (Σ indicates the sum of all k, j).

(3) Output teacher vector to output layer (d)
Enter That is, the output teacher vector corresponding to the input teacher vector is input to the output layer. (4) Learning by weighting.

The back propagation method, using known procedures, such as Kalman filter solution, based on the value of the sum of the error in the output o _k and the output teacher vector from the output layer (d), the coupling coefficient between layers And derive a new coupling coefficient. This corresponds to execution of the optimization processing. (5) Repeat. The value of the sum of errors of the output o _k and the output teacher vector from the output layer (d) is reached or a preset convergence value, or until a learning preset number of times (2) to (4) Repeat the process.

By performing the above processing, the coupling coefficient between the layers can be optimized, and the optimization processing of the neural network is performed. Next, a device according to a specific embodiment of the present invention and its operation will be described.

FIG. 9 shows a sensor time response correction device according to an embodiment of the present invention.
An input unit 30 for inputting necessary data, a display unit 40 for displaying necessary data, a storage unit 50 for storing teacher data and the like, and a control unit 20 for controlling operations of the neural network 1 and each unit. have.

The input unit 30 can be realized by an input device such as a mouse and a keyboard.
The storage unit 30 can be realized by a storage device such as a hard disk, and the neural network 1 and the control unit 20 function as a storage medium such as a ROM containing an operation program and a work area. This device can be realized by a single computer system such as a personal computer or a work station because the device can be realized by an electronic device such as a RAM that executes an operation program.

As shown in FIG. 10, the storage unit 50 is provided with a teacher data file 55 for storing teacher data. The teacher data file 55 includes data measured by the measurement system described above, specifically, discrete values of the input waveform f _i (t), discrete values of the response waveform g _i (t), and input waveforms. characteristic parameter of f _i (t), response waveform g
The feature parameters of _i (t) can be stored. In addition, although not shown, actually, a plurality of types of input and
A plurality of pairs of teacher data for the response waveform are stored.

The neural network 1 includes an input layer 11 having N (≧ 1) units, an intermediate layer 12 having K (≧ 1) units, and M (≧ 1) units. A three-layered neural network having an output layer 13 comprising: an N input and an M output, wherein each of the N inputs and the M outputs characterizes one or more of an output waveform from the sensor and an input waveform. (N or M) special amounts.

Here, the characteristic amount is a discrete value of the amplitude level of the time waveform, or the peak value, the rise rate,
It is at least one of a pulse half width, an integral amount (feature parameter) of a waveform amplitude value of a predetermined time width, and the like.

Then, one of the discrete values or one feature parameter is assigned to the input layer 11 and the output layer 13
Two units correspond. Therefore, for example, when ten types of discrete values are given to the input layer 11,
Is "10", and the input layer 1
When two types of characteristic parameters (for example, a peak value and a rise rate) are given to 1, the number of units constituting the input layer 11 is set to “2”.

The optimization of the neural network 1 is performed as described with reference to FIG. Specifically, one or more extraordinary amounts characterizing the output waveform from the sensor are used as input training data, and one or more extraordinary amounts characterizing the input waveform to the sensor are used as output teacher data. To the neural network 1 so as to adjust the coupling coefficient using a known processing procedure such as a back-propagation method or a Kalman filter solution so that the output of the neural network 1 becomes close to the output teacher data. It has become. In the present device, a three-layer neural network is shown, but it goes without saying that a neural network having two or more intermediate layers or a neural network having no intermediate layer may be used.

Next, the operation of the present apparatus will be described with reference to the flowchart of FIG. First, in step S1100, in the measurement system, a known input waveform f
_i (t) (i = 1, 2,...) is determined. For example, one or more types of time waveforms of the sensor input signal generated by the pulse generator 4 are determined. At this time, each input waveform f
The feature amount of _i (t) is grasped in advance.

Next, in step S1102, the input
Waveform f_iResponse waveform g to (t) _iMeasure (t)
You. At this time, the response waveform g_iBy grasping the feature of (t)
deep. Next, in step S1104, the input waveform f
_i(T) and response waveform g_i(T) features (discrete values,
When the (characteristic parameter) is input via the input unit 30,
By the control operation of the control unit 20, these known features are
It is stored in the teacher data file 55 in the storage unit 50.

Next, the processing in steps S1106 and S1108 will be described. In step S1106, the control unit 20 determines the input waveform f in the teacher data file 55.
The discrete value of _i (t) and the discrete value of the response waveform g _i (t) are given to the output layer 13 and the input layer 11 of the neural network 1, respectively.
The optimization processing of the neural network 1 is performed by the method described above. A plurality of pairs of discrete values of the input waveform f _i (t) and discrete values of the response waveform g _i (t) are stored in the teacher data file 55.
The control unit 20 supplies the plurality of pairs of data to the neural network 1 so as to reliably perform the optimization processing.

Then, in step S1108, a time response output waveform g 'for the unknown input waveform f' (t) is obtained.
When the discrete values of (t) are measured and the measurement results are provided via the input unit 30, the control unit 20 provides these discrete values to each unit of the input layer 11. As a result, a discrete value of the unknown input waveform f ′ (t) is output from each unit of the output layer 13, and the control unit 20 acquires this and displays it on the display unit 40.

As described above, steps S1100, S110
By the processes of 2, 1104, 1106, and 1108, the discrete value of the unknown input waveform is obtained from the discrete value of the known time response output waveform output from the sensor.

Next, the processing in steps S1110 and S1112 will be described. In step S1110, the control unit 20 determines the input waveform f in the teacher data file 55.
The characteristic parameter of _i (t) and the characteristic parameter of the response waveform g _i (t) are given to the output layer 13 and the input layer 11 of the neural network 1, respectively, and the neural network 1
Optimization processing. Note that a plurality of pairs of input waveforms f
The characteristic parameter of _i (t) and the characteristic parameter of response waveform g _i (t) are stored in the teacher data file 55, and the control unit 20 stores the plurality of pairs of data in the neural network 1
To ensure that the optimization process is performed.

Then, in step S1112, a time response output waveform g 'for the unknown input waveform f' (t) is obtained.
When the characteristic parameters of (t) are measured and the measurement result is provided via the input unit 30, the control unit 20 provides these characteristic parameters to each unit of the input layer 11. As a result, since the characteristic parameters of the unknown input waveform f ′ (t) are output from each unit of the output layer 13, the control unit 20 acquires this and displays it on the display unit 40.

As described above, steps S1100, S110
By the processes of 2, 1104, 1110, and 1112, the characteristic parameters of the unknown input waveform can be obtained from the characteristic parameters of the known time response output waveform output from the sensor.

Next, the processing in steps S1114 and S1116 will be described. In step S1114, the control unit 20 determines the input waveform f in the teacher data file 55.
_i The discrete values of the characteristic parameters and the response waveform g _i (t) of (t), respectively, the output layer 13 of the neural network 1, and supplied to the input layer 11, in the manner described above, the neural network 1 Perform optimization processing. A plurality of pairs of the characteristic parameters of the input waveform f _i (t) and the discrete values of the response waveform g _i (t) are stored in the teacher data file 55, and the control unit 20 transmits the plurality of pairs of data. Is given to the neural network 1 to ensure that the optimization process is performed.

Then, in step S1116, a time response output waveform g 'for the unknown input waveform f' (t) is obtained.
When the discrete values of (t) are measured and the measurement results are provided via the input unit 30, the control unit 20 provides these discrete values to each unit of the input layer 11. As a result, since the characteristic parameters of the unknown input waveform f ′ (t) are output from each unit of the output layer 13, the control unit 20 acquires this and displays it on the display unit 40.

As described above, steps S1100, S110
By the processes of 2, 1104, 1114, and 1116, the characteristic parameters of the unknown input waveform can be obtained from the discrete values of the known time response output waveform output from the sensor. FIGS. 12A and 12B show specific teacher data,
That is, discrete values used as teacher data when learning the neural network 1 are shown. Three types of circles shown in FIG. 12A indicate three types of input waveforms (input a) shown in FIG. ,
b, c) discrete values (solid lines are actually measured continuous values), FIG.
The three types of circles shown in (b) correspond to the discrete values (solid lines indicate actually measured continuous values) of the three types of response waveforms (responses a, b, and c) shown in FIG.

The discrete values of the three types of input / output data pairs (input a and response a, input b and response b, and input c and response c) shown in FIG. 12 are stored in the hierarchical neural network 1 shown in FIG. When used as teacher data for learning, the control unit 20 shifts the response waveform in the positive direction so that the amplitude value of the response waveform in FIG. In consideration of the correspondence between the waveform and the discrete value of the response waveform, the teacher data file 55 storing the data of the discrete value may be stored so as to be supplied to the neural network 1.

In the example shown in FIG. 12, ten discrete value data are given to the input layer 11 of the neural network 1 and 12 discrete value data are given to the output layer 13 of the neural network 1 as an example of the teacher data. Given data, learning (optimization processing) is performed.

As described in the flowchart of FIG. 11, in addition to the discrete values, the characteristic parameters (at least one of the peak value, the rise rate, the pulse half-width, and the integral amount of the specific time width) are set. Since the optimization process may be applied to the neural network 1, the learning teacher data pair provided to the input layer 11 and the output layer 13 may be in the form
“Discrete value, discrete value”, “feature parameter, feature parameter”, and “discrete value, feature parameter” can be considered.

Next, a result of optimizing the neural network 1 of the apparatus shown in FIG. 9 and estimating an unknown signal by this apparatus will be shown. The neural network 1 uses the discrete value pairs shown in FIG. 12 as teacher data, sets the number of units of the input layer, the hidden layer (the number of hidden layers is 1) and the number of units of the output layer to 12, 4, and 10, and solves the Kalman filter. , And the weighting coefficient alpha in the sigmoid function was set to 1, and the number of learnings was set to 1000 (times).

FIG. 13 is an example of a response waveform of the current probe 3 used for evaluating the estimation ability of the neural network 1 that has learned using the teacher data shown in FIG. In FIG. 13, three types of waveforms (response waveforms d, e, and f) are shown, the solid line is an actually measured value, and the circles are discrete values obtained by sampling the solid line at predetermined time intervals. The discrete values of these known response waveforms are given to the input layer 11 and the output layer 13 estimates and outputs the discrete values of the unknown input waveform.

FIG. 14 shows an unknown input waveform (input waveform d,
e and f) show the estimation results of the discrete values, and each estimation value (circled) is in good agreement with the actual measurement value shown by the solid line, and high-precision estimation can be realized. Recognize. Incidentally, as a result of evaluating the relative error of the estimated value with respect to the actually measured value (“(measured value−estimated value) / measured value” × 100 (%)), the waveform d was 4.5 (%) and the waveform e was 2.2. (%) And 5.0 (%) for the waveform f, confirming that practically sufficient accuracy can be ensured. Here, FIG. 13 shows a result of correcting the input waveform e using the conventional method. FIG.
In FIG. 3, the solid line is a correction waveform in consideration of the frequency characteristics of the current probe 3, and the dotted line is an actual input waveform. The correction waveform obtained in consideration of the frequency characteristics of the current probe 3 has a deviation from the actual input waveform. It can be seen that the estimation using the neural network according to the embodiment of the present invention is more accurate.

As described above, according to the embodiment of the present invention, the feature amounts (discrete values and feature parameters) of the input waveform and the response waveform of the sensor are used as teacher data, and the output side of the neural network and the input Side to perform optimization processing,
By giving the characteristic amount of the known response waveform to the input side of the neural network, an estimated value of the characteristic amount of the unknown response waveform was obtained from the output side of the neural network, and the measurement distortion error caused by the sensor was removed. Electromagnetic environment measurement becomes possible.

[0069]

As described above, according to the first or fifth aspect of the present invention, one or a plurality of extraordinary traces characterizing one or a plurality of known input waveforms to the sensor are provided. One or more traces characterizing one or more known time-response output waveforms corresponding to the one or more known input waveforms as teacher data to the output side of the neural network, Since the neural network optimization process is performed as teacher data to the input side of the network, the known time response output waveform corresponding to the unknown input waveform to the sensor is applied to the neural network after the optimization process. When one or more features that characterize are input, the output of the neural network corresponding to the input can be estimated as one or more features that characterize the unknown input waveform to the sensor. .

According to the second or sixth aspect of the present invention, one or more characteristic amounts of the input waveform to the sensor are discrete values of the amplitude level of the time waveform of the input waveform. In addition, one or more features that characterize the output waveform from the sensor, the output waveform, since the discrete value of the amplitude level of the time waveform, from the discrete value of the known time response output waveform, A discrete value of the unknown input waveform is obtained.

Further, according to the third or seventh aspect of the present invention, one or more characteristic amounts of the input waveform to the sensor include a peak value, a rise rate, and a pulse half width of the input waveform. , And at least one or more of the integrals of the waveform amplitude values of the predetermined time width, and one or more characteristic amounts that characterize the output waveform from the sensor have a peak value of the output waveform. , Rise rate, pulse half width, and at least one of the integrals of the waveform amplitude values of the predetermined time width, the characteristics of the unknown input waveform are calculated from the characteristic parameters of the known time response output waveform. The parameters are obtained.

Further, according to the invention according to claim 4 or claim 8, one or more characteristic amounts that characterize the input waveform to the sensor include a peak value of the input waveform,
Rise rate, pulse half width, and at least one or more of the integral amount of the waveform amplitude value of the time width defined in advance, one or more features that characterize the output waveform from the sensor, Since the output waveform is a discrete value of the amplitude level of the time waveform, the characteristic parameter of the unknown input waveform can be obtained from the discrete value of the known time response output waveform.

[Brief description of the drawings]

FIG. 1 is an explanatory diagram of an example of a pulse waveform that is an electromagnetic interference wave.

FIG. 2 is an explanatory diagram of a measured waveform in which a pulse waveform is received by a biconical antenna.

FIG. 3 is an explanatory diagram showing a measurement technique for measuring a characteristic amount of a response output waveform when an input waveform is given to a current probe 3 (sensor).

FIG. 4 is an explanatory diagram showing a measurement method for measuring a characteristic amount of a response output waveform when an input waveform is given to a current probe 3 (sensor).

FIG. 5 is an explanatory diagram showing a measuring method for measuring a characteristic amount of a response output waveform when an input waveform is given to a current probe 3 (sensor).

FIG. 6 is an explanatory diagram of another measurement system.

FIG. 7 is an explanatory diagram of the operation of the neural network.

FIG. 8 is an explanatory diagram of a configuration example of a hierarchical neural network.

FIG. 9 is a block diagram of an apparatus according to an embodiment of the present invention.

FIG. 10 is an explanatory diagram of a teacher data file.

FIG. 11 is a flowchart for explaining the operation of the apparatus.

FIG. 12 is an explanatory diagram of a specific example of teacher data.

FIG. 13 is an explanatory diagram for explaining an estimation result according to the embodiment of the present invention.

FIG. 14 is an explanatory diagram for explaining an estimation result according to the embodiment of the present invention.

FIG. 15 is an explanatory diagram of a correction processing result by a conventional method.

[Description of Signs] 1 ... neural network, 2 ... oscilloscope, 3 ... current probe, 4 ... pulse generator, 5a ... connector, 5b ...
Connector, 6 conducting wire, 7 conducting wire, 8 inverted cone antenna, 9 receiving antenna, 10 radio wave absorber, 11 input layer, 12 intermediate layer, 13 output layer, 2
0: control unit, 30: input unit, 40: display unit, 50: storage unit, 55: teacher data file

Claims (8)

[Claims] 1. One or more features that characterize an unknown input waveform to an EMI evaluation sensor (“sensor”) from one or more features that characterize a known time response output waveform. A method for estimating a characteristic amount, wherein one or more characteristic amounts characterizing an output waveform from the sensor are received on an input side, and one or more characteristic amounts characterizing an input waveform to the sensor are determined. Accepted on the output side, a neural network is configured to represent the relationship between input and output, and one or more features that characterize one or more known input waveforms to the sensor, As the teacher data to the output side of the neural network, one or more features that characterize one or more known time response output waveforms corresponding to the one or more known input waveforms, As training data to the input side of the neural network, the neural network performs optimization processing of the neural network, and the neural network after the optimization processing corresponds to a known input waveform corresponding to an unknown input waveform to the sensor. One or more features that characterize the time response output waveform are input, and the output of the neural network corresponding to the input is one or more features that characterize an unknown input waveform to the sensor. A sensor time response correction method characterized by estimating. 2. The output waveform from the sensor according to claim 1, wherein the one or more characteristic amounts characterizing the input waveform to the sensor are discrete values of the amplitude level of a time waveform of the input waveform. The sensor time response correction method, characterized in that one or a plurality of extraordinary amounts characterizing the above is a discrete value of an amplitude level of a time waveform of the output waveform. 3. The method according to claim 1, wherein the one or more characteristic amounts characterizing the input waveform to the sensor include a peak value, a rise rate, a pulse half width, and a predetermined time width of the input waveform. At least one or more of the integration amounts of the waveform amplitude values, and one or a plurality of extraordinary amounts characterizing the output waveform from the sensor, the peak value of the output waveform, the rise rate,
A sensor time response correction method, which is at least one of a pulse half width and an integral amount of a waveform amplitude value having a predetermined time width. 4. The method according to claim 1, wherein the one or more characteristic amounts characterizing the input waveform to the sensor include a peak value, a rise rate, a pulse half-width, and a predetermined time width of the input waveform. At least one or more of the integral amounts of the waveform amplitude values, and one or more characteristic amounts characterizing the output waveform from the sensor are discrete values of the amplitude level of the time waveform of the output waveform. A sensor time response correction method. 5. One or more features that characterize an unknown input waveform to an EMI evaluation sensor (“sensor”) from one or more features that characterize a known time response output waveform. An apparatus for estimating a characteristic amount, wherein one or more characteristic amounts characterizing an output waveform from the sensor are received at an input side, and one or more characteristic amounts characterizing an input waveform to the sensor are obtained. A neural network configured to accept on the output side and represent the relationship between input and output, and one or more first features that characterize one or more known input waveforms to the sensor; And storage means for storing one or more second features that characterize one or more known time response output waveforms corresponding to the one or more known input waveforms; And a control unit for performing a correction process of a response output from the control unit. The control unit acquires the first and second extraordinary amounts stored in the storage unit, and obtains each extraordinary amount. The output data and the input data of the neural network are given to each of the output side and the input side as teacher data, and the neural network performs the optimization process. For the neural network after the optimization process, an unknown input waveform to the sensor is provided. One or more features that characterize a known time response output waveform corresponding to the input, and the output of the neural network corresponding to the input, one or more features that characterize an unknown input waveform to the sensor. A sensor time response correction device for estimating and outputting as a plurality of extraordinary amounts. 6. The neural network according to claim 5, wherein the extraordinary amount received on the input side is a discrete value of the amplitude level of a time waveform of the output waveform from the sensor, and is received on the output side. A sensor time response correction device, wherein the amount is configured to be a discrete value of an amplitude level of a time waveform of an input waveform to the sensor. 7. The neural network according to claim 5, wherein the trace amount received on the input side is a peak value, a rise rate, a pulse half width, and a peak value of an output waveform from the sensor.
At least one of the integration amounts of the waveform amplitude values of the predetermined time width, and the feature amount received on the output side is the peak value of the input waveform to the sensor, the rise rate, the pulse half width, and A sensor time response correction device configured to be at least one of an integration amount of a waveform amplitude value having a predetermined time width. 8. The neural network according to claim 5, wherein the extraordinary amount received on the input side is a discrete value of the amplitude level of the time waveform of the output waveform from the sensor, and is received on the input side. The amount is configured to be at least one of the peak value of the input waveform to the sensor, the rising rate, the pulse half width, and the integral amount of the waveform amplitude value of the predetermined time width. A sensor time response correction device characterized by the above-mentioned.