JP2842473B2 - Signal frequency tracking method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an input signal such as a sound wave .
Using the distribution of the frequency space of the input signal strength that is obtained by performing frequency analysis to, characteristics relate signal frequency tracking method for measuring continuously the frequency of an unknown narrowband signal temporally is there.

[0002]

2. Description of the Related Art FIG. 2 is a functional block diagram of a signal frequency tracking device for implementing a signal frequency tracking method using a conventional neural network. With reference to FIG. 2, theory Aquiraz Ru conventional signal frequency tracking method. A sound wave as an input signal is received by the acoustic sensor 1, and the acoustic signal is subjected to FFT (Fast Fourier Transformation , high speed).
(Fourier transform) device 2. In the FFT device 2,
A discrete Fourier transform is performed every predetermined time S seconds, a signal component for each predetermined analysis width Δf is calculated, and its absolute value is obtained.
(F) is obtained. Then, the signal intensity distribution X (f) is sent to the integrator 10. Also, the signal intensity distribution X for each frequency
By storing (f) for a fixed time in the past by the storage device 3, the signal intensity distribution X in a two-dimensional space of time × frequency
(T, f) is obtained. This signal intensity distribution X (t, f)
Are generated every predetermined time T seconds and sent to the neural networks 4-1 to 4-3.

Each neural network 4-i (i = 1)
3) are learned in advance so as to output a large value if there is a linear pattern in the input. this is,
For example, it can be realized by learning with an error back propagation method (back propagation method) as a teacher signal with 1 as an input for a linear pattern and 0 for a random input. The following reference describes the neural network and the error backpropagation method. Literature; Hideki Aso, "Neural Network Information Processing", (1988), Industrial Books. Here, the degree to which the frequency of a signal changes over time is called a gradient. Neural networks 4-1 to 4-
3 detects signals by sharing the gradients in different ranges. For example, the neural network 4-1 detects a signal having a gradient in the range of -0.3 to -0.1 (Hz / sec).
The neural network 4-3 detects a signal having a gradient of 0.1 to +0.1 (Hz / sec) and a gradient of +0.1 to +0.3 (Hz / sec).

Now, each of the neural networks 4-1 to 4-1
4-3 shows a signal intensity distribution X in the time × frequency space.
From (t, f), a cutout of a rectangular small portion centered on a certain frequency is input. If a narrow band signal having a frequency close to the frequency is present, a linear pattern emerges in the small portion, and the neural network corresponding to the gradient shows a large value output. This is repeated for many frequencies, and the output value is sent to the event detection device 5. The event detection device 5 detects a local maximum point on the frequency axis of the output of each of the neural networks 4-1 to 4-3. If the output of any of the neural networks 4-1 to 4-3 shows a local maximum value at a certain frequency f ^* , and if the value exceeds a predetermined threshold, a narrow frequency having a frequency close to f ^* is obtained. It is determined that a band signal exists. This frequency f ^* is sent to the frequency precision device 20. On the other hand, the integrator 10 temporally integrates the signal strength for each frequency received from the FFT device 2. This result is sent to the frequency measurement device 20.

[0005] The frequency precise measuring device 20 receives the frequency f ^* from the event detecting device 5 and sets the maximum value of the integral value of the signal strength received from the integrating device 10 in a section having a predetermined width around the frequency f ^*. To detect. The frequency indicating the maximum value is used as a precise estimated value of the signal frequency ( this is called a precisely measured frequency) f ^** , and the tracking device 30
Send to The tracking device 30 receives the precise measurement frequency f ^** from the frequency precise measurement device 20 every predetermined time T seconds, and tracks this. As a tracking method, for example, there is a method of inputting a precise measurement frequency f ^** as an observation value to a Kalman filter and estimating a signal frequency. The state estimation using the Kalman filter is described in the following literature. Literature: Tanihagi, "Theory of Digital Signal Processing 3 Estimation and Adaptive Signal Processing", 1986, Corona, P.S. 97-1
14

[0006]

However, in the conventional signal frequency tracking method , as described below, the length of the integration time used in the integrator 10 is appropriately determined for signals having various gradients. There was a problem that it was not possible. 3 (a) and 3 (b) are diagrams showing a problem of the conventional signal frequency tracking method , FIG. 3 (a) shows a case where the signal has no gradient, and FIG. Has a gradient. Figure 3 (a), (b) in the (A) the time × signal intensity distribution on the frequency space, (b) the integration results of the integration time length is short, and (c) if the integration time is long This is the integration result of. As shown in FIG. 3 (a), since the noise has the property that averages more integrating over long time, the effect it is noise short integration time (a) a long integration time than (c) Reduced. That is, the use of a longer integration time results in higher signal frequency estimation accuracy. However, when the signal strength is sufficiently large compared to the noise strength, that is, when the S / N (signal-to-noise ratio) is high, the accuracy cannot be improved even if the integration time is lengthened, and the signal tracking ability is rather deteriorated. However, since the purpose of using a neural network for signal detection and tracking is to catch a weak signal, here S
/ N is considered to be sufficiently low.

On the other hand, FIG. 3 (b), those exemplified for the case where the signal has a slope. In this case, since the frequency of the signal changes with time, if the integration time is long, the peak frequency of the signal strength at the current time and the frequency of the peak of the past signal are shifted by a gradient, so that the current time In the frequency near the frequency of the peak, the peak of the past signal is reflected. Therefore, the localization of the signal is reduced. Therefore, the estimation accuracy of the signal frequency is higher in the case of using the short integration time (a) than in the case of using the long integration time (c). Well, Oite to the conventional method,
Consider how much the integration time should be set in order to cope with both the case where the signal has a gradient and the case where the signal does not have a gradient. If a long integration time is used , the frequency can be accurately obtained for a signal having no gradient. However, the accuracy of a signal having a gradient deteriorates due to a decrease in localization. Conversely, if a shorter integration time is used, the degradation of accuracy can be prevented for signals with gradients, but the accuracy is improved for signals without gradients (although longer integration times will improve). Is suppressed. As mentioned above,
In the conventional signal frequency tracking method, the integration time for precisely measuring the signal frequency must be optimally determined only for a specific gradient signal, and satisfactory accuracy cannot be obtained for a signal having a certain gradient. Was.

[0008]

To solve the previous SL problems SUMMARY OF THE INVENTION The first aspect of the present invention, the signal frequency tracking method
In, a frequency analysis process of performing a frequency analysis on the input signal to obtain a signal intensity distribution, a storage process of accumulating the signal intensity distribution over a certain period of time, and for each frequency analyzed by the frequency analysis process, The signal intensity distribution accumulated by the storage processing of a small portion limited to the vicinity of the frequency is input to a plurality of neural networks, and each of these neural networks is used for signal time.
Responsible for the different gradients frequency against by each of these neural networks, and a pattern detecting process of outputting large numbers when a linear pattern indicated by the signal with a gradient of a particular frequency is present for the time in the input Execute. Then, regarding the outputs of the plurality of neural networks, the presence of a signal is determined by detecting a point that has a maximum point in the output for each frequency and that has a value whose output value exceeds a predetermined threshold. The frequency of the signal with respect to time is changed from the gradient of the neural network in charge of outputting the largest value for the frequency to be used as the rough estimate of the frequency of the signal with the frequency of the local maximum point as the rough estimate of the frequency of the signal. Is executed. Further, for a plurality of predetermined integration times, an integration process for obtaining an integrated value by integrating for each frequency of the signal intensity distribution stored by the storage process, and the integration based on a gradient calculated by the event detection process. An integration time selection process of selecting which integration value is to be adopted among the integration values of a plurality of integration times by the process;
With respect to the integrated value integrated by the integration time adopted by the integration time selection processing, a local maximum point near a rough estimated value of the frequency of the signal detected by the event detection processing is searched for, and the frequency of the signal is calculated. The frequency precise measurement process for obtaining a precise frequency that is a more accurate estimation value, and the frequency precise measurement process from the frequency analysis process are repeated at predetermined time intervals, and the precise frequency obtained by the frequency precise measurement process is calculated based on the time. And a tracking process for sequentially tracking.

A second invention relates to a signal frequency tracking method.
A frequency analysis process of performing a frequency analysis on an input signal to obtain a signal intensity distribution, a storage process of accumulating the signal intensity distribution over a certain period of time, and for each frequency analyzed by the frequency analysis process, A signal intensity distribution accumulated by the storage processing of a small portion limited to the vicinity of the frequency is input to a plurality of neural networks, and each of the neural networks is responsible for a different gradient of the frequency of the signal with respect to time. Each of these neural networks executes a pattern detection process of outputting a large numerical value when a linear pattern represented by a signal having a specific frequency gradient with respect to time during input exists. Then, regarding the outputs of the plurality of neural networks, the presence of a signal is determined by detecting a point that has a maximum point in the output for each frequency and that has a value whose output value exceeds a predetermined threshold. The frequency of such a local maximum is used as a rough estimate of the frequency of the signal, and the maximum value for the frequency used as the estimate is output from the gradient of the neural network in charge of the signal.
For an event detection process of calculating a frequency gradient with respect to time, and for a plurality of predetermined frequency gradients with respect to time,
For each frequency of the signal strength distribution, based on the signal strength of the past frequency shifted from the current frequency by the amount of change in the frequency of the gradient from the past time to the current time, integrating at a predetermined time And an integration process for obtaining an integral value. Further, an integration gradient selection process of selecting a gradient closest to an estimated value of the gradient by the event detection process from among a plurality of gradients used in the integration process and employing an integration value by the integration process corresponding to the selected gradient. Searching for a local maximum point near a rough estimate of the frequency of the signal detected by the event detection process with respect to the integration value selected by the integration gradient selection process, and obtaining a more accurate estimate of the frequency of the signal. A frequency precise measurement process for obtaining a precise frequency, and a frequency tracking process for repeating the frequency precise measurement process from the frequency analysis process at predetermined time intervals, and temporally tracking the precise frequency obtained by the frequency precise measurement process. And processing.

[0010]

According to the first aspect of the present invention, since the signal frequency tracking method is configured as described above, a signal having a gradient of a specific frequency is input during input by a neural network which is responsible for a plurality of different signal gradients. A large numerical value is output when the linear pattern shown exists. The event detection process detects the presence of a signal by detecting the point of the output of each of the plurality of neural networks which is the maximum point in the output for each frequency and whose output value has a value exceeding a predetermined threshold value. It is determined that the frequency of the local maximum is used as a rough estimate of the frequency of the signal, and the frequency of the signal is calculated from the gradient of the neural network that outputs the largest value with respect to the frequency used as the estimate. Calculate the gradient. By the integration process, for a plurality of predetermined integration times, an integration value is obtained by integrating for each signal intensity distribution frequency stored by the storage process, and based on the gradient calculated by the event detection process by the integration time selection process, By selecting which one of the integration values obtained in a plurality of integration times by the integration process is to be adopted, the integration value integrated at the optimum integration time considering the influence of noise and the locality of the signal Is selected.

According to the second aspect of the present invention, by the integration process, for each of the plurality of frequency gradients, for each frequency of the signal intensity distribution, the current amount of change in frequency corresponding to the gradient from the past time to the current time is used. An integrated value is obtained in an integration time of a predetermined time based on the signal intensity of the past frequency shifted from the frequency. As a result, the peaks of the current and past signals can be matched, and signal localization is maintained. Then, according to the gradient of the signal calculated by the tracking processing, an integrated value integrated with the optimal gradient is selected. Therefore, the above problem can be solved.

[0012]

EXAMPLES First Embodiment FIG. 1 is a functional block diagram of a signal frequency tracking apparatus for carrying out a signal frequency tracking method of a first embodiment of the present invention, common to elements in Figure 2 the elements that have been with common reference numerals are <br/>. The difference between the signal frequency tracking method of the first embodiment and the conventional method is that the output of the integrator integrated at the optimum integration time according to the gradient of the signal from the outputs of the plurality of integrators having different integration times. Is determined by the frequency measuring device. This signal frequency tracking device has an acoustic sensor 1 that receives an input signal. The output side of the acoustic sensor 1, FF T equipment 2 is connected. On the output side of the FFT device 2, a storage device 3
And an integrator 50-1 for integrating over a long time, and an integrator 50-2 for integrating over a short time. At the output side of the storage device 3, neural networks 4-1 to 4- corresponding to the frequency gradient with respect to time are provided.
3 are connected. Each neural network 4-1
The event detection device 5 is connected to the output side of ４−4-3. Event detection device 5及 beauty integrator 50-1, 50
The -2 output side is connected to an integration time selection device 51. The output side of the event detection device 5及 beauty integration time selection unit 51, the frequency Seihaka device 20 is connected. On the output side of the frequency measuring device 20, a tracking device 30 is provided.
Is connected. Next , with reference to FIG. 1 , among the processing (I) to (VIII) of the signal frequency tracking method of the first embodiment .
The contents will be described.

(I) Frequency Analysis Processing An acoustic signal is received by the acoustic sensor 1 and sent to the FFT device 2. In the FFT apparatus 2, at a predetermined time interval (S seconds), a predetermined frequency analysis width (Δf Hz)
Performs a fast discrete Fourier transform, and obtains an absolute value for each frequency to obtain a signal intensity distribution. This is expressed as ｛X (f) | f
= F ₁ , f ₂ , ..., f _N }.

(II) Storage Process The storage device 3 stores a predetermined number (r times) of past times of the signal intensity distribution received from the FFT device 2. Thereby, the signal intensity distribution on the time × frequency space 上 の X (t,
f) | t = t ₀ − (r−1) S,..., t ₀ −S, t ₀ ;
f = f ₁ , f ₂ ..., f _N } are obtained. From this signal intensity distribution, each frequency f _n = f _1, ..., for f _N, f _n
と す る X (t, f) | t = t ₀ − (r−
1) S,..., T ₀ −S, f = f _n−ω ,..., F _{n + ω} ｝ (here, the size of the small section is set to 2ω + 1 times the frequency analysis width) and the neural network 4-1 Send to ~ 4-3.

(III) Pattern Detection Processing The neural networks 4-1 to 4-3 share input signals having different gradients. For example, the neural network 4-1 has a gradient of -0.3 to -0.1 (H
(z / sec) in advance. Similarly, the neural network 4-2
Is -0.1 to +0.1, neural network 4-3
Is responsible for signals with a gradient of +0.1 to +0.3. As a learning method, for example, a back propagation method is used to output 1 if a signal of a gradient in charge is being input (if there is a linear pattern in the input), and 0 if not. To learn. Small sections of the signal intensity distribution on the time × frequency space sent from the storage device 3 are input to these neural networks 4-1 to 4-3. In the neural networks 4-1 to 4-3, when a signal is present in the input and the gradient is within the range of the neural network, the output value shows a large value. If, Oh
If a narrow band signal having a frequency close to a certain frequency f ^* exists, a linear pattern emerges in the small section, and the neural network corresponding to the gradient shows a large value output. This is repeated for many frequencies, and all of the neural networks 4-1 to 4-1
The output of 4-3 is sent to the event detection device 5.

(IV) Event Detection Processing In the event detection device 5, each neural network 4
A local maximum point on the frequency axis of the output of -1 to 4-3 is detected. If any of the neural networks 4-1 to 4-1
If the output of 4-3 shows a local maximum value for a certain frequency f ^* and the value exceeds a predetermined threshold value, it is determined that there is a narrow band signal having a frequency close to f ^* . This frequency f ^* is set as a rough estimated frequency of the signal, and is sent to the frequency precision device 20. In addition, the gradient of this signal is calculated for each of the neural networks 4-1 to 4-
3 is estimated by comparing the output values. As the method for approximate estimation frequency f ^*, the most significant value searching for neural network 4-k to be output, the interval [df _k-1 ~df k gradients said neural network 4-k is responsible ] (Df _k-1 + d
Let f _k , / 2 be the estimated value of the gradient ψ ₀ . The estimated value ψ _{0 of the} gradient is sent to the integration time selection device 51.

(V) Integral Processing On the other hand, the integrators 50-1 and 50-2 receive the signal strength X (f) for each frequency from the FFT apparatus 2 and temporally integrate each frequency. For example, the integration time U
In order to perform integration in (seconds), the signal intensities for the past v times are added as v = int (U / S) (S is a time interval of fast Fourier transform, int () represents an integer by rounding). . That is, the output ｛Y _U (t,
f) | f = f ₁ ,..., f _N } is calculated by the following equation (1). Y _U (t, f) = X (t, f) + X (t-1, f) +... + X (t− (v−1) S, f) (1) Here, integration time U is Describe the method to determine properly. A signal whose absolute value of the gradient is ψ (Hz / sec) indicates a frequency change of ψ · U (Hz) while the time U elapses. Therefore, ψ · U is compared with 分析 · U <Δf
As far as U increases, noise is suppressed and the measurement accuracy of the signal is improved, but if ψ · U> Δf, the locality of the signal is reduced. That is, the optimal integration time U is Δf
/ Ψ.

[0018] In the first embodiment, charge signals of the gradient of the maximum absolute value 0.1 of the integration time in the integrator 50-1 U ₁ the neural networks 4-2 (Hz / sec) in accordance with U ₁ = Δf / 0.1 (second). Further, since the absolute value of the gradient of the signal handled by the neural networks 4-1 and 4-3 is a maximum of 0.3 (Hz / sec),
The integration time U ₂ in the integrator 50-2 is calculated as U ₂ = Δf /
0.3 (second).

[0019] In (VI) integration time selection process integration time selection unit 51 takes accept a gradient [psi ₀ of the signal detected by the event detection unit 5, the integration time which is suitable for the measurement of the signal with the gradient [psi ₀ select. Specifically, if the gradient ψ ₀ is within the range of responsibility −0.1 to +0.1 (Hz / sec) of the neural network 4-2, the result of the integrator 50-1 is adopted, and so on. If not, the result of the integrator 50-2 is adopted. The adopted signal intensity distribution is sent to the frequency precise measurement device 20.

[0020] (VII) Frequency Seihaka processing frequency Seihaka apparatus at 20 takes accept a rough estimated frequency f ^* of the signal from the event detection unit 5, selected by the integration time selected device 51 signal strength distribution {Y _Ui0 (f) | f =
f _0−ε ,..., f _{0 + ε} ｝ (here, the width of the range is 2ε + 1
In this case, the frequency f ^** giving the maximum value to Y _Ui0 (f) is defined as the precise measurement frequency of the signal. The measured frequency f ^** is sent to the tracking device 30.

(VIII) Tracking Process The tracking device 30 temporally tracks the precise measurement frequency f ^** sent from the frequency precise measurement device 20 . As described above, in the first embodiment, the result obtained by integrating the signal intensity distribution using the long and short integration times is selected and used by the integration time selection device 51. Therefore, a signal having a small gradient requires a long integration time.
The frequency of a signal having a large gradient can be precisely measured based on the result of integration using a short integration time. Therefore,
The frequency can be accurately estimated.

[0022] Second Embodiment FIG. 4 is a functional block diagram of a signal frequency tracking apparatus for carrying out a signal frequency tracking method of a second embodiment of the present invention, in common with the elements in the figures 1 the element that is with common reference numerals are <br/>. The signal frequency tracking method of the second embodiment is
The difference from the first embodiment is that a plurality of integrations with gradients are performed, and the integration result is selected according to the gradient of the detected signal. This signal frequency tracking device has an acoustic sensor 1. On the output side of the acoustic sensor 1,
FFT unit 2 is connected to the further output of the FFT unit 2, a storage device 3及 beauty beveled integrator 60-1～
60-3 are connected. On the output side of the storage device 3,
Neural networks 4-1 to 4-3 according to the frequency gradient with respect to time are connected. The output side of the neural network 4-1 to 4-3, the event detector 5 is connected. The output side of the event detection device 5及 beauty beveled integrator 60-1 to 60-3, the integration slope selection unit 61 is connected. Event detection device 5
The output side of the 及 beauty integral gradient selection unit 61, the frequency Seihaka device 20 is connected. A tracking device 30 is connected to an output side of the frequency precise measurement device 20.

Next, referring to FIG. 4, the signal frequency tracking method of the present second embodiment Ru theory Aquiraz. As in the first embodiment, after the acoustic signal is received by the acoustic sensor 1, F
The fast Fourier transform is performed by the FT device 2, converted into an absolute value, and the storage device 3 stores the signal intensity distribution received from the FFT device 2 for a certain number of past times (r times). Further, similarly to the first embodiment, the presence of a signal is detected by the neural networks 4-1 to 4-3 and the event detection device 5. Then, the following processes (i) to
(Iv) is performed.

(I) Integral Processing The integrators with gradients 60-1 to 60-3 receive the signal intensity X (f) for each frequency from the FFT device 2 and perform temporal integration for each frequency. As the integration time U,
Using a long integration time, for example, U ₁ described in the first embodiment
(Seconds). Integrators 60-1 to 60 with gradients
-3 is responsible for different gradients, and integrates past signal strengths while shifting them on the frequency axis by the gradients in charge. For example, in the integrator with a gradient in charge of the signal of the gradient φ (Hz / sec), the output 時刻 Z _φ (t, f) at time t
| F = f ₁ , f ₂ ,..., F _N } is calculated by the following equation (2). Zφ (t, f) = X (t, f) + X (tS, f−Sφ) +... + X (t− (v−1) S, f− (v−1) Sφ) (2) here The time interval S (seconds) for performing the Fourier transform and the cumulative number v = U / S are the same as in the first embodiment. In the second embodiment, the gradient is set to three sections [−0.3, −−] using three neural networks 4-1 to 4-3.
0.1], [-0.1, +0.1] , [+0.1, +
0.3], and the three gradient integrators 60-1 to 60-3 respectively use the center of the three sections -0.2, 0.0, and +0.2 (Hz). / Sec) is used as the gradient φ.

(Ii) Integration Time Gradient Selection Process The integration gradient selection device 61 receives the gradient φ ₀ of the signal from the event detection device 5 and performs integration with the gradient closest to the gradient φ ₀ to obtain a gradient integrated. Apparatus 60-1 to 60
Select from -3. For example, if the assigned range -0.3 to 0.1 in this gradient phi ₀ neural networks 4-1
(Hz / sec), the result of integration with a slope of -0.2 (Hz / sec) is adopted. The adopted signal strength distribution is sent to the frequency precise measurement device 20.

The (iii) The the frequency Seihaka processing frequency Seihaka device 20, as in the first embodiment, taken only <br/> receive a rough estimate frequency f ^* of the signal from the event detection unit 5, the integration time The frequency f ^** that gives the maximum value to the signal intensity distribution selected by the gradient selection device 61 is defined as the precise measurement frequency of the signal. The measurement frequency f ^** is the tracking device 30
Sent to

(Iv) Tracking Process The tracking device 30 temporally tracks the precise measurement frequency f ^** sent from the frequency precise measurement device 20 . As described above, according to this second embodiment, a long with the integration time integrating the signal intensity distribution while shifting by the slope on the frequency axis, so selected and used that results in the integration gradient selection unit 61, Even for a signal having a gradient, a long integration time can be used without deteriorating the limit of the signal frequency. Therefore, the precise measurement frequency can be obtained with high accuracy.

Third Embodiment FIG. 5 is a functional block diagram of a signal frequency tracking device for performing a signal frequency tracking method according to a third embodiment of the present invention, which is common to the elements in FIG. the element that is with common reference numerals are <br/>. The signal frequency tracking method of the third embodiment is
The difference from the first embodiment is that the signal intensity distribution is integrated with a long integration time and a short integration time, respectively, and the integration result is selected based on the frequency gradient predicted by the tracking device 30. This signal frequency tracking device has an acoustic sensor 1. An FFT device 2 is connected to an output side of the acoustic sensor 1. The output side of the FFT device 2 is connected to the storage device 3 , an integrator 50-1 for integrating over a long time, and an integrator 50-2 for integrating over a short time. Neural networks 4-1 to 4-3 according to the frequency gradient with respect to time are connected to the output side of the storage device 3. The output side of the neural network 4-1 to 4-3, the event detector 5 is connected. The tracking device 30 is connected to the output side of the event detection device 5. Tracking device 3 0及 beauty integrator 50-1, 50
The -2 output side is connected to an integration time selection device 51. The output side of the integration time selection device 51 is connected to the frequency precise measurement device 20. Next, referring to FIG. 5,
The signal frequency tracking method of the third embodiment Ru theory Aquiraz. First
Similarly to the embodiment, after the acoustic signal is received by the acoustic sensor 1, the presence of the signal is detected by the event detecting device 5, and the signals are integrated by the integrating devices 50-1 and 50-2 using the long and short integration time. Integrate the respective intensity distributions. That
Thereafter, the following processes (a) to (c) are performed.

(A) Tracking Process The tracking device 30 estimates the frequency f and the gradient ψ _{1 of the} signal using a Kalman filter that uses the approximate frequency of the signal received from the event detection device 5 as an observed value. The estimated value f of the obtained frequency is
And the gradient estimated value ψ ₁ is sent to the integration time selection device 51.

[0030] In (b) integration time selection process integration time selection unit 51, the estimated value [psi ₁ gradient from the tracking device 30 takes accept, select the integration time which is suitable for the measurement of the signal with the gradient [psi ₁ , And sends the signal intensity distribution to the frequency measurement device 20.

[0031] In (c) Frequency Seihaka processing frequency Seihaka device 20, along with receiving the estimated value f of the frequency from the tracking device 30 takes accept the selected signal intensity distribution integration time selection unit 51, a first As in the embodiment , the frequency at which the signal intensity distribution has the maximum value in a predetermined range centered on the estimated value of the frequency f is defined as the precise measurement frequency. As described above, according to the third embodiment, even in the case where frequency fine measurement is performed after tracking, as in the first embodiment, the signal intensity distribution is calculated using long and short integration times. Since the integration result can be selected and used for precise measurement, the frequency at each time point of the signal being tracked can be obtained with high accuracy. The present invention is not limited to the above-described embodiment, and various modifications are possible. For example, there are the following modifications.

(1) Each block other than the acoustic sensor 1 in FIGS . 1 , 4 and 5 may be realized by an electric circuit, or a part or all of them may be realized by a computer program. Is also good. (2) In the first to third embodiments, the input signal is a sound wave, but the present invention is not limited to this. For example, when the input signal electric wave, FIG. 1, functions similarly by using a receiving antenna instead of the acoustic sensors 1 in any of FIGS. (3) The number of neural networks 4-1 to 4-3 shown in the first to third embodiments is an example. In the first embodiment, three or more neural networks (“positive gradient” and “negative gradient” in two neural networks)
If the absolute value does not change, the integration time may be constant, and three or more are required when there is a difference between the absolute values of the shared gradients.) The neural network in the second embodiment is Two or more neural networks (because this is effective when there are a plurality of shared gradients), one or more neural networks in the third embodiment (the tracking device
Since the slope from the history of past Ibehento it is estimated by 30, if the neural network is the gradient of the plurality of types is estimated without sharing gradient), or may be any number. (4) the tracking device 30 in FIG. 4 is connected to the output side of the event detection device 5, the output side of the tracking device 30, in the configuration to connect the integrated gradient selection unit 61 and the frequency Seihaka device 20 Is also good. In this case , the integration gradient selection device 61 uses the gradient integration devices 60-1 to 60-1 based on the frequency gradient tracked by the tracking device 30.
The integrated value of 60-3 is selected, and the frequency precise measurement device 20 obtains the precisely measured frequency from the estimated value of the frequency tracked by the tracking device 30 and the integrated values of the gradient integrated devices 60-1 to 60-3. , to the precise measurement frequency and tracking the results of precise measurement, such frequency.

[0033]

As described in detail above, according to the first and third aspects, the integration time is selected according to the gradient of the signal frequency, so that the accuracy of tracking the frequency of the signal is improved. I do. According to the second and fourth aspects of the present invention, since the integration is performed while being shifted according to the gradient of the signal frequency, the accuracy of tracking the frequency of the signal is improved.

[Brief description of the drawings]

FIG. 1 is a functional block diagram of a signal frequency tracking device for implementing a signal frequency tracking method according to a first embodiment of the present invention.

FIG. 2 is a functional block diagram of a signal frequency tracking device for implementing a conventional signal frequency tracking method.

FIG. 3 is a diagram showing a problem of a conventional signal frequency tracking method.

FIG. 4 is a functional block diagram of a signal frequency tracking device for implementing a signal frequency tracking method according to a second embodiment of the present invention.

FIG. 5 is a functional block diagram of a signal frequency tracking device for implementing a signal frequency tracking method according to a third embodiment of the present invention.

[Explanation of symbols]

Reference Signs List 1 acoustic sensor 2 FFT device 3 storage device 4-1 4-2, 4-3 neural network 5 event detecting device 20 frequency measuring device 30 tracking device 50-1, 50-2 integrating device 51 integration time selecting device 60- 1,60-2,60-3 Integrator 61 Integral gradient selector

──────────────────────────────────────────────────続 き Continuation of the front page (56) References JP-A-58-139084 (JP, A) JP-A-2-205790 (JP, A) JP-A-4-276523 (JP, A) JP-A-5-205 273326 (JP, A) (58) Fields investigated (Int. Cl. ⁶ , DB name) G01H 3/08 G01R 23/02 G01R 23/16 G01S 3/80

Claims (4)

(57) [Claims] 1. A frequency analysis process for obtaining a signal intensity distribution by performing frequency analysis on an input signal; a storage process for accumulating the signal intensity distribution over a predetermined time; and each frequency analyzed by the frequency analysis process. Each time, a small portion limited to the vicinity of the frequency is stored by the storage process.
The input signal strength distributions are input to a plurality of neural networks, each of which is responsible for a different gradient of the frequency of the signal with respect to time , and each of these neural networks has a function of time versus time during input .
A pattern detection process of outputting a large numerical value when a linear pattern indicated by a signal having a gradient of a specific frequency is present, and for the outputs of the plurality of neural networks, a maximum point is obtained in the output for each frequency. The presence of a signal is determined by detecting a point having a value whose output value exceeds a predetermined threshold, and the frequency of such a local maximum point is used as a rough estimate of the frequency of the signal, An event detection process for calculating a frequency gradient with respect to the time of the signal from a gradient in charge of the neural network that outputs the largest value with respect to the frequency to be the estimated value; An integration process for obtaining an integrated value by integrating for each frequency of the intensity distribution; and an integration process based on the gradient calculated by the event detection process. An integration time selection process of selecting which integration value is to be used among the integration values obtained by the plurality of integration times, and an integration value integrated by the integration time adopted by the integration time selection process. Searching for a local maximum in the vicinity of a rough estimate of the frequency of the signal detected by the event detection process, and determining a precise frequency that is a more accurate estimate of the frequency of the signal; It said frequency Seihaka processing repeatedly at a predetermined time, the signal frequency tracking method, characterized in that the tracking process temporally tracking the precise measurement frequency determined by the frequency Seihaka process, run from. 2. A frequency analysis process for obtaining a signal intensity distribution by performing frequency analysis on an input signal; a storage process for accumulating the signal intensity distribution for a predetermined time; and each frequency analyzed by the frequency analysis process. Each time, a small portion limited to the vicinity of the frequency is stored by the storage process.
The input signal strength distributions are input to a plurality of neural networks, each of which is responsible for a different gradient of the frequency of the signal with respect to time , and each of these neural networks has a function of time versus time during input .
A pattern detection process of outputting a large numerical value when a linear pattern indicated by a signal having a gradient of a specific frequency is present, and for the outputs of the plurality of neural networks, a maximum point is obtained in the output for each frequency. The presence of a signal is determined by detecting a point having a value whose output value exceeds a predetermined threshold, and the frequency of such a local maximum point is used as a rough estimate of the frequency of the signal, Event detection processing for calculating the gradient of the frequency with respect to the time of the signal from the gradient in charge of the neural network that outputs the largest value with respect to the frequency to be the estimated value, and the gradient of a plurality of predetermined frequencies with respect to the time For each frequency of the signal intensity distribution, the frequency is shifted from the current frequency by the amount of change in the frequency of the gradient from the past time to the current time. Based on the signal strength of the previous frequency, integrating processing for integrating at a predetermined time to obtain an integrated value; and selecting a gradient closest to an estimated value of the gradient by the event detection process from among a plurality of gradients used in the integration process. And an integration gradient selection process that employs an integration value obtained by the integration process corresponding to the selected gradient, and an integration value that is detected by the event detection process with respect to the integration value selected by the integration gradient selection process. Searching for a local maximum point in the vicinity of the rough estimated value of the frequency, a frequency precise measurement process for obtaining a precise frequency that is a more accurate estimate of the frequency of the signal, and performing the frequency precise measurement process from the frequency analysis process for a predetermined time. repeated every, a tracking process temporally tracking the precise measurement frequency determined by the frequency Seihaka processing, signal frequency tracking direction, characterized in that to perform . 3. A frequency analysis process for obtaining a signal intensity distribution by performing frequency analysis on an input signal; a storage process for accumulating the signal intensity distribution for a predetermined time; and each frequency analyzed by the frequency analysis process. Each time, a small portion limited to the vicinity of the frequency is stored by the storage process.
The signal strength distribution is input to a neural network, and the neural network outputs a large numerical value when a linear pattern indicated by a signal is present during the input.Pattern detection processing, and for the output of the neural network, The presence of a signal is determined by detecting a point where the maximum value is present in the output for the frequency and the output value has a value exceeding a predetermined threshold value, and the frequency of such a maximum point is determined as the frequency of the signal. and event detection process to rough estimates, the repetition from the frequency analysis process for each of the event detection process for a predetermined time, as well as temporally track an estimate of the frequency of the obtained signals by the event detection process,
A tracking process of calculating the frequency slope of the relative time of the signal, and the integration process of obtaining the integrated and the integrated value for each frequency of their respective pre-serial signal intensity distribution for a plurality of predetermined integration time, calculated by the tracking processing was based on the slope, the the integration time selection process for selecting whether to employ any of the integral value of the integration values obtained in a plurality of integration times by integration processing, the integration value adopted by the integration time selected process On the other hand, a frequency precise measurement process of searching for a local maximum point near the estimated value of the frequency obtained as a result of the tracking by the tracking process and obtaining a precise frequency which is a more accurate estimate of the frequency of the signal is executed. A signal frequency tracking method, characterized in that: 4. A frequency analysis process for performing a frequency analysis on an input signal to obtain a signal intensity distribution, a storage process for accumulating the signal intensity distribution over a certain period of time, and each frequency analyzed by the frequency analysis process Each time, a small portion limited to the vicinity of the frequency is stored by the storage process.
The signal strength distribution is input to a neural network, and the neural network outputs a large numerical value when a linear pattern indicated by a signal is present during the input.Pattern detection processing, and for the output of the neural network, The presence of a signal is determined by detecting a point that has a maximum point in the output with respect to the frequency and the output value has a value exceeding a predetermined threshold, and the frequency of such a maximum point is determined by the frequency of the signal. and event detection process to rough estimate of the repetition from the frequency analysis process for each of the event detection process for a predetermined time, as well as temporally track an estimate of the frequency of the obtained signals by the event detection process ,
A tracking process of calculating a frequency gradient of a signal with respect to time; and, for a predetermined plurality of frequency gradients with respect to time, for each frequency of the signal intensity distribution, a change in frequency of the gradient from a past time to a current time. Based on the signal intensity of the past frequency shifted from the current frequency by an amount, integrating in a predetermined time to obtain an integral value, and calculating by the tracking process among a plurality of gradients used in the integration process. Select the one that is closest to the gradient,
An integration gradient selection process that employs an integration value obtained by the integration process corresponding to the selected gradient, and an integration value that is obtained as a result of tracking by the tracking process with respect to the integration value that is adopted by the integration gradient selection process. A signal frequency tracking method comprising: searching for a local maximum point in the vicinity of an estimated value; and performing a frequency precise measurement process for obtaining a precise frequency that is a more accurate estimated value of the frequency of the signal.