JP2784400B2 - 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 a method for analyzing the frequency of a narrow-band signal whose characteristics are unknown by using the distribution in the frequency space of the input signal strength obtained by analyzing the frequency of an input signal such as a sound wave. And a signal frequency tracking method for measuring continuously.

[0002]

2. Description of the Related Art FIG. 2 is a block diagram showing a conventional signal frequency tracking device. The signal frequency tracking device illustrated in FIG. 2 is a device that tracks the frequency of a narrow band signal in which the frequency in an input signal changes with time. For example, the signal sensor includes a sound wave input signal S _in as an input. Have. The output side of the acoustic sensor 1 is connected to an FFT device 2 of a frequency analyzer that performs high-speed discrete Fourier transform, and the output side of the FFT device 2 is connected to a storage device 3. On the output side of the storage device 3, k neural networks 4-1, 4-2,
.., 4-k are connected in parallel. The output side of the k neural networks 4-1, 4-2,..., 4-k is connected to an event detection device 5, and the output side of the event detection device 5 is connected to a track management device 6. The track management device 6 is connected to a frequency prediction device 7 that transmits and receives frequencies to and from the track management device 6.

Next, while explaining the operation of the signal frequency tracking device shown in FIG. 2, a conventional signal frequency tracking method will be described. A sound wave of an input signal S _in is input to the acoustic sensor 1, and the acoustic sensor 1 An acoustic signal s1 is generated and sent to the FFT device 2. The FFT device 2 performs a high-speed discrete Fourier transform on the audio signal s1 every predetermined time S, and obtains an absolute value of the audio signal s1 for each sample frequency. As a result, a signal strength distribution s2 _in the frequency space of the input signal Sin is obtained. The storage device 3 accumulates the signal intensity distribution in the frequency space obtained by the FFT device for a certain period of time in the past, and stores the time x
A signal intensity distribution s3 on the frequency space is created. This signal intensity distribution s3 is created every predetermined time T seconds and sent to the neural networks 4-1, 4-2,..., 4-k. Each neural network 4-1, 4-2, ..., 4
For -k, a cutout of a rectangular small section centered on a certain sample frequency from the signal intensity distribution s3 in the time × frequency space is commonly input. Each of the neural networks 4-1, 4-2,..., 4-k is previously learned so as to output a large value if a linear pattern exists in the inputted small section. For this learning, for example, a back propagation method is used.
Here, the degree of the temporal change of the frequency of the signal is referred to as a gradient. Neural networks 4-1 and 4
,..., 4-k share and detect signals having gradients in different ranges. For example, neural networks 4-1 gradient detects a signal in the range of df ₀ ~df ₁ (Hz / sec), neural networks 4-2 gradient d
A signal in the range of f _{1 to} df ₂ (Hz / sec) is detected. The reason why the gradient is distributed and detected in this way is that if one neural network attempts to cover a wide range of gradients, the detection capability is reduced.

If a narrow band signal in the input signal exists at a frequency close to the sample frequency in the signal intensity distribution s3, a linear pattern emerges in the small section. Output a large value. This operation is repeatedly performed for many sample frequencies, and the result S4 is sequentially sent to the event detection device 5. In the event detection device 5, each of the neural networks 4-1, 4-2,.
A local maximum point on the frequency axis of the output of -k is detected. If the neural networks 4-1, 4-2, ..., 4-
If any of the outputs k shows a local maximum value at a certain frequency f and the value exceeds a predetermined threshold value, it is detected that an event has occurred at that frequency f.
That is, the presence of a narrow band signal in the input signal is confirmed. The frequency f at which the event is detected is sent to the track management device 6. The track management device 6 detects a temporal connection of the frequency f of the event every predetermined time T seconds (this is referred to as a track). In this method, a predicted frequency φ obtained from the past history of the frequency in the track is received from the frequency prediction device 7 and compared with a frequency f of a newly obtained event, and a difference between the two is determined by a predetermined value. If this is not the case, the event is integrated into the track. At this time, the track information T
r is rewritten and output as follows.

[0005] (A) Describe that the event belongs. (A) A value obtained by dividing the difference between the previous frequency of the track and the frequency f of the current event by the time difference T from the previous event is recorded as a track gradient. (C) The frequency f of the event is recorded as a new frequency of the track. If the event does not integrate with any existing track, a new track is formed. At this time, the frequency f of the event is directly used as the track frequency, and the gradient of the track is unknown. If no event is integrated into a certain track, the track is held as it is (this case is called temporary search). This track is deleted when the temporary search has been performed a predetermined number of times in succession (this case is referred to as search). Finally, the frequency prediction device 7 calculates a predicted frequency φ from which an event is expected to occur after T seconds, which is the next processing cycle in the track management device 6, from the past history of the track. For example, if the track frequency is f and the gradient is df, the predicted frequency φ after T seconds is estimated by the following equation (1). φ = f + T · df (1) However, if df is unknown, it is estimated by equation (2). φ = f (2) By repeating the above processing, the ever-changing frequency of the narrow-band signal whose characteristic is unknown can be continuously measured over time.

[0006]

However, the conventional signal frequency tracking method has the following problems. The predicted frequency φ of the track shown in the above equation (1)
May not be obtained. For example, since there is only one data at the time of the first search, information on the predicted frequency φ cannot be obtained clearly. Also, until a certain time has passed after the initial search,
The accuracy of the prediction frequency φ is low, and there is a problem that a signal is easily lost. In addition, in order to prevent a loss of search, if a reliable predicted frequency φ is not obtained, a method of increasing the integration window width may be considered.However, there is a side effect that the probability of integrating an erroneous event increases. Performance was not obtained.

[0007]

According to a first aspect of the present invention, in order to solve the above-mentioned problems, a frequency analysis is performed on an input signal, and the analysis result is accumulated for a certain period of time to obtain a time-frequency two-dimensional space. Processing for obtaining the signal intensity distribution of, and inputting a section limited to the vicinity of the sample frequency of the signal intensity distribution to the neural network for each sample frequency, and when a linear pattern of the signal intensity distribution exists in the section Processing to output a large value to the neural network, and detecting, as an event, a point having a value exceeding a predetermined threshold at a local maximum point in the frequency space at the output of the neural network, thereby narrowing the narrow band signal in the input signal. And determining that a sample frequency at which the event is detected as the frequency of the narrowband signal. It repeats every interval T, and when the frequency of the newly detected event and the predicted frequency of the narrowband signal estimated from the past event occurrence situation are close, the new and old events are The following method is employed in a signal frequency tracking method for determining and integrating signals based on band signals and tracking the narrow band signal whose frequency changes with time.

That is, according to the signal frequency tracking method of the present invention, a plurality of neural networks each outputting a large numerical value with respect to the linear pattern indicated by the specific narrow-band signal, the gradient of which is a degree of frequency change with time. Using
A first gradient estimating process for estimating the gradient of the narrowband signal by comparing output values of the neural networks, and a past history of an estimated value of a frequency of the narrowband signal obtained at each of the predetermined times T; A second gradient estimating process of estimating the gradient of the narrowband signal from the change during the latest predetermined time P among the above. If the predetermined time P has not elapsed since the event was detected for the first time, the result of the first gradient estimation process is selected, and after the predetermined time P has elapsed, the second gradient estimation process is performed. A gradient selection process for selecting a result of the process, and a predicted frequency of the narrowband signal from the estimated value of the gradient of the narrowband signal and the estimated value of the frequency of the narrowband signal obtained as a result of the gradient selection process. And a prediction frequency calculation process for estimating, and the integration is determined using the prediction frequency.

A second invention performs the first gradient estimation process and the second gradient estimation process in the first invention, and further includes a result of the first gradient estimation process or the second gradient estimation process. Performing a gradient selection process of selecting the one with higher accuracy, and a predicted frequency calculation process in the first invention,
The integration frequency is determined using the predicted frequency. In a third aspect based on the signal frequency tracking method according to the second aspect, the gradient selection processing is performed by estimating the prediction frequency using the gradient estimated value obtained by the first gradient estimation processing. A process of evaluating the accuracy of the estimated value of the gradient based on the proximity between the frequency and the frequency of the newly detected event, and using the estimated value of the gradient obtained by the second gradient estimation process to calculate the predicted frequency. A process of estimating the accuracy of the estimated value of the gradient based on the proximity between the estimated frequency and the frequency of the newly detected event is performed, and a process of selecting an estimated value with a higher accuracy evaluation is performed. A fourth invention performs a process of estimating an S / N ratio of the narrowband signal in the input signal, and performs the gradient selection process in the signal frequency tracking method of the second invention by estimating the S / N ratio. If the value is lower than a preset threshold, the result of the first gradient estimation processing is selected, and the S / S
If the estimated value of the N ratio is higher than the threshold, the result of the second gradient estimation process is selected.

[0010]

According to the first aspect of the present invention, the signal frequency tracking method is configured as described above, so that the input signal is subjected to frequency analysis, and the result is accumulated at regular time intervals, and is stored in the time × frequency space. Is obtained. By the first gradient estimation process, the gradient of the narrowband signal is estimated from the signal intensity distribution in a two-dimensional space of time × frequency at fixed time intervals.
That is, the gradient of the event is estimated. On the other hand, in the second gradient estimation process, the gradient of the track is obtained as an estimated value of the gradient of the band signal from the past history of the frequency of the narrowband signal. As a result of the gradient selection processing, the result of the first gradient estimation processing is selected while the predetermined time P has not elapsed since the event was first detected, and after the time P has elapsed, the second gradient estimation processing is performed.
Is selected. The predicted frequency calculation process estimates the predicted frequency of the narrowband signal from the estimated value of the gradient of the narrowband signal and the estimated value of the frequency of the narrowband signal obtained as a result of the selection in the gradient selection process. New events are integrated into the track using the estimated prediction frequency.

[0011] According to the second aspect of the present invention, a higher-precision one of the results of the first gradient estimating process or the second gradient estimating process in the first aspect is selected by the gradient selecting process. The predicted frequency calculation process estimates the predicted frequency of the narrowband signal from the estimated value of the temporal gradient of the frequency of the narrowband signal and the estimated value of the frequency of the narrowband signal obtained as a result of the selection of the gradient selection process. New events are integrated into the track using the estimated prediction frequency. According to the third aspect, in the gradient selection processing according to the second aspect, the predicted frequency is estimated using the estimated value of the gradient of the event obtained by the first gradient estimation processing, and the predicted frequency is newly detected. The accuracy of the estimated value of the gradient is evaluated based on the proximity to the frequency of the event. or,
A prediction frequency is estimated using the gradient of the track obtained by the second gradient estimation processing, and the accuracy of the estimated value of the gradient is evaluated based on the proximity between the predicted frequency and the frequency of the newly detected event. You. Then, the estimated value having the higher accuracy evaluation is selected.

According to the fourth aspect, the S / N ratio of the narrow band signal in the input signal is estimated. In the gradient selection processing according to the second aspect, the estimated value of the S / N ratio is set to a predetermined threshold value. If lower, the result of the first gradient estimating process is selected, and if higher, the result of the second gradient estimating process is selected. That is, in the first to fourth inventions, for each event, the output of each neural network corresponding to the frequency of the event is compared, and based on which gradient the neural network in charge of shows the largest output,
The slope of the narrowband signal is estimated and recorded as an estimate of the slope of the event. If the predicted frequency of the track cannot be obtained in the course of the frequency tracking or if the accuracy of the predicted frequency is low, the predicted frequency of the track is calculated using the gradient of the event integrated in the track. The accuracy of the estimated value of the event gradient obtained by comparing the outputs of the neural networks deteriorates when the S / N ratio of the signal is high to some extent. This is because the output of the neural network reaches the upper limit and a plurality of neural networks output values arranged side by side. Rather, for a signal with a barely detectable S / N ratio,
Since only one neural network that includes the true gradient of the signal within the target range shows a response, the accuracy of gradient estimation is high. Therefore, the prediction frequency calculated based on the estimated value of the gradient has high accuracy.

On the other hand, the gradient of the track obtained by tracking and the predicted frequency based on the track have low accuracy immediately after the start of detection and when the accumulation of data is not sufficient. However, since the S / N ratio of the signal at the time of the initial search is not usually very high, the estimation accuracy of the gradient of the event is high in this situation. As described above, the method of obtaining the predicted frequency from the estimated value of the gradient of the event and the method of obtaining the predicted frequency by tracking have characteristics that complement each other. Therefore, when the predicted frequency of the track cannot be obtained based on the tracking, or when the accuracy is considered to be low, the predicted frequency is obtained based on the estimated value of the gradient of the event. Therefore, the above problem can be solved.

[0014]

EXAMPLES First Embodiment FIG. 1 is a block diagram illustrating a signal frequency tracking apparatus of the first embodiment of the present invention, it is denoted by the same reference numerals to elements common to FIG. In the present embodiment, when the occurrence of an event is detected for the first time, the output of a plurality of neural networks is compared to determine the neural network that outputs the largest value, and the estimated value of the gradient of the narrowband signal to be tracked is calculated. Ask. That is, the gradient of the event is obtained. A predicted frequency used for integration is estimated based on the obtained estimated value of the gradient. At other times, the signal frequency tracking device estimates the gradient of the narrow band signal by calculating the gradient in the track from the degree of the latest two changes in the past history of the frequency of the event, and calculates the gradient of the narrow band signal from the gradient of the track. The frequency tracking is performed by estimating the predicted frequency in the next tracking cycle of the narrowband signal. The signal frequency tracking device shown in FIG. 1 includes an acoustic sensor 1 for inputting a sound wave as an input signal S _{in as} in FIG. The output side of the acoustic sensor 1 is connected to an FFT device 2 of a frequency analyzer that performs high-speed discrete Fourier transform, and the output side of the FFT device 2 is connected to a storage device 3. On the output side of the storage device 3, K neural networks 4-1, 4-2,
.., 4-k are connected in parallel. The output side of the K neural networks 4-1, 4-2,..., 4-k is connected to an event detection device 15. The event detection device 15 detects the occurrence of an event and performs a first gradient estimation process for estimating the gradient in each event, as in the related art.

The output side of the event detection device 15 is connected to the track management device 16. Truck management device 16
Has the function of performing integration and performing a second gradient estimating process to determine the gradient of the track from the past history of the frequency of the track and to obtain an estimated value of the gradient of the narrowband signal. A frequency estimating device 17 that transmits and receives frequencies to and from the track management device 16 is connected to the track management device 16. The frequency prediction device 17 performs a predicted frequency calculation process. Further, the signal frequency tracking device is provided with a gradient selecting device 18. The gradient selection device 18 is connected to the output side of the event detection device 15 and between the track management device 16 and the frequency prediction device 17.
And an estimated value of each gradient from the event detection device 15 and sends it to the frequency prediction device 17.

Next, the operation of the signal frequency tracking device of FIG. 1 will be described. The acoustic sensor 1 receives an input signal S _in of a sound wave including a narrowband signal, converts the signal into an acoustic signal s1, and sends the acoustic signal s1 to the FFT device 2. The FFT device 2 performs a high-speed discrete Fourier transform on the acoustic signal s1 at a constant sampling time T to obtain a frequency analysis result s2. When formulating the frequency analysis result s2, it can be expressed by the following (3). {X (f) | f = f ₁ , f ₂ ,..., F _N } (3) The storage device 3 stores the frequency analysis result s2 shown in (3) for a fixed time in the past. Thus, a signal intensity distribution s3 in the time × frequency space shown in the following (4) is obtained. {X (t, f) | t = t 1, t 2, ..., t M; f = f 1, f 2, ..., f N} ··· (4) (4) signal shown in the distribution strength s3 Is created every predetermined time T seconds. From the signal intensity distribution s3, the storage device 3 obtains, for each sample frequency f _n = f ₁ , f ₂ ,..., F _N , a predetermined width 2 centered on f _n shown in the following (5).
The W + 1 subsection is divided into K neural networks 4-
.., 4-k. {X (t, f) | t = t ₁ , t ₂ , ..., t _M ; f = f _nW , ..., f _{n + W} } (5) Each neural network 4-1, 4-2, …, 4-k
Share input signals of different slopes. For example, a neural network 4-1, the gradient of the frequency from the input is to detect signals in the range of df ₀ ~df ₁ (Hz / sec), previously subjected to learning in advance back propagation method or the like. Similarly, the scope of the neural network 4-2 slope df ₁ ~df ₂ (Hz / sec), ..., the range of the signal of the neural network 4-k is _{df k-1 ~df k (Hz} / sec) In charge of detection. The processing and learning method in the neural network are described in detail in the following document. Literature; Hideki Aso, "Neural Network Information Processing"
First edition (1989) Sangyo Tosho Co., Ltd., P.9-18,50-54 Neural networks 4-1,4-2, ..., 4-k
Is configured to output a numerical value of 0 to 1 to the event detection device 15. If there is a gradient in charge in a small section of the input signal distribution intensity s3, the maximum value of 1 is output. When a gradient close to the assigned gradient exists, a value close to 1 is output, when a far gradient exists, a value close to 0 is output, and when there is no gradient in a small section, 0 is output. . If a narrow band signal exists at a frequency close to the sample frequency in the signal intensity distribution s3, a linear pattern emerges in the small section. Neural networks 4-1 and 4
−2,..., 4-k detect small sections, and the neural network in charge of the gradient close to the gradient of the small section outputs a large value. This operation is repeatedly performed for many sample frequencies, and the result s4 is sequentially sent to the event detection device 15.

The event detecting device 15 searches for a point which is a local maximum point and whose value exceeds a predetermined threshold value in the output with respect to the frequency of each of the neural networks 4-1, 4-2,..., 4-k. Detect points as events. Then, the frequency of the maximum point at this time is defined as the frequency f of the event. Also, the temporal change of the frequency f of the event, that is,
The gradient df _e of events, each neural network 4
.., 4-k are estimated by comparing the output values. As the estimation method, for the event of the frequency f, and detects a neural network for outputting the largest value, the center value of the slope segment df _k-1 ~df k in charge of the neural network (df _k-1 ~ df _k ) /
With 2 the estimated value df _e of the slope of the events. The frequency f of the event is sent to the frequency prediction device 17 and the track management device 16. Also, the estimated value df of the gradient of the event
_e is sent to the gradient selector 18. Truck management device 1
No. 6 compares the frequency f of the detected event with each track being tracked. If the difference between the frequency f of the event and the predicted frequency φ of the track supplied from the frequency prediction device 17 is equal to or smaller than the integration window width, the event is integrated into the track. At this time, information about the track is described below (A)
Update as in (c). (A) State that the event belongs. (B) the difference between the previous frequency and the frequency f of this event in the track, records the value obtained by dividing the time difference between the previous event as the gradient df _t tracks. (C) The frequency f of the event is recorded as a new frequency of the track.

Note that if the event is not integrated into any known track, the track manager 16 creates a new track. The event frequency f is used as it is as the frequency of the new track. Since respect gradient df _t track it is necessary to detect at least two events can not estimate the slope of the narrowband signal when forming a new track. In this case, the slope df _t tracks recorded to be unknown. Truck management device 16
It sends the frequency f of the track to the frequency estimating device 17, and sends the correspondence information s16 slope df _t and tracks and events of the track to the slope selection unit 18. Recording the estimated value df _e of the slope of the events from the gradient selector 18 in the event detector 15. Also, the gradient selection device 18
Also it receives gradient df _t from one track management device 16. Here, If the slope df _t tracks is given, and sends a gradient df _t to the frequency estimating device 17. However, since in the case of the track immediately after the detection slope df _t track is unknown, until a predetermined time P from the first probe, with reference to the events belonging to the track at this time, the estimated value of the slope of the events the df _e supplied to the frequency estimating device 17 as an alternative to gradient df _t of the tracks.

The frequency predicting device 17 calculates a predicted frequency φ of each track in the next processing cycle (after T seconds) and gives it to the track managing device 16. The predicted frequency φ is obtained by the following equation (6), where f is the frequency of the track and df is the gradient. The predicted frequency φ is sent to the track management device 16. φ = f + T · df (6) As described above, in the present embodiment, the gradient selecting device 18 is provided, and the neural networks 4-1 and 4-
2, ..., it has a configuration to calculate the 4-k with reference to the track predicted frequency estimates df _e of the slope of the events obtained by the comparison of the output values of phi. For this reason, it is possible to eliminate the disadvantage that the gradient cannot be obtained at the time of the first search and the prediction accuracy cannot be obtained and the tracking is not stable, which is observed in the conventional signal frequency tracking method.

Second Embodiment This embodiment estimates the predicted frequencies using the estimated value of the gradient of the event and the gradient of the track, and determines how close each predicted frequency is to the frequency of the newly detected event. The estimation accuracy is evaluated with. In the present embodiment, a prediction frequency with higher accuracy is selected and used. FIG. 3 is a block diagram of a signal frequency tracking apparatus according to a second embodiment of the present invention. Elements common to FIGS. 1 and 2 are denoted by the same reference numerals. This signal frequency tracking device includes an acoustic sensor 1 similar to that of the first embodiment and an FF
It includes a T device 2, a storage device 3, K neural networks 4-1, 4-2,..., 4-k, and an event detection device 15, which are connected in the same manner as in the first embodiment. ing. The signal frequency tracking device of FIG. 3 is newly provided with two frequency prediction devices 20 and 21 and a predicted frequency selection device 22. On the output side of the event detection device 15, a track management device 16, a frequency prediction device 20, and a predicted frequency selection device 22 are connected. The frequency prediction device 21 is connected to the output side of the track management device 16, and the output side of the frequency prediction device 21 is connected to the predicted frequency selection device 22. The output side of the predicted frequency selection device 22 is connected to the track management device 16. Each of the frequency predictors 20 and 21 estimates the predicted frequency using the estimated value of the gradient of the event or the gradient in the track, and the predicted frequency selector 22 obtains the accuracy and selects the one with the higher accuracy to determine the track. It is configured to perform processing to supply to the management device 16.

Next, the operation of the signal frequency tracking device of FIG. 3 will be described. The same operation as in the first embodiment is performed from the time when the acoustic sensor 1 receives the input signal S _in of the sound wave to the time when the event is detected by the event detection device 15. Estimate df _e of the frequency f and its gradient obtained event event detection device 15 is fed to a frequency predicting unit 20.
The frequency prediction device 20 obtains a predicted frequency φ _e of the narrowband signal after a predetermined time T seconds by the following equation (7). φ _e = f + T · df _e (7) On the other hand, the track management device 16 has a track frequency f (same as an event frequency) and a gradient df, as in the conventional and first embodiments. _t is calculated and sent to the frequency prediction device 21. Frequency prediction apparatus 21 estimates a predicted frequency phi _t narrowband signal after T seconds by the following equation (8).

Φ _t = f + T · df _t (8) The predicted frequency selecting device 22 determines which one of the predicted frequencies φ _e and φ _t obtained by the two methods has higher accuracy as the predicted frequency φ And sends it to the track management device 16. Here, the accuracy is evaluated as follows. The predicted frequency selecting device 22 receives the predicted frequency φ input for each processing cycle.
_e , φ _t , φ _e (1), φ _e (2), ..., and φ _t
The values of (1), φ _t (2),... Are recorded, and the predicted frequency φ predicted last time is the frequency f of the event captured this time.
Select the person who was closer to. That is, in certain processing cycle n, when it becomes the following equation (9) phi _e (n) is adopted as the predicted frequency phi, phi _t (n) is employed otherwise. | Φ _e (n−1) −f | <| φ _t (n−1) −f | (9) The track management device 16 uses the predicted frequency φ sent from the predicted frequency selection device 22 As in the first embodiment or the related art, the event and the track are integrated. As described above, in this embodiment, predicted frequency phi _e based on the estimated value df _e of the slope of the frequency of the event in the frequency predictor 20
It estimates the gradient of the track at a frequency prediction apparatus 21 df _t
Estimating a predicted frequency phi _t based on, it sends predicted frequency selection device 22 each predicted frequency phi _e, selected as predicted frequency phi a higher accuracy of the phi _t to the track management device 16. Therefore, the predicted frequency φ can be obtained with higher accuracy under a wider range of conditions than in the past, and stable frequency tracking can be performed.

Third Embodiment In this embodiment, the S / N ratio of a narrow band signal in an input signal is estimated, and either the estimated value of the event gradient or the value of the track gradient is determined according to the level. Is selected, a predicted frequency of the narrowband signal is obtained using the selected value, and signal frequency tracking is performed. FIG. 4 is a block diagram of a signal frequency tracking device according to a third embodiment of the present invention.
Elements common to FIG. 3 are denoted by common reference numerals. This signal frequency tracking device is a device for implementing the signal frequency tracking method of the present embodiment, and includes the same acoustic sensor 1, FFT device 2, storage device 3, and K neural circuits as in the first embodiment. .., 4-k
, An event detection device 15 and a track management device 16
And a frequency estimating device 17. Further, unlike the first embodiment, a signal precise measuring device 30 and a gradient selecting device 31 are provided. On the output side of the acoustic sensor 1 is an FFT device 2
Is connected to the output side of the FFT device 2, the storage device 3 and the signal precise measurement device 30. On the output side of the storage device 3, K neural networks 4-1, 4-2,
.., 4-k are connected. The output side of the K neural networks 4-1, 4-2,..., 4-k is connected to an event detection device 15. Event detection device 1
The output side of 5 is connected to the signal measuring device 30 and the gradient selecting device 31, and the output side of the signal measuring device 30 is connected to the track management device 16 and the gradient selecting device 31. The output side of the gradient selection device 31 is connected to the frequency prediction device 17,
The output side of the frequency prediction device 17 is connected to the track management device 16. The output side of the track management device 16 is connected to the gradient selection device 31 and the frequency prediction device 17.

Next, the operation of the signal frequency tracking device of FIG. 4 will be described. The acoustic sensor 1 receives the input signal S _in of the received sound wave.
Is converted to an acoustic signal s1, and the FET device converts the signal intensity distribution s2 shown in the following (10) on the frequency space every predetermined time S, and this is stored in the storage device 3 and the signal precise measurement device 30.
Send to {X (f, t) | f = f ₁ , f ₂ ,..., F _N } (10) After inputting the signal distribution intensity s 2 into the storage device 3, an event is detected by the event detection device 15. Up to this, the same operation as in the first and second embodiments is performed. The signal precise measurement device 30 measures a more accurate estimated value f _{s of the} signal frequency and the S / N ratio of the signal for each detected event. Here, the signal precise measurement device 30 receives the signal intensity distribution s2 from the FET device 2 and averages a predetermined time period (for example, K × S seconds) for each sample frequency for a value ΔX ₁ ( f, t) | f = f ₁ , f ₂ ,..., f _N } are calculated in advance. That is, for each sample frequency, the following (11)
Is calculated in advance. X ₁ (f, t) = (X (f, t) + X (f, t−S) X (f, t−2S) +... + X (f, t− (K−1) S)) / K) (11) Subsequently, the signal precise measurement device 30 receives the frequency f _n of the event from the event detection device 15. Average value ｛X
The frequency having the largest value in a predetermined range centered on the frequency f _n in ₁ (f, t)｝ is determined as the precisely measured frequency f _s . For example, assuming that a predetermined range is a one-sided ε sample frequency, the signal precision measuring device 30 calculates ΔX ₁ (f, t) | f = f
_{n-ε, ..., f n} -1, f n, f n + 1, ..., in the f n _{+ ε},} obtaining the frequency f _n0 showing the largest value as Seihaka frequency f _s. The measured frequency f _s is sent to the track management device 16 instead of the frequency f of the event. Further, the signal precise measurement device 30 obtains the estimated value L of the S / N ratio of the signal from the following equation (12), and sends it to the gradient selection device 31. L = ₁₀ log ₁₀ (X ₁ (f _n0 , t) / σ) (12) where σ is a value {X ₁ (f, t) | f = f ₁ , f ₂ ,
.., F _N }.

The gradient selecting device 31 is used for the event detecting device 1.
An estimate df _e of the slope of the events from 5 receives the slope df _t tracks from the track management unit 16 determines whether to choose. At this time, the signal precise measurement device 30
Estimate L of the obtained S / N ratio is adopted estimate df _e of the slope of the event if the value lower than a predetermined value, employing the slope df _t tracks is higher. Note that this threshold is both slope df _e with signals of various S / N ratio, df _t
Is set in advance by a simulation experiment for comparing the accuracy of. The frequency predicting device 17 includes the gradient selecting device 3
Next processing cycle using the gradient selected in 1 (after T seconds)
Is calculated and sent to the track management device 16. Track management device 16 receives the precise measurement frequency f _s of the event from the signal precise measurement device 30, and sends the slope df _t track the slope selection unit 31. Other operations of the track management device 16 are the same as those in the related art.

As described above, in this embodiment, the signal precise measuring device 30 and the gradient selecting device 31 are provided, and the signal precise measuring device 30 obtains the estimated value L of the S / N ratio of the signal. the estimated value L is lower if the event of the gradient df _e, by selecting the slope df _t track the higher frequency estimating device 17
Send to Then, the frequency prediction device 17 calculates the predicted frequency φ using the selected gradient. Therefore, it is possible to accurately obtain the predicted frequency φ for a signal having a wider S / N ratio than in the related art, and perform stable tracking. 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 except the acoustic sensor 1 in FIGS. 1, 3 and 4 may be realized by an electric circuit,
Some or all of them may be realized by a computer program. (2) In the first to third embodiments, a sound wave is used as the input signal S _in , but the present invention is not limited to the sound wave and may be a radio wave or the like. For example, when the input signal S _in is a radio wave, the reception sensor is used in place of the acoustic sensor 1 so that
The same effect as that of the embodiment is exhibited.

[0027]

As described above in detail, according to the first aspect, the first gradient estimating process for estimating the gradient of the narrowband signal by comparing the output values of a plurality of neural networks, A second gradient estimating process for estimating the gradient of the narrowband signal from the change during the latest predetermined time P in the past history of the estimated value of the frequency of the signal, and a predetermined time P from the time when the event is first detected; While the time has not elapsed, the result of the first gradient estimation process is selected, and after the time elapses, a gradient selection process of selecting the result of the second gradient estimation process is performed, and the result of the gradient selection process is obtained. A predicted frequency calculation process for estimating a predicted frequency of the narrowband signal from the estimated value of the gradient of the narrowband signal and the estimated value of the frequency of the narrowband signal is performed. Therefore, the problem that a gradient cannot be obtained at the time of the first search is solved, and stable narrow-band signal frequency tracking can be performed. According to the second and third inventions, a gradient selection process for selecting a higher accuracy one of the results of the first gradient estimation process or the second gradient estimation process in the first invention, and a predicted frequency calculation process Therefore, the prediction frequency can be accurately obtained under a wider range of conditions than in the past, and stable frequency tracking can be performed. According to the fourth aspect, the processing for estimating the S / N ratio of the narrow band signal in the input signal is performed. In the gradient selection processing in the second aspect, the estimated value of the S / N ratio is set to a predetermined threshold value. If the value is lower than the threshold, the result of the first gradient estimation process is selected, and if the estimated value of the S / N ratio is higher than the threshold, the second
Since the result of the gradient estimation process is selected, a predicted frequency can be obtained with high accuracy for a signal having a wider S / N ratio than in the past, and stable tracking can be performed.

[Brief description of the drawings]

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

FIG. 2 is a block diagram showing a conventional signal frequency tracking device.

FIG. 3 is a block diagram of a signal frequency tracking device according to a second embodiment of the present invention.

FIG. 4 is a block diagram of a signal frequency tracking device 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 to 4-k neural network 15 event detection device 16 track management device 17, 20, 21 frequency prediction device 18 gradient selection device 22 prediction frequency selection device 30 signal precise measurement device 31 gradient Selection device

──────────────────────────────────────────────────続 き Continuing on the front page (72) Inventor: Satoshi Mizota 1-7-12 Toranomon, Minato-ku, Tokyo Oki Electric Industry Co., Ltd. (72) Inventor Shunji Ozaki 1-7-12 Toranomon, Minato-ku, Tokyo Oki Electric Industry Co., Ltd. (56) References JP-A-58-139084 (JP, A) JP-A-7-19947 (JP, A) JP-A-5-273326 (JP, A) (58) Field (Int.Cl. ⁶ , DB name) G01H 3/08 G01H 17/00 G01S 3/80

Claims (4)

(57) [Claims] 1. An input signal is subjected to frequency analysis, and the analysis result is accumulated over a certain period of time.
A process of obtaining a signal intensity distribution in a dimensional space; and inputting a section of the signal intensity distribution limited to the vicinity of the sample frequency to the neural network for each sample frequency, where a linear pattern of the signal intensity distribution is input to the section. Processing to output a large value to the neural network when present, and detecting a point having a value exceeding a predetermined threshold at a local maximum point in the frequency space in the output of the neural network as an event in the input signal. The process of determining the presence of a narrowband signal and estimating the sample frequency at which the event was detected as the frequency of the narrowband signal is repeated for each predetermined time T, and the frequency of the newly detected event And the predicted frequency of the narrowband signal estimated from the past event occurrence situation is close to the new and old In a signal frequency tracking method for determining and integrating events based on the same narrow-band signal and tracking the narrow-band signal having a temporally changing frequency, a gradient, which is a degree of temporal change in frequency, is a specific gradient. A first method for estimating the gradient of the narrowband signal by using a plurality of neural networks each outputting a large numerical value with respect to the linear pattern indicated by the narrowband signal and comparing output values of the respective neural networks. A gradient estimating process, and estimating the gradient of the narrowband signal from a change during a latest predetermined time P in a past history of the estimated value of the frequency of the narrowband signal obtained at each of the predetermined times T And selecting the result of the first gradient estimating process while the predetermined time P has not elapsed since the event was detected for the first time. After the interval P has elapsed, a gradient selection process for selecting the result of the second gradient estimation process, and an estimated value of the gradient of the narrowband signal and a frequency of the narrowband signal obtained as a result of the gradient selection process A prediction frequency calculation process of estimating a prediction frequency of the narrowband signal from an estimated value, and determining the integration using the prediction frequency. 2. A higher-precision one of a first gradient estimation process and a second gradient estimation process according to claim 1, and a result of the first gradient estimation process or the second gradient estimation process is selected. A signal frequency tracking method, comprising: performing a gradient selection process to perform the prediction frequency calculation process according to claim 1; and performing the integration determination according to claim 1 using the prediction frequency. 3. The gradient selecting process includes estimating the predicted frequency using an estimated value of the gradient obtained by the first gradient estimating process, and comparing the predicted frequency with a frequency of a newly detected event. A process of estimating the accuracy of the estimated value of the gradient based on the degree of proximity, and estimating the predicted frequency by using the estimated value of the gradient obtained by the second gradient estimating process. A process of evaluating the accuracy of the estimated value of the gradient based on the proximity to the frequency of the event, and a process of selecting an estimated value with a higher evaluation of the accuracy. Frequency tracking method. 4. A process for estimating an S / N ratio of the narrow-band signal in the input signal, wherein the gradient selection process is performed when the estimated value of the S / N ratio is lower than a preset threshold. 3. The signal according to claim 2, wherein the result of the first gradient estimating process is selected, and the result of the second gradient estimating process is selected if the estimated value of the S / N ratio is higher than the threshold value. Frequency tracking method.