JP5487062B2 - Noise removal device - Google Patents

Description

  The present invention relates to a noise removal device for removing impulse magnetic noise included in an input signal acquired by a magnetic sensor.

  Due to the high sensitivity and miniaturization of magnetic sensors, the environment in which the magnetic sensor is used is diversified, and the influence of magnetism generated in the installation environment on the output signal is increasing.

  In particular, when a magnetic sensor is installed in a device, the influence of impulse magnetic noise generated in synchronization with the operation of other electronic devices in the device is significant, and noise that is tens of times the detection level of the magnetic sensor is generated. Yes.

  Conventionally, techniques as described in Patent Documents 1 and 2 are used as a method for removing noise. In Patent Literature 1, when speech is sometimes input in an environment where environmental noise exists, the noise spectrum update unit is based on the output of the Fourier analysis unit and the previous noise spectrum in the noise spectrum storage unit. The current noise spectrum is obtained, the previous noise spectrum in the noise spectrum storage unit and the obtained current noise spectrum are switched based on the determination of the noise signal determination unit, and output as an estimated noise spectrum, and the noise spectrum storage unit The previous noise spectrum is updated to the current noise spectrum, and the noise is removed from the speech with the superimposed noise.

  Further, in Patent Document 2, an acoustic feature amount of an input signal is extracted for each frame, and a noise model parameter is forward-processed by parallel processing and with respect to a time axis by using a probability model of a clean speech signal and a silence signal. As well as the reverse direction. Then, for each frame, a non-speech state / speech probability and a ratio of the speech probability to the non-speech state probability are calculated, and a speech interval estimation is performed by comparing the speech probability ratio with a threshold. Furthermore, using the parameters of each probability model and the non-speech state / speech probability, a frequency response filter that removes a noise signal is generated, the frequency response filter is converted into an impulse response filter, and the impulse response is converted to the input signal. The filter is convoluted to generate and output a noise-removed audio signal.

  FIG. 4 is a diagram illustrating an example of an operation flow of the conventional noise removing device disclosed in Patent Document 1. In FIG. In FIG. 4, local time data of about 10 ms to 20 ms is cut out from the input waveform on which noise is superimposed, and a window function is applied by the windowing unit 21 to maintain data continuity. Subsequently, frequency analysis is performed by the FFT analysis unit 22, and the analysis result is decomposed into an amplitude spectrum and a phase spectrum. The amplitude spectrum is output to the voice section detection unit 23, the noise spectrum update unit 24, and the noise subtraction unit 25. On the other hand, the phase spectrum is output to the IFFT synthesis unit 27 for re-synthesis of the speech waveform.

  The voice section detection unit 23 discriminates a voice section with a voice signal on which the input noise signal is superimposed and a noise section with only a noise signal without voice and outputs them to the noise spectrum update unit 24. The noise spectrum updating unit 24 estimates a noise signal in the voice section and the noise section. The noise subtracting unit 25 performs processing for subtracting the noise signal estimated by the noise spectrum updating unit 24 for each frequency from the amplitude spectrum on which the noise signal from the FFT analyzing unit is superimposed. The negative coefficient zero unit 26 sets the amplitude to zero for a frequency at which the amplitude becomes negative as a result of the subtraction in the noise subtraction unit 25. The IFFT synthesis unit 27 performs synthesis processing and inverse Fourier transform on the amplitude spectrum and phase spectrum from which the noise signal has been removed, and further reproduces and outputs a continuous waveform by the synthesis unit 28 based on addition of windowed waveforms. The waveform of the audio signal from which noise has been removed can be obtained.

JP-A-8-160994 JP 2009-210647 A

  The above-described conventional noise removal apparatus has an advantage in audio signal processing with a sampling interval of about 10 ms to 20 ms. However, when the sampling interval becomes large, the delay and the increase in the amount of memory associated with performing FFT analysis and IFFT synthesis become a problem. That is, once converted to the frequency domain, noise estimation and removal is performed with the frequency component, and re-converted and synthesized again into the time domain, for example, a signal exceeding a sampling interval of 100 ms, such as a low frequency signal of 50 Hz or less. In this case, a longer data processing time is required. For this reason, it has been difficult to achieve both real-time performance and noise removal performance.

  Therefore, an object of the present invention is to provide a noise removing device that can remove impulse magnetic noise without impairing real-time performance.

  In order to solve the above-described problems, the noise removal device of the present invention performs simple processing in the time axis region for low-frequency observation signals, and can remove impulse magnetic noise without impairing real-time characteristics. is there.

According to the present invention, there is provided a noise removing device for removing the impulse magnetic noise from an input signal mixed with the impulse magnetic noise, wherein the impulse magnetic noise held in a memory is generated together with the generation of the impulse magnetic noise. A prediction value calculation unit that calculates an approximate prediction value obtained by approximating past data that is an input signal until noise is mixed with a primary or secondary approximation formula, and a window function calculation unit that calculates a window function coefficient value. A noise estimation interpolation signal is calculated by multiplying the noise estimation signal obtained by subtracting the approximate prediction value from the input signal and the window function coefficient value , and subtracting the noise estimation interpolation signal from the input signal, thereby generating the impulse shape A noise removal apparatus characterized by removing magnetic noise is obtained.

According to the present invention, the window function calculator, as a starting point the occurrence of the impulse magnetic noise, said noise removal and outputs the window function coefficient value that varies continuously from 1 to 0 A device is obtained.

  In the present invention, a noise estimation interpolation signal is calculated by performing a prediction value calculation and a window function calculation starting from the generation of the impulse magnetic noise, and the impulse magnetic noise is reduced by subtracting this signal from the input signal. It can be removed. In the present invention, since the frequency component analysis is unnecessary, it is possible to provide a noise removing device that does not impair the real-time property.

The figure which shows the noise removal apparatus by this invention. FIG. 2A is a diagram illustrating a result of removing impulse magnetic noise, FIG. 2A is a frequency-power spectrum characteristic diagram, and FIG. 2B is a time-magnetic level characteristic diagram. The figure which shows the window function coefficient output value by a window function calculation part. The figure which shows an example of the operation | movement flow of the conventional noise removal apparatus.

  Embodiments of the present invention will be described below with reference to the drawings.

  FIG. 1 is a diagram showing a noise removing apparatus according to the present invention. The noise removal apparatus 1 of the present invention will be described below. An impulse generation trigger signal 12 is inputted as impulse magnetic noise to the input signal 11 detected by the magnetic sensor, and sampling is performed. The predicted value calculation unit 13 uses the past data until the impulse generation trigger signal 12 is input, and sequentially calculates the regression linear coefficients a and b by the following (Equation 1). (Equation 1) is a mathematical formula for calculating a regression linear coefficient. At the same time when the impulse generation trigger signal 12 is input, it is fixed to the latest regression linear coefficient, and is configured to output an approximate predicted value corresponding to the number of calculated data over time from the input of the impulse generation trigger signal 12. Approximate prediction for 11 is performed. In (Expression 1), N is the number of calculated data, x (i) is a data number in which the oldest past data defined by N is zero, and y (i) is an input signal corresponding to x (i). The approximate predicted value output from the predicted value calculation unit 13 is subtracted from the input signal 11 to output a noise estimation signal.

  In the present embodiment, the predicted value calculation unit 13 approximates the regression line. In this embodiment, it is assumed that the generation time of the impulse magnetic noise is short with respect to the low frequency signal. This is because even if linear approximation is performed, the influence on the input signal is very small. In addition to the approximation by the regression line, the effect of the present invention can be obtained by an approximate expression of a second order.

  As the number of calculated data with the passage of time from the input of the impulse generation trigger signal 12 increases, the noise estimation signal obtained by subtracting the approximate prediction value increases the delay error due to the time difference with respect to the input signal 11. In order to suppress an increase in the delay error, it is preferable to multiply the noise estimation signal by the window function coefficient calculated by the window function calculation unit 14 and interpolate the noise estimation signal with the window function. FIG. 3 is a diagram illustrating a window function coefficient output value by the window function calculation unit. The window function calculation unit 14 calculates the window function coefficient from the following (Equation 2) so that it changes from 1 to 0 during the impulse generation trigger signal 12 input, and the coefficient as shown in FIG. 3 is output. It is configured. The window function coefficient output value W (i) for the i-th data of the window width WN is expressed as (Equation 2). Here, the weight determination coefficient COE is a coefficient that changes the window function coefficient output value W (i) with respect to the number of elapsed data.

  The noise estimation interpolation signal obtained by interpolating the noise estimation signal obtained by subtracting the change due to the approximate prediction value from the input signal 11 and the window function coefficient output value calculated by the window function calculation unit 14 is obtained as the input signal 11. By subtracting from, it is possible to obtain the output signal 15 from which the impulse magnetic noise has been removed.

  The window function calculation unit 14 is configured to output zero when the impulse generation trigger signal 12 is not input, and therefore the input signal 11 is output to the output signal 15 without compensation.

  In this embodiment, the input signal sampling interval is 100 ms, the regression line coefficient calculation data number N in (Expression 1) is 2 to 100, the window width WN in (Expression 2) is 11 to 61, and the weight determination coefficient COE. 1 to 32, the input signal before noise removal and the output signal after noise removal were subjected to frequency analysis, the frequency-power spectrum characteristics were measured, and the noise removal level was calculated. The noise removal level is a value obtained by subtracting the maximum value of the output signal from the maximum value of the input signal mixed with the impulse generation trigger signal at 10 Hz. Table 1 shows the calculation result of the noise removal level under each condition.

  As can be seen from Table 1, according to the noise removal apparatus of the present invention, it can be confirmed that a noise removal effect of about −15 dB to −30 dB is obtained from the input signal before noise removal. Note that the same level of noise removal effect was obtained at frequencies other than 10 Hz.

  In the noise removal apparatus of the present invention, if the sampling interval of the input signal is less than 50 ms, an accurate signal cannot be reproduced, and if it is greater than 500 ms, the number of data increases more than necessary and signal processing takes time. 500 ms is desirable. When the number N of regression line coefficient calculation data in (Equation 1) is 100 or more, the prediction value calculation time becomes long, and therefore the interpolation error increases due to the influence of the low frequency signal, and the implementation effect of the present invention cannot be obtained. 2-100 are desirable. Further, the window width WN in (Equation 2) is preferably 21 to 61. When the window width WN is outside the range of 21 to 61, as can be seen from Table 1, the noise removal effect cannot be obtained. When the weight determination coefficient COE exceeds 32, it becomes closer to simple linear approximation interpolation and the interpolation error increases, and the effect of the present invention cannot be obtained.

  Further, the frequency when the sampling interval of the input signal is 100 ms, the number N of calculated regression line coefficient data in (Equation 1) is 5, the window width WN is 41 in (Equation 2), and the weight determination coefficient COE is 10. -The magnetic level was measured as a power spectrum characteristic and an amplitude characteristic in the time domain. FIG. 2 is a diagram showing a result of removing impulse-like magnetic noise, FIG. 2A is a frequency-power spectrum characteristic diagram, and FIG. 2B is a time-magnetic level characteristic diagram.

  As apparent from FIG. 2 (a), according to the noise removal apparatus of the present invention, a high noise removal level is obtained at all measured frequencies after noise removal. Further, as apparent from FIG. 2B, as a result of measuring the magnetic level before and after the magnetic noise removal, a noise removal effect for the impulse magnetic noise is obtained, and no signal delay occurs on the time axis. It could be confirmed.

  The noise removal apparatus of the present invention may be configured as a program on a computer, or at least a part may be configured as hardware. Needless to say, the present invention is not limited to the above-described embodiment, and the design values of the components can be changed according to the purpose and required performance.

DESCRIPTION OF SYMBOLS 1 Noise removal apparatus 11 Input signal 12 Impulse generation trigger signal 13 Predicted value calculation part 14 Window function calculation part 15 Output signal 21 Windowing part 22 FFT analysis part 23 Voice area detection part 24 Noise spectrum update part 25 Noise subtraction part 26 Negative coefficient Zeroing unit 27 IFFT synthesis unit 28 Synthesis unit by addition of windowed waveform

Claims (2)

A noise removing device that removes the impulse magnetic noise from an input signal in which impulse magnetic noise is mixed. A prediction value calculation unit that calculates an approximate prediction value obtained by approximating the past data that is an input signal with a primary or secondary approximation formula; and a window function calculation unit that calculates a window function coefficient value , the approximate prediction value A noise estimation interpolation signal is calculated by multiplying the noise estimation signal subtracted from the input signal by the window function coefficient value, and the impulse magnetic noise is removed by subtracting the noise estimation interpolation signal from the input signal. A noise removal device characterized by the above. The window function calculator, as a starting point the occurrence of the impulse magnetic noise, noise removal device according to claim 1, characterized in that outputs the window function coefficient value that varies continuously from 1 to 0.

Images