JP2019017406A - Noise removal device of kinetic recall output signal of time series data, noise removal method, program, analysis device of respiration, heart beat and voice by cluster analysis, analysis method and program - Google Patents

Abstract

To achieve a noise removal with high accuracy.SOLUTION: Input data from a sensor is recalled at real time at weight of teacher data which is reference and which has been created in advance by machine learning, to acquire data from which noise is completely removed. Time series data for unit time acquired from a sensor 1 is subjected to machine learning in advance, for determining which state data in the unit time is in. Therefore, teacher data which is reference for allowing kinetic recall of information being an object of noise removal and a division weight 7 are created, and the division weight 7 is used, to acquire data having no noise by recall at real time of the data input from the sensor 1. Then, the data recalled at real time is classified into clusters to analyze a user's state. In the cluster analysis, the data is classified by numerical values into normal/abnormal state, noise/voice, heart beat/sleep and the like, and determination with high accuracy is performed by a threshold which is a determination reference determined in advance, and an analysis result is displayed on a display.SELECTED DRAWING: Figure 5

Description

  The present invention relates to a noise removal apparatus, a noise removal method, a program, and a cluster analysis for breathing, heartbeat, speech, etc., an analysis apparatus, an analysis method, etc. It is about the program.

  As this type of method, there is a specific frequency detection technique using a band-pass filter. In noise removal, there is a noise removal technique such as a hearing aid that matches the opposite phases with the in-phase phase shifted by 180 degrees.

  In Patent Document 1, an input signal is divided into a signal in a high frequency band and a signal in a low frequency band by a band pass filter and a second low pass filter. The first calculation unit obtains the level change amount of the supplied signal. The second calculation unit extracts a noise component from the supplied signal. The third calculation unit and the fourth calculation unit generate an attenuation signal for attenuating noise from the extracted noise component. The first multiplication unit multiplies the attenuation signal and the high frequency band signal to attenuate the noise. The second multiplication unit multiplies the attenuation signal and the low frequency band signal to attenuate the noise. It has been proposed that the adding unit adds the high frequency band signal from which noise has been removed and the low frequency band signal and outputs the result to the outside.

  In Patent Document 2, in a hearing aid, a receiving array including a plurality of channel microphones is arranged on a frame of a spectacles-type wearing tool or a support beam of a headphone-type wearing tool. The output of each microphone is wavelet transformed to obtain the amplitude and phase spectrum, and the following operation is performed for each frequency band. The phase difference between each channel is multiplied by an arbitrary coefficient to expand the difference, and the number of channels is increased to an arbitrary size by interpolating the amplitude and phase between each channel. These outputs are multiplied by an arbitrary weight function and phased and added, and wavelet inverse transformation is performed to obtain a conventional amplifier input. It has been proposed that by increasing the phase difference between each channel and multiplying the number of channels, the directivity width is reduced and the external noise removal performance is improved.

JP2004-096185 JP2002-095084

  Conventionally, noise removal using analog and digital filters has been the main method for removing noise from raw data entering the time series from the sensor. It is possible to obtain numerical analysis such as frequency analysis by FFT (Fast Fourier Transform), waveform differential value, integral value, and calculation of the number of crossovers at a certain threshold in the digital analysis of the digital filter by computer. The accuracy was limited.

  According to the first invention of the present application, input data from a sensor is divided into data per unit time, a learning unit that generates teacher data and division weights from the divided data, and stores the data, and input data from the sensor Is divided into data per unit time, and a noise removal apparatus is provided that includes a recall unit that calculates input data by recall using the divided weights for the divided data.

  According to the second aspect of the present application, the learning unit includes a sensor that converts input data such as sound and vibration into an electrical signal, a sensor amplifier that amplifies the electrical signal from the sensor, and the electrical signal as analog data. Performs machine learning on analog / digital converters that convert data into digital data, bandpass filters that cut specific frequencies, and data that is divided per unit time, and generates and stores reference teacher data and division weights There is provided a noise removal device comprising a machine learning device.

According to the third invention of the present application, the recalling unit includes a sensor that converts input data such as sound and vibration into an electrical signal, a sensor amplifier that amplifies the electrical signal from the sensor,
An analog / digital converter that converts electrical signals from analog data to digital data;
There is provided a noise removal device including a band-pass filter that cuts a specific frequency, and a recollecting unit that calculates input data by recollection using the divided weights for data divided per unit time.

  According to the fourth invention of the present application, there is provided a noise removal device, a determination device that classifies the data from which noise has been removed into a hyperdimensional cluster according to the number of parameters, and a display device that displays the determination result. Is provided.

  According to the fifth invention of the present application, a learning process for dividing input data from a sensor into data per unit time, generating teacher data and divided weights from the divided data, and storing the data, and input data from the sensor Is divided into data per unit time, and a recalling process is provided which includes a recalling step of calculating input data by recalling the divided data using the split weights.

  According to the sixth invention of the present application, there is a noise removal method, a determination method for classifying the data from which noise has been removed into a hyperdimensional cluster according to the number of parameters, and a display method for displaying the determination result An analysis method is provided.

  According to the seventh invention of the present application, the computer divides input data from the sensor into data per unit time, generates teacher data and divided weights from the divided data, and stores the learning unit, the sensor A program is provided for functioning as a noise removal device comprising a recall unit that divides input data into data per unit time and calculates the input data by recall using the divided weights for the divided data Is done.

  According to the eighth invention of the present application, there is provided a program for causing a computer to function as a noise removal device, wherein the determination unit classifies the data from which noise has been removed into a superdimensional cluster according to the number of parameters; A program for causing an analyzer to function as a display device that displays a result of the determination is provided.

  In the past, the fact that frequency analysis, which has a little difficulty in noise removal and FFT accuracy, has been perfect, and now it is possible to grasp the state that is assumed in advance and the frequency analysis with high accuracy. Therefore, even a very small heartbeat waveform can separate a noise waveform with a small amplitude similar to that.

  With existing hearing aids, it is difficult to distinguish between speech and noise, and speech cannot be extracted, making it very difficult to use. By using the technique according to the present invention, not only can noise be removed, but voice can be extracted and amplified. It becomes closer to the human ear.

It is a block diagram which shows the process of the machine learning for the noise removal of sensor input data. It is a flowchart figure which shows the process of the machine learning for the noise removal of sensor input data. It is a block diagram which shows the process of removing noise from the data containing noise by real-time recall. It is a flowchart figure which shows the process of removing noise from the data containing noise by real-time recall. It is a block diagram which shows the method of classifying the data recalled in real time and analyzing a user's state. It is a flowchart figure which shows the method of classifying the data recalled in real time and analyzing a user's state. It is an example of the processing screen of the program concerning the present invention. It is a flowchart figure of the program which concerns on this invention. This is an example of a screen for machine learning of data such as noise, turning over, and nobody. It is a flowchart figure of the screen which carries out machine learning of data, such as noise, turning over, and nobody.

  Example 1 is a monitoring device. In human monitoring devices, breathing and heart rate are measured by a flat sensor placed between the back and bed or seat when sleeping or driving a car. Can be measured without being disturbed by user behavior or automobile noise. There is no restriction on the form of the sensor, and the optimum form for measurement can be selected. In the pet watching device, it is possible to detect an abnormality by accurately measuring the breathing, heartbeat, etc. of the pet during sleep.

  FIG. 1 shows a machine learning process for denoising sensor input data. Here, noise is not only physical and electronic noise but also logical noise, that is, when heartbeat and breathing information is detected at the same time, breathing is considered as noise if only heartbeat is to be detected. .

  In the first embodiment, when the time series data of a certain unit time (1 second in the present embodiment) is sampled (10 times / one second in the present embodiment), the state of the data of the one second is determined in advance. Let me learn machine. Machine learning is performed by shifting 10 times of data every second, and 600 types of machine learning are performed per minute.

  In the machine learning data, a discrimination code is assigned by a human hand for each item to be classified, such as noise, heartbeat, breathing, voice, and abnormal state. Prepare the necessary amount of data for machine learning and repeat machine learning.

  The sensor 1 uses a sensor capable of detecting information to be machine learning target. The information includes information such as sound, vibration, temperature, humidity, atmospheric pressure, and illuminance. Data such as lung sounds, muscle movements, and heartbeats within a predetermined time (for example, 60 seconds) is acquired for each normal / abnormal state of the user, voice / noise, wakefulness / sleep, and the like.

  FIG. 2 is a flowchart showing a machine learning process for removing noise from sensor input data. Data acquired by the sensor 1 (step 11) is amplified by the sensor amplifier 2 (step 12), analog data is converted into digital data by the analog / digital converter 3 (step 13), and then machine learning is performed by the bandpass filter 4. A specific frequency that is not used in the process is physically passed (step 14). The input data 5 is set in the machine learning device (step 15), and machine learning is performed for recognizing target information by dividing every unit time (step 16).

  For example, machine learning is performed by dividing data measured for one minute by unit time and dividing it into 10 data per second and 600 data per minute. Any unit time can be set as the unit time. The machine learning 6 by the machine learning device is performed by class classification based on a super-dimensional multilayer perceptron theory, and an optimal division weight is learned (step 17).

  In the present invention, as a result of the machine learning 6, the machine learning 6 is performed on the data divided every unit time, and reference teacher data and division weights 7 are created (FIGS. 1 and 2). As a result, it is possible to create the teacher data and the division weight 7 as a reference that makes it possible to dynamically recall information to be subjected to noise removal. The created teacher data and the divided weight 7 are once stored in a storage device (not shown).

  3 and 4 show a process of removing noise from the data containing noise by the real-time recall 8. [0026] By using the teacher data and the division weight 7 which are the reference created in [0025] and recalling the data inputted from the sensor 1 in real time, the recalled data 9 without noise is obtained.

  This will be specifically described with reference to the flowchart of FIG. First, the data acquired by the sensor 1 (step 21) is amplified by the sensor amplifier 2 (step 22), the analog data is converted into digital data by the analog / digital converter 3 (step 23), and then the band pass filter 4 is used. A specific frequency not used for machine learning is physically passed (step 24). Divide digital data into the same unit time used in machine learning. In the example of [0024], the data measured for 1 minute is divided by the unit time and divided into 10 data for 1 second and 600 data for 1 minute and set (step 25).

  Next, the machine-learned teacher data closest to the acquired data is recalled (step 27). In this case, the divided data comes in 600 times per minute. The number of the acquired data is determined by comparing with the machine-learned data generated in advance by shifting the acquired data every second. If the acquired data is data close to the number 100 assigned to the teacher data, 100 is recognized, and if the acquired data is close to 200, 200 is recognized. The method of recognizing the closest number of teacher data is performed by reading out the teacher data created in [0025] and the division weight 7 from the storage device and recalling the acquired data for each unit time using the division weight 7. . If the value of the acquired data is 95, it is closest to the teacher data having a value of 100, so that 100 teacher data is recalled for 95 acquired data. The recalled teacher data is arranged in the chronological order of the acquired data.

  In this example, the above process is performed on 600 pieces of data per minute. The number of data to be processed can be arbitrarily set. This processing can be performed in real time. For this reason, noise is removed from data including noise in real time.

  In the configuration from the sensor 1 to the machine learning 6 in the present invention, the same device may be used in the machine learning step and the real-time recall step, or different devices may be used (FIGS. 2 and 4).

  FIG. 5 and FIG. 6 show a method of classifying the data obtained by the real-time recall 8 into the cluster table 10 and analyzing the state of the user. The cluster in FIG. 5 is expressed in two dimensions for easy understanding. The actual cluster is hyperdimensional according to the number of parameters. The cluster concept here is used for facilitating understanding of the invention, and the actual analysis result 11 is processed inside the analysis apparatus without displaying the cluster.

  This will be specifically described with reference to the flowchart of FIG. Since steps 31 to 37 are the same as the recall stage, the description is omitted. The data from which noise is removed is classified into hyperdimensional clusters according to the number of parameters. For example, conditions such as the degree of irregularity of breathing, the speed of breathing, and the number of heartbeats are classified into clusters according to the number of parameters (step 38).

  In the cluster analysis, numerical values are classified into normal / abnormal, noise / voice, heartbeat / sleep, etc., and a high-accuracy judgment is made based on a threshold value based on a predetermined criterion (step 39).

  In the present invention, the data is divided into unit time, and the data input from the sensor and the teacher data prepared in advance are recalled. Therefore, the cluster is not only a specific unit time but also the next unit time, The state of the recalled data according to the passage of unit time such as the next unit time can be comprehensively determined according to the passage of time. For this reason, a user's state can be grasped very accurately. While breathing for 1 minute, it is possible to grasp a normal state such as normal breath 15, turn over detection 16, apnea determination 17, noise 18, 19 in which there is no response, or a state in which no one is present.

  The determination is performed by setting a threshold value for the parameter. The degree to which the input data from the sensor is far from the teacher data is set in advance. If the distance from the teacher data is large, the determination is abnormal.

  The user's condition is comprehensively determined and the analysis result is displayed on a display device (not displayed). The analysis result display method may be displayed on the display with characters, illustrations, or the like, or may be expressed with sound, voice, vibration, or the like.

You may acquire temperature, humidity, atmospheric pressure, and illuminance as parameters. By measuring breathing, heartbeat, etc., temperature, humidity, barometric pressure, and illuminance at the same time, the user's condition due to the influence of weather can be grasped, and it can be used for meteorological medicine. Temperature, humidity, barometric pressure, and illuminance are acquired as parameters, and the relationship between sleep, heartbeat regularity, and irregularity is machine-learned. The recall method is the same as the method described in [0020] to [0030].
The weather data may be obtained from statistical analysis.

  7 and 8 are examples of processing screens for programs according to the present invention. The horizontal axis represents time, and the vertical axis represents the movement of the heart and lung muscles. The difference in waveform from FIG. 1 is due to the difference in program and unit time. In FIG. 7, 12 represents the state of the heartbeat when sleeping, and 13 represents the movement of the chest muscles. The flowchart of FIG. 8 represents setting of unit time for data acquisition (steps 41 and 42) and a graph after processing (step 43).

  FIG. 9 and FIG. 10 are examples of a program screen for machine learning of data such as normal breathing, turning over, noise, and nobody. The flowchart of FIG. 10 shows that machine learning is performed for conditions such as normal breathing, turning over, apnea determination, and noise for analysis (step 51) (from step 52 to step 56).

Example 2 is a hearing aid. Pick up the voice with a microphone even in a noisy place, and separate and extract the necessary voice information from it.
Conventional hearing aids are noisy and difficult to use. On the other hand, the human ear can distinguish human voice from noise. This is done by machine learning in the hearing aid.

In the prior art, since the hearing aid was intended to reduce noise by enhancing any sound among the acquired sounds, the noise could not be completely removed. According to the present invention, unnecessary sound can be completely removed as noise, and only desired sound can be reproduced. As teacher data, sounds that need to be heard, such as one's own voice and relative's voice, are sampled and machine-learned to create reference teacher data and division weights. On the other hand, by recalling the sound input from the sensor in real time, only the sound that needs to be heard can be taken out and reproduced, and a hearing aid free from noise can be provided to the elderly and the like.
Since each figure of Example 2 is the same as that of Example 1, each figure of Example 1 is used.

The present invention can be applied to a monitoring device, a medical device, a vehicle driver state detector, and the like.
In addition to health management in individual homes, the monitoring device can be used for personal health management by installing sensors in nursing homes and hospitals.
In addition, it is useful for telemedicine because you can check your health.
Furthermore, the watching device is useful for watching pets with few veterinarians.

1. Sensor 2. Sensor amplifier 3. Analog / digital converter 4. Band pass filter 5. Input data 6. Machine learning 7. Divided weight 8. Real time recall 9. Post recall data 10. Cluster table 11. Analysis results 12.1 minute scale 13.3 seconds Machine learning scale 14.1 hour scale 15. Normal breath 16. Rolling detection 17. Apnea judgment 18. Noise 19. No response

Claims (8)

A learning unit that divides input data from the sensor into data per unit time, generates teacher data and division weights from the divided data, and stores the data;
Recalling unit that divides input data from a sensor into data per unit time, and calculates input data by recollection using the divided weight for the divided data;
A noise elimination device composed of The learning unit
A sensor that converts input sound, vibration, and other data into electrical signals;
A sensor amplifier that amplifies the electrical signal from the sensor;
An analog / digital converter that converts electrical signals from analog data to digital data;
A bandpass filter that cuts off a specific frequency;
A machine learning device that performs machine learning on data divided every unit time, generates reference teacher data and division weights, and stores them;
The noise removal apparatus according to claim 1, further comprising: The recall part is
A sensor that converts input sound, vibration, and other data into electrical signals;
A sensor amplifier that amplifies the electrical signal from the sensor;
An analog / digital converter that converts electrical signals from analog data to digital data;
A bandpass filter that cuts off a specific frequency;
A recall unit for calculating input data by recall using the divided weights for data divided every unit time; and
The noise removal apparatus according to claim 1, further comprising: The noise removal device according to claim 1,
A determination device that classifies the data from which noise has been removed into a hyperdimensional cluster according to the number of parameters;
A display device for displaying the determined result;
An analysis device further comprising: A learning step of dividing input data from the sensor into data per unit time, generating teacher data and divided weights from the divided data, and storing them,
Recall process of dividing input data from the sensor into data per unit time, and calculating the input data by recall using the divided weight for the divided data;
A noise removal method comprising: The noise removal method according to claim 5, comprising:
A determination method for classifying the noise-removed data into hyperdimensional clusters according to the number of parameters,
A display method for displaying the judgment result;
An analysis method further comprising: Computer
A learning unit that divides input data from the sensor into data per unit time, generates teacher data and division weights from the divided data, and stores the data;
Recalling unit that divides input data from a sensor into data per unit time, and calculates input data by recollection using the divided weight for the divided data;
A program for functioning as a noise removal device composed of Computer
A program for causing a noise removal device according to claim 7 to function.
A determination unit that classifies the data from which noise is removed into a hyperdimensional cluster according to the number of parameters;
A display for displaying the result of the determination;
A program for functioning as an analysis device further comprising: