KR102170733B1 - Method for denoising of periodic vital signal and apparatus thereof - Google Patents

Abstract

An embodiment of the present invention is a noise removal method capable of effectively removing noise included in a biosignal having a first feature point and a second feature point within a period, and more specifically, Disclosed is a method of effectively removing noise by inputting a biosignal as time series data (time series data).

Description

BACKGROUND OF THE INVENTION 1. Field of the Invention Method for denoising of periodic vital signal and apparatus thereof

The present invention relates to a method and apparatus for removing noise from a periodic bio-signal, and more specifically, effectively removing noise from a bio-signal having a certain period among bio-signals that can be measured by a device from a person or an animal. It relates to a method and an apparatus thereof.

The business of measuring human bio-signals and providing customized medical services with information obtained from bio-signals is widely known. In order for the above-described customized medical service to be effectively provided to a visitor who visits a clinic or hospital, it is obvious that a bio-signal that can effectively represent the visitor's disease and disease information by minimizing noise must be measured.

However, the types of bio-signals that can be measured from humans vary, and noise is inevitably added to the bio-signals according to the adjacent organ or measurement environment in the process of measuring the bio-signals for a specific organ or region. A method of effectively removing noise from a biological signal and minimizing the loss of unique characteristic information of a biological signal has been studied for a long time.

As an example of a method of removing noise from a biosignal, there is a method of removing only a specific frequency component or leaving only a specific frequency component using a frequency filter. While the above method has the advantage that only the necessary frequency components can be left, a phase shift occurs for all or part of the biosignal, so even if it is restored to a time domain signal, it is In addition to the fact that information about the time point cannot be accurately known, there is a limitation in that it is logically impossible to remove noise if the frequency of the noise falls within the frequency range of the biological signal.

In addition, as another example of a method of removing noise from a bio-signal, there is a method of using a denoising autoencoder. The automatic noise reduction encoder is a module (or program) that operates with an algorithm made using the structure of an artificial neural network (ANN), and has the characteristic of learning so that the input and output of the artificial neural network implemented internally are the same. The automatic noise reduction encoder includes an artificial neural network composed of an input layer, a hidden layer, and an output layer, and is trained to reduce the difference between input and output, and training is repeated. As you do, you will grasp the characteristics of the input. After training for a certain bio-signal is completed, the automatic noise-reduction encoder outputs a bio-signal from which noise is removed when receiving a bio-signal mixed with noise.

However, since the automatic noise reduction encoder receives a parameter at a fixed point in time as an input and outputs an output value corresponding to that time by the characteristic operating based on a general artificial neural network, a biosignal having a time-series characteristic is subjected to specific preprocessing. There is a limitation that it cannot be input to the automatic noise reduction encoder without going through the process. In general, this is because the biosignal is not simply data that changes with the passage of time, but is expressed in a form in which the signal value of a certain past point affects the signal value of the future point of time rather than the point of time.

1. US Patent Publication No. 5339818 (registered on August 23, 1994) 2. US Patent Publication No. 9767557 (registered on September 19, 2017) 3. US Patent Publication No. 2017-0273632 (published on September 28, 2017)

1.Innovative continuous non-invasive cufless blood pressure monitoring based on photoplethysmography technology. Intensive Care Med, 39:1618-1625. Ruiz_rodriguez, et. al. (2013). 2. Improving photoplethysmographic measurements under motion artifacts using artificial neural network for personal healthcare. IEEE Transactions on Instrumentation and Measurement, doi: 10.1109/TIM. 2018.2829488 Roy et al. (2018).

The technical problem to be solved by the present invention is to provide a method and apparatus capable of effectively removing noise contained in a biosignal having at least two or more characteristic points within one cycle with a certain period among biosignals. In having to.

A method according to an embodiment of the present invention for solving the above technical problem includes a sampling step of obtaining a plurality of sampled signals by sampling a biosignal of a user having a predetermined period based on a preset sampling rate; A group separation step of separating the sampled signals into a learning group and a test group, and adding noise to the sampling signals included in the learning group; The noise-added sampling signals are arranged in the order of signal generation time, and are arranged in a denoising autoencoder module that sequentially performs encoding and decoding based on an artificial neural network (ANN). A learning step of learning the automatic noise reduction encoder module by inputting a sampling signal in a time series; And a test input step of removing noise added to the sampling signals included in the test group by adding noise to the sampling signals included in the test group, and inputting them to the automatic noise reduction encoder module in time series.

An apparatus according to another embodiment of the present invention for solving the above technical problem includes: a sampling unit for obtaining a plurality of sampled signals by sampling a biosignal of a user having a predetermined period based on a preset sampling rate; A group separation unit for separating the sampled signals into a learning group and a test group, and adding noise to the sampling signals included in the learning group; The sampling signal to which the noise is added is arranged in the order of signal generation time, and the alignment is performed in a denoising autoencoder module that sequentially performs encoding and decoding based on an artificial neural network (ANN). A module learning unit for learning the automatic noise reduction encoder module by inputting the sampled signal in a time series; And a test input unit configured to remove noise added to the sampling signal included in the test group by adding noise to the sampling signal included in the test group and inputting it to the automatic noise reduction encoder module in time series.

An embodiment of the present invention provides a computer-readable recording medium storing a program for implementing a method of removing noise from a periodic bio-signal.

According to the present invention, it is possible to effectively remove noise included in a biosignal having at least two or more characteristic points within one period and having a certain period among biosignals.

1 is a block diagram showing an example of an apparatus for removing noise from a bio-signal according to the present invention.
2 is a diagram schematically illustrating an example of a biosignal of a user that is a target sampled by a sampling unit.
3 is a diagram for explaining a process in which the module learning unit learns the automatic noise reduction encoder module.
4 to 8 are diagrams schematically showing noise added to a sampling signal included in a test group.
9 is a diagram schematically showing the operating characteristics of a bidirectional cyclic neural network automatic encoder.
10 is a diagram schematically showing the noise reduction performance according to the bidirectional cyclic neural network noise reduction automatic encoder.
11 is a diagram schematically illustrating an example of constructing a signal from which noise is removed through a sliding window method.
12 is a flowchart illustrating an example of a method of removing noise from a periodic bio-signal according to the present invention.

Since the present invention can apply various transformations and have various embodiments, specific embodiments are illustrated in the drawings and will be described in detail in the detailed description. Effects and features of the present invention, and a method of achieving them will be apparent with reference to the embodiments described later in detail together with the drawings. However, the present invention is not limited to the embodiments disclosed below and may be implemented in various forms.

Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings, and when describing with reference to the drawings, the same or corresponding components are assigned the same reference numerals, and redundant descriptions thereof will be omitted. .

In the following embodiments, terms such as first and second are used for the purpose of distinguishing one component from other components, not for a limiting meaning.

In the following embodiments, a singular expression includes a plurality of expressions unless the context clearly indicates otherwise.

In the following embodiments, the terms include or have means that the features or elements described in the specification are present, and do not preclude the possibility of adding one or more other features or components in advance.

When a certain embodiment can be implemented differently, a specific process order may be performed differently from the described order. For example, two processes described in succession may be performed substantially simultaneously, or may be performed in an order opposite to the described order.

1 is a block diagram showing an example of an apparatus for removing noise from a bio-signal according to the present invention.

Referring to FIG. 1, an apparatus 10 for removing noise from a biological signal according to the present invention includes a sampling unit 110, a group separation unit 130, a module learning unit 150, a test input unit 170, and a test verification unit. It can be seen that it includes 190 and a database 195. Depending on the embodiment, the test and verification unit 190 and the database 195 may be omitted.

The sampling unit 110 obtains a plurality of sampled signals by sampling the user's bio-signals based on a preset sampling rate. Various signals may be included in the user's bio-signals sampled by the sampling unit 110, and in order to maximize a learning effect according to a circulatory neural network to be described later, a signal that satisfies a plurality of preset conditions is preferable.

For example, the user's bio-signals sampled by the sampling unit 110 must have a certain period. The fact that the user's bio-signal has a certain period means that the part of the bio-signal that can be regarded as a feature point appears every predetermined time, so the bio-signal damaged in the process of removing noise and removing noise by using these characteristics It is easy to recover.

Here, the predetermined time may be a fixed point of time or a predetermined range of time. For example, a biosignal with a period of 10 seconds is a signal capable of removing noise by the present invention, and a biosignal with a period between 8 seconds and 12 seconds that changes with time is also a basic condition for effective noise reduction. Considered to be satisfied, noise reduction can be achieved by the present invention.

In addition, the user's bio-signal may include a first feature point and a second feature point for each period. In the bio-signal including the first feature point and the second feature point per period, it is preferable that the time interval between the first feature point and the second feature point is within a preset range.

2 is a diagram schematically illustrating an example of a biosignal of a user that is a target sampled by a sampling unit.

More specifically, FIG. 2 shows an example of a PPG signal according to a photoplethysmogram test, and the PPG signal refers to a systolic peak and a dicrotic notch as a first feature point and a second feature point. It is included as. The signal according to FIG. 2 is a signal segment obtained by only one cycle from the PPG signal according to the optical volumetric pulse wave test, and the PPG signal of an actual user shows a pattern in which the signal segment shown in FIG. 2 is continuously repeated.

The Systolic Peak 210 is the highest point in the PPG signal, and after the PPG signal reaches the Systolic Peak 210, it does not reach a point higher than the Systolic Peak 210 until the end of the corresponding period.

Diastolic Peak (250) is a characteristic point of the PPG signal observed after the Systolic Peak (210) appears, and after reaching the Systolic Peak (210), the PPG signal continues to descend to the Dicrotic notch (230), and then the Diastolic Peak (250). It has a signal characteristic that temporarily rises until it reaches ).

The reason that the PPG signal has the above signal characteristics is that when the ventricle contracts, when the pressure of the aorta is lower than the pressure of the aorta, the blood cannot open the meniscus, so pressure is applied to the atrium as well, and when the ventricle is relaxed, the valve This is because, as it is momentarily closed, a dicrotic notch must appear due to recoil. There are only differences between x(280) and y(290) depending on the measurement object. In the waveform of the PPG signal, the waveform of the PPG signal is Systolic. Peak 210, Dicrotic notch 230, Diastolic Peak 250 are sequentially observed, and Systolic Peak 210, Dicrotic notch 230, Diastolic Peak 250 are the first feature points or first features in the present invention. 2 It can be a feature point. Here, x(280) denotes a PPG signal value at the Systolic Peak 210, and y(290) denotes a PPG signal value at the Dicrotic notch 230 or Diastolic Peak 250.

In addition, the width 270, which is a temporal difference between the Systolic Peak 210 and the Dicrotic notch 230, is used for signal restoration when either of the Systolic Peak 210 and the Dicrotic notch 230 is not observed by noise. It means a possible time interval, and becomes an important factor in determining the user's bio-signals sampled by the sampling unit 110. By the sampling unit 110, the time interval between the Systolic Peak 210 and the Dicrotic notch 230 is limited to a value within a preset range.

In FIG. 2, for convenience of explanation, the PPG signal according to the optical volumetric pulse wave test is exemplarily described, but as described above, as a periodic biosignal including a first feature point and a second feature point, a first feature point It is obvious that any biological signal can be a sampling target if the time interval between the second feature points is within the range preset by the sampling unit 110. For example, in the biological signals that can be sampled by the sampling unit 110, electrocardiogram test ECG: electrocardiogram, seismic beat curve test (seismocardiogram), impedance cardiogram, cardiac ballistic test (BCG: ballistocardiogram), branching volume All signals according to a plethysmogram, an ultrasound, and an electroephalography may be included.

The sampling unit 110 obtains a plurality of sampled signals by sampling the bio-signals at a preset sampling rate. Here, the sampling rate may be 125Hz. The plurality of sampled signals denote each signal segment constituting a bio-signal. For example, if the total length of the bio-signal is 10 seconds and the sampling rate is 125 Hz, the number of signal segments of the bio-signal is 1250. The length and sampling rate of the bio-signals as described above are not fixed values and may vary according to embodiments. Hereinafter, the sampled signal and the signal segment will be used with the same meaning.

The group separation unit 130 divides the sampled signals into a learning group and a test group, and adds noise to the sampling signals included in the learning group. The learning group includes signal segments for training the neural network included in the automatic noise canceling encoder described later. The test group includes signal segments to test whether the training is actually valid after the neural network included in the automatic noise reduction encoder is trained by the signal segments included in the learning group.

The group separation unit 130 additionally undergoes an amplitude-normalized process by subtracting the minimum value of the signal belonging to the group and dividing it by the maximum value of the signal belonging to the group for the sampling signals included in the learning group. Can be processed. This normalization process can greatly improve the learning efficiency of the automatic noise reduction encoder.

Various noises may be included in noise that the group separation unit 130 adds to the sampled signals included in the learning group. As an example, the group separation unit 130 may add random noise to the sampled signals included in the learning group. As another example, the group separation unit 130 is white noise (white noise). noise), slipping noise, and saturation noise may be added to the sampled signals included in the learning group. According to an embodiment, a noise in which two or more of white noise, slope noise, and saturated noise are compositely superimposed may be treated the same as random noise as a combined noise. Sampled signals included in the learning group may be copied before noise is added, stored in the database 195, and then transmitted to the module learning unit 150 by a call to the module learning unit 150 to be described later.

The module learning unit 150 arranges the sampling signal to which noise is added in the order of signal generation time, and the sampling arranged in a noise reduction automatic encoder module that sequentially performs encoding and decoding based on an artificial neural network. By inputting signals in time series, the automatic noise reduction encoder module is trained. The module learning unit 150 physically or logically includes an automatic noise canceling encoder module, or is connected to the automatic noise canceling encoder module by wire or wirelessly.

Here, the noise reduction automatic encoder module is a state in which when the same signal is input to the input and output as described in the background art, the characteristics of the signal are trained by the ANN included in the automatic noise reduction encoder module, and the training is completed (learned state). It refers to a module that has a function of outputting a signal with no noise when the signal with noise is input to the input of the automatic noise reduction encoder.

3 is a diagram for explaining a process in which the module learning unit learns the automatic noise reduction encoder module.

Referring to FIG. 3, it can be seen that the artificial neural network included in the automatic noise reduction encoder module is composed of an input layer 330, a forward layer 350, a backward layer 370, and an output layer 390. A segment of a bio-signal is input to the input layer 330 in time series, and in FIG. 3, such a time series input operation is expressed through time steps 310.

The module learning unit 150 arranges the sampling signals to which noise is added in the order of signal generation time, and inputs them in a time series to the input layer 330 of the artificial neural network of the automatic noise reduction encoder module to provide the forward layer 350 and the backward layer. Through 370, the sampling signal added to the learning group is trained.

For example, if the biosignal is sampled at 125Hz by the sampling unit 110 and the signal segments included in the learning group are sampled at a length corresponding to 2 seconds of the user's biosignal, the number of signal segments becomes 250, and the most past time point. From the signal segment at time t-249, the signal segment at time t at the most recent time point is sequentially input to the input layer 330. Due to the basic operating characteristics of the automatic noise reduction encoder module, the signal segments (250) of the learning group before noise is mixed by the group separation unit 130 are input to the output layer 390 in time series.

In FIG. 3, the forward layer 350 and the backward layer 370 each consist of 40 nodes, and while training signal segments to which noise input in time series to the input layer 330 is added, the output layer ( Like the signal segments input to 390), signal segments from which noise has been removed are iteratively trained to be output. In FIG. 3, since the number of nodes of the forward layer 350 and the backward layer 370 is not limited to 40, it may be 20 or 60.

The Forward Layer 350 and the Backward Layer 370 constitute a hidden layer (not shown) to perform training to distinguish between noise and non-noise bio-signals among signal segments input in time series. According to FIG. 3, signal segments at various viewpoints are inputted in time series to the input layer 330 and output layer 390 with or without noise added and trained, so that the cyclic neural network is not a simple artificial neural network (ANN). It works as the principle of (RNN). RNN is an artificial neural network with a memory function added, and in that it can learn multiple signal segments inputted in time series order rather than a signal segment at one time point, the signal value at the past time point affects the signal value at the future time point. It has a characteristic that is very suitable for processing periodic bio-signals.

In FIG. 3, the number of nodes of the input layer 330 and the output layer 390 is limited to 250, but this is only an exemplary value, and according to an embodiment, the nodes of the input layer 330 and the output layer 390 The number can be less or more than 250.

The module learning unit 150 inputs signal segments to which noise has been added to the automatic noise reduction encoder, and ends the iterative learning process for signal segments included in the learning group when a signal segment from which noise has been removed by a certain level or more is output. Is done.

The test input unit 170 adds noise to the sampling signal included in the test group, and inputs the signal segments to which the noise is added to the automatic noise reduction encoder module in time series, so that the noise added to the sampling signal included in the test group is Remove. Specifically, when noise is added to the sampling signal included in the test group, the test input unit 170 timely inputs the signal segments to which the noise is added to the automatic noise reduction encoder module, thereby removing noise from the automatic noise reduction encoder module. Signal segments are output.

4 to 8 are diagrams schematically showing noise added to a sampling signal included in a test group.

Prior to describing FIGS. 4 to 8, it is assumed that the sampling signal included in the test group is a signal measured when a user performs a PPG test.

FIG. 4 is a diagram schematically illustrating one selected PPG segment when a PPG signal is sampled at a specific sampling rate and divided into a plurality of PPG segments. Referring to FIG. 4 with reference to FIG. 2, it can be seen that FIG. 4 shows an original PPG segment including two systolic peaks and a total of three dicrotic notches. I can.

5 schematically shows the result of applying white noise to the original signal PPG segment. More specifically, FIG. 5 shows a state in which a waveform having a large vibration rate is applied to the original signal PPG segment of FIG. 4 according to the inherent characteristics of white noise, and is changed to a signal segment having a different aspect from that of the original signal PPG segment.

6 schematically shows the result of applying the slope noise to the original signal PPG segment. More specifically, FIG. 6 schematically shows the original signal PPG segment in which the section showing the contraction peak and the aortic depression disappeared due to the inherent characteristics of the slope noise.

7 schematically shows the result of applying saturation noise to the original signal PPG segment. More specifically, FIG. 7 schematically shows the original signal PPG segment maintaining the saturation value by rapidly increasing to a saturation value after the second aortic indentation due to the inherent characteristics of saturation noise. After reaching the saturation value in FIG. 7, information on the original signal PPG segment is completely lost.

8 schematically shows the result of applying the composite noise to the original signal PPG segment. More specifically, the complex noise of FIG. 8 refers to noise in which at least two or more of white noise, slope noise, and saturation noise described through FIGS. 5 to 7 are mixed, and characteristics of noises included in the complex noise Accordingly, a completely different waveform from that of the original PPG segment is displayed.

The test input unit 170 generates a signal segment to which noise is added by adding at least one of white noise, slope noise, and saturation noise to the signal segment as shown in FIG. The input layer 330 of the automatic noise canceling encoder module learned by the module learning unit 150 is inputted in time series.

As an optional embodiment, if the test input unit 170 determines that noise added to the sampling signal included in the test group has been removed by exceeding a preset ratio by the learned noise reduction automatic encoder module, the test input unit 170 Noise may be removed by additionally receiving the sampled target biological signal and inputting the received target biological signal to the learned noise reduction automatic encoder module in time series.

In this optional embodiment, the fact that the noise added to the sampling signal included in the test group is removed by the learned noise reduction automatic encoder module exceeding a preset ratio is determined by the module learning unit 150. This means that it has been properly trained. Here, the preset ratio is a value that is previously stored in the test input unit 170 or stored in the database 195 and transmitted to the test input unit 170 in the process of the test input unit 170 making a logical decision. The preset ratio may be increased or decreased depending on how strictly the level of learning of the automatic noise reduction encoder module is set.

In addition, the target biological signal is actually noisy after the noise reduction automatic encoder module, which has been trained by the module learning unit 150, is validated by the test input unit 170 according to the above-described process. It means a biosignal that has been previously stored in the database 195 in order to remove or has been newly received after validity has been verified. The target biological signal is appropriately sampled through the sampling unit 110 and then inputted in time series to the automatic noise reduction encoder module, so that noise can be sufficiently removed.

The test and verification unit 190 outputs an output signal by the noise reduction automatic encoder module receiving the signal segment to which noise is added by the test input unit 170 as an input value, and the sampling signal included in the test group and the output signal are The test process is verified through a method of determining the degree of identity. The output signal output from the automatic noise reduction encoder module is a signal segment corresponding to the input signal segment and has the same length of time, and the test and verification unit 190 is the RMSE between the sampling signal and the output signal included in the test group. The verification process can be performed by calculating Root Mean Square Error) or correlation and comparing it with a preset value. According to an embodiment, the output signal output from the automatic noise reduction encoder module may be a signal obtained by combining signal segments.

As described above, when the signal segments included in the test group are input to the automatic noise reduction encoder module in time series by the test input unit 170, the test verification unit 190 removes the noise included in the signal segments. It verifies how well it has been removed, and performs a function to output the verified result to the user. The administrator who manages the apparatus 10 for removing noise from the bio-signal according to the present invention observes the result verified by the test and verification unit 190, and is included in each layer of the neural network included in the noise reduction automated encoder module. The number of nodes can be adjusted, or parameters for improving the learning algorithm of the module learning unit 150 can be more precisely adjusted.

The database 195 performs communication with the sampling unit 110, the group separation unit 130, the module learning unit 150, the test input unit 170, and the test verification unit 190, while learning a noise reduction automatic encoding module. It performs the function of temporarily storing and returning various data transmitted and received in the process of testing and verifying the signal segments included in the test group.

As described above, the noise canceling apparatus 10 according to the present invention is a device that effectively uses the conventionally known ANN-based noise canceling automatic encoder module, and has a first feature point and a second feature point within a certain period. Noise of the biosignal including the can be removed. More specifically, the present invention is a configuration that allows the operation of the automatic noise reduction encoder module to be performed by the RNN without limiting the operation to the ANN, and the biosignal is not input to the input layer of the neural network at once, but is measured in a time series data format. A device having a structure of inputting at least one or more samples for each time step is proposed.

In particular, when the size of the noise included in the bio-signal is large, the feature points of the bio-signal may be buried. A waveform feature accentuation that can be reconstructed by predicting where such feature points occur using temporal information. There is an advantage that the same function can be used in the automatic noise reduction encoder incorporating an RNN as in the present invention.

Hereinafter, an additional embodiment of the noise removal apparatus 10 according to the present invention will be described.

As an optional embodiment, if the biological signal sampled by the sampling unit 110 is a PPG signal, the first feature point may be an aortic dent, and the second feature point may be a contraction peak. According to this optional embodiment, when the bio-signal is a PPG signal, the aortic depression is considered to be positioned earlier in time than the contraction peak, and as the learning and testing steps of the circulatory neural network of the automatic noise reduction encoder module are performed, the test While observing the signal segment from which the noise has been removed by the verification unit 190, if the constriction peak is not found in a specific period and only the aortic dent exists, the signal component corresponding to the constriction peak is omitted in the process of removing the noise. It becomes easy to grasp that. That is, in the present optional embodiment, in the case of executing the noise reduction algorithm with the first feature point and the second feature point as contraction peak and aortic depression, respectively, when the result is not satisfactory, additional noise is removed. It may be one method performed to obtain a signal.

As another optional embodiment, the module learning unit 150 not only inputs signal segments arranged in chronological order into the input layer 330 of the automatic noise reduction encoder module in time series, but also inputs them in a counter time series. Through this, it is also possible to learn the automatic noise reduction encoder module. In that signal segments are not simply input in one direction, but signal segments in both directions are input to the input layer 330, the device implemented according to this optional embodiment is a bidirectional cyclic neural network automatic noise reduction encoder (BRDAE: Bidirectional Recurrent Denoising AutoEncoder).

9 is a diagram schematically showing the operating characteristics of a bidirectional cyclic neural network automatic encoder.

First, in the case of a Recurrent Denoising AutoEncoder (RDAE), signal segments arranged in chronological order are input to the automatic noise reduction encoder in the order of 0 to i (time series) and learned. In this case, 0 in the signal segment is considered to mean the most advanced time point.

Subsequently, in the case of the bidirectional cyclic neural network automatic noise reduction encoder according to the present exemplary embodiment, signal segments arranged in chronological order are not only input to the automatic noise reduction encoder in the order of 0 to i (time series), It is input and learned in the order from i to 0 (anti-time series).

As described above, according to the bidirectional cyclic neural network noise reduction automatic encoder, not only information that is subsequently input but also information that is input in the opposite direction can be applied and used. From the viewpoint of removing the noise of the biological signal, if a large noise is located at the earliest part of the signal segment, the cyclic neural network noise reduction automatic encoder removes the noise of the remaining parts included in the signal segment by this noise. Performance can decrease dramatically. This is because the noise located at the front disturbs the time sequence information of the signal segment, making it impossible to grasp the characteristic points of the signal segment.

As an example, when a very large noise is mixed from the first sample to the systolic peak in the signal segment of the optical volume pulse wave, the contraction peak itself cannot be identified, and the time heat information before and after the contraction peak is meaningless. As such, it may become impossible to reconstruct the contraction peak with the existing unidirectional circulatory neural network noise reduction automatic encoder.

In the above situation, according to the bidirectional circulatory neural network denoising automatic encoder, as the circulatory neural network learns signal segments in both directions, when a certain signal segment is temporally forward, a certain time elapses after the constriction peak appears, In addition to the appearance of the dent, the information that the constriction peak appears after the appearance of the aortic dent when in the reverse direction can be considered in restoring and reconstructing the signal segment. In the above-described embodiment, for convenience of description, the signal segment is described as a PPG segment, and the first and second feature points are described as constriction peaks and aortic dents. It will be apparent to a person skilled in the art that it can be a segment of a signal.

10 is a diagram schematically showing the noise reduction performance according to the bidirectional cyclic neural network noise reduction automatic encoder.

Referring to FIG. 10, an output signal 1010 from which noise is removed is a signal according to a PPG that vibrates irregularly, and an output signal 1030 before noise is removed is a complex noise similar to FIG. You can see that this is an added form of signal.

Referring to FIG. 10, there is little external similarity between the output signal 1010 from which noise is removed and the output signal 1030 before noise is removed. However, noise in the output signal 1030 before noise is removed. As the characteristic points of the removed output signal 1010 are not only included, but also the noise is removed from the output signal 1030 before the noise is removed according to the learned information of the bidirectional cyclic neural network noise reduction automatic encoder, finally An output signal 1010 from which noise has been removed can be obtained.

The test verification unit 190 may verify similarity between the output signal 1010 from which noise is removed and the output signal 1030 before noise is removed according to the present optional embodiment through various methods. For example, if the biosignal is a PPG signal according to the optical volumetric pulse wave test, the test verification unit 190 calculates a heart rate (HR) based on the PPG signal from which noise has been removed, and calculates a heart rate according to the calculated PPG signal. By comparing the heart rate with the heart rate according to the previously stored electrocardiogram (ECG) and calculating the correlation and RMSE (Root Mean Square Error), it is possible to verify the noise removal rate for the biosignal of the present invention. .

Denoising Method None Wavelet DAE RDAE BRDAE Correlation 0.408±0.301 0.406±0.304 0.558±0.238 0.620±0.254 0.686±0.147 RMSE 14.8±15.7 14.4±15.0 6.2±4.9 7.5±10.7 6.1±8.2

Table 1 is a table showing the results of the test verification unit 190 removing noise from the PPG signal to which noise was added according to various noise removal methods.

Referring to Table 1, the method according to the present invention (RDAE, BRDAE) is more effective than the case in which noise of the PPG signal is removed by no noise reduction (None) or by a known method (Wavelet, DAE: Denoising AutoEncoder). Accordingly, it can be seen that the case of removing the noise has a higher correlation and a lower RMSE. That is, when a periodic bio-signal is denoised by using information from a past point of time like a circulatory neural network, the noise reduction effect due to the denoising is superior to that of conventionally known methods. In addition, it can be seen that when noise is removed by using a bidirectional non-unidirectional cyclic neural network noise reduction automatic encoder, a higher correlation and lower RMSE are shown than when noise is removed using a unidirectional cyclic neural network noise reduction automatic encoder. .

As another alternative embodiment from the above, the test input unit 170 is included in the test group through a sliding window method based on an input sampling rate lower than the sampling rate used for sampling by the sampling unit 110. The sampled signal is inputted in a time series to the automatic noise reduction encoder module, and the test verification unit 190 superimposes the restoration based on the input sampling rate, and compares the generated output signal with the sampling signal included in the test group to be identical. Can be verified.

11 is a diagram schematically illustrating an example of constructing a signal from which noise is removed through a sliding window method.

Even through the above-described bidirectional circulatory neural network noise reduction automatic encoder, there may be a case in which noise for a biosignal is not effectively removed. In this case, the signal segments are overlapped in a sliding window method and input to the input layer of the neural network of the automatic noise reduction encoder module, and the output signal is calculated by averaging after overlapping the output according to the input. High noise reduction effect can be expected.

Referring to FIG. 11, a PPG signal of 125 Hz is input to a bidirectional circulatory neural network automatic noise reduction encoder (BRDAE) in units of 2 seconds (250 samples) at a time, and a sliding window is 50 samples, thereby removing the noise. The signal can be acquired. Since the input signal segment is 250 samples and the sliding window is 50 samples, the overlapping portion between the input signal segments corresponds to a maximum of 200 samples, and noise is removed by averaging the overlapping portion. The signal can be obtained.

11 is a method of taking an average value of only the overlapping portions of the segments 1110 from which noise has been removed through a sliding window method by a bidirectional cyclic neural network noise reduction automatic encoder, so that the biosignal according to the average result 1130 is The resulting process is schematically shown. In FIG. 11, the biosignal according to the average result 1130 is a PPG signal according to the input signal, but if a biosignal has a certain period and has a first feature point and a second feature point per cycle, the biosignal is not a PPG signal. It will be apparent to those skilled in the art that a method of calculating a bio-signal from which noise has been removed can also be applied through a sliding window method.

12 is a flowchart illustrating an example of a method of removing noise from a periodic bio-signal according to the present invention.

Since the method according to FIG. 12 can be implemented by the device 10 for removing the noise of the periodic bio-signal of FIG. 1, it will be described with reference to FIG. 1, and hereinafter, the method described with reference to FIGS. 1 to 11 Duplicate description will be omitted.

First, the sampling unit 110 obtains a plurality of sampled signals by sampling periodic bio-signals having a first feature point and a second feature point (S1210).

The group separation unit 130 divides the sampled signals into a learning group and a test group, and adds noise to the signals included in the learning group (S1230).

The module learning unit 150 trains the bio-signals by inputting signals included in the learning group in a time-sequential manner to the automatic noise reduction encoder module (S1250).

The test input unit 170 adds noise to the signal included in the test group, and inputs it to the automatic noise reduction encoder module in time series (S1270).

The test verification unit 190 compares the output signal output from the automatic noise reduction encoder module with the sampling signal included in the test group to verify the identity (S1290).

The embodiment according to the present invention described above may be implemented in the form of a computer program that can be executed through various components on a computer, and such a computer program may be recorded in a computer-readable medium. In this case, the medium is a magnetic medium such as a hard disk, a floppy disk, and a magnetic tape, an optical recording medium such as a CD-ROM and a DVD, a magneto-optical medium such as a floptical disk, and a ROM. A hardware device specially configured to store and execute program instructions, such as, RAM, flash memory, and the like.

Meanwhile, the computer program may be specially designed and configured for the present invention, or may be known and usable to those skilled in the computer software field. Examples of the computer program may include not only machine language codes produced by a compiler but also high-level language codes that can be executed by a computer using an interpreter or the like.

The specific implementations described in the present invention are examples and do not limit the scope of the present invention in any way. For brevity of the specification, descriptions of conventional electronic configurations, control systems, software, and other functional aspects of the systems may be omitted. In addition, the connection or connection members of the lines between the components shown in the drawings exemplarily represent functional connections and/or physical or circuit connections, and in an actual device, various functional connections that can be replaced or additionally Connections, or circuit connections, may be represented. In addition, if there is no specific mention such as “essential” or “importantly”, it may not be an essential component for the application of the present invention.

In the specification of the present invention (especially in the claims), the use of the term “above” and a similar reference term may correspond to both the singular and the plural. In addition, when a range is described in the present invention, the invention to which an individual value falling within the range is applied (unless otherwise stated), and each individual value constituting the range is described in the detailed description of the invention. Same as Finally, if there is no explicit order or contrary to the steps constituting the method according to the present invention, the steps may be performed in a suitable order. The present invention is not necessarily limited according to the order of description of the steps. The use of all examples or illustrative terms (for example, etc.) in the present invention is merely for describing the present invention in detail, and the scope of the present invention is limited by the above examples or illustrative terms unless limited by the claims. It does not become. In addition, those skilled in the art can recognize that various modifications, combinations, and changes may be configured according to design conditions and factors within the scope of the appended claims or their equivalents.

Claims (19)

A sampling step of obtaining a plurality of sampling signals by sampling a biosignal of a user having a predetermined period based on a preset sampling rate;
A group separation step of separating the sampling signals into a learning group and a test group, and adding noise to the sampling signals included in the learning group;
The noise-added sampling signals are arranged in the order of signal generation time, and are arranged in a denoising autoencoder module that sequentially performs encoding and decoding based on an artificial neural network (ANN). A learning step of learning the automatic noise reduction encoder module by inputting a sampling signal in a time series; And
Periodic living body including a test input step of removing noise added to the sampling signal included in the test group by adding noise to the sampling signal included in the test group, and inputting it to the automatic noise reduction encoder module in time series. How to denoise the signal. The method of claim 1,
The above method,
A method of removing noise from a periodic bio-signal further comprising a test verification step of verifying the identity of the output signal output from the noise-removing automatic encoder module and the sampling signal included in the test group. The method of claim 1,
The noise is,
A method for removing noise from a periodic bio-signal, characterized in that at least one of white noise, slloping noise, and saturation noise The method of claim 1,
The biosignal is
A method for removing noise from a periodic bio-signal, characterized in that it is a PPG (photoplethysmogram) signal calculated as a result of performing an optical volume pulse wave test on the user. The method of claim 4,
The biosignal includes a first feature point and a second feature point within a period,
The first feature point is a systolic peak,
The second feature point is a method for removing noise from a periodic bio-signal, characterized in that the aortic notch. The method of claim 4,
The biosignal includes a first feature point and a second feature point within a period,
The first feature point is an aortic notch,
The second feature point is a method for removing noise from a periodic bio-signal, characterized in that a systolic peak. The method of claim 1,
The learning step,
In addition to inputting the sorted sampling signal in a time series, and inputting the sorted sampling signal in a reverse time series to learn the noise reduction automatic encoder module, the noise of the periodic bio-signal is reduced. How to remove. The method of claim 2,
The test input step,
Inputting the sampling signal included in the test group into the automatic noise reduction encoder module in time series through a sliding window method based on an input sampling rate lower than the sampling rate,
The test verification step,
A method of removing noise from a periodic biosignal for verifying identity by comparing the output signal generated by overlapping restoration based on the input sampling rate with a sampling signal included in the test group. The method of claim 1,
The test input step,
If it is determined that the noise added to the sampling signal included in the test group has been removed by exceeding a preset ratio by the learned noise reduction automatic encoder module, the target biological signal sampled in the sampling step is additionally received,
And removing noise by inputting the received target biological signal to the learned automatic noise reduction encoder module in a time series to remove noise. A computer-readable recording medium storing a program for executing the method according to any one of claims 1 to 9. A sampling unit for obtaining a plurality of sampling signals by sampling a biosignal of a user having a certain period based on a preset sampling rate;
A group separation unit for separating the sampling signals into a learning group and a test group, and adding noise to the sampling signals included in the learning group;
The sampling signal to which the noise is added is arranged in the order of signal generation time, and the alignment is performed in a denoising autoencoder module that sequentially performs encoding and decoding based on an artificial neural network (ANN). A module learning unit for learning the automatic noise reduction encoder module by inputting the sampled signal in a time series; And
Periodic biosignal comprising a test input unit for removing noise added to the sampling signal included in the test group by adding noise to the sampling signal included in the test group and inputting it to the automatic noise reduction encoder module in time series. Device to remove noise from The method of claim 11,
The device,
An apparatus for removing noise from a periodic bio-signal further comprising a test verification unit for verifying the identity of the output signal output from the automatic noise reduction encoder module and the sampling signal included in the test group. The method of claim 11,
The noise is,
An apparatus for removing noise from a periodic bio-signal, characterized in that at least one of white noise, sloping noise, and saturation noise. The method of claim 11,
The biosignal is
An apparatus for removing noise from a periodic bio-signal, characterized in that it is a PPG (photoplethysmogram) signal calculated as a result of performing an optical volumetric pulse wave test on the user. The method of claim 14,
The biosignal includes a first feature point and a second feature point within a period,
The first feature point is a systolic peak,
The second characteristic point is an aortic notch. The method of claim 14,
The biosignal includes a first feature point and a second feature point within a period,
The first feature point is an aortic notch,
The device for removing noise from a periodic bio-signal, characterized in that the second feature point is a systolic peak. The method of claim 11,
The module learning unit,
In addition to inputting the sorted sampling signal in a time series, and inputting the sorted sampling signal in a reverse time series to learn the noise reduction automatic encoder module, the noise of the periodic bio-signal is reduced. Device to remove. The method of claim 12,
The test input unit,
Inputting the sampling signal included in the test group into the automatic noise reduction encoder module in time series through a sliding window method based on an input sampling rate lower than the sampling rate,
The test verification unit,
An apparatus for removing noise from a periodic biosignal for verifying identity by comparing the output signal generated by overlapping restoration based on the sampling rate with a sampling signal included in the test group. The method of claim 18,
The test input unit,
When it is determined that noise added to the sampling signal included in the test group has been removed by exceeding a preset ratio by the learned noise reduction automatic encoder module, the sampled target biological signal is additionally received from the sampling unit,
A device for removing noise from a periodic bio-signal, characterized in that the received target bio-signal is inputted to the learned automatic noise-reduction encoder module in time series to remove noise.

Images