KR102370622B1 - Signal compression system and method based on deep learning - Google Patents

Abstract

a measuring device for recording the measuring signal; and performing pre-processing to generate a regular signal, generate a compressed signal from the regular signal using a weight, generate a reconstructed signal from the compressed signal using a weight, and a difference between the normal signal and the reconstructed signal is provided in advance Provided is a signal compression system comprising an auto-encoder device that learns weights to satisfy a threshold range.

Description

Signal compression system and method based on deep learning

The present invention relates to a signal compression system and method based on deep learning, and more particularly, to a signal compression system and method for compressing and restoring a signal using one or more encoding blocks and one or more decoding blocks.

In addition, the present invention relates to a signal compression system and method for improving the security and battery efficiency of a signal transmitted through a network by using a weight learned in a signal compression and decompression process.

Signal compression technology in a signal measuring device is a very important technology to reconsider the efficiency of the battery. In the case of IoT or wearable devices, signals are measured or communicated through various sensors, and at this time, the battery provided in the device has a limited capacity, and the time to measure the signal, the type of measurable signal, and the acquired signal It acts as a constraint on the amount of information to analyze or visualize. Therefore, the battery efficiency of the device that measures the signal is very important. Conventionally, in order to increase battery efficiency, the capacity is increased by increasing the size of the battery itself, or a plurality of batteries are connected in parallel and used. However, this method affects the heat generation, size, charging time, etc. of the battery, and thus there is a limit to be applied to various devices. In addition, in the case of existing devices for measuring bio-signals, there is a disadvantage in that high-quality and large-capacity original signals cannot be transmitted to other devices through communication due to limited battery capacity.

On the other hand, under the existing Personal Information Protection Act, a protection target is defined as an element that can identify an individual, and in general, elements that can identify an individual through technology are limited to the iris, fingerprint, and the like. However, with the development of technology, as individuals can be identified through high-accuracy signal analysis, the types and types of information to be protected as personal information are diversifying.

In the past, simple encryption techniques using random numbers have been recommended and used for personal information protection, but these conventional techniques generate random numbers within a short time and continuously attempt password recovery as computer performance improves. The possibility of being hacked in this way is increasing. Accordingly, a new encryption technology is required, and in this context, block chain is attracting attention as a technology to protect personal information. There is a fatal flaw that it is difficult. The disadvantage of such a block chain is that it is difficult to satisfy the provision of Article 36 of the Personal Information Protection Act that personal information can be freely deleted at the user's will.

Accordingly, there is a need for a method of efficiently compressing information in order to increase the efficiency of communication with the battery, and a new type of security method that can identify an individual and is easy to delete is required.

Technical problem to be solved by the present invention is to provide a signal compression system and method for compressing and restoring a signal using one or more encoding blocks and one or more decoding blocks.

In addition, another technical problem to be solved by the present invention is to provide a signal compression system and method for improving the security and battery efficiency of a signal transmitted through a network by using a weight learned in a signal compression and restoration process.

One aspect of the present invention, a measurement device for recording a measurement signal; and performing pre-processing on the measurement signal to generate a normal signal, generate a compressed signal from the normal signal using a weight set to compress the normal signal, and generate a restored signal from the compressed signal using the weight and an auto-encoder device for learning the weight so that a difference between the normal signal and the reconstructed signal satisfies a predetermined threshold range.

In addition, the auto-encoder device, the collection unit for collecting the measurement signal; a preprocessor for generating a normal signal by performing preprocessing on the measurement signal; an encoder unit for generating a compressed signal from the regular signal using one or more encoding blocks to which a preset weight is applied; a decoder unit for generating a reconstructed signal from the compressed signal using one or more decoding blocks to which the weight is applied; and a learning unit that generates difference information by comparing the normal signal and the reconstructed signal, and learns the weight so that the difference information satisfies a threshold range provided to indicate a criterion for the difference between the normal signal and the reconstructed signal. may include more.

In addition, the encoder unit generates a first extracted signal from which a feature of the regular signal is extracted by applying the weight to the regular signal, and reduces the first extracted signal according to the extracted feature to generate a first reduced signal and applying the weight to the normal signal to generate a second extracted signal from which the features of the normal signal are extracted, and to convert the second extracted signal to a preset size to generate a first converted signal, A first encoded signal may be generated by adding the first reduced signal and the first converted signal.

In addition, when a plurality of encoding blocks are provided, the encoder unit generates a third extracted signal from which a characteristic of the first encoded signal is extracted by applying the weight to the first encoded signal, and the third extracted signal A second reduced signal is generated by reducing according to the extracted features, and a fourth extracted signal from which the features of the regular signal are extracted by applying the weight to the regular signal is generated, and the fourth extracted signal is set in advance It is possible to generate a second converted signal by converting to a magnitude, and to generate a second encoded signal by adding the second reduced signal and the second converted signal.

In addition, the decoder unit generates a fifth extracted signal from which features of the compressed signal are extracted by applying the weight to the compressed signal, and expands the fifth extracted signal according to the extracted features to generate a first enlarged signal and generating a sixth extracted signal from which the characteristics of the compressed signal are extracted by applying the weight to the compressed signal, and converting the sixth extracted signal to a preset size to generate a third converted signal, A first decoded signal may be generated by adding the first enlarged signal and the third converted signal.

In addition, when a plurality of decoding blocks are provided, the decoder unit generates a seventh extracted signal from which a characteristic of the first decoded signal is extracted by applying the weight to the first decoded signal, and generates the seventh extracted signal. Generates a second enlarged signal by reducing according to the extracted features, applying the weight to the compressed signal to generate an eighth extracted signal from which the features of the compressed signal are extracted, and setting the eighth extracted signal in advance It is possible to generate a fourth converted signal by converting the signal to a magnitude, and to generate a second decoded signal by adding the second enlarged signal and the fourth converted signal.

In addition, the preprocessor may extract one or more divided signals appearing at the same time interval from the measurement signal, and perform normalization such that the one or more divided signals satisfy the same magnitude range to generate a normal signal.

In addition, when the measurement signal is a biological signal that is measured and recorded so as to determine an arrhythmia from a living body, a classification model is generated by learning the distribution of the compressed signal generated according to the type of arrhythmia, and based on the classification model, the It may further include an arrhythmia classification device for generating arrhythmia information by determining an arrhythmia according to the distribution of the compressed signal.

In addition, the arrhythmia classification device determines arrhythmia information for one or more compressed signals generated based on the measurement signal, respectively, and measures the arrhythmia information representing the highest frequency according to the arrhythmia information for the one or more compressed signals. It can be determined by the arrhythmia judgment result for the signal.

In addition, a smart device that stores a weight set to compress the regular signal, receives the compressed signal, provides one or more decoding blocks to which the weight is applied, and generates a restored signal from the compressed signal. can

Also, when the weight is changed by the auto encoder device, the smart device may receive the changed weight.

Another aspect of the present invention, an arrhythmia classification method by an arrhythmia classification system using a compressed signal based on deep learning, the steps of: recording a measurement signal; generating a normal signal by performing pre-processing on the measurement signal; generating a compressed signal from the normal signal using a weight set to compress the normal signal; generating a restored signal from the compressed signal using the weight; and learning the weight so that a difference between the normal signal and the reconstructed signal satisfies a predetermined threshold range.

According to one aspect of the present invention described above, by providing a signal compression system and method based on deep learning, it is possible to compress and restore a signal using one or more encoding blocks and one or more decoding blocks.

In addition, according to one aspect of the present invention described above, by providing a signal compression system and method based on deep learning, the security and battery efficiency of a signal transmitted through a network using a weight learned in the process of compression and restoration of a signal is improved. can be improved

1 is a schematic diagram of a signal compression system according to an embodiment of the present invention;
2 is a schematic diagram showing another embodiment of a signal compression system according to an embodiment of the present invention.
3 is a control block diagram of the auto encoder device of FIG.
4 is a control block diagram of the arrhythmia classification device of FIG.
FIG. 5 is a block diagram illustrating a process in which the auto-encoder device of FIG. 3 compresses and restores a measurement signal.
6 is a block diagram illustrating a process of learning weights in the learning unit of FIG. 3 .
7 is a block diagram illustrating a process of classifying an arrhythmia in the controller of FIG. 4 .
8 is a block diagram illustrating a process in which the smart device of FIG. 2 generates a restoration signal.
9 is a schematic diagram illustrating an embodiment of compression and decompression performed in the auto-encoder apparatus of FIG. 1 .
Fig. 10 is a schematic diagram showing an embodiment of a signal compressed by the auto-encoder device of Fig. 1;
11 is a schematic diagram illustrating an embodiment of an encoding block provided in the encoder unit of FIG. 3 .
12 is a schematic diagram illustrating an embodiment of a decoding block provided in the decoder unit of FIG. 3 .
13 is a flowchart of a signal compression method according to an embodiment of the present invention.

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS [0010] DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS [0010] DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS [0010] Reference is made to the accompanying drawings, which show by way of illustration specific embodiments in which the present invention may be practiced. These embodiments are described in sufficient detail to enable those skilled in the art to practice the present invention. It should be understood that the various embodiments of the present invention are different but need not be mutually exclusive. For example, certain shapes, structures, and characteristics described herein with respect to one embodiment may be implemented in other embodiments without departing from the spirit and scope of the invention. In addition, it should be understood that the location or arrangement of individual components within each disclosed embodiment may be changed without departing from the spirit and scope of the present invention. Accordingly, the following detailed description is not intended to be taken in a limiting sense, and the scope of the present invention, if properly described, is limited only by the appended claims, along with all scope equivalents to those claimed. Like reference numerals in the drawings refer to the same or similar functions throughout the various aspects.

Hereinafter, preferred embodiments of the present invention will be described in more detail with reference to the drawings.

1 is a schematic diagram of a signal compression system according to an embodiment of the present invention;

The signal compression system 1 may include a measuring device 100 and an auto-encoder device 200 .

The measurement device 100 may record a measurement signal. In this case, the measurement signal may include a measurement signal measured from a living body, and the measurement signal may include vibration and electrical signals appearing from the ground, metal, living body, machinery, electronic components, and the like.

Accordingly, the measurement apparatus 100 may record the measured measurement signal to indicate a change in the waveform over time, and in this case, the measurement signal may be measured so that a plurality of waveforms appear at the same time.

The measurement device 100 may transmit the recorded measurement signal to the auto encoder device 200 .

The auto-encoder 200 may generate a regular signal by performing pre-processing on the measurement signal, and the auto-encoder 200 may generate a compressed signal from the regular signal using a weight set to compress the regular signal. can In addition, the auto-encoder 200 may generate a reconstructed signal from the compressed signal by using the weight, and the auto-encoder 200 sets the weight so that the difference between the normal signal and the reconstructed signal satisfies a threshold range provided in advance. can learn

In this regard, the auto-encoder device 200 may receive and collect a measurement signal measured by the measurement device 100 , and the auto-encoder device 200 performs pre-processing on the collected measurement signal to generate a regular signal. can do.

To this end, the auto-encoder device 200 may extract one or more divided signals appearing at the same time interval from the measurement signal, and the auto-encoder device 200 performs normalization so that the one or more divided signals satisfy the same size range. A regular signal can be generated.

At this time, the auto encoder device 200 may set a time interval for extracting the divided signal so that the waveform of one cycle appearing in the measurement signal appears in one divided signal, and accordingly, the auto encoder device 200 is the measurement signal One or more split signals can be extracted from

For example, the auto-encoder apparatus 200 extracts information existing in a window area provided in advance from the signal, and when the extraction of information is completed, moves the window area to a subsequent area of the extracted information, and moves A split signal can be extracted using a sliding window technique for extracting information existing in the window region.

In addition, the auto-encoder device 200 may perform normalization so that the magnitude ranges of values appearing from one or more extracted divided signals are the same. When there is a value out of the size range of , the corresponding value may be substituted with the closest value within the size range of the split signal.

Here, the auto-encoder apparatus 200 may set the size range of the value generating the normal signal so that the magnitude of each value according to time of the measurement signal does not become a meaningless value due to another value.

Accordingly, the normal signal may be understood as a plurality of values arranged over time, and in this case, the plurality of values listed may be understood as values normalized based on the measurement signal.

The auto-encoder apparatus 200 may provide one or more encoding blocks to which a preset weight is applied, and the auto-encoder apparatus 200 may generate a compressed signal from a regular signal using one or more encoding blocks.

Here, the encoding block may be provided to extract features of the regular signal by applying a weight to the regular signal, and the encoding block may be provided to reduce the regular signal based on the features extracted from the weighted regular signal. .

Accordingly, the auto-encoder device 200 may generate an extracted signal from which the characteristics of the normal signal are extracted by applying a weight to the normal signal, and the auto-encoder device 200 reduces the extracted signal by reducing the extracted characteristics according to the extracted characteristics. signal can be generated.

In this case, when a plurality of encoding blocks are provided, the auto-encoder device 200 may input a signal processed by one or more encoding blocks to a connected encoding block, and the auto-encoder device 200 provides a feature from the input signal. It is possible to increase the compression ratio of the measurement signal by repeating the process of extracting and reducing the .

In this regard, extracting the features of the regular signal can be understood as applying 1D-CNN (1-Dimension Convolution Neural Network) to the regular signal, and reducing the regular signal can be understood as inputting the regular signal to the pooling layer. can Accordingly, the weight can be understood as meaning a filter of 1D-CNN.

In addition, the encoding block may be provided to extract features of the regular signal by applying a weight to the regular signal, and the encoding block may be provided to convert to a size set in advance based on the features extracted from the weighted regular signal. there is.

Accordingly, the auto-encoder device 200 may generate an extracted signal from which the characteristics of the regular signal are extracted by applying a weight to the regular signal, and the auto-encoder device 200 converts the extracted signal to a preset size A conversion signal can be generated.

At this time, when a plurality of encoding blocks are provided, the auto-encoder device 200 may input a normal signal generated by performing pre-processing to each encoding block, and the auto-encoder device 200 may input the input to each encoding block. It can be controlled to repeat the process of extracting a feature from a normal signal and transforming the size.

In this regard, the size at which the normal signal from which the features are extracted is converted may be set to be the same as the size of the compressed signal in which the measurement signal is compressed, and the size of the compressed signal may be set to be the same as the size of the normal signal.

Accordingly, the encoding block may be provided such that a signal obtained by extracting features from the regular signal and reducing the signal and a signal converted to a size set in advance by extracting features from the regular signal are added. In this case, the auto-encoder apparatus 200 may generate an encoding signal by adding the reduced signal and the converted signal.

In this regard, when a plurality of encoding blocks are provided, the auto-encoder device 200 sums a signal in which a feature is extracted from a regular signal and reduced to a signal converted to a preset size by extracting a feature from the regular signal and a preset size. Thereafter, the auto encoder device 200 may repeat the process of summing the signal converted to a preset size by extracting features from the summed signal and extracting features from the reduced signal and the normal signal. .

For example, the auto-encoder device 200 may generate a first extracted signal from which features of the regular signal are extracted by applying a weight to the regular signal, and the auto-encoder device 200 applies a weight to the first extracted signal and extracts the features. It is possible to generate a first reduced signal by reducing according to .

In addition, the auto-encoder device 200 may generate a second extracted signal from which the characteristics of the regular signal are extracted by applying a weight to the regular signal, and the auto-encoder device 200 sets the second extracted signal to a preset size. may be converted to to generate a first converted signal.

Accordingly, the auto-encoder 200 may generate a first encoded signal by adding the first reduced signal and the first converted signal, and then, the auto-encoder 200 applies a weight to the first encoded signal to A third extracted signal from which features of the first encoded signal are extracted may be generated, and the auto-encoder apparatus 200 may reduce the third extracted signal according to the extracted features to generate a second reduced signal.

In addition, the auto-encoder device 200 may generate a fourth extracted signal from which the characteristics of the regular signal are extracted by applying a weight to the regular signal, and the auto-encoder device 200 sets the fourth extracted signal to a preset size. can be converted to to generate a second converted signal.

In this case, when the weights set in each encoder block are the same, the auto-encoder 200 may generate the first converted signal and the second converted signal to appear as the same signal, but the auto-encoder 200 is each When the weights set in the encoder block of are different, the first transformed signal and the second transformed signal may be generated to appear as different signals.

Accordingly, the auto-encoder 200 may generate a second encoded signal by adding the second reduced signal and the second converted signal, and the auto-encoder 200 repeats this process to increase the compression ratio of the regular signal. can be raised

As such, the auto-encoder device 200 may generate a compressed signal by extracting features from the signal processed by the encoding block, in this case, the auto-encoder device 200 may include one or more 1D signals in the signal processed by the encoding block -CNN can be applied to generate a compressed signal.

For example, the auto encoder device 200 may be provided with five encoding blocks. In this case, in the first encoding block, features are extracted from the regular signal, and features are extracted from the reduced signal and the regular signal in advance. Signals converted to a set size can be summed, and the auto-encoder device 200 extracts features from the summed signal in the second to fifth encoding blocks, extracts features from the reduced signal and the regular signal, and sets it in advance It is possible to sum the signals converted to the size of Accordingly, the auto-encoder device 200 may generate a compressed signal by inputting the signals processed by the five encoding blocks to one or more 1D-CNNs.

The auto-encoder apparatus 200 may provide one or more decoding blocks to which weights are applied, and the auto-encoder apparatus 200 may generate a reconstructed signal from the compressed signal by using the one or more decoding blocks.

Here, the decoding block may be provided to extract a feature of the compressed signal by applying a weight used in the encoding block to the compressed signal, and the decoding block may expand the compressed signal based on the feature extracted from the weighted compressed signal. catalog may be prepared.

Accordingly, the auto-encoder apparatus 200 may generate an extracted signal from which the characteristics of the compressed signal are extracted by applying a weight to the compressed signal, and the auto-encoder apparatus 200 expands and enlarges the extracted signal according to the extracted characteristics. signal can be generated.

At this time, when a plurality of encoding blocks are provided, the auto-encoder 200 may provide the same number of decoding blocks, and in this case, the auto-encoder 200 is a signal processed by one or more decoding blocks. may be input to the connected decoding block, and the auto-encoder apparatus 200 may restore a compressed signal compressed from the measurement signal by repeating the process of extracting and enlarging a feature from the input signal.

In this regard, extracting the characteristics of the compressed signal can be understood as applying 1D-CNN (1-Dimension Convolution Neural Network) to the compressed signal, and expanding the compressed signal can be understood as inputting the compressed signal to the pooling layer. can

In addition, the decoding block may be provided to extract a feature of the compressed signal by applying a weight used in the encoding block to the compressed signal, and the decoding block may have a size set in advance based on the feature extracted from the weighted compressed signal. It may be arranged to convert.

Accordingly, the auto-encoder apparatus 200 may generate an extracted signal from which the characteristics of the compressed signal are extracted by applying a weight to the compressed signal, and the auto-encoder apparatus 200 converts the extracted signal to a preset size A conversion signal can be generated.

At this time, when a plurality of encoding blocks are provided, the auto-encoder 200 may provide the same number of decoding blocks, and the auto-encoder 200 applies a compressed signal compressed from the measurement signal to each decoding block. input, and the auto-encoder device 200 may control to repeat the process of extracting features from the compressed signal input to the decoding block and converting the size.

In this regard, the size at which the compressed signal from which the features are extracted is converted may be set to be the same as the size of the reconstructed signal reconstructed from the compressed signal, and the size of the reconstructed signal may be set to be the same as the size of the normal signal.

Accordingly, the decoding block may be provided such that a signal obtained by extracting a feature from the compressed signal and converting it to a preset size by extracting the feature from the enlarged signal and the compressed signal may be added. In this case, the auto-encoder 200 may generate a decoded signal by adding the enlarged signal and the converted signal.

In this regard, when a plurality of decoding blocks are provided, the auto-encoder apparatus 200 sums a signal converted to a predetermined size by extracting features from a compressed signal and extracting a feature from a compressed signal and a signal that is enlarged by extracting a feature from the compressed signal Afterwards, the auto encoder device 200 may repeat the process of summing the signal converted to a preset size by extracting features from the summed signal and extracting features from the enlarged signal and the compressed signal. .

For example, the auto-encoder 200 may generate a fifth extracted signal from which features of the compressed signal are extracted by applying a weight to the compressed signal, and the auto-encoder 200 may generate a fifth extracted signal from the extracted features. The first enlarged signal may be generated by enlarging according to the .

In addition, the auto-encoder device 200 may generate a sixth extracted signal from which features of the compressed signal are extracted by applying a weight to the compressed signal, and the auto-encoder device 200 sets the sixth extracted signal to a preset size. can be converted to to generate a third converted signal.

Accordingly, the auto-encoder 200 may generate a first decoded signal by adding the first enlarged signal and the third converted signal, and then, the auto-encoder 200 applies a weight to the first decoded signal to A seventh extracted signal from which features of the first decoded signal are extracted may be generated, and the auto-encoder apparatus 200 may generate a second enlarged signal by enlarging the seventh extracted signal according to the extracted features.

In addition, the auto-encoder apparatus 200 may generate an eighth extracted signal from which characteristics of the compressed signal are extracted by applying a weight to the compressed signal, and the auto-encoder apparatus 200 sets the eighth extracted signal to a preset size. , to generate a fourth converted signal.

In this case, when the weights set in each decoder block are the same, the auto-encoder device 200 may generate the third transformed signal and the fourth transformed signal to appear as the same signal, but the auto-encoder device 200 is each When the weights set in the decoder block of are different, the third transformed signal and the fourth transformed signal may be generated to appear as different signals.

Accordingly, the auto-encoder device 200 may generate a second decoded signal by adding the second enlarged signal and the fourth converted signal, and in this case, the auto-encoder device 200 is provided in the auto-encoder device 200 . The same number of decoding blocks as the encoding blocks may be provided.

In this way, the auto-encoder device 200 may generate a restored signal by extracting features from the signal processed by the decoding block, at this time, the auto-encoder device 200 may include one or more 1D signals in the signal processed by the decoding block. -CNN can be applied to generate a restored signal.

For example, the auto-encoder apparatus 200 may provide 5 decoding blocks when 5 encoding blocks are provided. In this case, a feature extracted from a compressed signal in the first decoding block and an enlarged signal and features extracted from the compressed signal, and the signal converted to a preset size can be summed, and the auto-encoder device 200 extracts features from the summed signal in the second to fifth decoding blocks and adds the enlarged signal and A feature may be extracted from the compressed signal and the signals converted to a preset size may be summed. Accordingly, the auto-encoder device 200 may generate a restored signal by inputting the signals processed by the five decoding blocks to one or more 1D-CNNs.

The auto-encoder device 200 may generate difference information by comparing the normal signal and the reconstructed signal, and the auto-encoder device 200 sets a threshold range provided to indicate a criterion for the difference between the normal signal and the reconstructed signal as the difference information Weights can be learned to satisfy

To this end, the auto-encoder apparatus 200 may repeat the process of generating a compressed signal from a normal signal using different weights, generating a reconstructed signal from the compressed signal, and comparing the normal signal and the reconstructed signal several times, and thus Accordingly, the auto-encoder apparatus 200 may set a weight according to difference information having the smallest difference between a normal signal and a reconstructed signal among the difference information satisfying the threshold range to be used in the encoding block and the decoding block.

Accordingly, the compressed signal may be generated only by the decoding block according to the generated weight, and through this, the auto encoder device 200 transmits the compressed signal to an external device using a wired or wireless network. Even if the compressed signal is leaked by a hacker in the process, it is difficult to generate a restored signal from the compressed signal. may appear.

In this regard, the auto-encoder apparatus 200 is trained so that the compressed and restored information is most similar to the original information, and accordingly, an auto-encoder technique for compressing the original information or restoring the compressed information. is available.

Meanwhile, the auto-encoder is known as a type of unsupervised learning in which an input layer and an output layer have the same size.

Performance evaluation of the auto encoder device 200 may be performed according to a process described below, and the following is an example of performing performance evaluation of the auto encoder device 200 .

In this regard, Table 1 shows the number of information for each arrhythmia type collected to perform performance evaluation of the auto-encoder device 200, and for this purpose, the measuring device 100 of the signal compression system 1 is measured from the living body It can be understood that the biosignal is recorded as a measurement signal.

In this case, the measurement signals in Table 1 are 48 pieces of information extracted from ECG signals measured continuously for 24 hours from 47 subjects.

In the performance evaluation of the auto-encoder device 200, weights are learned using 80% of information among the retained information, the learned weights are verified using 10% of information, and the auto-encoder device using the remaining 10% of information. The process of evaluating the performance of (200) can be performed.

In this case, in the performance evaluation of the auto-encoder apparatus 200, cross-validation may be performed by repeating the performance evaluation 10 times so that information used in the verification and evaluation process is not duplicated.

In addition, the performance evaluation of the auto-encoder device 200 may be performed in a 32-fold compression model and a 64-fold compression model that are provided to enable compression of 32% and 64% of the measurement signal, and in this regard, Equation 1 is a calculation formula of a performance evaluation index for signal compression and restoration of a measurement signal.

Here, O_i may mean a regular signal, and R_i may mean a restored signal. Also, O_avg may mean an average value of one or more normal signals.

In addition, the PRDN may represent an index for quantifying the difference between the normal signal and the reconstructed signal, and a lower PRDN may mean that the reconstructed signal is reconstructed to be the same as the normal signal. Accordingly, PRDN can be understood as meaning difference information.

In addition, the SNR may mean an index comparing the relative strength of signal-to-noise, and a lower SNR may mean that noise is not mixed in the reconstructed signal.

On the other hand, the performance evaluation result of the 32x compression model shows that the average PRDN is 12.67% and the SNR can be 43.43dB. In addition, the performance evaluation result of the 64x compression model may show that the average PRDN is 22.98% and the SNR is 31.49dB.

Table 2 is a table showing the performance evaluation results for the 32x compression model and the 64x compression model.

2 is a schematic diagram showing another embodiment of a signal compression system according to an embodiment of the present invention.

The signal compression system 1 may further include an arrhythmia classification device 300 and a smart device 400 .

The arrhythmia classification apparatus 300 may generate a classification model by learning the distribution of the compressed signal generated according to the type of arrhythmia, and the arrhythmia classification apparatus 300 based on the generated classification model, the arrhythmia according to the distribution of the compressed signal can be determined to generate arrhythmia information.

In this case, the measurement apparatus 100 may record the biological signal measured from the living body as the measurement signal. Here, the biosignal may mean a signal measured to detect arrhythmias, and the biosignal may mean an electrocardiogram (ECG) that appears as a waveform of a minute current generated in the heart.

To this end, the measuring device 100 may be a device capable of recording a waveform of a minute current measured from one or more electrodes attached to the user's body, and accordingly, the measuring signal measured from the measuring device 100 is may include P, Q, R, S and T waves.

In this case, the P wave may be generated by contraction of the atria present in the heart, and the Q, R, S, and T waves may be generated by the ventricles present in the heart.

The arrhythmia classification device 300 may receive a compressed signal from the auto encoder device 200 .

Accordingly, the arrhythmia classification apparatus 300 may generate a classification model by learning the distribution of the compressed signal generated according to the type of arrhythmia, and the arrhythmia classification apparatus 300 based on the generated classification model, the distribution of the compressed signal The arrhythmia may be determined according to the arrhythmia and arrhythmia information may be generated.

Here, the arrhythmia information may include a normal pattern indicating a normal pulse, an abnormal impulse pattern indicating a state in which an abnormal pulse is detected, and a conduction disturbance pattern indicating a state in which some or all of the measurement signal is not detected.

In addition, the abnormal impulse pattern is an early atrial contraction pattern indicating a state of premature atrial contraction (PAC), in which the atria of the heart contract irregularly compared to a normal pulse, and a ventricle of the heart irregularly compared to a normal pulse. It may further include a premature ventricular contraction pattern indicating a contracting premature ventricular contraction (PVC: Premature Ventricular Contraction) state.

In addition, the conduction disturbance pattern may further include a left-brain blocking pattern indicating a left bundle branch block (LBBB) state of conduction disturbance and a right-brain blocking pattern indicating a right bundle branch block (RBBB) state of conduction disturbance. there is.

The arrhythmia classification apparatus 300 may generate a classification model by combining one or more convolutional neural networks (CNN) and one or more bidirectional long short-term memories (BLSTM).

Here, the bidirectional long and short-term memory has a forward layer and a reverse layer, performs operations on the same information through the forward layer and the reverse layer, respectively, and combines information output from the forward layer and the reverse layer to perform learning. can be understood as

Accordingly, the arrhythmia classification apparatus 300 can extract the features of the compressed signal using a synthetic product neural network, and then, the arrhythmia classification apparatus 300 learns the compressed signal from which the features are extracted using the bidirectional long and short-term memory. A classification model can be created.

At this time, the arrhythmia classification apparatus 300 may determine the arrhythmia information for one or more compressed signals generated based on the measurement signal, respectively, and the arrhythmia classification apparatus 300 determines the highest value according to the arrhythmia information for one or more compressed signals. Arrhythmia information indicating the frequency may be determined as an arrhythmia determination result with respect to the measurement signal.

In this regard, the arrhythmia classification apparatus 300 may learn a classification model using a supervised learning technique for receiving arrhythmia information from a user with respect to the distribution of one or more compressed signals.

Accordingly, the arrhythmia classification apparatus 300 may classify the compressed signal into at least one of a normal pattern, an abnormal impulse pattern, and a conduction disturbance pattern based on the classification model.

Also, based on the classification model, the arrhythmia classification apparatus 300 may classify the compressed signal representing the abnormal impulse pattern into at least one of an early atrial contraction pattern and an early ventricular contraction pattern, and the arrhythmia classification apparatus 300 may Based on the classification model, the compressed signal representing the conduction disturbance pattern may be classified into at least one of a left-angle block pattern and a right-angle block pattern.

Accordingly, the arrhythmia classification apparatus 300 may classify a plurality of compressed signals generated from one measurement signal as arrhythmia information, respectively, and the arrhythmia information that appears with the highest frequency among the plurality of arrhythmia information is an arrhythmia for the corresponding measurement signal. information can be judged.

Alternatively, the arrhythmia classification apparatus 300 may set a threshold number of times to determine the arrhythmia information on the measurement signal from the arrhythmia information on the plurality of compressed signals. Each of the generated compressed signals may be classified as arrhythmia information, and in this case, arrhythmia information exceeding a threshold number may be determined as arrhythmia information for the corresponding measurement signal. In this case, the arrhythmia classification apparatus 300 may set different threshold times for different arrhythmia information.

The arrhythmia classification apparatus 300 may perform performance evaluation according to the process described below, and the following is an example of performing performance evaluation of the arrhythmia classification apparatus 300 .

An embodiment of the performance evaluation of the arrhythmia classification device 300 may be understood as using information used in the performance evaluation of the auto-encoder device 200 .

The performance evaluation of the arrhythmia classification device 300 is to learn the classification model using 80% of the information possessed, verify the learned classification model using 10% of the information, and use the remaining 10% of the information to evaluate the arrhythmia. A process of evaluating the performance of the classification apparatus 300 may be performed.

In this case, the performance evaluation of the arrhythmia classification apparatus 300 may perform cross-validation by repeating the performance evaluation 10 times so that the information used in the verification and evaluation process is not duplicated.

In addition, the performance evaluation of the arrhythmia classification apparatus 300 may be performed using a compressed signal compressed by a 32-fold compression model and a compressed signal compressed by a 64-fold compression model.

Accordingly, the performance evaluation results of the arrhythmia classification apparatus 300 using the compressed signal compressed by the 32-fold compression model and the performance evaluation result of the arrhythmia classification apparatus 300 using the compressed signal compressed by the 64-fold compression model are shown in Table 3 can be checked in

The smart device 400 may store a weight set to compress a regular signal, the smart device 400 may receive a compressed signal, and the smart device 400 may prepare one or more decoding blocks to which the weight is applied. , a reconstructed signal may be generated from the compressed signal.

Here, the smart device 400 may receive a decoding block used in the auto encoder device 200 .

In this regard, when the weight set in the auto encoder device 200 is changed by the auto encoder device 200 , the smart device 400 may receive the changed weight from the auto encoder device 200 , at this time , the smart device 400 may convert the compressed signal into a reconstructed signal by using the decoding block to which the changed weight is applied.

Accordingly, the smart device 400 can generate a restored signal by receiving the compressed signal generated by the auto encoder device 200 by the weighted decoding block, so the signal compression system 1 is transmitted through the network An effect of reducing the amount of information may occur, and also, since the signal compression system 1 cannot obtain the weight set in the auto-encoder device 200 even when a packet of the compressed signal is intercepted, high-strength security effect can be obtained.

To this end, the signal compression system 1 may further include an encoding device for generating a compressed signal from the measurement signal measured by the measurement device 100 using the encoding block to which the weight set in the auto encoder device 200 is applied. may be

In this case, the smart device 400 may generate a restored signal by receiving the compressed signal generated by the encoding device.

On the other hand, the signal compression system 1 may learn individualized weights for each user, and in this case, the signal compression system 1 based on the learned weights, a regular signal and a restored signal generated from the user's measurement signal The user can be distinguished by comparing the difference information according to the reference difference information set in advance.

To this end, the auto-encoder device 200 may learn a personalized weight from the user's individual measurement signal measured by the measurement device 100 , and in this case, the auto-encoder device 200 provides the user's individual regular signal and the restored signal. A weight may be learned according to reference difference information provided to indicate a criterion for the difference of .

Accordingly, the auto-encoder device 200 generates a normal signal and a reconstructed signal from the measurement signal measured by the user, and generates difference information from the generated normal signal and the generated reconstructed signal, so that the reference difference information and the difference information are When it is determined that the measurement signal is the same, it may be determined that the user who has measured the measurement signal is the same person as the user who provided the measurement signal used to generate the weight based on the reference difference information.

In this way, the auto-encoder device 200 can use the compressed signal from the measurement signal measured by the user and the weight learned by the auto-encoder device 200 as personal information.

In this way, the compressed signal may be generated only by the decoding block according to the generated weight, and through this, the auto-encoder device 200 transmits the compressed signal to an external device using a wired or wireless network. Even if the compressed signal is leaked by a hacker in the process, there is a difficulty in generating a restored signal from the compressed signal. there is.

3 is a control block diagram of the auto encoder device of FIG.

The auto encoder apparatus 200 may include a collection unit 210 , a preprocessor 220 , an encoder unit 230 , a decoder unit 240 , a learning unit 250 , and a communication unit 260 .

The collection unit 210 may receive and collect the measurement signal measured by the measurement device 100 .

The preprocessor 220 may perform preprocessing on the collected measurement signal to generate a normal signal.

To this end, the pre-processing unit 220 may extract one or more divided signals appearing at the same time interval from the measurement signal, and the pre-processing unit 220 performs normalization so that the one or more divided signals satisfy the same magnitude range to obtain a normal signal. can create

In this case, the preprocessor 220 may set a time interval for extracting the divided signal so that a waveform of one period appearing in the measurement signal appears in one divided signal, and accordingly, the preprocessor 220 may A split signal can be extracted.

In addition, the preprocessor 220 may perform normalization so that the magnitude range of values appearing from the one or more extracted divided signals becomes the same. When a value out of the range exists, the corresponding value may be substituted with the closest value within the amplitude range of the split signal.

Here, the preprocessor 220 may set a range of values for generating the normal signal so that the magnitude of each value according to time of the measurement signal does not become a meaningless value due to another value.

The encoder unit 230 may provide one or more encoding blocks to which a preset weight is applied, and the encoder unit 230 may generate a compressed signal from a regular signal using one or more encoding blocks.

In this case, when a plurality of encoding blocks are provided, the encoder unit 230 may input a signal processed by one or more encoding blocks to a connected encoding block, and the encoder unit 230 extracts features from the input signal. and repeating the reduction process to increase the compression ratio of the measurement signal.

In addition, when a plurality of encoding blocks are provided, the encoder unit 230 may input a normal signal generated by preprocessing to each encoding block, and the encoder unit 230 may input the normal signal input to the encoding block. It can be controlled to repeat the process of extracting features and transforming the size.

In this regard, when a plurality of encoding blocks are provided, the encoder unit 230 sums the reduced signal by extracting features from the regular signal and the signal converted to a preset size by extracting features from the regular signal. Thereafter, the encoder unit 230 may repeat the process of summing the signals, in which features are extracted from the summed signal and the features are extracted from the reduced signal and the normal signal, and converted to a preset size.

Accordingly, the encoder unit 230 may generate a compressed signal by extracting features from the signal processed by the encoding block. At this time, the encoder unit 230 may add one or more 1D-CNN to the signal processed by the encoding block. can be applied to generate a compressed signal.

The decoder unit 240 may provide one or more decoding blocks to which weights are applied, and the decoder unit 240 may generate a reconstructed signal from the compressed signal by using the one or more decoding blocks.

At this time, when a plurality of encoding blocks are provided, the decoder unit 240 may provide the same number of decoding blocks. In this case, the decoder unit 240 connects signals processed by one or more decoding blocks. It may be input to the decoding block, and the decoder unit 240 may extract a feature from the input signal and repeat the process of expanding it to restore the compressed signal compressed from the measurement signal.

In addition, when a plurality of encoding blocks are provided, the decoder unit 240 may provide the same number of decoding blocks, and the decoder unit 240 may input a compressed signal compressed from the measurement signal to each decoding block. In addition, the decoder unit 240 may control to repeat the process of extracting a feature from the compressed signal input to the decoding block and converting the size.

In this regard, when a plurality of decoding blocks are provided, the decoder unit 240 sums up a signal converted to a size set in advance by extracting features from the compressed signal and extracting features from the compressed signal with the enlarged signal. Thereafter, the decoder unit 240 may repeat the process of summing the signal converted to a preset size by extracting features from the summed signal and extracting features from the enlarged signal and the compressed signal.

Accordingly, the decoder unit 240 may extract features from the signal processed by the decoding block to generate a reconstructed signal, and in this case, the decoder unit 240 may include one or more 1D-CNN in the signal processed by the decoding block. can be applied to generate a restored signal.

The learning unit 250 may generate difference information by comparing the normal signal and the reconstructed signal, and the learning unit 250 satisfies a threshold range provided to indicate a criterion for the difference between the normal signal and the reconstructed signal. You can learn weights to do this.

To this end, the learning unit 250 generates a compressed signal from the regular signal using different weights in the encoder unit 230 , and generates a reconstructed signal from the compressed signal in the decoder unit 240 to combine the normal signal and the reconstructed signal. The comparison process can be controlled to be repeated several times. Accordingly, the learning unit 250 sets a weight according to the difference information with the smallest difference between the regular signal and the reconstructed signal among the difference information satisfying the threshold range between the encoding block and the decoding block. can be set to be used in

Accordingly, the compressed signal may be generated only by the decoding block according to the generated weight, and through this, the auto encoder device 200 transmits the compressed signal to an external device using a wired or wireless network. Even if the compressed signal is leaked by a hacker in the process, there is a difficulty in generating a restored signal from the compressed signal. there is.

The communication unit 260 may receive a measurement signal from the measurement device 100 through a wired or wireless network, and the communication unit 260 transmits the compressed signal generated by the encoder unit 230 to the smart device ( 400), and the communication unit 260 may transmit the weight generated by the learning unit 250 to the smart device 400 through a wired or wireless network.

4 is a control block diagram of the arrhythmia classification device of FIG.

The arrhythmia classification apparatus 300 may include a receiver 310 , a classification model generator 320 , a controller 330 , and an output unit 340 .

The receiver 310 may receive a compressed signal or a weight generated by the auto encoder device 200 .

The classification model generator 320 may generate a classification model by learning the distribution of the compressed signal generated by the auto-encoder apparatus 200 according to the type of arrhythmia.

The classification model generator 320 may generate a classification model by combining one or more convolutional neural networks (CNNs) and one or more bidirectional long short-term memories (BLSTM).

Here, the classification model generation unit 320 may extract features of the compressed signal using a convolutional neural network, and then, the classification model generation unit 320 generates the compressed signal from which the features are extracted using a bidirectional long and short-term memory. It can be trained to create a classification model.

At this time, the classification model generation unit 320 may determine the arrhythmia information for one or more compressed signals generated based on the measurement signal, respectively, and the classification model generation unit 320 according to the arrhythmia information for the one or more compressed signals. The arrhythmia information representing the highest frequency may be determined as the arrhythmia determination result for the measurement signal.

In this regard, the classification model generator 320 may learn the classification model by using a supervised learning technique for receiving arrhythmia information from a user with respect to the distribution of one or more compressed signals.

The controller 330 may determine an arrhythmia according to a distribution of the generated compressed signal based on the generated classification model and generate arrhythmia information.

In this case, the controller 330 may classify the compressed signal into at least one of a normal pattern, an abnormal impulse pattern, and a conduction disturbance pattern based on the classification model.

Also, based on the classification model, the controller 330 may classify the compressed signal representing the abnormal impulse pattern into at least one of an early atrial contraction pattern and an early ventricular contraction pattern, and the controller 330 is configured to Accordingly, the compressed signal representing the conduction disturbance pattern may be classified into at least one of a left-angle blocking pattern and a right-angle blocking pattern.

The output unit 340 may output the arrhythmia information on the display, and the output unit 340 may output the arrhythmia information to an external device through a wired or wireless network.

FIG. 5 is a block diagram illustrating a process in which the auto-encoder device of FIG. 3 compresses and restores a measurement signal.

Referring to FIG. 5 , the collection unit 210 may receive and collect a measurement signal measured by the measurement device 100 , and the preprocessor 220 performs pre-processing on the collected measurement signal to generate a regular signal. can do.

Accordingly, the encoder unit 230 may provide one or more encoding blocks to which a preset weight is applied, and the encoder unit 230 may generate a compressed signal from the regular signal using one or more encoding blocks.

In this regard, when a plurality of encoding blocks are provided, the encoder unit 230 sums the reduced signal by extracting features from the regular signal and the signal converted to a preset size by extracting features from the regular signal. Thereafter, the encoder unit 230 may repeat the process of summing the signals, in which features are extracted from the summed signal and the features are extracted from the reduced signal and the normal signal, and converted to a preset size.

Accordingly, the encoder unit 230 may generate a compressed signal by extracting features from the signal processed by the encoding block. At this time, the encoder unit 230 may add one or more 1D-CNN to the signal processed by the encoding block. can be applied to generate a compressed signal.

Meanwhile, the decoder unit 240 may provide one or more decoding blocks to which weights are applied, and the decoder unit 240 may generate a reconstructed signal from the compressed signal by using the one or more decoding blocks.

In this regard, when a plurality of decoding blocks are provided, the decoder unit 240 sums up a signal converted to a size set in advance by extracting features from the compressed signal and extracting features from the compressed signal with the enlarged signal. Thereafter, the decoder unit 240 may repeat the process of summing the signal converted to a preset size by extracting features from the summed signal and extracting features from the enlarged signal and the compressed signal.

Accordingly, the decoder unit 240 may extract features from the signal processed by the decoding block to generate a reconstructed signal, and in this case, the decoder unit 240 may include one or more 1D-CNN in the signal processed by the decoding block. can be applied to generate a restored signal.

6 is a block diagram illustrating a process of learning weights in the learning unit of FIG. 3 .

Referring to FIG. 6 , the learning unit 250 may generate difference information by comparing the normal signal and the reconstructed signal, and the learning unit 250 is a threshold provided to indicate a criterion for the difference between the normal signal and the reconstructed signal. Weights can be learned so that the range difference information satisfies.

To this end, the learning unit 250 generates a compressed signal from the regular signal generated by the preprocessing unit 220 using different weights in the encoder unit 230 , and a reconstructed signal from the compressed signal in the decoder unit 240 . By generating , the process of comparing the normal signal and the reconstructed signal can be controlled to repeat several times. Accordingly, the learning unit 250 determines the difference between the normal signal and the reconstructed signal with the smallest difference among the difference information satisfying the threshold range. It is possible to set the weight according to the to be used in the encoding block and the decoding block.

Accordingly, the compressed signal may be generated only by the decoding block according to the generated weight, and through this, the auto encoder device 200 transmits the compressed signal to an external device using a wired or wireless network. Even if the compressed signal is leaked by a hacker in the process, there is a difficulty in generating a restored signal from the compressed signal. there is.

7 is a block diagram illustrating a process of classifying an arrhythmia in the controller of FIG. 4 .

Referring to FIG. 7 , the communication unit 260 of the auto encoder device 200 may transmit the compressed signal generated by the encoder unit 230 to the smart device 400 through a wired or wireless network, and the communication unit 260 ) may transmit the weight generated by the learning unit 250 to the smart device 400 through a wired or wireless network.

Accordingly, the receiving unit 310 may receive a compressed signal or a weight generated by the auto-encoder device 200, and the controller 330 based on the generated classification model, according to the distribution of the generated compressed signal, arrhythmia can be determined to generate arrhythmia information.

To this end, the classification model generation unit 320 may generate a classification model by learning the distribution of the compressed signal generated by the auto-encoder device 200 according to the type of arrhythmia.

Meanwhile, the output unit 340 may output the arrhythmia information on the display, and the output unit 340 may output the arrhythmia information to an external device through a wired or wireless network.

8 is a block diagram illustrating a process in which the smart device of FIG. 2 generates a restoration signal.

Referring to FIG. 8 , the collection unit 210 may receive and collect a measurement signal measured by the measurement device 100 , and the preprocessor 220 performs pre-processing on the collected measurement signal to generate a regular signal. can do.

Accordingly, the encoder unit 230 may generate a compressed signal from the regular signal by using one or more encoding blocks, and in this case, the learning unit 250 is provided to indicate a reference for the difference between the regular signal and the reconstructed signal. Weights can be learned so that the difference information satisfies a threshold range.

The smart device 400 may receive and store the weight set to compress the regular signal from the communication unit 260 of the auto-encoder device 200 , and the smart device 400 may store the weights set to compress the regular signal, and the smart device 400 is the communication unit 260 of the auto-encoder device 200 . A compressed signal may be received from

Accordingly, the smart device 400 may generate one or more decoding blocks to which weights are applied to generate a restored signal from the compressed signal.

9 is a schematic diagram illustrating an embodiment of compression and decompression performed in the auto-encoder apparatus of FIG. 1 .

Referring to FIG. 9 , an embodiment showing the structure of the encoder unit 230 including five encoder blocks to which a regular signal is input and two 1D-CNNs for performing an operation on a signal output from the fifth encoder block can be confirmed, and an embodiment of the decoder unit 240 including five decoder blocks to which a compressed signal is input and two 1D-CNNs for performing an operation on a signal output from the fifth decoder block can be confirmed. .

Accordingly, the encoder unit 230 may generate a compressed signal from the regular signal using one or more encoding blocks, and the decoder unit 240 may generate a reconstructed signal from the compressed signal using one or more decoding blocks. can

In this regard, referring to FIG. 10 , an embodiment showing a process in which an electrocardiogram signal is compressed by an auto encoder device can be confirmed.

At this time, the electrocardiogram signal is input to the 1D-CNN, the characteristics of each electrocardiogram signal are extracted, and accordingly, the type of arrhythmia can be distinguished based on the shape of each signal from which the characteristics are extracted.

In addition, it can be confirmed that the size of the x-axis indicating the size of the signal is reduced by performing Reshaping on the signal from which the feature is extracted. It can be seen that is compressed to 400.

In this way, it can be confirmed that the ECG signal is compressed 32 times to 25 in size by passing through 5 encoding blocks from the size of 800.

11 is a schematic diagram illustrating an embodiment of an encoding block provided in the encoder unit of FIG. 3 .

Referring to FIG. 11 , it is possible to confirm an encoding block in which a signal calculated by two 1D-CNNs and a pooling layer and a reshaped signal through one 1D-CNN are summed. Here, Reshape may be understood as converting a signal to a preset size.

In this way, the auto-encoder apparatus 200 may add a reduced signal obtained by extracting features from the regular signal and a signal converted to a preset size by extracting features from the regular signal.

In addition, when a plurality of encoding blocks are provided, the auto-encoder 200 may add a signal that is reduced by extracting features from a regular signal and a signal converted to a size set in advance by extracting features from the regular signal and Thereafter, the auto-encoder apparatus 200 may repeat the process of summing the signal converted to a preset size by extracting features from the summed signal and extracting features from the reduced signal and the normal signal.

12 is a schematic diagram illustrating an embodiment of a decoding block provided in the decoder unit of FIG. 3 .

Referring to FIG. 12 , an encoding block in which a signal calculated by two 1D-CNNs and an Upsampling Layer and a signal calculated by 1D-CNN and the reshaped signal are summed can be confirmed. Here, Reshape may be understood as converting a signal to a preset size.

In this way, the auto-encoder apparatus 200 may add a signal that is enlarged by extracting features from the compressed signal and a signal that is converted to a predetermined size by extracting features from the compressed signal.

In addition, when a plurality of decoding blocks are provided, the auto-encoder device 200 may add a signal converted to a predetermined size by extracting a feature from the compressed signal and extracting a feature from the compressed signal and a signal that is enlarged by extracting a feature from the compressed signal. Thereafter, the auto-encoder apparatus 200 may repeat the process of summing a signal converted to a preset size by extracting a feature from the summed signal and extracting the feature from the enlarged signal and the compressed signal.

13 is a flowchart of a signal compression method according to an embodiment of the present invention.

Since the signal compression method according to an embodiment of the present invention proceeds on substantially the same configuration as the signal compression system 1 shown in FIG. 1 , the same reference numerals refer to the same components as the signal compression system 1 of FIG. 1 . , and repeated descriptions will be omitted.

The signal compression method includes the steps of recording a measurement signal (600), generating a regular signal (610), generating a compressed signal (620), generating a restored signal (630), and learning a weight ( 640) may be included.

Recording the measurement signal 600 may measure the user's measurement signal. Here, the measurement signal may mean a signal measured to detect an arrhythmia, and the measurement signal may mean an electrocardiogram displayed as a waveform of a minute current generated in the heart.

In step 610 of generating the normal signal, the normal signal may be generated by performing pre-processing on the measurement signal.

To this end, the step 610 of generating the regular signal may extract one or more divided signals appearing at the same time interval from the measurement signal, and the step 610 of generating the regular signal is that the one or more divided signals have the same magnitude range. A normal signal can be generated by performing normalization to satisfy it.

In step 620 of generating the compressed signal, the compressed signal may be generated from the regular signal using a weight set to compress the regular signal.

In this regard, the auto-encoder apparatus 200 may provide one or more encoding blocks to which a preset weight is applied, and the step of generating a compressed signal ( 620 ) is a compressed signal from a regular signal using one or more encoding blocks. can create

At this time, in the step of generating the compressed signal ( 620 ), when a plurality of encoding blocks are provided, a reduced signal by extracting features from a regular signal and a signal converted to a preset size by extracting features from the regular signal After that, the step of generating a compressed signal ( 620 ) is a process of summing the signal converted to a preset size by extracting features from the summed signal and extracting features from the reduced signal and the normal signal and Can be repeated.

Accordingly, the step of generating the compressed signal ( 620 ) may generate a compressed signal by extracting features from the signal processed by the encoding block, and in this case, the step of generating the compressed signal ( 620 ) is processed by the encoding block Compressed signal can be generated by applying one or more 1D-CNN to the compressed signal.

In step 630 of generating the reconstructed signal, a reconstructed signal may be generated from the compressed signal using a weight set to compress the compressed signal.

In this regard, the auto-encoder device 200 may provide one or more decoding blocks to which a preset weight is applied, and the step 630 of generating a restored signal is a restored signal from a compressed signal using one or more decoding blocks. can create

At this time, in the step 630 of generating the restored signal, when a plurality of decoding blocks are provided, a signal converted to a preset size by extracting a feature from the compressed signal and extracting a feature from the compressed signal and an enlarged signal from the compressed signal can be summed, and then, in the step 630 of generating a restored signal, a process of summing the signal converted to a preset size by extracting features from the summed signal and extracting features from the enlarged signal and the compressed signal Can be repeated.

Accordingly, the step 630 of generating the reconstructed signal may generate a reconstructed signal by extracting features from the signal processed by the decoding block, and in this case, the step 630 of generating the reconstructed signal is processed by the decoding block One or more 1D-CNNs may be applied to the obtained signal to generate a reconstructed signal.

The step of learning the weight 640 may generate difference information by comparing the normal signal and the reconstructed signal, and the step 640 of learning the weight is a threshold provided to indicate a reference for the difference between the normal signal and the reconstructed signal. Weights can be learned so that the range difference information satisfies.

To this end, in the step 640 of learning the weights, the process of generating a compressed signal from a regular signal using different weights, generating a reconstructed signal from the compressed signal, and comparing the normal signal and the reconstructed signal may be repeated several times, , accordingly, the step of learning the weights 640 may set the encoding block and the decoding block to use a weight according to difference information having the smallest difference between the normal signal and the reconstructed signal among the difference information satisfying the threshold range.

Although the above has been described with reference to the embodiments, those skilled in the art will understand that various modifications and changes can be made to the present invention without departing from the spirit and scope of the present invention as set forth in the following claims. will be able

1: signal compression system
100: measuring device
200: auto encoder device
300: arrhythmia classification device
400: smart device

Claims (12)

a measuring device for recording the measuring signal; and
A normal signal is generated by performing pre-processing on the measurement signal, a compressed signal is generated from the normal signal using a weight set to compress the normal signal, and a restored signal is generated from the compressed signal using the weight. and an auto-encoder device for learning the weight so that a difference between the normal signal and the restored signal satisfies a threshold range provided in advance,
The auto-encoder device,
When the signal measured from the measuring device is an individual measured signal measured by the user,
Further learning by generating a personalized weight through the reference difference information prepared in advance in the individual measurement signal,
Generating an individual normal signal and a restored signal from the individual measurement signal, comparing the individual's normal signal and the restored signal to generate individual difference information,
When the individual difference information is the same as the reference difference information prepared in advance,
A signal compression system for discriminating a user by determining that the user who has measured the individual measurement signal is the same person as the user who provided the individual measurement signal used to generate the personalized weight based on the reference difference information.
According to claim 1, wherein the auto encoder device,
a collection unit for collecting a measurement signal;
a preprocessor for generating a normal signal by performing preprocessing on the measurement signal;
an encoder unit for generating a compressed signal from the regular signal using one or more encoding blocks to which a preset weight is applied;
a decoder unit for generating a reconstructed signal from the compressed signal using one or more decoding blocks to which the weight is applied; and
A learning unit that generates difference information by comparing the normal signal and the reconstructed signal, and learns the weight so that the difference information satisfies a threshold range provided to indicate a criterion for a difference between the normal signal and the reconstructed signal comprising, a signal compression system.
The method of claim 2, wherein the encoder unit,
applying the weight to the normal signal to generate a first extracted signal from which the characteristics of the normal signal are extracted, and reducing the first extracted signal according to the extracted characteristics to generate a first reduced signal;
applying the weight to the normal signal to generate a second extracted signal from which the characteristics of the normal signal are extracted, and converting the second extracted signal to a preset size to generate a first converted signal,
and summing the first reduced signal and the first transformed signal to generate a first encoded signal.
The method of claim 3, wherein the encoder unit,
When a plurality of encoding blocks are provided,
generating a third extracted signal from which features of the first encoded signal are extracted by applying the weight to the first encoded signal, and reducing the third extracted signal according to the extracted features to generate a second reduced signal;
generating a fourth extracted signal from which features of the regular signal are extracted by applying the weight to the regular signal, and converting the fourth extracted signal to a preset size to generate a second converted signal,
and summing the second reduced signal and the second transformed signal to generate a second encoded signal.
The method of claim 2, wherein the decoder unit,
generating a fifth extracted signal from which features of the compressed signal are extracted by applying the weight to the compressed signal, and expanding the fifth extracted signal according to the extracted features to generate a first enlarged signal;
generating a sixth extracted signal from which features of the compressed signal are extracted by applying the weight to the compressed signal, and converting the sixth extracted signal to a preset size to generate a third converted signal,
A signal compression system for generating a first decoded signal by adding the first enlarged signal and the third converted signal.
The method of claim 5, wherein the decoder unit,
When a plurality of decoding blocks are provided,
applying the weight to the first decoded signal to generate a seventh extracted signal from which features of the first decoded signal are extracted, and reducing the seventh extracted signal according to the extracted features to generate a second enlarged signal;
applying the weight to the compressed signal to generate an eighth extracted signal from which features of the compressed signal are extracted, and converting the eighth extracted signal to a preset size to generate a fourth converted signal,
and generating a second decoded signal by adding the second enlarged signal and the fourth converted signal.
According to claim 2, wherein the pre-processing unit,
Extracting one or more divided signals appearing at the same time interval from the measurement signal,
and performing normalization so that the one or more divided signals satisfy the same magnitude range to generate a normal signal.
According to claim 1,
When the measurement signal is a biological signal that is measured and recorded so as to determine an arrhythmia from a living body,
An arrhythmia classification device for generating a classification model by learning the distribution of the compressed signal generated according to the type of arrhythmia, and determining the arrhythmia according to the distribution of the compressed signal, based on the classification model, further comprising an arrhythmia classification device, signal compression system.
The method of claim 8, wherein the arrhythmia classification device,
Arrhythmia information for one or more compressed signals generated based on the measurement signal is determined, respectively, and arrhythmia information indicating the highest frequency according to the arrhythmia information for the one or more compressed signals is determined as an arrhythmia determination result for the measurement signal which is a signal compression system.
The method of claim 1,
A smart device that stores a weight set to compress the regular signal, receives the compressed signal, and provides one or more decoding blocks to which the weight is applied, further comprising a smart device for generating a restored signal from the compressed signal, Signal compression system.
The method of claim 10, wherein the smart device,
When the weight is changed by the auto-encoder device, the changed weight is transmitted.
In a signal compression method by a signal compression system using a compressed signal based on deep learning,
recording the measurement signal;
generating a normal signal by performing pre-processing on the measurement signal;
generating a compressed signal from the normal signal using a weight set to compress the normal signal;
generating a restored signal from the compressed signal using the weight; and
learning the weight so that a difference between the normal signal and the restored signal satisfies a threshold range provided in advance,
The signal compression system,
When the signal measured from the measurement device is an individual measurement signal measured from the user,
Further learning by generating a personalized weight through the reference difference information prepared in advance in the individual measurement signal,
Generating an individual normal signal and a restored signal from the individual measurement signal, comparing the individual's normal signal and the restored signal to generate individual difference information,
When the individual difference information is the same as the reference difference information prepared in advance,
A method for discriminating a user by determining that the user who has measured the individual measurement signal is the same person as the user who provided the individual measurement signal used to generate the personalized weight based on the reference difference information.

Images