Dog biometric identification device using electrocardiogram and Method thereof

Abstract

A dog biometric recognition device using an electrocardiogram includes an electrocardiogram measuring device for measuring the electrocardiogram of a dog, a measuring unit for generating an electrocardiogram signal, a preprocessing unit for removing noise from and normalizing the electrocardiogram signal, an R-peak extraction unit for detecting an R-peak in the preprocessed electrocardiogram signal, a segmentation unit for performing segmentation of the electrocardiogram signal by setting a window centered on the detected R-peak, a predetermined first time before and a predetermined second time after the R-peak, and a classifier including a neural network for classifying a pre-entered dog individual using the segmented electrocardiogram signals as input.

Description

Technology Field

[0001] The present invention relates to a dog biometric recognition device using an electrocardiogram and a dog biometric recognition method using an electrocardiogram. More specifically, the invention relates to a dog biometric recognition device that distinguishes individuals using an artificial intelligence neural network algorithm based on an electrocardiogram signal generated by measuring the electrocardiogram of a dog, and a dog biometric recognition method using the same. Background Technology

[0002] Recently, the number of households owning pets has been increasing rapidly. This trend is attributed to rising awareness of animal welfare, as well as the global increase in the aging population and single-person households. However, as the demand for pets rises, the abandonment of pets is also surging, leading to a global crisis of abandoned animals.

[0003] According to research by the Animal Freedom Solidarity in Korea, approximately 110,000 animals were transferred to animal shelters in 2022 due to their owners' irresponsible or intentional abandonment. Of these, dogs accounted for 71.3% of the total, and about 44.1% died in shelters through euthanasia or natural causes. Furthermore, approximately 30 billion won is spent annually on the care of lost and abandoned animals, resulting in massive financial losses.

[0004] In line with this trend, the national pet registration system is being expanded and implemented to prevent animal abandonment. However, pet registration primarily relies on identification tags, which are susceptible to easy removal, damage, or loss. Another method involves implanting microchips into the animal's body, but this approach also presents challenges such as high costs, side effects, and resistance to having the device implanted in the neck.

[0005] Meanwhile, although biometric technologies utilizing electrocardiograms (ECGs) are being researched as non-invasive biometric technologies, most studies focus on humans rather than animals; consequently, recognition rates for animals, particularly dogs, are not high, and there are no examples of appropriate models being developed. Prior art literature

[0006] Republic of Korea Published Patent Application No. 10-2021-0075545 The problem to be solved

[0007] Accordingly, the technical problem of the present invention is conceived from this point, and the objective of the present invention is to provide a dog biometric recognition device using an electrocardiogram that can implement a dog biometric recognition model using an electrocardiogram with only a simple architecture, short training time, and minimal hardware.

[0008] Another objective of the present invention is to provide a method for recognizing a dog's biometrics using an electrocardiogram. means of solving the problem

[0009] A dog biometric recognition device using an electrocardiogram according to one embodiment for realizing the above-described objective of the present invention comprises an electrocardiogram measuring device for measuring the electrocardiogram of a dog, a measuring unit for generating an electrocardiogram signal, a preprocessing unit for removing noise from and normalizing the electrocardiogram signal, an R-peak extraction unit for detecting an R-peak in the preprocessed electrocardiogram signal, a segmenting unit for performing segmentation of the electrocardiogram signal by setting a window centered on the detected R-peak, a first time earlier and a second time later than the R-peak, and a classifier including a neural network for classifying a pre-entered dog individual using the segmented electrocardiogram signals as input.

[0010] In one embodiment of the present invention, the neural network of the classifier may include an input layer that takes the divided electrocardiogram signals as inputs, a convolution layer that performs a one-dimensional convolution operation as a hidden layer, a Long Short-Term Memory (LSTM) layer, and an output layer.

[0011] In one embodiment of the present invention, the calculation of the convolution layer may include the operation of multiplying and adding input data using a moving kernel filter with the following mathematical formula.

[0012] [Mathematical Formula]

[0013] (X(t), W(t), and Y(t) represent the input, kernel filter, and feature map generated from the convolution operation, respectively)

[0014] In one embodiment of the present invention, in the R-peak detection unit, after calculating the average amplitude of the detected R-peak candidates, the R-peak candidates having a value of 80% or less of the average amplitude value may be excluded by considering them as noise.

[0015] In one embodiment of the present invention, in the dividing unit, the first time may be 250 ms and the second time may be 500 ms. The divided electrocardiogram signal may have a size of 200 × 1 and the divided electrocardiogram signal having a size of 200 × 1 may be input to the neural network of the classifier.

[0016] A dog biometric recognition device using an electrocardiogram according to one embodiment for realizing the purpose of the present invention described above includes an electrocardiogram measuring device for measuring the electrocardiogram of a dog, a measuring unit for generating an electrocardiogram signal, a preprocessing unit for removing noise and normalizing the electrocardiogram signal, a dividing unit for blindly dividing the preprocessed electrocardiogram signal into 1-second lengths, and a classifier including a neural network for classifying a pre-entered dog individual by taking the divided electrocardiogram signals as input.

[0017] In one embodiment of the present invention, the dividing unit can perform wavelet transform on the divided electrocardiogram signals.

[0018] In one embodiment of the present invention, the neural network of the classifier may include an input layer that takes the divided electrocardiogram signals as inputs, a convolution layer that performs a one-dimensional convolution operation as a hidden layer, a Long Short-Term Memory (LSTM) layer, and an output layer.

[0019] A dog biometric recognition device using an electrocardiogram according to an embodiment for realizing the purpose of the present invention described above includes a measurement terminal comprising an electrocardiogram measurement device that measures the electrocardiogram of a dog and generates an electrocardiogram signal, and a computing device that receives the electrocardiogram signal, recognizes the corresponding individual information of the dog, and performs dog biometric recognition. The computing device includes a communication module that receives the electrocardiogram signal from the measurement terminal and transmits data regarding the dog biometric recognition result to the measurement terminal; a learning module that trains an artificial intelligence model using the electrocardiogram signal, the dog biometric recognition result, the individual information of the dog, and data generated during the process of performing dog biometric recognition; a storage module that stores the artificial intelligence model, the electrocardiogram signal, the dog biometric recognition result, the individual information of the dog, and the data generated during the process of performing dog biometric recognition; and a control module that inputs the electrocardiogram signal into the pre-trained artificial intelligence model and outputs a biometric recognition result.

[0020] In one embodiment of the present invention, the artificial intelligence model may include an input layer that takes the divided electrocardiogram signals as inputs, a convolution layer that performs a one-dimensional convolution operation as a hidden layer, a Long Short-Term Memory (LSTM) layer, and an output layer.

[0021] A method for biometric recognition of a dog using an electrocardiogram according to one embodiment for realizing the purpose of the present invention described above comprises: a step of obtaining an electrocardiogram signal that generates an electrocardiogram signal by measuring the electrocardiogram of a dog through an electrocardiogram measuring device; a preprocessing step of removing noise and normalizing the electrocardiogram signal using a band-pass filter; an R-peak extraction step of extracting an R-peak from the preprocessed electrocardiogram signal; a segmentation step of segmenting by setting a window 250ms ahead and 500ms behind the R-peak centered on the detected R-peak; and a classification step of performing biometric recognition by inputting the segmented electrocardiogram signal into a neural network that classifies a dog individual that has been input in advance.

[0022] In one embodiment of the present invention, the classification step comprises: a convolution operation step for performing a 1D convolution operation on input data; a max-pooling step for performing max-pooling to reduce the dimensionality of a generated feature map by half; an LSTM step for passing through two LSTM (Long Short Term Memory Network) layers to extract sequential information of an electrocardiogram signal based on a dimensionality-reduced feature map; a post-processing step for flattening the output of the LSTM step into a vector and then passing through two dense layers; and an output step for outputting a dog biometric recognition result at the output layer where the post-processed data is finally output.

[0023] A method for biometric recognition of a dog using an electrocardiogram according to one embodiment for realizing the purpose of the present invention described above includes: a step of obtaining an electrocardiogram signal that generates an electrocardiogram signal by measuring the electrocardiogram of a dog through an electrocardiogram measuring device; a preprocessing step of removing noise and normalizing the electrocardiogram signal using a band-pass filter; a segmentation step of blindly segmenting the preprocessed electrocardiogram signal in 1-second intervals; and a classification step of performing biometric recognition by inputting the segmented electrocardiogram signal into a neural network that classifies a dog individual that has been input in advance. Effects of the invention

[0024] According to embodiments of the present invention, a dog biometric device using an electrocardiogram can implement a dog biometric model using an electrocardiogram with only a simple architecture, a short training time, and minimal hardware. In particular, it can be usefully utilized in small wearable devices, etc.

[0025] However, the effects of the present invention are not limited to the above effects and may be extended in various ways without departing from the spirit and scope of the present invention. Brief explanation of the drawing

[0026] FIG. 1 is a diagram showing the configuration of a dog biometric recognition device using an electrocardiogram according to one embodiment of the present invention. FIG. 2 is a diagram showing the configuration of a dog biometric recognition device using an electrocardiogram according to another embodiment of the present invention. FIG. 3 is a diagram showing the configuration of a dog biometric recognition device using an electrocardiogram according to one embodiment of the present invention. FIG. 4 is a flowchart illustrating a method for recognizing a dog's biometric data using an electrocardiogram according to one embodiment of the present invention. Figure 5 is a flowchart showing in detail the classification steps of the dog biometric recognition method using the electrocardiogram of Figure 4. FIG. 6 is a flowchart illustrating a method for recognizing a dog's biometric data using an electrocardiogram according to another embodiment of the present invention. FIG. 7 is a block diagram illustrating the operation of a dog biometric recognition device using an electrocardiogram according to one embodiment of the present invention. Figure 8 (a) is a graph showing an example of an R-peak-based split signal, and (b) is a graph showing an example of a blind split signal. Figure 9 is a diagram showing the characteristics of normal electrocardiogram signals of dogs in a table. FIG. 10 is a diagram illustrating the architecture of a deep neural network model of a classifier of a dog biometric recognition device using an electrocardiogram according to one embodiment of the present invention. Figure 11 is a diagram showing an example of the output of the 1D CNN of the neural network model of Figure 9. Figure 12 is a diagram showing the intermediate output of the LSTM of the neural network model of Figure 9. Figure 13 is a diagram showing the confusion matrix scheme in table form. Figure 14 is a diagram showing the identification accuracy results in a table. Specific details for implementing the invention

[0027] Hereinafter, preferred embodiments of the present invention will be described in more detail with reference to the drawings.

[0028] The present invention is capable of various modifications and may take various forms, and specific embodiments are illustrated in the drawings and described in detail in the text. However, this is not intended to limit the invention to the specific disclosed forms, and it should be understood that the invention includes all modifications, equivalents, and substitutions that fall within the spirit and scope of the invention.

[0029] FIG. 1 is a diagram showing the configuration of a dog biometric recognition device using an electrocardiogram according to one embodiment of the present invention.

[0030] Referring to FIG. 1, the dog biometric recognition device using the electrocardiogram includes a measurement unit (110), a preprocessing unit (120), an R-peak extraction unit (130), a segmentation unit (140), and a classifier (150).

[0031] The above measurement unit (110) includes an electrocardiogram (ECG) measuring device and can generate an electrocardiogram signal by measuring the electrocardiogram of a dog through the electrocardiogram measuring device. An ECG is a graph showing the electrical changes of the heart that occur during the contraction and relaxation of the heart muscle during a heartbeat. The electrical changes of the heart are caused by potential differences resulting from the depolarization and repolarization of heart cells. The electrocardiogram signal is a one-dimensional analog signal that measures the action current generated in the myocardium according to the beating of the dog's heart using the standard 12-lead method, and the measured electrocardiogram signal can be represented by recording the change in amplitude (V) during the measurement time (Sec).

[0032] The above electrocardiogram measuring device can be implemented using various known methods to measure electrocardiograms, for example, a method of measuring pressure applied to electrode pads, a method of measuring microcurrents generated during the depolarization process of the myocardium, a non-contact method using photoplethysmography sensors, etc., and the electrocardiogram can be measured in various ways, but is not limited to the methods exemplified.

[0033] The above preprocessing unit (120) can perform preprocessing to analyze the electrocardiogram signal. For example, the above preprocessing unit (120) can remove noise from the measured electrocardiogram signal using a band-pass filter and normalize it.

[0034] The R-peak extraction unit (130) can extract R-peaks from the preprocessed electrocardiogram signal. To extract features to be used as input to a classifier for ECG-based biometric recognition, the start and end points of a single heartbeat must first be detected in the continuous electrocardiogram signal. In particular, the reference feature extraction method is a method of extracting features based on the morphological features of a reference point, and the R-peaks of the electrocardiogram signal can be used to diagnose atrial fibrillation (AFib). The R-peak extraction unit (130) can perform 4-level DWT to detect R-peaks in the electrocardiogram signal.

[0035] The above DWT is performed using the following [Equation 1].

[0036] [Mathematical Formula 1]

[0037] (Here, ф, j, and k represent the mother wavelet, scale parameter, and translation parameter, respectively)

[0038] Based on this, a continuous wavelet transform (CWT) is performed on the input signal x(t). At this time, the wavelet coefficient γ can be defined as follows [Equation 2].

[0039] [Mathematical Formula 2]

[0040] (ф, j, and k represent the mother wavelet, scale parameter, and translation parameter, respectively. Based on this, the Continuous Wavelet Transform (CWT) wavelet coefficient γ for the input signal x(t) is defined as above.)

[0041] Subsequently, unwanted noise can be removed from the signal using SWT, and then the R-peak can be found using QRS complex detection and additional algorithms. SWT is generally denoted by the coefficient cj,k, as shown in [Equation 3] below.

[0042] [Mathematical Formula 3]

[0043] R-peaks can be detected by applying a 3-level SWT after using the Daubechies 3 (db3) wavelet function.

[0044] At this time, although noise was removed from the measured electrocardiogram signal using a band-pass filter, some spike-shaped noise remained, leading to instances where it was incorrectly detected as an R-peak. To resolve this issue, the average amplitude of the detected R-peak candidates was calculated, and all peak points having a value less than 80% of the average value—which is considered noise rather than an R-peak—were removed. When the value was set higher than 80%, R-peaks were considered noise, and when it was set lower than 80%, the noise was not removed; therefore, the optimal value was set to 80%.

[0045] The above-mentioned splitting unit (140) splits the electrocardiogram signal to include a single heartbeat. For example, it performs splitting of the electrocardiogram signal by setting a window 250ms before and 500ms after the R peak detected above.

[0046] Considering that the PR interval varies inversely to the heart rate within the range of 60 to 130 ms, a length of 400 ms is considered optimal to include a single heartbeat electrocardiogram signal.

[0047] The classifier (150) can classify pre-input dog individuals from the preprocessed and R-peak-centered electrocardiogram signals. In the case of electrocardiogram-based biometric recognition using electrocardiogram signals, individual features are extracted from time-varying waveforms formed by small electrical signals generated during heartbeats and used for identification. The classifier (150) can distinguish dog individuals from electrocardiogram signals and perform dog biometric recognition using a pre-trained artificial intelligence algorithm. The specific operation algorithm of the classifier (150) will be explained next.

[0048] After preprocessing for noise reduction and segmentation of the electrocardiogram signals, a dog biometric model was created using a 1D_CNN LSTM neural network (see Fig. 10). The neural network is composed of a deep learning architecture consisting of one input layer, seven hidden layers, and one output layer. The input dimension of the neural network can have a size of 200 × 1 to receive R-peak based window segmentation signals as input data.

[0049] A 1D convolution operation is performed on the input data. The calculation of the convolution layer may include multiplying and adding the input data using a moving kernel filter as shown in [Equation 4].

[0050] [Mathematical Formula 4]

[0051] In the above [Equation 4], X(t), W(t), and Y(t) represent the input, kernel filter, and feature map generated from the convolution operation, respectively. After numerous experiments for optimal performance, the kernel filter size was selected as 33.

[0052] In addition, to reduce the input data size to 1 / 4, a stride of 4 and zero padding were implemented before passing it to the next layer.

[0053] In this embodiment, a 1D CNN structure connected to an LSTM (Long short-term memory) is used, which can maintain time-series characteristics without requiring additional image conversion.

[0054] The detailed structure of the neural network is as follows.

[0055] First, a 1D CNN (1D Convolutional Neural Network) was constructed to extract features and generate feature maps while maintaining the format of the electrocardiogram time series data. Then, max-pooling was performed to reduce the dimensionality of the generated feature maps by half. Figure 11 shows a portion of the 3D schematic diagram of the 1D_CNN and feature maps using 200 inputs configured through R-peak-based segmentation. Specifically, this involves convolving 33 values—equivalent to the kernel filter size—with a stride of 4 from the 200 single heart rate input data samples segmented based on R-peaks. Subsequently, the intermediate outputs passing through the 1D_CNN layers are represented as 128 line graphs in the form of 25 data streams. Based on the reduced dimensionality of the feature maps, two LSTM layers were added to extract sequential information of the electrocardiogram signals. This enables the LSTM layers to perform sequential learning. Time series feature maps were extracted and reduced using 1D CNN and max pooling.

[0056] Figure 12 shows the intermediate output of an LSTM represented by a line graph of 256 lines consisting of 25 data streams, using features obtained from the 1D CNN of Figure 11 as input. The output of the LSTM is flattened to convert the data into a one-dimensional vector and passed as input to a subsequent dense layer. After passing through two additional dense layers, the final biometric recognition result is obtained in an output layer composed of 33 neurons. A softmax function was used in the output layer to calculate the probability of the predicted dog identification result for each input value. After the 1D convolution was completed, batch normalization was performed, and dropout was applied at a rate of 0.3 to all layers except the input and output layers. For the LSTM layer, the cost is calculated in the final step, and the intermediate output of each step is fed to the next layer. The weights of all layers were initialized using an initialization method.

[0057] Meanwhile, the above neural network was trained using an R-peak based window partitioning dataset. The training process was performed for 150 epochs with a batch size of 32 and a learning rate of 0.0001. The Adam optimizer was used for parameter tuning to minimize the loss function.

[0058] According to the present invention, a dog biometric recognition model using an electrocardiogram can be implemented with a simple architecture, low training time, and minimal hardware. In particular, since the electrocardiogram signal can satisfy all of the universality, uniqueness, collectibility, circumvention, acceptability, and permanence requirements for biometric identification, it can be easily used for individual dog identification.

[0059] FIG. 2 is a diagram showing the configuration of a dog biometric recognition device using an electrocardiogram according to another embodiment of the present invention.

[0060] Referring to FIG. 2, the dog biometric device using the electrocardiogram is substantially the same as the dog biometric device using the electrocardiogram of FIG. 1, except that instead of finding the R peak in the electrocardiogram signal and dividing it into a certain range window, it performs blind division. Therefore, repetitive descriptions are brief or omitted.

[0061] The above-described biometric device using an electrocardiogram includes a measurement unit (210), a preprocessing unit (220), a dividing unit (240), and a classifier (250).

[0062] The above measurement unit (210) includes an electrocardiogram (ECG) measurement device and can generate an electrocardiogram signal by measuring an electrocardiogram signal through the electrocardiogram measurement device. The above measurement unit (210) may be substantially the same as the measurement unit (110) of FIG. 1, and a detailed description is omitted.

[0063] The above preprocessing unit (220) can perform preprocessing to analyze the electrocardiogram signal. The above preprocessing unit (220) may be substantially the same as the preprocessing unit (120) of FIG. 1, and a detailed description is omitted.

[0064] The above-mentioned splitting unit (240) blindly splits the preprocessed electrocardiogram signal to include at least one single heartbeat. For example, the electrocardiogram signal can be split into a range of 1 second. Since the heart rate of an adult dog is in the range of 70 to 160 bpm and the heart rate of a puppy is in the range of 70 to 200 bpm, when splitting into a range of 1 second, at least one single heartbeat can be included as an analysis target. To remove the phase difference between randomly selected signals, a wavelet transform can be performed.

[0065] Figure 8(b) shows an example of a blindly divided electrocardiogram signal. Since the other components of the division unit (240) are substantially the same as those of the division unit (140) in Figure 1, a redundant description is omitted.

[0066] The classifier (250) can classify a number of individuals that have been input in advance from the preprocessed and segmented electrocardiogram signals. The classifier (150) can distinguish individuals from electrocardiogram signals and perform biometric recognition of the individuals using a pre-learned artificial intelligence algorithm, and the specific operation algorithm may be substantially the same as the classifier (150) of FIG. 1.

[0067] At this time, the neural network input dimension can have a size of 512 × 1 to receive blind-based segmented electrocardiogram signals as input data.

[0068] According to the present embodiment, since it does not require an algorithm to find a separate reference point in the electrocardiogram signal, relatively low computational cost and fast processing are possible.

[0069] FIG. 3 is a diagram showing the configuration of a dog biometric recognition device using an electrocardiogram according to one embodiment of the present invention.

[0070] Referring to FIG. 3, the dog biometric recognition device using the electrocardiogram includes a computing device (10) and a measurement terminal (20).

[0071] The above computing device (10) may be configured to include a communication module (11), a learning module (13), a storage module (15), and a control module (17).

[0072] The communication module (11) transmits and receives at least one piece of information or data with at least one electrocardiogram measuring device (20). Additionally, the communication module (11) may also perform communication with other servers or devices, and transmits and receives wireless signals in a communication network according to wireless internet technologies.

[0073] Wireless internet technologies include, for example, WLAN (Wireless LAN), Wi-Fi (Wireless-Fidelity), Wi-Fi (Wireless Fidelity) Direct, DLNA (Digital Living Network Alliance), WiBro (Wireless Broadband), WiMAX (World Interoperability for Microwave Access), HSDPA (High Speed ​​Downlink Packet Access), HSUPA (High Speed ​​Uplink Packet Access), LTE (Long Term Evolution), LTE-A (Long Term Evolution-Advanced), etc., and the computing device (10) transmits and receives data according to at least one wireless internet technology within a range that includes internet technologies not listed above.

[0074] For short-range communication, short-range communication can be supported using at least one of the following technologies: Bluetooth™, RFID (Radio Frequency Identification), Infrared Data Association (IrDA), Ultra Wideband (UWB), ZigBee, Near Field Communication (NFC), Wireless-Fidelity (Wi-Fi), Wi-Fi Direct, and Wireless Universal Serial Bus (Wireless USB). Such short-range wireless communication networks can support wireless communication between the computing device (10) and at least one electrocardiogram measuring device (20). In this case, the short-range wireless communication network may be a short-range wireless personal area network.

[0075] The above learning module (130) trains an artificial intelligence model using an input dataset. At this time, there may be at least one artificial intelligence model, but is not limited thereto.

[0076] The above storage module (150) stores various data (information), artificial intelligence models, etc. that are received by the computing device (10) or generated or acquired (measured) by the computing device (10), and stores at least one process necessary to calculate a value for individual identification from the electrocardiogram signal of a dog that is a biometric recognition target. Here, at least one artificial intelligence model may be stored.

[0077] The control module (17) controls the operation to produce an output value for object recognition from a dog electrocardiogram signal through all components based on at least one process stored in the storage module (15).

[0078] Specifically, the control module (170) collects a plurality of electrocardiogram signals, checks the labeled output value for each of the collected plurality of electrocardiogram signals to generate a dataset, and then performs training on an artificial intelligence model using the generated dataset.

[0079] Subsequently, when a specific number of electrocardiogram signals are input, the control module (170) uses a pre-trained artificial intelligence model to produce and provide an output value corresponding to the input electrocardiogram signals, i.e., a biometric recognition result through a classifier.

[0080] FIG. 4 is a flowchart illustrating a method for recognizing a dog's biometric data using an electrocardiogram according to one embodiment of the present invention.

[0081] Referring to FIG. 4, a dog biometric recognition method using conduction includes a step of acquiring an electrocardiogram signal (S100), a preprocessing step (S200), an R-peak extraction step (S300), a segmentation step (S4000), and a classification step (S500).

[0082] In the step of acquiring the above electrocardiogram signal (S100), the electrocardiograms of the dogs are measured through an electrocardiogram measuring device to generate an electrocardiogram signal.

[0083] In the above preprocessing step (S200), the electrocardiogram signal can be normalized and noise removed using a band-pass filter.

[0084] In the above R-peak extraction step (S300), an R-peak can be extracted from the preprocessed electrocardiogram signal. The process for extracting the R-peak is omitted as it has been described above in the description of FIG. 1.

[0085] In the above division step (S4000), the division can be performed by setting windows before and after the detected R-peak by a predetermined amount of time relative to the R-peak.

[0086] In the above classification step (S500), the divided electrocardiogram signals are used as input to classify the number of individuals entered in advance using a neural network algorithm. The specific process of the neural network is omitted as it has been described above in the description of FIG. 1.

[0087] Figure 5 is a flowchart showing in detail the classification steps of the dog biometric recognition method using the electrocardiogram of Figure 4.

[0088] Referring to FIGS. 4 and 5, the classification step (500) may include a convolution operation step (S510), a max pooling step (S520), an LSTM step (S530), a post-processing step (S540), and an output step (S550).

[0089] In the above convolution operation step (S510), a feature map is generated by performing a 1D convolution operation on the input data.

[0090] In the above max pooling step (S520), max pooling is performed to reduce the dimension of the generated feature map by half.

[0091] In the above LSTM step (S530), the data passes through two LSTM (Long Short Term Memory Network) layers to extract sequential information of the electrocardiogram signal based on the dimensionality-reduced feature map. Through this, the LSTM layers can perform sequential learning. LSTM is a deep learning model developed for the long-term dependency problem of RNN (Recurrent Neural Networks), and as a variation of the existing RNN model, it is a representative improved model that alleviates the gradient vanishing problem of RNN. This model includes a memory cell in the hidden node, and this memory cell is a gate that stores a value in storage, outputs it, and adjusts the forgotten value.

[0092] In the above post-processing step (S540), the output of the LSTM is flattened to convert the data into a one-dimensional vector and is passed as the input to a subsequent dense layer. After passing through two additional dense layers, the final biometric recognition result can be obtained in the output layer.

[0093] In the output step (S550) above, the final biometric recognition result is output, and the biometric recognition of the subject of the electrocardiogram measurement can be completed. For example, the output values ​​of output nodes corresponding to pre-entered dog recognition information are calculated using a softmax function, and the corresponding dog recognition information is output as a recognition result, thereby completing the dog biometric recognition using the electrocardiogram.

[0094] FIG. 6 is a flowchart illustrating a method for recognizing a dog's biometric data using an electrocardiogram according to another embodiment of the present invention.

[0095] Referring to FIG. 6, a dog biometric recognition method using an electrocardiogram includes a step of acquiring an electrocardiogram signal (S10), a preprocessing step (S20), a segmentation step (S40), and a classification step (S50).

[0096] The above-described dog biometric recognition method using an electrocardiogram is substantially the same as the dog biometric recognition method using an electrocardiogram described in FIGS. 4 and 5, except that it performs blind division instead of finding the R peak and dividing the window of a certain range, so a repetitive explanation is omitted.

[0097] In the above splitting step (S40), the preprocessed electrocardiogram signal is blind-divided to include at least one single heartbeat. For example, the electrocardiogram signal can be divided into a range of 1 second.

[0098] FIG. 7 is a block diagram illustrating the operation of a dog biometric device using an electrocardiogram according to an embodiment of the present invention. FIG. 8 (a) is a graph showing an example of an R-peak-based segmented signal, and (b) is a graph showing an example of a blind segmented signal. FIG. 9 is a table showing the characteristics of a normal electrocardiogram signal of a dog. FIG. 10 is a diagram illustrating the architecture of a neural network model of a classifier of a dog biometric device using an electrocardiogram according to an embodiment of the present invention.

[0099] The contents shown in FIGS. 7 to 10 are identical to the specific descriptions of the dog biometric recognition device and method using an electrocardiogram described in FIGS. 1 to 6, so a repetitive description is omitted.

[0100] Figure 11 is a diagram showing an example of the output of the 1D CNN of the neural network model of Figure 9.

[0101] Referring to Figure 11, a portion of the 3D schematic of a 1D_CNN and a feature map using 200 inputs configured through R-peak-based segmentation is shown. That is, it is a process of convolving 33 values, equal to the kernel filter size, out of 200 single heart rate input data samples segmented based on R-peaks, with a stride of 4.

[0102] Figure 12 is a diagram showing the intermediate output of the LSTM of the neural network model of Figure 9.

[0103] Referring to Fig. 12, the intermediate output passing through the 1D CNN layer of Fig. 11 is represented in the form of 128 line graphs in the form of 25 data streams. Two LSTM layers were added to extract sequential information of the electrocardiogram signal based on the dimensionality reduction of the feature map. Through this, the LSTM layers can perform sequential learning. Time-series feature maps were extracted and reduced through the 1D CNN and max pooling.

[0104] This shows the intermediate output of an LSTM, represented as a line graph of 256 lines consisting of 25 data streams, using features obtained from a 1D CNN as input. The LSTM output is flattened to convert the data into a one-dimensional vector and passed as input to a subsequent dense layer. After passing through two additional dense layers, the final biometric recognition result is obtained in an output layer composed of 33 neurons. A softmax function was used in the output layer to calculate the probability of the predicted dog identification result for each input value. Batch normalization was performed after the 1D convolution was completed, and dropout was applied at a rate of 0.3 to all layers except the input and output layers. For the LSTM layer, the cost is calculated in the final step, and the intermediate output of each step is fed to the next layer. The weights of all layers were initialized using an initialization method.

[0105] Figure 13 is a table showing the confusion matrix scheme. Figure 14 is a table showing the identification accuracy results.

[0106] Referring to Figures 13 and 14, for performance evaluation, the model was verified by comparing accuracy, equal error rate (EER), and receiver operating characteristic curve (ROC) based on the performance evaluation confusion matrix of Figure 13.

[0107] Referring to FIG. 14, three additional neural network models for comparing the identification accuracy of the proposed recognition model according to the present embodiment and performance in six experimental scenarios are shown. The accuracy was calculated using [Equation 5] below based on FIG. 13.

[0108] [Mathematical Formula 5]

[0109] As shown in Figure 14, the excellence of the model was proven by confirming that the proposed recognition model demonstrates uniform and excellent accuracy regardless of the type of R-peak splitting, blind splitting, or data setter.

[0110] According to embodiments of the present invention, a dog biometric device using an electrocardiogram can implement a dog biometric model using an electrocardiogram with only a simple architecture, a short training time, and minimal hardware. In particular, it can be usefully utilized in small wearable devices, etc.

[0111] Although the invention has been described with reference to the above embodiments, those skilled in the art will understand that various modifications and changes can be made to the invention without departing from the spirit and scope of the invention as described in the following claims. Explanation of the symbols

[0112] 100: Dog biometric device using electrocardiogram 110: Measurement unit 120: Preprocessing unit 130: R-peak extraction unit 140: Splitting unit 150: Classifier

Claims

Claim 1 A dog biometric recognition device using an electrocardiogram, comprising: an electrocardiogram measuring device for measuring an electrocardiogram of a dog; a measuring unit for generating an electrocardiogram signal; a preprocessing unit for removing noise from and normalizing the electrocardiogram signal; an R-peak extraction unit for detecting an R-peak in the preprocessed electrocardiogram signal; a segmentation unit for performing segmentation of the electrocardiogram signal by setting a window centered on the detected R-peak, a first time earlier and a second time later than the R-peak, and a classifier including a neural network for classifying a pre-entered dog individual using the segmented electrocardiogram signals as input. Claim 2 A dog biometric recognition device using an electrocardiogram according to claim 1, wherein the neural network of the classifier comprises an input layer that takes the divided electrocardiogram signals as inputs, a convolution layer that performs a one-dimensional convolution operation as a hidden layer, a Long Short-Term Memory (LSTM) layer, and an output layer. Claim 3 A dog biometric device using an electrocardiogram, characterized in that, in paragraph 2, the calculation of the convolution layer includes the operation of multiplying and adding input data using a moving kernel filter with the following mathematical formula. [Mathematical Equation](X(t), W(t), and Y(t) represent the input, kernel filter, and feature map generated from the convolution operation, respectively) Claim 4 A dog biometric recognition device using an electrocardiogram according to claim 1, characterized in that, in the R-peak extraction unit, after calculating the average amplitude of the detected R-peak candidates, the R-peak candidates having a value of 80% or less of the average amplitude value are considered as noise and excluded. Claim 5 A dog biometric recognition device using an electrocardiogram according to claim 1, wherein in the dividing unit, the first time is 250ms and the second time is 500ms, the divided electrocardiogram signal has a size of 200 × 1, and the divided electrocardiogram signal having a size of 200 × 1 is input to the neural network of the classifier. Claim 6 A dog biometric recognition device using an electrocardiogram, comprising: an electrocardiogram measuring device for measuring an electrocardiogram of a dog; a measuring unit for generating an electrocardiogram signal; a preprocessing unit for removing noise and normalizing the electrocardiogram signal; a dividing unit for blindly dividing the preprocessed electrocardiogram signal into 1-second lengths; and a classifier including a neural network for classifying a pre-entered dog individual using the divided electrocardiogram signals as input. Claim 7 A biometric device using an electrocardiogram according to claim 1, wherein the dividing unit performs a wavelet transform on the divided electrocardiogram signals. Claim 8 delete Claim 9 A dog biometric recognition device using an electrocardiogram, characterized by comprising: a measurement terminal including an electrocardiogram measurement device that measures an electrocardiogram of a dog and generates an electrocardiogram signal; a computing device that receives the electrocardiogram signal, recognizes corresponding dog information, and performs dog biometric recognition, wherein the computing device receives the electrocardiogram signal from the measurement terminal and transmits data regarding the dog biometric recognition result to the measurement terminal; a learning module that trains an artificial intelligence model using the electrocardiogram signal, the dog biometric recognition result, the dog information, and data generated during the process of performing dog biometric recognition; a storage module that stores the artificial intelligence model, the electrocardiogram signal, the dog biometric recognition result, the dog information, and data generated during the process of performing dog biometric recognition; and a control module that divides the electrocardiogram signal, inputs the divided electrocardiogram signals into the pre-trained artificial intelligence model, and outputs a biometric recognition result. Claim 10 A dog biometric recognition device using an electrocardiogram according to claim 9, characterized in that the artificial intelligence model comprises an input layer that takes the segmented electrocardiogram signals as inputs, a convolution layer that performs a one-dimensional convolution operation as a hidden layer, a Long Short-Term Memory (LSTM) layer, and an output layer. Claim 11 A method for dog biometric recognition using an electrocardiogram, comprising: a step of obtaining an electrocardiogram signal that generates an electrocardiogram signal by measuring the electrocardiogram of a dog through an electrocardiogram measuring device; a preprocessing step of removing noise and normalizing the electrocardiogram signal using a bandpass filter; an R-peak extraction step of extracting an R-peak from the preprocessed electrocardiogram signal; a segmentation step of segmenting by setting a window 250ms ahead and 500ms behind the R-peak centered on the detected R-peak; and a classification step of performing biometric recognition by inputting the segmented electrocardiogram signal into a neural network that classifies a pre-entered dog. Claim 12 In claim 11, the classification step comprises: a convolution operation step that performs a 1D convolution operation on input data; a max-pooling step that performs max-pooling to reduce the dimensionality of the generated feature map by half; an LSTM step that passes through two LSTM (Long Short Term Memory Network) layers to extract sequential information of the electrocardiogram signal based on the dimensionality-reduced feature map; a post-processing step that flattens the output of the LSTM step and passes through two dense layers to convert it into a vector; and an output step in which the post-processed data finally outputs a dog biometric recognition result at an output layer. Claim 13 A method for dog biometric recognition using an electrocardiogram, comprising: a step of obtaining an electrocardiogram signal that generates an electrocardiogram signal by measuring the electrocardiogram of a dog through an electrocardiogram measuring device; a preprocessing step of removing noise and normalizing the electrocardiogram signal using a bandpass filter; a segmentation step of blindly segmenting the preprocessed electrocardiogram signal in 1-second intervals; and a classification step of performing biometric recognition by inputting the segmented electrocardiogram signal into a neural network that classifies pre-entered dog entities.

Images