CN112307996A - Fingertip electrocardiogram identity recognition device and method - Google Patents
Fingertip electrocardiogram identity recognition device and method Download PDFInfo
- Publication number
- CN112307996A CN112307996A CN202011226668.XA CN202011226668A CN112307996A CN 112307996 A CN112307996 A CN 112307996A CN 202011226668 A CN202011226668 A CN 202011226668A CN 112307996 A CN112307996 A CN 112307996A
- Authority
- CN
- China
- Prior art keywords
- layer
- fingertip
- neural network
- ecg signal
- convolutional neural
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 238000000034 method Methods 0.000 title claims abstract description 65
- 238000013528 artificial neural network Methods 0.000 claims abstract description 18
- 238000007781 pre-processing Methods 0.000 claims abstract description 18
- 230000003750 conditioning effect Effects 0.000 claims abstract description 16
- 238000004891 communication Methods 0.000 claims abstract description 13
- 238000013508 migration Methods 0.000 claims abstract description 13
- 230000005012 migration Effects 0.000 claims abstract description 13
- 238000013527 convolutional neural network Methods 0.000 claims description 88
- 238000012545 processing Methods 0.000 claims description 46
- 230000006870 function Effects 0.000 claims description 35
- 230000003044 adaptive effect Effects 0.000 claims description 28
- 238000012549 training Methods 0.000 claims description 22
- 238000010586 diagram Methods 0.000 claims description 18
- 239000011159 matrix material Substances 0.000 claims description 16
- 238000005516 engineering process Methods 0.000 claims description 13
- 238000013526 transfer learning Methods 0.000 claims description 10
- 125000004122 cyclic group Chemical group 0.000 claims description 9
- 238000013519 translation Methods 0.000 claims description 9
- 230000006978 adaptation Effects 0.000 claims description 8
- 238000013139 quantization Methods 0.000 claims description 6
- 238000005070 sampling Methods 0.000 claims description 6
- 230000008859 change Effects 0.000 claims description 5
- 238000010276 construction Methods 0.000 claims description 5
- 238000010606 normalization Methods 0.000 claims description 5
- 238000011176 pooling Methods 0.000 claims description 5
- 230000008014 freezing Effects 0.000 claims description 3
- 238000007710 freezing Methods 0.000 claims description 3
- 238000013507 mapping Methods 0.000 claims description 3
- 230000009466 transformation Effects 0.000 claims description 3
- 230000008929 regeneration Effects 0.000 claims description 2
- 238000011069 regeneration method Methods 0.000 claims description 2
- 230000008569 process Effects 0.000 abstract description 16
- 210000000038 chest Anatomy 0.000 description 19
- 210000000115 thoracic cavity Anatomy 0.000 description 9
- 238000004364 calculation method Methods 0.000 description 8
- 230000007246 mechanism Effects 0.000 description 7
- 238000013135 deep learning Methods 0.000 description 6
- 238000005259 measurement Methods 0.000 description 5
- 238000006243 chemical reaction Methods 0.000 description 4
- 230000000694 effects Effects 0.000 description 4
- 230000010355 oscillation Effects 0.000 description 4
- 238000012546 transfer Methods 0.000 description 4
- 230000003321 amplification Effects 0.000 description 3
- 230000008901 benefit Effects 0.000 description 3
- 238000011161 development Methods 0.000 description 3
- 230000018109 developmental process Effects 0.000 description 3
- 238000001914 filtration Methods 0.000 description 3
- 230000033001 locomotion Effects 0.000 description 3
- 238000003199 nucleic acid amplification method Methods 0.000 description 3
- 238000004458 analytical method Methods 0.000 description 2
- 238000013473 artificial intelligence Methods 0.000 description 2
- 230000005540 biological transmission Effects 0.000 description 2
- 238000013461 design Methods 0.000 description 2
- 230000008034 disappearance Effects 0.000 description 2
- 230000002401 inhibitory effect Effects 0.000 description 2
- 238000003062 neural network model Methods 0.000 description 2
- 238000005457 optimization Methods 0.000 description 2
- 238000012360 testing method Methods 0.000 description 2
- 239000000853 adhesive Substances 0.000 description 1
- 230000001070 adhesive effect Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000008030 elimination Effects 0.000 description 1
- 238000003379 elimination reaction Methods 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 230000004927 fusion Effects 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 230000010354 integration Effects 0.000 description 1
- 239000002184 metal Substances 0.000 description 1
- 230000002093 peripheral effect Effects 0.000 description 1
- 238000007493 shaping process Methods 0.000 description 1
- 239000007787 solid Substances 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/02—Preprocessing
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
- G06F17/10—Complex mathematical operations
- G06F17/14—Fourier, Walsh or analogous domain transformations, e.g. Laplace, Hilbert, Karhunen-Loeve, transforms
- G06F17/148—Wavelet transforms
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F21/00—Security arrangements for protecting computers, components thereof, programs or data against unauthorised activity
- G06F21/30—Authentication, i.e. establishing the identity or authorisation of security principals
- G06F21/31—User authentication
- G06F21/32—User authentication using biometric data, e.g. fingerprints, iris scans or voiceprints
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/02—Preprocessing
- G06F2218/04—Denoising
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Artificial Intelligence (AREA)
- Pure & Applied Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Optimization (AREA)
- Mathematical Analysis (AREA)
- Computational Mathematics (AREA)
- Evolutionary Computation (AREA)
- Computer Security & Cryptography (AREA)
- Biophysics (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Computational Linguistics (AREA)
- Biomedical Technology (AREA)
- Health & Medical Sciences (AREA)
- Computer Hardware Design (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Signal Processing (AREA)
- Algebra (AREA)
- Databases & Information Systems (AREA)
- Measurement And Recording Of Electrical Phenomena And Electrical Characteristics Of The Living Body (AREA)
Abstract
The invention discloses a fingertip electrocardiogram identity recognition device and method. The device comprises a fingertip electrocardiosignal acquisition module, a signal conditioning circuit, a controller and a Bluetooth communication module, wherein the output end of the fingertip electrocardiosignal acquisition module is electrically connected with the input end of the signal conditioning circuit, the output end of the signal conditioning circuit is electrically connected with the controller, and the controller is also electrically connected with the Bluetooth communication module. The method comprises the steps of collecting fingertip ECG signals of a human fingertip, preprocessing the fingertip ECG signals, inputting the preprocessed fingertip ECG signals into a fingertip electrocardio-identity recognition model, carrying out identity recognition on the fingertip electrocardio-identity recognition model, outputting corresponding identity information, and carrying out depth model migration twice on the fingertip electrocardio-identity recognition model based on a depth convolution neural network GoogLeNet to obtain the fingertip electrocardio-identity recognition model. The invention collects the fingertip ECG signal to identify the identity, thereby not only ensuring the stability of the electrocardio collection process, but also realizing convenient and comfortable identification under a small sample.
Description
Technical Field
The invention relates to the technical field of modern identity recognition, in particular to a fingertip electrocardiogram identity recognition device and method.
Background
With the gradual and deep development of the Internet of Things (IoT) and the Artificial Intelligence (AI), many lawless persons apply the IoT and the AI to modern information security systems, which leads to the increasing rampant of more and more intelligent counterfeiting technologies and theft technologies such as fake faces, fake fingerprints, voice imitation, etc., and this undoubtedly increases the potential danger of the information security systems. The development of the sensor technology promotes the gradual application of the electrocardio-acquisition chip which has small volume, easy integration and low power consumption and does not need conductive adhesive, provides technical support for developing a wearable or fingertip Electrocardiosignal (ECG) acquisition system, and provides wide application prospect for an ECG-based identity recognition system. Because the ECG living body acquisition requirement and the hidden inherent characteristic enable the ECG living body acquisition requirement and the hidden inherent characteristic to have very high resistance to external attack, the fingertip electrocardio-based identity recognition technology has wider application prospect. However, there are technical challenges, mainly expressed in the following areas:
the first challenge comes from the physical implementation of the fingertip ecg acquisition system. As a time varying signal, the ECG is extremely unstable during acquisition, causing a significant amount of noise interference.
The second challenge comes from the accurate identification of the fingertip-ecg system. On one hand, the algorithm applicability challenges: compared with the ECG acquired by chest standard leads, the amplitude of ECG components of signals acquired by fingertips is weaker, the interference of motion artifacts and the like is more frequent and even dominates, most of the existing ECG identification methods mostly adopt chest acquisition ECG signals with abundant data quantity as data sets for model training, and the chest acquisition ECG signals are rapidly transferred from an identification system acquired by the chest to an identification system acquired by the fingertips, so that the practicability and the adaptability of the ECG identification system still face challenges; on the other hand, the absence of fingertip data samples: the brisk development of Deep Learning (DL) enables more and more experts in various fields at home and abroad to see the strong advantages of the Deep Learning in feature self-Learning, and also provides strong technical support for fingertip electrocardiogram identity recognition, but the strong and sufficient feature self-Learning still depends on a large data sample, and how to solve the problem that the Deep neural network is still very reluctant to rely on the large computing capacity of the large data in feature self-Learning under the condition of insufficient sample data (the existing network database has less ECG data acquired by wearable and fingertip-type acquisition equipment).
In view of the above, an identity recognition device and method based on fingertip electrocardiograph acquisition are provided, which can realize efficient and accurate identification of individual identity while satisfying stable fingertip electrocardiograph acquisition, and have wide practical application value.
Disclosure of Invention
The invention provides a fingertip ECG identification device and a method, aiming at solving the technical problems that the existing ECG identification system needs to collect thoracic ECG signals, and the collection process is complicated and inconvenient.
In order to solve the problems, the invention adopts the following technical scheme:
the invention discloses a fingertip electrocardio identity recognition device which comprises a fingertip electrocardiosignal acquisition module, a signal conditioning circuit, a controller and a Bluetooth communication module, wherein the output end of the fingertip electrocardiosignal acquisition module is electrically connected with the input end of the signal conditioning circuit, the output end of the signal conditioning circuit is electrically connected with the controller, and the controller is also electrically connected with the Bluetooth communication module.
According to the scheme, fingertip electrocardiosignals of a user are rapidly acquired through a fingertip electrocardiosignal acquisition module, are subjected to differential amplification and filtering through a signal conditioning circuit, are sent to an ADC (analog-to-digital converter) of a controller MCU (micro control unit) for analog-to-digital acquisition and conversion, are sent to a mobile phone end through a Bluetooth communication module in a UART (universal asynchronous receiver/transmitter) mode, and the mobile phone end acquires fingertip electrocardiosignal data of the person to be identified. Meanwhile, the mobile phone end accesses the cloud platform through the wifi/4G wireless transmission module, uploads fingertip electrocardio data of the current user and downloads a recognition result of the fingertip electrocardio-based identity recognition model in the cloud platform. The method only needs to collect fingertip electrocardiogram data, and is more convenient and comfortable compared with the method for collecting thoracic electrocardiogram data.
The invention relates to a fingertip electrocardio identity recognition method, which comprises the following steps:
collecting fingertip ECG signals of a human fingertip, preprocessing the fingertip ECG signals, inputting the preprocessed fingertip ECG signals to a fingertip electrocardiogram identity recognition model, carrying out identity recognition on the fingertip electrocardiogram identity recognition model, and outputting corresponding identity information;
the construction method of the fingertip electrocardio identity recognition model comprises the following steps:
s1: establishing a deep convolutional neural network GoogLeNet;
s2: constructing a deep convolutional neural network GoogLeNet-I1 capable of identifying individual identity according to a chest ECG signal based on the deep convolutional neural network GoogLeNet, preprocessing large sample chest ECG signal data acquired by standard leads in a database (a Physionet network reference library) by using a finetune technology of transfer learning, and training the deep convolutional neural network GoogLeNet-I1;
s3: a deep convolution neural network GoogLeNet-I2 capable of identifying individual identities according to fingertip ECG signals is constructed on the basis of a trained deep convolution neural network GoogLeNet-I1, a small sample fingertip ECG signal data in a database (a Check Young biological signals Here initial database) is preprocessed by using a migration learning finetune technology, then the deep convolution neural network GoogLeNet-I2 is trained, and the trained deep convolution neural network GoogLeNet-I2 is a fingertip electrocardiogram identity identification model.
Preferably, the method for constructing the deep convolutional neural network *** net-I1 in the step S2 includes the following steps:
n1: modifying an inclusion structure of a deep convolution neural network GoogLeNet to obtain an inclusion-I1 structure, wherein the inclusion-I1 structure comprises an input layer, a processing layer and a connecting layer, the processing layer comprises a first processing branch, a second processing branch, a third processing branch and a fourth processing branch which are connected between the input layer and the connecting layer in parallel, the first processing branch comprises a 1 × 1 convolution unit, the second processing branch comprises a 1 × 1 convolution unit and a 3 × 3 convolution unit which are sequentially and fully connected, the third processing branch comprises a 1 × 1 convolution unit, a 2 × 2 convolution unit and a 3 × 3 convolution unit which are sequentially and fully connected, and the fourth processing branch comprises a 3 × 3 maximal pooling unit and a 1 × 1 convolution unit which are sequentially and fully connected;
n2: migrating the front 10 layers of the deep convolutional neural network GoogLeNet, and keeping the network layer parameters and the state of an optimizer of the front 10 layers unchanged by freezing;
n3: adding a Batch Normalization layer between the 39 th layer and the 40 th layer, between the 110 th layer and the 111 th layer, and between the 139 th layer and the 140 th layer of the deep convolutional neural network GoogLeNet;
n4: and sequentially adding a Dropout layer and a Fully connected layer between the 142 th layer and the 143 th layer of the deep convolutional neural network GoogleNet to construct a new deep convolutional neural network GoogleNet-I1, wherein the Dropout layer of the 144 th layer, the Fully connected layer of the 145 th layer, the Dropout layer of the 146 th layer, the Fully connected layer of the 147 th layer, the Softx layer of the 148 th layer and the Output layer of the 149 th layer are sequentially arranged from the last 6 layers of the new deep convolutional neural network GoogleNet-I1 from front to back.
The traditional inclusion mechanism consists of 3 groups of convolution kernels and a pooling unit, receives an input image from a previous layer together, processes the input image in parallel, and splices output results according to channels. The invention improves the increment mechanism on the existing basis, and aims at the problem that the large convolution (5 multiplied by 5) limits the total calculation capacity of the model, the invention uses two layers of small convolutions to simulate the original convolution structure, thereby inhibiting the great increase of the calculation amount; in addition, a small convolution of 1 × 1 is added before the original convolution of 3 × 3, and the effect of reducing the dimensionality and depth of the image space is achieved.
The deep convolutional neural network GoogLeNet is taken as a champion of ImageNet challenge race in 2014, has reasonable network depth, few parameters and high calculation performance, and is widely applied to classification/identification tasks. The GoogLeNet of the deep convolutional neural network also converts the full-connection layer and the convolutional layer into sparse connection, replaces the sparse connection with a global average pool, and clusters the sparse matrix into a denser sub-matrix, so that the overfitting phenomenon is avoided, and the calculation performance is also improved. According to the method, a deep convolutional neural network GoogleNet is used as a reference model, and an ECG identity recognition model collected by a thoracic cavity is reconstructed and trained.
Because the first few layers of each network are general characteristics of the current task, the effect of direct migration is better, and the deep convolutional neural network GoogLeNet adds an addition mechanism from the 11 th layer to perform multi-resolution learning. The first 10-level structure of the original *** lenet is thus migrated directly and its network-level parameters and optimizer states are frozen and kept unchanged.
In model training, the parameters and the optimizer state of each layer are continuously updated along with iteration, and if a certain layer has deviation, serious deviation of internal covariates is caused, so that the data characteristic distribution is continuously changed. In order to solve the problem, a Batch Normalization layer is added after the 39 th, 110 th and 139 th layers of the original deep convolutional neural network GoogLeNet to avoid the problem of gradient disappearance; a Dropout-Fully connected (Dropout-Fc) combination is added after the 142 th layer of the original deep convolutional neural network GoogLeNet to reduce the over-fitting/under-fitting problem.
The initial construction of the GoogLeNet-I1 is completed through the steps, then the finetune technology of transfer learning is utilized, namely the current network structure is kept unchanged, the model is trained through designing new training parameters, so that the weight parameters of the model and the state of an optimizer are adjusted, the identification accuracy and the equal error rate are used as measurement indexes, and the constructed GoogLeNet-I1 network is trained on the basis of the ECG identity identification model of thoracic cavity acquisition.
Preferably, the control deep convolutional neural network *** net-I1 has a characteristic rejection degree of the Dropout layer at layer 144 of 50% and a characteristic rejection degree of the Dropout layer at layer 146 of 10%.
Preferably, the method for constructing the deep convolutional neural network *** net-I2 in the step S3 includes the following steps:
the network structure of two Dropout layers and two Fully connected layers in the last 6 layers of the deep convolutional neural network GooglLeNet-I1 is modified into an adaptive layer structure by the following method: the Dropout layers at the 144 th layer and the 146 th layer of the GoogleLeNet-I1 are modified into two Dropout structures connected in parallel and an adaptive loss function, and the Fully connected layers at the 145 th layer and the 147 th layer of the GoogleLeNet-I1 are modified into two Fully connected structures connected in parallel and an adaptive loss function;
the adaptive layer structure performs countermeasure learning from a source domain to a target domain, and constructs an adaptive loss function to measure an adaptive error from the source domain to the target domain, so that deep countermeasure migration learning of a new task is realized, the source domain is a chest ECG signal, the target domain is a fingertip ECG signal, and the adaptive loss function corresponding to the adaptive layer structure is as shown in formula (1):
since the conventional loss function is a conventional classification loss on the source domain, which measures the difference between the predicted result and the true result, the conventional loss function is as in equation (2):
the loss function of the whole network is calculated based on the formulas (1) and (2) as follows:
where Φ (-) represents the mapping of the source/target domain to the regeneration kernel hilbert space, Θ represents all the weights and bias parameters of the deep convolutional neural network, J (-) represents the cross entropy, K represents the total number of network layers,
a data set representing all annotations in the source domain, which is the chest ECG signal,
data sets representing all labels in a target domain, the target domain being a fingertip ECG signal, n1 representing the adaptation layer of the source domain, n2 representing the adaptation layer of the target domain, DsData set representing a source domain, DtData set, λ, representing a target domainIndicating the degree of network adaptation (the invention adjusts lambda epsilon [0, 1)]Performing feature learning);
and constructing a new deep convolutional neural network GoogLeNet-I2.
Although the "finetune" advantage is clear, it does not require reconfiguration of the network structure, it also has certain disadvantages: it is not possible to deal with the situation where the training data and the test data are distributed differently. Specifically, in the model migration of the invention, the image classification task of the original deep convolutional neural network *** net and the signal classification of the current deep convolutional neural network *** net-I1 do not follow the same data distribution, and therefore, the invention proposes the concept of the adaptive layer of the deep convolutional neural network. The invention aims to realize the self-adaptation from a source domain (based on a chest ECG signal acquired by a chest) to a target domain (based on a fingertip ECG signal acquired by a fingertip) by adding a network self-adaptation layer, so that the data distribution of the source domain and the target domain is closer, and the better learning of key characteristics is realized. Since the first layers of the network will learn the general characteristics of the current task, and the last layers are new training using the training data of the new task, the migration and adaptation tasks of the last layers of the model should be considered. From the theoretical analysis of deep learning, the previous layer of the classifier, namely the features, is just the work to be completed by transfer learning by adding self-adaptation to the features.
Preferably, the method of preprocessing the chest ECG signal is the same as the method of preprocessing the fingertip ECG signal, and the preprocessing method is as follows:
m1: processing the original ECG signal by adopting a cyclic shift denoising algorithm based on a wavelet hard threshold value to obtain a clean ECG signal;
m2: the fixed window length is 3f, the ECG signal with the fixed window length is randomly intercepted and used as a short-period ECG signal, and f is the sampling frequency of the ECG signal;
m3: and carrying out generalized S transformation on the short-period ECG signal, constructing a corresponding frequency domain instantaneous track structure characteristic map group, and taking the frequency domain instantaneous track structure characteristic map group as the input of the deep convolutional neural network.
As a non-stable time-varying weak signal, electrocardio is very easy to be interfered by various aspects such as instrument, human body movement and the like in the acquisition process, so that the components of the electrocardio signal are weak, and the electrocardio signal is mixed with a noise frequency band, thereby seriously influencing the subsequent characteristic learning, capturing and representing of a time sequence signal.
On the basis of the denoising, a sliding window with the length of 3f is adopted to randomly slide the denoised ECG signal, and a short-period clean electrocardiosignal with the window length is intercepted. And analyzing the time-frequency joint characteristics of the ECG time sequence signals to obtain a sufficient frequency domain instantaneous track structure characteristic diagram which is used as the input of a deep convolution neural network model to provide sufficient two-dimensional image characteristics for identity recognition.
Preferably, the step M1 includes the steps of:
m11: performing 8 times of cyclic translation processing on the original ECG signal to change the position of a singular point in the original ECG signal;
m12: decomposing the original ECG signal by discrete wavelet transform;
m13: performing threshold quantization processing on the wavelet coefficient through a hard threshold function in a wavelet domain, and performing inverse discrete wavelet transform according to an estimated wavelet coefficient obtained after quantization processing to obtain a reconstructed ECG signal;
m14: the reconstructed ECG signal is subjected to 8 inverse cyclic shifts to obtain a clean ECG signal.
The translation changes the position of a signal singular point in the whole signal section, the translation invariance of the signal is kept by the operation of inverse translation, the whole operation reduces the process of waveform oscillation, and the Gibbs oscillation is weakened or even eliminated, so that the original clean signal is better approximated.
Preferably, the step M3 includes the steps of:
m31: the generalized S-transform is applied to the short-cycle ECG signal as follows:
the window function of the generalized S transform is：
Wherein z (t) is a short-period ECG electrocardiosignal, t is time, tau is a time shift factor, f is the sampling frequency of the ECG signal, h is a window width parameter of a window function, and w is an amplitude parameter of the window function;
obtaining a complex matrix through the above processing
The complex matrix comprises a real part and an imaginary part,
wherein, N is 1.5f, M is 3 f;
m32: and drawing a frequency domain instantaneous track structure characteristic diagram corresponding to each line of data according to each line of data of the complex matrix to obtain M frequency domain instantaneous track structure characteristic diagrams, wherein the M frequency domain instantaneous track structure characteristic diagrams form a frequency domain instantaneous track structure characteristic diagram group corresponding to the short-period ECG signal and serve as the input of the deep convolutional neural network.
Each column of the complex matrix reflects the "instantaneous frequency characteristic" of the current time point.
Preferably, the method for drawing the frequency domain instantaneous trajectory structural feature map corresponding to a certain column of data in step M32 is as follows: and establishing a rectangular coordinate system by taking the real part alpha as an X axis and the imaginary part beta as a Y axis, marking the positions of the line of data on the rectangular coordinate system, drawing a track, and reducing the image pixels to 224 multiplied by 224.
The frequency domain instantaneous track structure characteristic diagram reflects the frequency characteristic information of the time sequence signal at each time point and the change trend of the time points before and after the time point, and the frequency characteristic information is used as the input of the deep convolutional neural network.
Preferably, the
The setting of w and h directly affects the shape of g (t), and thus the capture and identification of key featuresAnd (4) precision. The invention takes the identification precision Acc as an objective function, adjusts the parameters w and h and obtains the optimal parameter set table
The invention has the beneficial effects that: (1) an electrocardio acquisition system architecture for acquiring fingertip electrocardiosignals of a person to be identified is constructed based on the internet of things technology, so that the stability and the signal quality of the electrocardio acquisition process are ensured, and the convenience and the comfort of the acquisition process are also considered. (2) The method comprises the steps of constructing and training a fingertip electrocardio identity recognition model by adding a deep convolution neural network in transfer learning, carrying out model transfer twice based on the deep convolution neural network GoogLeNet, carrying out model improvement and optimization from multiple levels such as a network structure, an optimization algorithm, a loss function and the like, obtaining an electrocardio identity recognition model based on thoracic cavity collection through one transfer, namely the deep convolution neural network GoogLeNet-I1, determining an electrocardio identity recognition model based on fingertip collection through the second transfer, namely the deep convolution neural network GoogLeNet-I2, and realizing recognition under a small sample without manual feature extraction.
Drawings
FIG. 1 is a schematic structural view of an embodiment;
FIG. 2 is a flow chart of an embodiment;
FIG. 3 is a flow chart of the construction of a apical electrocardio identity recognition model;
FIG. 4 is a schematic diagram of the structure of inclusion-I1;
FIG. 5 is a schematic diagram of the structure of the deep convolutional neural network GoogLeNet-I2.
In the figure: 1. the device comprises a fingertip electrocardiosignal acquisition module, a signal conditioning circuit, a controller, a Bluetooth communication module and a Bluetooth communication module.
Detailed Description
The technical scheme of the invention is further specifically described by the following embodiments and the accompanying drawings.
Example (b): the fingertip electrocardiograph identity recognition device of the embodiment is shown in fig. 1, and includes a fingertip electrocardiograph signal acquisition module 1, a signal conditioning circuit 2, a controller 3 and a bluetooth communication module 4, wherein an output end of the fingertip electrocardiograph signal acquisition module 1 is electrically connected with an input end of the signal conditioning circuit 2, an output end of the signal conditioning circuit 2 is electrically connected with the controller 3, and the controller 3 is further electrically connected with the bluetooth communication module 4.
According to the scheme, fingertip electrocardiosignals of a user are rapidly acquired through a fingertip electrocardiosignal acquisition module, are subjected to differential amplification and filtering through a signal conditioning circuit, are sent to an ADC (analog-to-digital converter) of a controller MCU (micro control unit) for analog-to-digital acquisition and conversion, are sent to a mobile phone end through a Bluetooth communication module in a UART (universal asynchronous receiver/transmitter) mode, and the mobile phone end acquires fingertip electrocardiosignal data of the person to be identified. Meanwhile, the mobile phone end accesses the cloud platform through the wifi/4G wireless transmission module, uploads fingertip electrocardio data of the current user and downloads a recognition result of the fingertip electrocardio-based identity recognition model in the cloud platform. The method only needs to collect fingertip electrocardiogram data, and is more convenient and comfortable compared with the method for collecting thoracic electrocardiogram data.
The PS25253 sensor in the EPIC sensor of the plessey company is used as a fingertip ECG signal acquisition module to acquire a fingertip ECG signal, and the fingertip ECG signal is converted into an electric signal. PS25253 has an ultrahigh impedance solid state, can be used as a dry contact type ECG sensor, is convenient and quick to measure, introduces little noise interference and is stable in acquisition. In addition, two metal dry electrodes are used as measuring electrodes, and no gel or matrix is needed to enhance signal conduction. The whole process not only ensures the stability of the electrocardio acquisition process and the accuracy of signal acquisition, but also considers the comfort of the acquisition process.
Because the electric signal obtained by the direct difference of the sensor is very weak, and the acquisition process is interfered by individuals, instruments and other aspects, a large amount of noise signals are mixed in the acquired original signal, so that the signal acquired by the fingertip electrocardiosignal acquisition module in the previous step is subjected to difference amplification and shaping filtering processing, and the noise interference is preliminarily eliminated.
And sending the filtered electric signal into an ADC (analog to digital converter) of the MCU (micro control unit) to perform digital to analog conversion data, thereby finishing the whole signal conditioning process. The controller MCU not only controls the normal work of peripheral circuits, but also carries out A/D conversion on the signals after noise elimination, simultaneously designs and processes related circuits such as a power module of a system, a Bluetooth communication module and the like, and transmits the signals to the mobile phone end in real time.
The method for identifying the fingertip electrocardiogram identity of the embodiment, as shown in fig. 2, includes the following steps:
collecting fingertip ECG signals of a human fingertip, preprocessing the fingertip ECG signals, inputting the preprocessed fingertip ECG signals to a fingertip electrocardiogram identity recognition model, carrying out identity recognition on the fingertip electrocardiogram identity recognition model, and outputting corresponding identity information;
as shown in fig. 3, the fingertip electrocardiogram identity recognition model is constructed as follows:
s1: establishing a deep convolutional neural network GoogLeNet;
s2: constructing a deep convolutional neural network GoogLeNet-I1 capable of identifying individual identities according to chest ECG signals based on the deep convolutional neural network GoogLeNet-I1, preprocessing large sample chest ECG signal data acquired by standard leads in a Physionet network reference library by using a finetune technology of transfer learning, and then training the deep convolutional neural network GoogLeNet-I1 (namely keeping the network structure of the deep convolutional neural network GoogLeNet-I1 unchanged, training the deep convolutional neural network GoogLeNet-I1 by designing new training parameters, so as to adjust the weight parameters of the deep convolutional neural network GoogLeNet-I1 and the state of an optimizer, and training the constructed deep convolutional neural network GoogLeNet-I1 by taking identification accuracy and error rate as net measurement indexes);
s3: constructing a deep convolution neural network GoogLeNet-I2 capable of identifying the identity of an individual according to a fingertip ECG signal based on the trained deep convolution neural network GoogLeNet-I1, by using a finetune technology of transfer learning, preprocessing 63 small sample fingertip ECG signal data in a Check young biological signals Here initiative database, and then training a deep convolutional neural network GoogLeNet-I2 (namely keeping a structure of the deep convolutional neural network GoogLeNet-I2 unchanged, setting training parameters to train the deep convolutional neural network GoogLeNet-I2, so as to adjust weight parameters of the deep convolutional neural network GoogLeNet-I2 and a state of an optimizer, and training the constructed deep convolutional neural network GoogLeNet-I2 by combining recognition accuracy and error rate as measurement indexes), wherein the trained deep convolutional neural network GoogLeNet-I2 is a fingertip electrocardio identity recognition model.
The method for constructing the deep convolutional neural network GoogleLeNet-I1 in the step S2 comprises the following steps:
n1: modifying an inclusion structure of a deep convolutional neural network *** net to obtain an inclusion-I1 structure, as shown in fig. 4, the inclusion-I1 structure includes an input layer, a processing layer, and a connection layer, the processing layer includes a first processing branch, a second processing branch, a third processing branch, and a fourth processing branch connected in parallel between the input layer and the connection layer, the first processing branch includes a 1 × 1 convolution unit, the second processing branch includes a 1 × 1 convolution unit and a 3 × 3 convolution unit that are sequentially and fully connected, the third processing branch includes a 1 × 1 convolution unit, a 2 × 2 convolution unit, and a 3 × 3 convolution unit that are sequentially and fully connected, and the fourth processing branch includes a 3 × 3 maximal pooling unit and a 1 × 1 convolution unit that are sequentially and fully connected;
n2: migrating the front 10 layers of the deep convolutional neural network GoogLeNet, and keeping the network layer parameters and the state of an optimizer of the front 10 layers unchanged by freezing;
n3: adding a Batch Normalization layer between the 39 th layer and the 40 th layer, between the 110 th layer and the 111 th layer, and between the 139 th layer and the 140 th layer of the deep convolutional neural network GoogLeNet;
n4: sequentially adding a Dropout layer and a Fully connected layer between the 142 th layer and the 143 th layer of the deep convolutional neural network GoogleNet to construct a new deep convolutional neural network GoogleNet-I1, wherein the Dropout layer of the 144 th layer, the Fully connected layer of the 145 th layer, the Dropout layer of the 146 th layer, the Fully connected layer of the 147 th layer, the Softy layer of the 148 th layer and the Output layer of the 149 th layer are sequentially arranged from the last 6 layers of the new deep convolutional neural network GoogleNet-I1 from front to back;
the control deep convolutional neural network *** lenet-I1 has a characteristic rejection degree of Dropout layer 144 of 50% and a characteristic rejection degree of Dropout layer 146 of 10%.
The traditional inclusion mechanism consists of 3 groups of convolution kernels and a pooling unit, receives an input image from a previous layer together, processes the input image in parallel, and splices output results according to channels. The invention improves the increment mechanism on the existing basis, and aims at the problem that the large convolution (5 multiplied by 5) limits the total calculation capacity of the model, the invention uses two layers of small convolutions to simulate the original convolution structure, thereby inhibiting the great increase of the calculation amount; in addition, a small convolution of 1 × 1 is added before the original convolution of 3 × 3, and the effect of reducing the dimensionality and depth of the image space is achieved.
The deep convolutional neural network GoogLeNet is taken as a champion of ImageNet challenge race in 2014, has reasonable network depth, few parameters and high calculation performance, and is widely applied to classification/identification tasks. The GoogLeNet of the deep convolutional neural network also converts the full-connection layer and the convolutional layer into sparse connection, replaces the sparse connection with a global average pool, and clusters the sparse matrix into a denser sub-matrix, so that the overfitting phenomenon is avoided, and the calculation performance is also improved. According to the method, a deep convolutional neural network GoogleNet is used as a reference model, and an ECG identity recognition model collected by a thoracic cavity is reconstructed and trained.
Because the first few layers of each network are general characteristics of the current task, the effect of direct migration is better, and the deep convolutional neural network GoogLeNet adds an addition mechanism from the 11 th layer to perform multi-resolution learning. The first 10-level structure of the original *** lenet is thus migrated directly and its network-level parameters and optimizer states are frozen and kept unchanged.
In model training, the parameters and the optimizer state of each layer are continuously updated along with iteration, and if a certain layer has deviation, serious deviation of internal covariates is caused, so that the data characteristic distribution is continuously changed. In order to solve the problem, a Batch Normalization layer is added after the 39 th, 110 th and 139 th layers of the original deep convolutional neural network GoogLeNet to avoid the problem of gradient disappearance; a Dropout-Fully connected (Dropout-Fc) combination is added after the 142 th layer of the original deep convolutional neural network GoogLeNet to reduce the over-fitting/under-fitting problem.
The initial construction of the GoogLeNet-I1 is completed through the steps, then the finetune technology of transfer learning is utilized, namely the current network structure is kept unchanged, the model is trained through designing new training parameters, so that the weight parameters of the model and the state of an optimizer are adjusted, the identification accuracy and the equal error rate are used as measurement indexes, and the constructed GoogLeNet-I1 network is trained on the basis of the ECG identity identification model of thoracic cavity acquisition.
The method for constructing the deep convolutional neural network GoogleLeNet-I2 in the step S3 comprises the following steps:
although the "finetune" advantage is clear, it does not require reconfiguration of the network structure, it also has certain disadvantages: it is not possible to deal with the situation where the training data and the test data are distributed differently. Specifically, in the model migration of the invention, the image classification task of the original deep convolutional neural network *** net and the signal classification of the current deep convolutional neural network *** net-I1 do not follow the same data distribution, and therefore, the invention proposes the concept of the adaptive layer of the deep convolutional neural network. The invention aims to realize the self-adaptation from a source domain (based on a chest ECG signal acquired by a chest) to a target domain (based on a fingertip ECG signal acquired by a fingertip) by adding a network self-adaptation layer, so that the data distribution of the source domain and the target domain is closer, and the better learning of key characteristics is realized. Since the first layers of the network will learn the general characteristics of the current task, and the last layers are new training using the training data of the new task, the migration and adaptation tasks of the last layers of the model should be considered. From the theoretical analysis of deep learning, the previous layer of the classifier, namely the features, is just the work to be completed by transfer learning by adding self-adaptation to the features.
The network structure of two Dropout layers and two Fully connected layers in the last 6 layers of the deep convolutional neural network GooglLeNet-I1 is modified into an adaptive layer structure by the following method: the Dropout layers at the 144 th layer and the 146 th layer of the GoogleLeNet-I1 are modified into two Dropout structures connected in parallel and an adaptive loss function, and the Fully connected layers at the 145 th layer and the 147 th layer of the GoogleLeNet-I1 are modified into two Fully connected structures connected in parallel and an adaptive loss function;
the adaptive layer structure performs countermeasure learning from a source domain to a target domain, and constructs an adaptive loss function to measure an adaptive error from the source domain to the target domain, so that deep countermeasure migration learning of a new task is realized, the source domain is a chest ECG signal, the target domain is a fingertip ECG signal, and the adaptive loss function corresponding to the adaptive layer structure is as shown in formula (1):
since the conventional loss function is a conventional classification loss on the source domain, which measures the difference between the predicted result and the true result, the conventional loss function is as in equation (2):
the loss function of the whole network is calculated based on the formulas (1) and (2) as follows:
where Φ (-) denotes the mapping of the source/target domain to the Regenerated Kernel Hilbert Space (RKHS), Θ denotes all the weights and bias parameters of the deep convolutional neural network (which are targets for learning), J (-) denotes the cross entropy, K denotes the total number of network layers,
a data set representing all annotations in the source domain, which is the chest ECG signal,
data sets representing all labels in a target domain, the target domain being a fingertip ECG signal, n1 representing the adaptation layer of the source domain, n2 representing the adaptation layer of the target domain, DsData set representing a source domain, DtThe data set of the target domain is represented, and the lambda represents the degree of network self-adaptation adjustment (the invention adjusts the lambda belongs to 0, 1]Performing feature learning);
a new deep convolutional neural network GoogLeNet-I2 was constructed, as shown in FIG. 5.
The method of preprocessing the chest ECG signal is the same as the method of preprocessing the fingertip ECG signal, and the preprocessing method is as follows:
m1: processing the original ECG signal by adopting a cyclic shift denoising algorithm based on a wavelet hard threshold value to obtain a clean ECG signal;
m2: the fixed window length is 3f, the ECG signal with the fixed window length is randomly intercepted and used as a short-period ECG signal, and f is the sampling frequency of the ECG signal;
m3: and carrying out generalized S transformation on the short-period ECG signal, constructing a corresponding frequency domain instantaneous track structure characteristic map group, and taking the frequency domain instantaneous track structure characteristic map group as the input of the deep convolutional neural network.
As a non-stable time-varying weak signal, electrocardio is very easy to be interfered by various aspects such as instrument, human body movement and the like in the acquisition process, so that the components of the electrocardio signal are weak, and the electrocardio signal is mixed with a noise frequency band, thereby seriously influencing the subsequent characteristic learning, capturing and representing of a time sequence signal.
On the basis of the denoising, a sliding window with the length of 3f is adopted to randomly slide the denoised ECG signal, and a short-period clean electrocardiosignal with the window length is intercepted. And analyzing the time-frequency joint characteristics of the ECG time sequence signals to obtain a sufficient frequency domain instantaneous track structure characteristic diagram which is used as the input of a deep convolution neural network model to provide sufficient two-dimensional image characteristics for identity recognition.
Step M1 includes the following steps:
m11: performing 8 times of cyclic translation processing on the original ECG signal to change the position of a singular point in the original ECG signal;
m12: decomposing the original ECG signal by discrete wavelet transform;
m13: performing threshold quantization processing on the wavelet coefficient through a hard threshold function in a wavelet domain, and performing inverse discrete wavelet transform according to an estimated wavelet coefficient obtained after quantization processing to obtain a reconstructed ECG signal;
m14: the reconstructed ECG signal is subjected to 8 inverse cyclic shifts to obtain a clean ECG signal.
The translation changes the position of a signal singular point in the whole signal section, the translation invariance of the signal is kept by the operation of inverse translation, the whole operation reduces the process of waveform oscillation, and the Gibbs oscillation is weakened or even eliminated, so that the original clean signal is better approximated.
Step M3 includes the following steps:
m31: the generalized S-transform is applied to the short-cycle ECG signal as follows:
the window function of the generalized S transform is:
wherein z (t) is a short-period ECG electrocardiosignal, t is time, tau is a time shift factor, f is the sampling frequency of the ECG signal, h is a window width parameter of a window function, and w is an amplitude parameter of the window function;
obtaining a complex matrix through the above processing
The complex matrix comprises a real part and an imaginary part,
wherein, N is 1.5f, M is 3 f;
m32: and drawing a frequency domain instantaneous track structure characteristic diagram corresponding to each line of data according to each line of data of the complex matrix to obtain M frequency domain instantaneous track structure characteristic diagrams, wherein the M frequency domain instantaneous track structure characteristic diagrams form a frequency domain instantaneous track structure characteristic diagram group corresponding to the short-period ECG signal and serve as the input of the deep convolutional neural network.
Each column of the complex matrix reflects the "instantaneous frequency characteristic" of the current time point.
The method for drawing the frequency domain instantaneous trajectory structural feature map corresponding to a certain row of data in the step M32 is as follows: and establishing a rectangular coordinate system by taking the real part alpha as an X axis and the imaginary part beta as a Y axis, marking the positions of the line of data on the rectangular coordinate system, drawing a track, and reducing the image pixels to 224 multiplied by 224.
The frequency domain instantaneous track structure characteristic diagram reflects the frequency characteristic information of the time sequence signal at each time point and the change trend of the time points before and after the time point, and the frequency characteristic information is used as the input of the deep convolutional neural network.
The embodiment designs an inclusion-I1 mechanism, and performs multi-scale and multi-resolution fusion on the image; based on a deep convolutional neural network GoogLeNet, by combining with finetune of transfer learning, overfitting/underfitting measures such as Dropout and the like are added, the primary transfer of the model is realized, an ECG identity recognition model trained by collecting ECG data of a chest of a large sample is created, and the model is marked as the deep convolutional neural network GoogLeNet-I1; then, the deep convolutional neural network GoogLeNet-I1 is migrated to fingertip electrocardio, a deep network adaptive layer is constructed, an adaptive measurement mode is designed, secondary migration of a model is realized, a deep convolutional neural network GoogLeNet-12 is created, self-adaptation from a source domain (ECG acquired by a chest cavity) to a target domain (ECG acquired by a fingertip) is realized, the requirement of a deep learning algorithm on the strong computing power of big data under small sample data is solved, the trained deep convolutional neural network GoogLeNet-I2 is stored for the subsequent identification requirement based on fingertip electrocardio, and the result of individual identification is rapidly output.
Claims (10)
1. The fingertip electrocardio identity recognition device is characterized by comprising a fingertip electrocardiosignal acquisition module (1), a signal conditioning circuit (2), a controller (3) and a Bluetooth communication module (4), wherein the output end of the fingertip electrocardiosignal acquisition module (1) is electrically connected with the input end of the signal conditioning circuit (2), the output end of the signal conditioning circuit (2) is electrically connected with the controller (3), and the controller (3) is also electrically connected with the Bluetooth communication module (4).
2. A fingertip electrocardio identity recognition method is characterized by comprising the following steps:
collecting fingertip ECG signals of a human fingertip, preprocessing the fingertip ECG signals, inputting the preprocessed fingertip ECG signals to a fingertip electrocardiogram identity recognition model, carrying out identity recognition on the fingertip electrocardiogram identity recognition model, and outputting corresponding identity information;
the construction method of the fingertip electrocardio identity recognition model comprises the following steps:
s1: establishing a deep convolutional neural network GoogLeNet;
s2: constructing a deep convolutional neural network GoogLeNet-I1 capable of identifying individual identity according to a chest ECG signal based on the deep convolutional neural network GoogLeNet, preprocessing large sample chest ECG signal data acquired by standard leads in a database by using a finetune technology of transfer learning, and then training the deep convolutional neural network GoogLeNet-I1;
s3: the method comprises the steps of constructing a deep convolution neural network GoogLeNet-I2 capable of identifying individual identities according to fingertip ECG signals based on a trained deep convolution neural network GoogLeNet-I1, preprocessing small sample fingertip ECG signal data in a database by using a migration learning finetune technology, and then training the deep convolution neural network GoogLeNet-I2, wherein the trained deep convolution neural network GoogLeNet-I2 is a fingertip electrocardiogram identity identification model.
3. The method for fingertip electrocardiographic identification according to claim 2, wherein the method for constructing the deep convolutional neural network *** lenet-I1 in the step S2 comprises the following steps:
n1: modifying an inclusion structure of a deep convolution neural network GoogLeNet to obtain an inclusion-I1 structure, wherein the inclusion-I1 structure comprises an input layer, a processing layer and a connecting layer, the processing layer comprises a first processing branch, a second processing branch, a third processing branch and a fourth processing branch which are connected between the input layer and the connecting layer in parallel, the first processing branch comprises a 1 × 1 convolution unit, the second processing branch comprises a 1 × 1 convolution unit and a 3 × 3 convolution unit which are sequentially and fully connected, the third processing branch comprises a 1 × 1 convolution unit, a 2 × 2 convolution unit and a 3 × 3 convolution unit which are sequentially and fully connected, and the fourth processing branch comprises a 3 × 3 maximal pooling unit and a 1 × 1 convolution unit which are sequentially and fully connected;
n2: migrating the front 10 layers of the deep convolutional neural network GoogLeNet, and keeping the network layer parameters and the state of an optimizer of the front 10 layers unchanged by freezing;
n3: adding a Batch Normalization layer between the 39 th layer and the 40 th layer, between the 110 th layer and the 111 th layer, and between the 139 th layer and the 140 th layer of the deep convolutional neural network GoogLeNet;
n4: and sequentially adding a Dropout layer and a Fully connected layer between the 142 th layer and the 143 th layer of the deep convolutional neural network GoogleNet to construct a new deep convolutional neural network GoogleNet-I1, wherein the Dropout layer of the 144 th layer, the Fully connected layer of the 145 th layer, the Dropout layer of the 146 th layer, the Fully connected layer of the 147 th layer, the Softx layer of the 148 th layer and the Output layer of the 149 th layer are sequentially arranged from the last 6 layers of the new deep convolutional neural network GoogleNet-I1 from front to back.
4. The method as claimed in claim 3, wherein the degree of feature rejection of the Dropout layer at the 144 th layer of the control deep convolutional neural network GoogleLeNet-I1 is 50%, and the degree of feature rejection of the Dropout layer at the 146 th layer is 10%.
5. The fingertip electrocardiographic identification method according to claim 3, wherein the method for constructing the deep convolutional neural network GoogLeNet-I2 in the step S3 comprises the following steps:
the network structure of two Dropout layers and two Fully connected layers in the last 6 layers of the deep convolutional neural network GooglLeNet-I1 is modified into an adaptive layer structure by the following method: the Dropout layers at the 144 th layer and the 146 th layer of the GoogleLeNet-I1 are modified into two Dropout structures connected in parallel and an adaptive loss function, and the Fully connected layers at the 145 th layer and the 147 th layer of the GoogleLeNet-I1 are modified into two Fully connected structures connected in parallel and an adaptive loss function;
the adaptive layer structure performs countermeasure learning from a source domain to a target domain, and constructs an adaptive loss function to measure an adaptive error from the source domain to the target domain, so that deep countermeasure migration learning of a new task is realized, the source domain is a chest ECG signal, the target domain is a fingertip ECG signal, and the adaptive loss function corresponding to the adaptive layer structure is as shown in formula (1):
due to the conventional loss function as in equation (2):
the loss function of the whole network is calculated based on the formulas (1) and (2) as follows:
where Φ (-) represents the mapping of the source/target domain to the regeneration kernel hilbert space, Θ represents all the weights and bias parameters of the deep convolutional neural network, J (-) represents the cross entropy, K represents the total number of network layers,
a data set representing all annotations in the source domain, which is the chest ECG signal,
data set representing all annotations in the target domain, the target domain being the fingertip ECG signal, n1 representing the self of the source domainAdaptation layer, n2 denotes the adaptation layer of the target domain, DsData set representing a source domain, DtRepresenting a data set of a target domain, and lambda represents the degree of network self-adaptation adjustment;
and constructing a new deep convolutional neural network GoogLeNet-I2.
6. The fingertip electrocardiographic identification method according to claim 2, 3, 4 or 5, wherein the method for preprocessing the chest ECG signal is the same as the method for preprocessing the fingertip ECG signal, and the preprocessing method is as follows:
m1: processing the original ECG signal by adopting a cyclic shift denoising algorithm based on a wavelet hard threshold value to obtain a clean ECG signal;
m2: the fixed window length is 3f, the ECG signal with the fixed window length is randomly intercepted and used as a short-period ECG signal, and f is the sampling frequency of the ECG signal;
m3: and carrying out generalized S transformation on the short-period ECG signal, constructing a corresponding frequency domain instantaneous track structure characteristic map group, and taking the frequency domain instantaneous track structure characteristic map group as the input of the deep convolutional neural network.
7. The fingertip electrocardiographic identification method according to claim 6, wherein the step M1 includes the following steps:
m11: performing 8 times of cyclic translation processing on the original ECG signal to change the position of a singular point in the original ECG signal;
m12: decomposing the original ECG signal by discrete wavelet transform;
m13: performing threshold quantization processing on the wavelet coefficient through a hard threshold function in a wavelet domain, and performing inverse discrete wavelet transform according to an estimated wavelet coefficient obtained after quantization processing to obtain a reconstructed ECG signal;
m14: the reconstructed ECG signal is subjected to 8 inverse cyclic shifts to obtain a clean ECG signal.
8. The fingertip electrocardiographic identification method according to claim 6, wherein the step M3 includes the following steps:
m31: the generalized S-transform is applied to the short-cycle ECG signal as follows:
the window function of the generalized S transform is:
wherein z (t) is a short-period ECG electrocardiosignal, t is time, tau is a time shift factor, f is the sampling frequency of the ECG signal, h is a window width parameter of a window function, and w is an amplitude parameter of the window function;
obtaining a complex matrix through the above processing
The complex matrix comprises a real part and an imaginary part,
wherein, N is 1.5f, M is 3 f;
m32: and drawing a frequency domain instantaneous track structure characteristic diagram corresponding to each line of data according to each line of data of the complex matrix to obtain M frequency domain instantaneous track structure characteristic diagrams, wherein the M frequency domain instantaneous track structure characteristic diagrams form a frequency domain instantaneous track structure characteristic diagram group corresponding to the short-period ECG signal and serve as the input of the deep convolutional neural network.
9. The method for fingertip electrocardiographic identification according to claim 8, wherein the method for drawing the frequency domain instantaneous trajectory structural feature map corresponding to a certain line of data in the step M32 is as follows: and establishing a rectangular coordinate system by taking the real part alpha as an X axis and the imaginary part beta as a Y axis, marking the positions of the line of data on the rectangular coordinate system, drawing a track, and reducing the image pixels to 224 multiplied by 224.
10. The fingertip electrocardiographic identification method according to claim 8, wherein the method is characterized in that
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011226668.XA CN112307996B (en) | 2020-11-05 | 2020-11-05 | Fingertip electrocardio identity recognition device and method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011226668.XA CN112307996B (en) | 2020-11-05 | 2020-11-05 | Fingertip electrocardio identity recognition device and method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN112307996A true CN112307996A (en) | 2021-02-02 |
| CN112307996B CN112307996B (en) | 2024-04-02 |
Family
ID=74326168
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202011226668.XA Active CN112307996B (en) | 2020-11-05 | 2020-11-05 | Fingertip electrocardio identity recognition device and method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112307996B (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113468988A (en) * | 2021-06-18 | 2021-10-01 | 南京润楠医疗电子研究院有限公司 | ECG signal-based identity recognition method under multi-pressure state |
| CN113486917A (en) * | 2021-05-17 | 2021-10-08 | 西安电子科技大学 | Radar HRRP small sample target identification method based on metric learning |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110458197A (en) * | 2019-07-11 | 2019-11-15 | 启东市知微电子科技有限公司 | Personal identification method and its system based on photoplethysmographic |
| US20200241646A1 (en) * | 2019-01-25 | 2020-07-30 | Tata Consultancy Services Limited | On-device classification of fingertip motion patterns into gestures in real-time |
| CN111709267A (en) * | 2020-03-27 | 2020-09-25 | 吉林大学 | Electroencephalogram signal emotion recognition method of deep convolutional neural network |
| CN111783558A (en) * | 2020-06-11 | 2020-10-16 | 上海交通大学 | Satellite navigation interference signal type intelligent identification method and system |
-
2020
- 2020-11-05 CN CN202011226668.XA patent/CN112307996B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20200241646A1 (en) * | 2019-01-25 | 2020-07-30 | Tata Consultancy Services Limited | On-device classification of fingertip motion patterns into gestures in real-time |
| CN110458197A (en) * | 2019-07-11 | 2019-11-15 | 启东市知微电子科技有限公司 | Personal identification method and its system based on photoplethysmographic |
| CN111709267A (en) * | 2020-03-27 | 2020-09-25 | 吉林大学 | Electroencephalogram signal emotion recognition method of deep convolutional neural network |
| CN111783558A (en) * | 2020-06-11 | 2020-10-16 | 上海交通大学 | Satellite navigation interference signal type intelligent identification method and system |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113486917A (en) * | 2021-05-17 | 2021-10-08 | 西安电子科技大学 | Radar HRRP small sample target identification method based on metric learning |
| CN113486917B (en) * | 2021-05-17 | 2023-06-02 | 西安电子科技大学 | Radar HRRP small sample target recognition method based on metric learning |
| CN113468988A (en) * | 2021-06-18 | 2021-10-01 | 南京润楠医疗电子研究院有限公司 | ECG signal-based identity recognition method under multi-pressure state |
| CN113468988B (en) * | 2021-06-18 | 2024-04-05 | 南京润楠医疗电子研究院有限公司 | Identification method under multiple pressure states based on ECG (electronic pulse-taking) signals |
Also Published As
| Publication number | Publication date |
|---|---|
| CN112307996B (en) | 2024-04-02 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111062250B (en) | Multi-subject motor imagery electroencephalogram signal identification method based on deep feature learning | |
| CN110826625B (en) | Finger gesture classification method based on surface electromyographic signals | |
| CN114533086B (en) | Motor imagery brain electrolysis code method based on airspace characteristic time-frequency transformation | |
| CN110969108A (en) | Limb action recognition method based on autonomic motor imagery electroencephalogram | |
| CN102930284A (en) | Surface electromyogram signal pattern recognition method based on empirical mode decomposition and fractal | |
| CN105769173A (en) | Electrocardiogram monitoring system with electrocardiosignal denoising function | |
| CN111597991A (en) | Rehabilitation detection method based on channel state information and BilSTM-Attention | |
| CN112307996B (en) | Fingertip electrocardio identity recognition device and method | |
| CN109299647B (en) | Vehicle control-oriented multitask motor imagery electroencephalogram feature extraction and mode recognition method | |
| CN103345641A (en) | Hand electromyographic signal motion recognition method based on wavelet entropy and support vector machine | |
| CN111476158A (en) | Multi-channel physiological signal somatosensory gesture recognition method based on PSO-PCA-SVM | |
| CN110786849B (en) | Electrocardiosignal identity recognition method and system based on multi-view discriminant analysis | |
| CN106599903B (en) | Signal reconstruction method for weighted least square dictionary learning based on correlation | |
| CN113111831A (en) | Gesture recognition technology based on multi-mode information fusion | |
| Shrivastwa et al. | A brain–computer interface framework based on compressive sensing and deep learning | |
| CN111053552B (en) | QRS wave detection method based on deep learning | |
| CN111772669A (en) | Elbow joint contraction muscle force estimation method based on adaptive long-time and short-time memory network | |
| CN115512440A (en) | Cross-user gesture recognition method based on multi-source domain self-adaption | |
| CN114707539A (en) | Hand joint angle estimation method, hand joint angle estimation device, storage medium and equipment | |
| CN111783669B (en) | Surface electromyogram signal classification and identification method for individual user | |
| CN113128384A (en) | Brain-computer interface software key technical method of stroke rehabilitation system based on deep learning | |
| CN117150292A (en) | Incremental learning-based gesture recognition model training method | |
| CN115329800B (en) | Convolutional neural network activity recognition method based on decoupling training | |
| CN111466930A (en) | Audio-visual evoked emotion recognition method and system based on electroencephalogram signals | |
| Krishnapriya et al. | Surface electromyography based hand gesture signal classification using 1d cnn |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |