CN109907733A - A kind of ECG signal analysis method towards abnormal heart rhythms classification - Google Patents
A kind of ECG signal analysis method towards abnormal heart rhythms classification Download PDFInfo
- Publication number
- CN109907733A CN109907733A CN201910285158.0A CN201910285158A CN109907733A CN 109907733 A CN109907733 A CN 109907733A CN 201910285158 A CN201910285158 A CN 201910285158A CN 109907733 A CN109907733 A CN 109907733A
- Authority
- CN
- China
- Prior art keywords
- ecg signal
- ecg
- lstm
- analysis method
- result
- 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.)
- Pending
Links
Landscapes
- Measurement And Recording Of Electrical Phenomena And Electrical Characteristics Of The Living Body (AREA)
Abstract
The present invention provides a kind of ECG signal analysis method towards abnormal heart rhythms classification, can use long-term dependence and local feature that SB-LSTM and TD-CNN excavate signal from ECG signal;Designed TAG, FAM, LAM can be finely adjusted resulting long-term dependence and local feature, to obtain accurate ECG entirety fluctuation model and local fluctuation model.Finally, the processing result of ECG signal section is obtained using FCN.This method sufficiently combine ECG signal whole fluctuation model and local fluctuation model information, can on large-scale dataset service performance it is more preferable.In addition, stable classification results can be obtained this method avoid difference in inter-individual difference caused by manual sort and individual.Particularly, this method without any expertise, without the various features of manual designs, without carrying out feature selecting processing, without individually constructing classifier, be typical end-to-end method, have the characteristics that easy to use, nicety of grading is high.
Description
Technical field
The present invention relates to biological healthy calculating fields, and in particular, to a kind of ECG letter towards abnormal heart rhythms classification
Number analysis method.
Background technique
Cardiac rhythm is generally divided into five classes: non-dystopy type (N:Non-ectopic), room property dystopy type (V:Ventricular
Ectopic), supraventricular dystopy type (S:Supraventricular ectopic), mixed type (F:Fusion), unknown type (Q:
Unknown).Every kind of subtype usually has different clinical manifestations, needs different treatment methods, therefore, to abnormal heart
The rhythm and pace of moving things carries out Accurate classification and is to provide the premise effectively treated.
Existing abnormal heart rhythms classification method is broadly divided into two major classes: 1) the artificial inspection method based on medical practitioner;
2) the Computer-Aided Classification method based on Feature Engineering.The former needs doctor to have clinical experience extremely abundant, and needs to consume
Take a large amount of time cost, efficiency is very low;Further, since ECG often shows difference in inter-individual difference and individual, therefore
The artificial usual precision of inspection method based on medical practitioner is lower.Although the latter by computer technology greatly reduce the time at
This, and ensure that the consistency of result, but it must design in advance several features according to expertise.Since ECG fluctuates mould
All there are greatest differences in vivo with a between individuals in formula, the feature of Uniting is difficult to accurately portray the ECG wave under all situations
Dynamic model formula, this limits the processing accuracy of the computer aided processing method based on Feature Engineering to a certain extent.
Summary of the invention
In view of the above problems, the present invention provides a kind of ECG signal analysis method towards abnormal heart rhythms classification, with solution
The lower problem of computer aided processing method precision certainly based on Feature Engineering.
The technical solution for the ECG signal analysis method that present invention kind is classified based on abnormal heart rhythms are as follows: one kind is towards different
The ECG signal analysis method of normal cardiac rhythm classification, comprising the following steps:
S1: pretreatment operation is carried out to ECG signal, obtains the ECG signal section of uniform format;
S2: it designs and the multi-layer biaxially oriented length of door Memory Neural Networks in short-term is paid attention to based on trend, dug from the result that S1 is exported
Dig the whole fluctuation model of ECG signal;
S3: two-dimensional convolution neural network of the design based on feature attention mechanism and position attention mechanism, the knot exported from S2
The localised waving mode of ECG signal is excavated in fruit;
S4: it is further processed using output result of the full Connection Neural Network to S3, uses full Connection Neural Network
The result of S3 output is fitted, and fitting result is mapped as probability distribution using softMax function, to obtain each
The processing result of ECG signal section.
Further, a kind of ECG signal analysis method towards abnormal heart rhythms classification, to ECG signal in the S1
Carrying out pretreatment operation includes: that ECG signal is normalized using Z-score algorithm first;Then in ECG signal
The peak R centered on, intercept an appropriate number of sampled point as ECG signal section corresponding to current heartbeat;Finally gained ECG is believed
The set of number section carries out up-sampling treatment, so that each cardiac rhythm type includes the example number of identical quantity.
Further, a kind of ECG signal analysis method towards abnormal heart rhythms classification excavates ECG letter in the S2
Number whole fluctuation model refer to: using noticing that Memory Neural Networks excavate ECG and believe in short-term for the multi-layer biaxially oriented length of door based on trend
Number long-term dependence, and according to each sampled point the location of in ECG and its around waveform it is long-term to what is excavated
Dependence is finely adjusted, to obtain accurate ECG Long-term Fluctuation mode.
Further, a kind of ECG signal analysis method towards abnormal heart rhythms classification, the multi-layer biaxially oriented length is in short-term
Memory Neural Networks structure is made of three layers of two-way LSTM;The result of upper one layer of LSTM, which exports, gives next layer equidirectional LSTM,
Each LSTM unit is made of four doors, it may be assumed that forgets that door, input gate, out gate and trend pay attention to door;Four doors phase interaction
With the update of co- controlling LSTM location mode and the output of result;Wherein, forget that door is used to lose before LSTM in state
Garbage;Which information input gate decision should currently add to LSTM state;Out gate determines that current cell should be by which
A little end-state of the information as LSTM;Trend pays attention to door according to the location of current sampling point and surrounding waveform to adjust
The weight of long-term dependence corresponding to current sampling point.
Further, a kind of ECG signal analysis method towards abnormal heart rhythms classification, it is hidden in each LSTM unit
Hiding neuron number is set as 16;L is set as 8;N` is set as the length that 56, L=N-N` is waveform around, and N, N` are sampling
Point.
Further, a kind of ECG signal analysis method towards abnormal heart rhythms classification excavates ECG letter in the S3
Number localised waving mode refer to: using based on the two-dimensional convolution neural network of feature attention mechanism and position attention mechanism excavate
Local feature in ECG signal, and carried out respectively according to the type of feature and its to characteristic value the location of in ECG signal
Fine tuning, to obtain accurate ECG localised waving mode.
Further, a kind of ECG signal analysis method towards abnormal heart rhythms classification, connection nerve entirely in the S4
Network is connected layer and constituted entirely by two, contained by neuron number be set to 32 and 5;Particularly, second connects layer and number entirely
Quantity according to concentration cardiac rhythm type is equal;One Softmax layers are added after the last one connects layer entirely, and network is exported
As a result it is converted into form of probability, so that it is determined that processing result.
The invention has the benefit that the ECG signal analysis method, can use SB-LSTM and TD-CNN from ECG
The long-term dependence and local feature of signal are excavated in signal.Particularly, designed TAG, FAM, LAM can be to resulting length
Phase dependence and local feature are finely adjusted, to obtain accurate ECG entirety fluctuation model and local fluctuation model.Most
Eventually, the processing result of ECG signal section is obtained using FCN.This method sufficiently combines whole fluctuation model and the part of ECG signal
Fluctuation model information, can on large-scale dataset service performance it is more preferable.In addition, this method avoid caused by manual sort
Difference in inter-individual difference and individual, can obtain stable classification results.Particularly, this method is known without any expert
Know, without the various features of manual designs, without carrying out feature selecting processing, without individually constructing classifier, be typically to hold
To end method, have the characteristics that easy to use, nicety of grading is high.
Detailed description of the invention
Fig. 1 is the structural schematic diagram of the abnormal heart rhythms disaggregated model the present invention is based on ECG;
Fig. 2 is a kind of ECG signal analysis method overall procedure towards abnormal heart rhythms classification in the embodiment of the present invention
Schematic diagram;
Fig. 3 is the structural schematic diagram of LSTM unit in TAG-SB-LSTM designed in the embodiment of the present invention;
Fig. 4 is the structural schematic diagram of convolutional layer in FAM-LAM-TD-CNN designed in the embodiment of the present invention.
Specific embodiment
Further describe technical solution of the present invention with reference to the accompanying drawing:
In order to enable those skilled in the art to better understand the solution of the present invention, below in conjunction with the embodiment of the present invention and
The technical solution in the present invention is clearly and completely described in its attached drawing.
A kind of ECG signal analysis method towards abnormal heart rhythms classification, utilizes TAG-SB-LSTM and FAM-LAM-
TD-CNN excavates long-term dependence and local feature from ECG signal respectively, and according to the location of sampled point and surrounding
Waveform is finely adjusted long-term dependence, is finely adjusted according to the type and location of extracting spy to characteristic value, from
And accurate ECG entirety fluctuation model and local fluctuation model are obtained, finally fully-connected network is recycled to determine the ECG signal section
Processing result.The integral frame structure of this method is as shown in Figure 1.
A kind of ECG signal analysis method towards abnormal heart rhythms classification, as shown in Figure 2, comprising the following steps:
S1: it the pretreatment operations such as is normalized, is segmented, up-sampling to ECG, obtaining the ECG signal section of uniform format.
In the embodiment of the present invention, ECG signal is usually expressed as difference in apparent inter-individual difference and individual, different
ECG signal amplitude range will affect the accuracy of processing result, for this purpose, being returned first using Z-socre method to ECG signal
One change processing, shown in following formula:
Wherein, μ is the average value of ECG signal sequence, and σ is the standard deviation of signal sequence, XiIt is i-th in ECG signal sequence
A signal value, Xnor_iIt is XiValue after normalized.Then the position at the peak R in ECG signal is determined using existing algorithm
(embodiment of the present invention carries out R blob detection using popular Pan-Tompkins algorithm), then centered on the moment where the peak R,
An appropriate number of sampled point is chosen to represent and work as ECG signal section corresponding to time heartbeat.In view of ECG signal in the embodiment of the present invention
Sample frequency be 360Hz, and human normal heart rate range be 60-100 beats/min, therefore centered on the peak R select 256 adopt
Sampling point not only can utmostly retain the main information when time heartbeat as ECG signal section corresponding to time heartbeat is worked as in this way, but also
Avoid the information for being mixed into adjacent cardiac.Finally, the cardiac rhythm type less to instance number carries out up-sampling operation, so that every kind
Cardiac rhythm includes the example of identical quantity.The embodiment of the present invention utilizes existing SMOTE (Synthetic Minority
Over-sampling Technique) technology realize the up-sampling operate.
S2: it designs and the multi-layer biaxially oriented length of door Memory Neural Networks (TAG-SB-LSTM:Trend in short-term is paid attention to based on trend
Attention Gate based Stacked Bidirectional Long Short-term Memory Network), from
The whole fluctuation model of ECG signal is excavated in the result of S1 output.
The long-term dependence in every ECG signal between sampled point is excavated in the embodiment of the present invention using TAG-SB-LSTM to close
System, to portray the whole fluctuation model of ECG signal section.Designed TAG-SB-LSTM network structure is by three layers of two-way LSTM structure
At.Wherein, the result of upper one layer of LSTM, which exports, gives next layer equidirectional LSTM.Internal structure such as Fig. 3 of each LSTM unit
Shown, it is made of four doors, it may be assumed that forgets that door, input gate, out gate and trend pay attention to door.Four door interactions, it is common to control
The update of LSTM location mode processed and the output of result.Wherein, forget the garbage before door is used to lose LSTM in state;
Which information input gate decision should currently add to LSTM state;Out gate determine current cell should using which information as
The end-state of LSTM;Trend pays attention to door according to the location of current sampling point and surrounding waveform to adjust current sampling point
The weight of corresponding long-term dependence.The input vector x of given t momentt, the output h of LSTMtAnd the state c of LSTMt
It is updated according to following formula:
ft=σ (Wxf·xt+Whf·ht-1+bf) (1)
it=σ (Wxi·xt+Whi·hr-1+bi)(2)
ot=Tt⊙σ(Wxo·x′t+Who·ht-1+bo) (3)
Tt=σ (Wxt·x′t+σ(Wrt·Rt)+bt). (4)
ct=ft⊙ct-1+it⊙Tt⊙tanh(Wxc·x′t+Whc·ht-1+bc) (5)
ht=ot⊙tanh(ct) (6)
Wherein, ft, it, ot, TtIt respectively represents and forgets that door, input gate, out gate, trend pay attention to door;W, b respectively represent power
Weight and bias term;σ indicates activation primitive;⊙ representing matrix is operated by element multiplication;RtIndicate current sampling point (i.e. xt) locating for
Position and surrounding shape information.Specifically, RtIt can be obtained by following process: give an ECG signal section x=x1, x2..., xN, cut
Input of the intermediate a sampled point of N ' as TAG-SB-LSTM is taken, x ' is denoted as.For t-th of element x in x ' 't, it opposite
Position and surrounding shape information can indicate are as follows:
Rt=[01..., 0t-1, xt..., xt+L, 0t+L+1..., 0N], (7)
Wherein, xt..., xt+LPart is x 'tSurrounding's waveform;L=N-N ' is the length of waveform around.Particularly,
xt..., xt+LNeutral element xt+L/2As x 'tItself, therefore, x 'tIn RtIn position can represent x 'tBelieve in entire ECG
Position in number section.RtIn several zeros, on the one hand can shield other apart from farther away sampled point to current sampling point
It influences, on the other hand can also guarantee that the position of each sampled point and surrounding shape information possess identical format, in order to
LSTM processing.Specifically when implementing, the hidden neuron number in each LSTM unit is set as 16;L is set as 8;N ' setting
It is 56;Finally, TAG-SB-LSTM is fused into a square to result with after in the forward direction result that each moment is exported by us
Battle array, shown in following formula:
Wherein, YI, f, YI, bRespectively indicate x 'iCorresponding forward direction result and backward result.
S3: two-dimensional convolution neural network (FAM-LAM-TD- of the design based on feature attention mechanism and position attention mechanism
CNN:Feature Attention Mechanism and Location Attention Mechanism based Two-
DimensionalConvolutional Neural Network), the local wave of ECG signal is excavated from the result that S2 is exported
Dynamic model formula.
FAM-LAM-TD-CNN neural network structure is used in the embodiment of the present invention, is further excavated from the output Y of S2
The localised waving mode of ECG signal.Designed FAM-LAM-TD-CNN network structure in the embodiment of the present invention are as follows: convolutional layer+
Maximum value pond layer+convolutional layer+average value pond layer.The structure of each convolutional layer as shown in figure 4, particularly, FAM and LAM's
Calculating process is specific as follows:
FAM: assuming thatAndRespectively indicate first of convolutional layer
In characteristic pattern, the i-th convolution kernel and convolution results corresponding with the convolution kernel.First by FlAndShape adjust separately as FL
=[f1..., fH×W×C],WhereinThen it calculates according to the following formulaFeature attention weight
WhereinAndFor transition matrix, it is used to Fl
AndIt is mapped toThen, we willAdjusting Shape beThen, withCorresponding power
The convolution results changed againIt can calculate are as follows:
Finally, by the convolution results of the resulting weight of first of convolutional layerIt may be expressed as:
Wherein, C " is the number of convolution kernel in first of convolutional layer.
LAM: the input of one side convolution operation is a matrix, and the i-th row of matrix corresponds to the in ECG signal section
I sampled point;On the other hand, convolution operation and pondization operation will not change the spatial relationship of characteristic pattern.Therefore, in characteristic pattern
Capable sequence can be with relative position of the characteristic feature in ECG signal.So we can basisRow belonging to middle feature
Characteristic value is adjusted, thus the simulation feature influence of present position to feature importance in ECG.
It is assumed thatIndicate withCorresponding ECG signal section.First willAdjusting Shape beThen position attention weight WLIt can be calculated as follows:
WhereinAndFor transition matrix, being used to willBeing mapped as length with S is H "
Vector, to make resulting WLIt can be corresponded to each other with every a line in feature.Why introduced in formula 13 original
ECG signal s is because in this case, model can preferably determine which position should have higher power in ECG
Weight.Finally, by position attention weight WLCharacteristic pattern adjustedIt can indicate are as follows:
Wherein,In i-th of element and WLIn i-th of element carry out product calculation.
In the specific implementation, the number of the convolution kernel of two convolutional layers is set as 32, and each convolution kernel is sized to 4 × 4,
And step-length is 2;The reception domain size of maximum value pond layer is also 4 × 4, but step-length is 4;The reception domain size of average value pond layer
And step-length is identical as maximum value pond layer.Finally, the output result of FAM-LAM-TD-CNN is subjected to flaky process, in order to
Subsequent fully-connected network processing.
S4: it is carried out using output result of the full Connection Neural Network (FCN:Fully Connected Network) to S3
It is further processed, finally obtains the processing result of each ECG signal section.
In the embodiment of the present invention, FCN is connected layer and constituted entirely by two, contained by neuron number be set to 32 and 5.It is special
Not, second to connect layer equal with the quantity of data set cardiac rhythm and pace of moving things type entirely.Then, add after the last one connects layer entirely
Add one Softmax layers, so that network output result is converted into form of probability, so that it is determined that original ECG signal section
Processing result.
In the entire neural network model of training, each heart murmur parting is encoded using one-hot format, is used
Classification intersects the loss that entropy function calculates network;Use LeakyRelu as the activation primitive of neuron in network;Use Adam
Optimizer updates the parameters in network;Various parameters in network, such as weight matrix, biasing are initialized as at random
Value;In order to avoid over-fitting, using decline learning rate technology, (learning rate initial value is 0.002 to the embodiment of the present invention, often
200 iteration, learning rate fall to original 90%), dropout technology (dropout layers be located at FAM-LAM-TD-CNN with
Between FCN, and conservation rate is set as 0.95) handling network;Data set is divided into instruction in 0.7: 0.1: 0.2 ratio
Practice collection, verifying collection, test set, training is using 128 examples as a mini-batch every time, and entire data set training 5 times, choosing
The model of the best performance on test set is selected as final model.
Claims (7)
1. a kind of ECG signal analysis method towards abnormal heart rhythms classification, it is characterised in that: the following steps are included:
S1: pretreatment operation is carried out to ECG signal, obtains the ECG signal section of uniform format;
S2: it designs and the multi-layer biaxially oriented length of door Memory Neural Networks in short-term is paid attention to based on trend, excavate ECG from the result that S1 is exported
The whole fluctuation model of signal;
S3: two-dimensional convolution neural network of the design based on feature attention mechanism and position attention mechanism, from the result that S2 is exported
Excavate the localised waving mode of ECG signal;
S4: it is further processed using output result of the full Connection Neural Network to S3, using full Connection Neural Network to S3
The result of output is fitted, and fitting result is mapped as probability distribution using softMax function, to obtain each ECG
The processing result of signal segment.
2. a kind of ECG signal analysis method towards abnormal heart rhythms classification according to claim 1, feature exist
In: carrying out pretreatment operation to ECG signal in the S1 includes: that ECG signal is normalized using Z-score algorithm first
Processing;Then it centered on the peak R in ECG signal, intercepts an appropriate number of sampled point and believes as ECG corresponding to current heartbeat
Number section;Up-sampling treatment finally is carried out to the set of gained ECG signal section, so that each cardiac rhythm type includes identical quantity
Example number.
3. a kind of ECG signal analysis method towards abnormal heart rhythms classification according to claim 1, feature exist
In: the whole fluctuation model that ECG signal is excavated in the S2 refers to: noticing that the multi-layer biaxially oriented length of door is remembered in short-term using based on trend
Recall the long-term dependence that neural network excavates ECG signal, and according to each sampled point the location of in ECG and its around
Waveform is finely adjusted the long-term dependence excavated, to obtain accurate ECG Long-term Fluctuation mode.
4. a kind of ECG signal analysis method towards abnormal heart rhythms classification according to claim 1, feature exist
In: the multi-layer biaxially oriented length, Memory Neural Networks structure is made of three layers of two-way LSTM in short-term;The result of upper one layer of LSTM exports
Give next layer equidirectional LSTM, each LSTM unit is made of four doors, it may be assumed that forget door, input gate, out gate and trend note
Meaning door;Four doors interaction, the update of co- controlling LSTM location mode and the output of result;Wherein, forget with
To lose the garbage before LSTM in state;Which information input gate decision should currently add to LSTM state;Output
Door determines current cell should be using which information as the end-state of LSTM;Trend pays attention to door according to locating for current sampling point
Position and surrounding waveform adjust the weight of long-term dependence corresponding to current sampling point.
5. a kind of ECG signal analysis method towards abnormal heart rhythms classification according to claim 4, feature exist
In: the hidden neuron number in each LSTM unit is set as 16;L is set as 8;It is around wave that N`, which is set as 56, L=N-N`,
The length of shape, N, N` are sampled point.
6. a kind of ECG signal analysis method towards abnormal heart rhythms classification according to claim 1, feature exist
In: the localised waving mode that ECG signal is excavated in the S3 refers to: using based on feature attention mechanism and position attention mechanism
Two-dimensional convolution neural network excavates the local feature in ECG signal, and the institute respectively according to the type of feature and its in ECG signal
The position at place is finely adjusted characteristic value, to obtain accurate ECG localised waving mode.
7. a kind of ECG signal analysis method towards abnormal heart rhythms classification according to claim 1, feature exist
In: full Connection Neural Network is connected layer and constituted entirely by two in the S4, contained by neuron number be set to 32 and 5;It is special
Not, second to connect layer equal with the quantity of data set cardiac rhythm and pace of moving things type entirely;One is added after the last one connects layer entirely
It is Softmax layers a, network output result is converted into form of probability, so that it is determined that processing result.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910285158.0A CN109907733A (en) | 2019-04-10 | 2019-04-10 | A kind of ECG signal analysis method towards abnormal heart rhythms classification |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910285158.0A CN109907733A (en) | 2019-04-10 | 2019-04-10 | A kind of ECG signal analysis method towards abnormal heart rhythms classification |
Publications (1)
Publication Number | Publication Date |
---|---|
CN109907733A true CN109907733A (en) | 2019-06-21 |
Family
ID=66969276
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910285158.0A Pending CN109907733A (en) | 2019-04-10 | 2019-04-10 | A kind of ECG signal analysis method towards abnormal heart rhythms classification |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109907733A (en) |
Cited By (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110393519A (en) * | 2019-08-19 | 2019-11-01 | 广州视源电子科技股份有限公司 | Analysis method, device, storage medium and the processor of electrocardiosignal |
CN111739034A (en) * | 2020-06-28 | 2020-10-02 | 北京小白世纪网络科技有限公司 | Coronary artery region segmentation system and method based on improved 3D Unet |
CN112307287A (en) * | 2020-11-11 | 2021-02-02 | 国网山东省电力公司威海供电公司 | Cloud edge cooperative architecture based power internet of things data classification processing method and device |
CN113057648A (en) * | 2021-03-22 | 2021-07-02 | 山西三友和智慧信息技术股份有限公司 | ECG signal classification method based on composite LSTM structure |
WO2021197135A1 (en) * | 2020-04-03 | 2021-10-07 | The University Of Hong Kong | Da-bd-lstm-dense-unet for liver lesion segmentation |
-
2019
- 2019-04-10 CN CN201910285158.0A patent/CN109907733A/en active Pending
Cited By (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110393519A (en) * | 2019-08-19 | 2019-11-01 | 广州视源电子科技股份有限公司 | Analysis method, device, storage medium and the processor of electrocardiosignal |
CN110393519B (en) * | 2019-08-19 | 2022-06-24 | 广州视源电子科技股份有限公司 | Electrocardiosignal analysis method and device, storage medium and processor |
WO2021197135A1 (en) * | 2020-04-03 | 2021-10-07 | The University Of Hong Kong | Da-bd-lstm-dense-unet for liver lesion segmentation |
CN111739034A (en) * | 2020-06-28 | 2020-10-02 | 北京小白世纪网络科技有限公司 | Coronary artery region segmentation system and method based on improved 3D Unet |
CN112307287A (en) * | 2020-11-11 | 2021-02-02 | 国网山东省电力公司威海供电公司 | Cloud edge cooperative architecture based power internet of things data classification processing method and device |
CN112307287B (en) * | 2020-11-11 | 2022-08-02 | 国网山东省电力公司威海供电公司 | Cloud edge cooperative architecture based power internet of things data classification processing method and device |
CN113057648A (en) * | 2021-03-22 | 2021-07-02 | 山西三友和智慧信息技术股份有限公司 | ECG signal classification method based on composite LSTM structure |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109907733A (en) | A kind of ECG signal analysis method towards abnormal heart rhythms classification | |
Yao et al. | A comprehensive survey on convolutional neural network in medical image analysis | |
CN109567793A (en) | A kind of ECG signal processing method towards cardiac arrhythmia classification | |
CN108090658A (en) | Arc fault diagnostic method based on time domain charactreristic parameter fusion | |
Cheruku et al. | Rst-batminer: A fuzzy rule miner integrating rough set feature selection and bat optimization for detection of diabetes disease | |
CN109117864A (en) | Coronary heart disease risk prediction technique, model and system based on heterogeneous characteristic fusion | |
CN109214250A (en) | A kind of static gesture identification method based on multiple dimensioned convolutional neural networks | |
CN107730003A (en) | One kind supports more high-precision NILM implementation methods of appliance type | |
CN101564328B (en) | Laptop artificial limb multi-movement-mode identifying method based on support vector data description | |
CN108319928A (en) | A kind of deep learning model and application based on Multi-objective PSO optimization | |
CN108805268A (en) | Deeply learning strategy network training method based on evolution algorithm | |
CN102622515A (en) | Weather prediction method | |
Wu et al. | A multi-objective particle swarm optimization for neural networks pruning | |
Deepika et al. | A empirical study on disease diagnosis using data mining techniques | |
Cao et al. | Improved crow search algorithm optimized extreme learning machine based on classification algorithm and application | |
CN111382840B (en) | HTM design method based on cyclic learning unit and oriented to natural language processing | |
Xu et al. | A fuzzy process neural network model and its application in process signal classification | |
Liu et al. | Kiwifruit leaf disease identification using improved deep convolutional neural networks | |
Niu et al. | A novel quantum neural network based on multi-level activation function | |
CN108171322A (en) | A kind of Learning Algorithm based on particle group optimizing | |
Khot et al. | A detailed study of interpretability of deep neural network based top taggers | |
Lin et al. | Few-shot learning approach with multi-scale feature fusion and attention for plant disease recognition | |
Cao et al. | The optimal multi-degree reduction of Ball Bézier curves using an improved squirrel search algorithm | |
Zhang et al. | Optimization and performance evaluation of deep learning algorithm in medical image processing | |
Ying et al. | Efficient multi-objective evolutionary neural architecture search for U-Nets with diamond atrous convolution and Transformer for medical image segmentation |
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 | ||
RJ01 | Rejection of invention patent application after publication | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190621 |