CN114366047B - Multi-task neural network pulse condition data processing method, system and terminal - Google Patents
Multi-task neural network pulse condition data processing method, system and terminal Download PDFInfo
- Publication number
- CN114366047B CN114366047B CN202210102125.XA CN202210102125A CN114366047B CN 114366047 B CN114366047 B CN 114366047B CN 202210102125 A CN202210102125 A CN 202210102125A CN 114366047 B CN114366047 B CN 114366047B
- Authority
- CN
- China
- Prior art keywords
- pulse
- data
- pulse condition
- extraction
- recognition 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.)
- Active
Links
Images
Classifications
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/02—Detecting, measuring or recording pulse, heart rate, blood pressure or blood flow; Combined pulse/heart-rate/blood pressure determination; Evaluating a cardiovascular condition not otherwise provided for, e.g. using combinations of techniques provided for in this group with electrocardiography or electroauscultation; Heart catheters for measuring blood pressure
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/48—Other medical applications
- A61B5/4854—Diagnosis based on concepts of traditional oriental medicine
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/72—Signal processing specially adapted for physiological signals or for diagnostic purposes
- A61B5/7235—Details of waveform analysis
-
- 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
- G06N3/044—Recurrent networks, e.g. Hopfield networks
-
- 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
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/12—Classification; Matching
Landscapes
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Molecular Biology (AREA)
- Theoretical Computer Science (AREA)
- Medical Informatics (AREA)
- Artificial Intelligence (AREA)
- Public Health (AREA)
- Animal Behavior & Ethology (AREA)
- Surgery (AREA)
- Veterinary Medicine (AREA)
- Heart & Thoracic Surgery (AREA)
- Pathology (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Physiology (AREA)
- Computational Linguistics (AREA)
- Computing Systems (AREA)
- Software Systems (AREA)
- Mathematical Physics (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Signal Processing (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Psychiatry (AREA)
- Alternative & Traditional Medicine (AREA)
- Cardiology (AREA)
- Image Analysis (AREA)
Abstract
According to the method, the system and the terminal for processing the pulse condition data of the multi-task neural network, the pulse rate identification result, the rhythm identification result, the pulse potential flow advantage identification result and the pulse potential tension identification result corresponding to the pulse condition data segment are obtained according to the pulse condition data segment based on the constructed multi-task pulse condition signal identification model. The scheme of the invention can subdivide the results of the pulse condition signals from a single element, not just output the pulse condition names, accords with the theory of traditional Chinese medicine, and can obtain the pulse rate identification result, the rhythm identification result, the pulse potential fluency identification result and the pulse potential tension identification result with high accuracy.
Description
Technical Field
The application relates to the technical field of data processing, in particular to a method, a system and a terminal for processing multitasking neural network pulse condition data.
Background
Regarding the types of pulse conditions, different factors in medical works such as "Nei Jing" (interior channel) and "F.lake Mai Xuan (frequency lake Mai Xuan Shen) are classified into different categories. Among the pulse conditions listed by the physician of the past generation, the later-clearing of the physician's peripheral sea divides the pulse conditions into four aspects of location, number, shape and potential, which is a highly accepted way of explanation at present. Specifically stated as "the person is in position," the size of the sink and float is also; the number of patients is slow, the number of nodes is fast; the shape is that the length, width, thickness, hardness and softness are also the wired surface of the abacus; the potential person can also hold up Shu Shensu and move forward and backward with the fluctuation. Many single pulses are not actually a simple one, but rather a plurality of single elements (bit attributes, number attributes, shape attributes, potential attributes) are recombined in different proportions.
For the identification of pulse condition data, the current wider processing mode is to identify the characteristic points of the time domain, the frequency domain or the time-frequency domain, and classify the pulse condition data by means of the characteristic value judgment mode of an expert system. Meanwhile, as the time-frequency domain characteristic points of part of pulse signals are not obvious or various expression modes exist, the identification mode of the characteristic points is not suitable for all signals. And the obtained classification result is only the names of pulse conditions, such as flat pulse, slippery pulse, wiry pulse, and the like, and lacks the subdivision content of the pulse conditions on the single pulse condition element.
Disclosure of Invention
In view of the above-mentioned drawbacks of the prior art, the present application aims to provide a method, a system and a terminal for processing data of a multi-task neural network, which solve the problem that in the prior art, the pulse data identification processing mode lacks the subdivision content of the pulse on a single pulse element.
To achieve the above and other related objects, the present application provides a method for processing data of a multitasking neural network pulse condition, including: acquiring pulse condition data segments to be identified; based on the constructed multi-task pulse condition signal identification model, acquiring a multi-element pulse condition identification result corresponding to the pulse condition data segment; wherein the multi-element recognition result includes: pulse rate recognition results, rhythm recognition results, pulse potential fluency recognition results, and pulse potential tone recognition results; the method for constructing the multi-task pulse condition signal identification model comprises the following steps: constructing a pulse condition signal identification frame model; training a pulse condition signal recognition frame model by using the training data set to obtain a multi-task pulse condition signal recognition model; and wherein the training data set comprises: a plurality of pulse condition data segment samples and a multi-element pulse condition recognition result corresponding to each pulse condition data segment sample.
In one or more embodiments of the present application, the pulse signal recognition framework model includes: the characteristic rough extraction structure is used for carrying out rough characteristic extraction on the input pulse condition data segments so as to output rough characteristic extraction data; the characteristic fine extraction structure is connected with the characteristic coarse extraction structure and is used for carrying out fine characteristic extraction on the coarse characteristic extraction data so as to output fine characteristic extraction data; the multi-element result output structure is connected with the characteristic fine extraction structure and is used for obtaining a multi-element identification result corresponding to the pulse condition data fragment based on the input fine characteristic extraction data; wherein the multi-element result output structure comprises: the pulse rate identification result output structure is used for extracting data based on the input fine features and outputting pulse rate identification results; a rhythm recognition result output structure for outputting a rhythm recognition result based on the input fine feature extraction data; the pulse potential flow utilization degree identification result output structure is used for extracting data based on the input fine characteristics and outputting pulse potential flow utilization degree identification results; and the pulse tension recognition result output structure is used for extracting data based on the input fine features and outputting pulse tension recognition results.
In one or more embodiments of the present application, the feature coarse extraction structure includes: two connected characteristic coarse sub extraction structures; wherein each feature coarse sub-extraction structure comprises: a convolution layer, a Dropout layer connected with the convolution layer and a Batchnormalization layer connected with the Dropout layer.
In one or more embodiments of the present application, the feature extraction structure includes: a branch feature extraction structure comprising: three branch structures corresponding to different convolution kernels are used for respectively outputting corresponding branch feature data according to the input coarse feature extraction data; wherein each branching structure comprises: the ResBlock1D structure is connected in three layers and is used for carrying out three residual calculation on the input coarse feature extraction data; the global average pooling layer is used for carrying out average processing on the data subjected to three times of residual error processing; and the fusion structure is connected with the branch feature extraction structure and is used for fusing the branch feature data output by the three branch structures and outputting fine feature extraction data.
In one or more embodiments of the present application, the activation function adopted by the pulse rate identification result output structure is a relu function; the activation function adopted by the rhythm recognition result output structure is a sigmoid function; the activation function adopted by the pulse potential flow advantage degree identification result output structure is a softmax function; and an activation function adopted by the pulse tension recognition result output structure is a sigmoid function.
In one or more embodiments of the present application, the multi-element result output structure includes: a pulse rate recognition result output structure for extracting pulse rate values in units of sub-minute data output based on the input fine features; a rhythm recognition result output structure for extracting data output rhythm classification results based on the input fine features; wherein the rhythm classification result comprises: corresponding to the results of rhythmic or dysrhythmic; the pulse potential flow advantage degree identification result output structure is used for extracting data based on the input fine features and outputting pulse potential flow degree three-classification results; wherein, the three classification results of the pulse potential fluency comprise: a result corresponding to one of a slippery pulse, a astringent pulse, and a non-slippery pulse; the pulse tension recognition result output structure is used for extracting data based on the input fine features and outputting pulse tension classification results; wherein the pulse tension classification result comprises: corresponding to a wiry pulse or a flat pulse.
In one or more embodiments of the present application, the pulse signal recognition framework model further includes: a loss weighting calculation layer connected to the multi-element result output structure, comprising: the first loss weighting calculation module is used for carrying out loss weighting calculation on the pulse rate identification result by adopting an MSE function; and the second loss weighting calculation module is used for carrying out loss weighting calculation on the rhythm recognition result, the pulse potential stream benefit recognition result and the pulse potential tension recognition result by adopting a weighted cross entropy loss function.
In one or more embodiments of the present application, the pulse condition data segment is obtained by resampling a pulse condition signal.
To achieve the above and other related objects, the present application provides a system for processing data of a multi-tasking neural network, comprising: the data acquisition module is used for acquiring pulse condition data fragments to be identified; the identification module is connected with the data acquisition module and is used for acquiring a multi-element pulse condition identification result corresponding to the pulse condition data segment based on the constructed multi-task pulse condition signal identification model; wherein the multi-element recognition result includes: pulse rate recognition results, rhythm recognition results, pulse potential fluency recognition results, and pulse potential tone recognition results; the method for constructing the multi-task pulse condition signal identification model comprises the following steps: constructing a pulse condition signal identification frame model; training a pulse condition signal recognition frame model by using the training data set to obtain a multi-task pulse condition signal recognition model; and wherein the training data set comprises: a plurality of pulse condition data segment samples and a multi-element pulse condition recognition result corresponding to each pulse condition data segment sample.
To achieve the above and other related objects, the present application provides a multi-task neural network pulse condition data processing terminal, including: one or more memories and one or more processors; the one or more memories are used for storing computer programs; the one or more processors are connected with the memory and are used for running the computer program to execute the multitasking neural network pulse condition data processing method.
As described above, the method, system and terminal for processing the pulse condition data of the multi-task neural network acquire the pulse rate identification result, the rhythm identification result, the pulse potential flow advantage identification result and the pulse potential tension identification result corresponding to the pulse condition data segment based on the constructed multi-task pulse condition signal identification model. The scheme of the invention can subdivide the results of the pulse condition signals from a single element, not just output the pulse condition names, accords with the theory of traditional Chinese medicine, and can obtain the pulse rate identification result, the rhythm identification result, the pulse potential fluency identification result and the pulse potential tension identification result with high accuracy.
Drawings
Fig. 1 is a flow chart of a method for processing data of a multitasking neural network according to an embodiment of the present application.
Fig. 2 is a schematic structural diagram of a pulse signal recognition framework model in an embodiment of the present application.
Fig. 3 is a schematic structural diagram of a feature rough extraction structure in an embodiment of the present application.
Fig. 4 is a schematic structural diagram of a feature extraction structure in an embodiment of the present application.
Fig. 5 shows a schematic structural diagram of the ResBlock1D structure in the embodiment of the present application.
FIG. 6 is a schematic diagram of a pulse signal recognition framework according to an embodiment of the present application.
FIG. 7 is a schematic diagram of a system for processing data of a multi-task neural network according to an embodiment of the present application.
Fig. 8 is a schematic structural diagram of a terminal of a data processing side of a multi-task neural network according to an embodiment of the present application.
Detailed Description
Other advantages and effects of the present application will become apparent to those skilled in the art from the present disclosure, when the following description of the embodiments is taken in conjunction with the accompanying drawings. The present application may be embodied or applied in other specific forms and details, and various modifications and alterations may be made to the details of the present application from a different perspective or perspective without departing from the spirit of the present application. It should be noted that, in the case of no conflict, the embodiments and features in the embodiments may be combined with each other.
The embodiments of the present application will be described in detail below with reference to the drawings so that those skilled in the art to which the present application pertains can easily implement the same. This application may be embodied in many different forms and is not limited to the embodiments described herein.
For the purpose of clarity of explanation of the present application, components not related to the explanation are omitted, and the same or similar components are given the same reference numerals throughout the specification.
Throughout the specification, when a component is said to be "connected" to another component, this includes not only the case of "direct connection" but also the case of "indirect connection" with other elements interposed therebetween. In addition, when a certain component is said to "include" a certain component, unless specifically stated to the contrary, it is meant that other components are not excluded, but other components may be included.
When an element is referred to as being "on" another element, it can be directly on the other element but be accompanied by the other element therebetween. When a component is stated to be "directly on" another component, it is stated that there are no other components between them.
Although the terms first, second, etc. may be used herein to describe various elements in some examples, these elements should not be limited by these terms. These terms are only used to distinguish one element from another element. Such as a first interface and a second interface, etc. Furthermore, as used herein, the singular forms "a", "an" and "the" are intended to include the plural forms as well, unless the context indicates otherwise. It will be further understood that the terms "comprises," "comprising," "includes," and/or "including" specify the presence of stated features, steps, operations, elements, components, items, categories, and/or groups, but do not preclude the presence, presence or addition of one or more other features, steps, operations, elements, components, items, categories, and/or groups. The terms "or" and/or "as used herein are to be construed as inclusive, or meaning any one or any combination. Thus, "A, B or C" or "A, B and/or C" means "any of the following: a, A is as follows; b, a step of preparing a composite material; c, performing operation; a and B; a and C; b and C; A. b and C). An exception to this definition will occur only when a combination of elements, functions, steps or operations are in some way inherently mutually exclusive.
The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. As used herein, the singular forms "a", "an" and "the" are intended to include the plural forms as well, unless the language clearly indicates the contrary. The meaning of "comprising" in the specification is to specify the presence of stated features, regions, integers, steps, operations, elements, and/or components, but does not preclude the presence or addition of other features, regions, integers, steps, operations, elements, and/or components.
Terms representing relative spaces such as "lower", "upper", and the like may be used to more easily describe the relationship of one component relative to another component illustrated in the figures. Such terms refer not only to the meanings indicated in the drawings, but also to other meanings or operations of the device in use. For example, if the device in the figures is turned over, elements described as "under" other elements would then be oriented "over" the other elements. Thus, the exemplary term "lower" includes both upper and lower. The device may be rotated 90 deg. or at other angles and the terminology representing relative space is to be construed accordingly.
Although not differently defined, including technical and scientific terms used herein, all terms have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The term addition defined in the commonly used dictionary is interpreted as having a meaning conforming to the contents of the related art document and the current hint, so long as no definition is made, it is not interpreted as an ideal or very formulaic meaning too much.
In view of the defects in the prior art, a pulse rate recognition result, a rhythm recognition result, a pulse potential flow proficiency recognition result and a pulse potential tension recognition result corresponding to the pulse condition data segment are obtained according to the pulse condition data segment based on the constructed multi-task pulse condition signal recognition model. The scheme of the invention can subdivide the results of the pulse condition signals from a single element, not just output the pulse condition names, accords with the theory of traditional Chinese medicine, and can obtain the pulse rate identification result, the rhythm identification result, the pulse potential fluency identification result and the pulse potential tension identification result with high accuracy.
The embodiments of the present invention will be described in detail below with reference to the attached drawings so that those skilled in the art to which the present invention pertains can easily implement the present invention. This invention may be embodied in many different forms and is not limited to the embodiments described herein.
Fig. 1 is a schematic flow chart of a method for processing pulse condition data of a multi-task neural network according to an embodiment of the invention.
The method comprises the following steps:
step S11: and acquiring pulse condition data fragments to be identified.
Optionally, the pulse condition data segment is data with a certain length; for example, since a plurality of pulse cycles are required for judging the pulse signal rhythm, the input length is set to 18 seconds.
Alternatively, the pulse signal is sampled by a device with a sampling rate set to 100sps and a channel number of 1, so the dimension of the acquired data is 1800,1.
Optionally, since the pulse rate of the pulse condition signal is a specific numerical value in units of sub-divisions, the value is between 40 and 200, and the pulse rate of the detected data is generally distributed in a middle position, the pulse rate results at two ends are less obtained, and the distribution is uneven; therefore, resampling of the pulse signal is required to obtain the pulse data segment. Preferably, the resampling parameter is set to 1-1.3, and the length of 1800samples is randomly truncated after the resampling is completed. By adopting the resampling mode, the robustness of pulse rate output can be greatly improved, the judgment results of rhythms and pulse potentials can not be influenced, and the multitask pulse signal recognition model can output the pulse rate recognition result, the rhythm recognition result, the pulse potential flow proficiency recognition result and the pulse potential tension recognition result at the same time.
Step S12: based on the constructed multi-task pulse condition signal identification model, a multi-element pulse condition identification result corresponding to the pulse condition data segment is obtained according to the pulse condition data segment.
In detail, the multi-element recognition result includes: pulse rate recognition results, rhythm recognition results, pulse potential fluency recognition results, and pulse potential tone recognition results. The method for constructing the multi-task pulse condition signal identification model comprises the following steps: constructing a pulse condition signal identification frame model; training a pulse condition signal recognition frame model by using the training data set to obtain a multi-task pulse condition signal recognition model; and wherein the training data set comprises: a plurality of pulse condition data segment samples and a multi-element pulse condition recognition result corresponding to each pulse condition data segment sample.
Optionally, as shown in fig. 2, the pulse signal identification framework model includes: a feature rough extraction structure 21 for performing rough feature extraction on the input pulse condition data segment to output rough feature extraction data; a fine feature extraction structure 22 connected to the coarse feature extraction structure for performing fine feature extraction on the coarse feature extraction data to output fine feature extraction data; a multi-element result output structure 23 connected to the feature fine extraction structure 22 for obtaining a multi-element recognition result corresponding to the pulse condition data segment based on the input fine feature extraction data; wherein the multi-element result output structure 23 includes: a pulse rate recognition result output structure 231 for outputting a pulse rate recognition result based on the input fine feature extraction data; a rhythm recognition result output structure 232 for outputting a rhythm recognition result based on the input fine feature extraction data; a pulse stream benefit recognition result output structure 233 for outputting a pulse stream benefit recognition result based on the input fine feature extraction data; and a pulse tension recognition result output structure 234 for outputting a pulse tension recognition result based on the input fine feature extraction data.
Optionally, as shown in fig. 3, the feature coarse extraction structure includes: two connected feature coarse sub-extraction structures 311 and 312; namely inputting pulse condition data fragments into the characteristic rough sub-extraction structure 311 to output first rough characteristic extraction data, and inputting the first rough characteristic extraction data into the characteristic rough sub-extraction structure 312 to output rough characteristic extraction data; wherein each feature coarse sub-extraction structure comprises: a convolution layer Conv for mapping data onto the feature space; the Dropout layer is connected with the convolution layer and is used for removing a training unit of the neural network from the network according to a certain probability, so that the activation value of the neuron is stopped, and the generalization of the model is stronger; preferably, the probability value of its removal is set to 0.4. And the Batchnormalization layer is connected with the Dropout layer and is used for normalizing the output value of the upper layer and solving the influence of the transmission offset and the increase of the input data.
Optionally, as shown in fig. 4, the feature extraction structure includes:
a branch feature extraction structure comprising: three branch structures 411, 412 and 413 corresponding to different convolution kernels, for outputting corresponding branch feature data according to the input coarse feature extraction data, respectively; the different convolution kernel sizes are set to extract features at different scales.
Wherein each branching structure comprises: the ResBlock1D structure is connected in three layers and is used for carrying out three residual calculation on the input coarse feature extraction data; the residual structure is used for preventing the gradient vanishing problem after the network is deepened; the global average pooling layer is connected with the ResBlock1D structure and is used for carrying out average processing on data subjected to three times of residual processing and reducing the dimension of characteristic output;
and the fusion structure 42 is connected with the branch feature extraction structure and is used for fusing the branch feature data output by the three branch structures and outputting fine feature extraction data.
Alternatively, the three branch structures may employ convolution kernels of sizes 3, 5, and 7, respectively.
Alternatively, as shown in fig. 5, the residual structure adopted by each ResBlock1D structure includes: each having two interconnected structures of a convolution layer, a Batchnormal layer, a relu function layer, connected in sequence, and a skip-connected residual structure layer that directly adds input to output.
Optionally, the activating function adopted by the pulse rate identification result output structure is a relu function; the activation function adopted by the rhythm recognition result output structure is a sigmoid function; the activation function adopted by the pulse potential flow advantage degree identification result output structure is a softmax function; and an activation function adopted by the pulse tension recognition result output structure is a sigmoid function.
Optionally, the multi-element result output structure includes: a pulse rate recognition result output structure for extracting pulse rate values in units of sub-minute data output based on the input fine features; a rhythm recognition result output structure for extracting data output rhythm classification results based on the input fine features; wherein the rhythm classification result comprises: corresponding to the results of rhythmic or dysrhythmic; for example, the outputs are either uniform or non-uniform; the pulse potential flow advantage degree identification result output structure is used for extracting data based on the input fine features and outputting pulse potential flow degree three-classification results; wherein, the three classification results of the pulse potential fluency comprise: a result corresponding to one of a slippery pulse, a astringent pulse, and a non-slippery pulse; for example, one of a slip, a astringency, and a no slip feature is output; the pulse tension recognition result output structure is used for extracting data based on the input fine features and outputting pulse tension classification results; wherein the pulse tension classification result comprises: corresponding to the result of wiry pulse or flat pulse; for example, output chords or flat.
Optionally, training the pulse condition signal recognition frame model by using the prepared data set, so that a computer automatically calculates parameters of the model to enable an output result to be similar to a marked result as far as possible, and training the model by using an Adam optimizer as an optimizer for model training;
Preferably, for the acquisition of the training data set, pulse signals with the time length of 20 seconds can be intercepted each time, and the data in the middle of 18 seconds can be randomly intercepted and taken as training data to prepare the training set. For setting the model training parameters, an Adam optimizer can be used as an optimizer for model training to train 700 epochs, the learning rate is initially set to be 0.0001, and the learning rate is reduced to be 0.1 times after 300 epochs.
Optionally, the pulse condition signal identification framework model further includes: the loss calculation layer is connected with the multi-element result output structure and is used for carrying out loss weighted calculation on the output multi-element identification result;
the loss calculation layer includes:
the first loss calculation module is used for calculating the loss of the pulse rate identification result by adopting an MSE function;
because the actual pulse condition data has the problem of data imbalance, for example, in the dimension of the rhythm, the data with the rhythm is more than the data with the rhythm, so that the model is enabled to pay attention to the class of few samples on the basis of accurate output, and the sample size weighting is carried out on the output loss value, so that a loss function of weighted cross entropy is formed; namely, the loss calculation layer further includes: the second loss calculation module is used for calculating losses of the rhythm recognition result, the pulse potential flow proficiency recognition result and the pulse potential tension recognition result by adopting a weighted cross entropy loss function;
I.e. the weighted cross entropy loss function is:
wherein omega (i) The value of the weight is the reciprocal of the class proportion.
In addition, to balance the loss values of the various parts, control parameters may also be added to the loss term of the model loss function.
I.e. model loss function:
wherein alpha is i For each control parameter.
In the preferred embodiment, the control parameters of the pulse rate recognition result, the rhythm recognition result, the pulse potential flow rate recognition result and the pulse potential tension recognition result are 0.5,1,8,4, respectively.
It should be noted that, if the combination of the loss weighting calculation and the signal resampling is used, the robustness of the model can be further improved, and the occurrence of the over-fitting phenomenon can be prevented. In order to better describe the multitasking neural network pulse condition data processing method, a specific embodiment is provided;
example 1: a method of multitasking neural network pulse condition data processing, the method comprising:
1. and (3) model building: constructing a pulse signal identification framework model, wherein a structural schematic diagram of the pulse signal identification framework model is shown in fig. 6;
wherein the pulse signal recognition frame model comprises:
the Conv+Dropout+BN structures are used for performing characteristic rough extraction, and comprise a convolution layer, a Dropout layer and a Batchnormalization layer. The data are mapped onto the feature space through the convolution layer, the Dropout layer is used for removing the training unit of the neural network from the network according to a certain probability, so that the activation value of the neuron of the training unit stops working, the model generalization which can be realized is stronger, and the removal probability value is set to be 0.4. The Batchnormalization normalizes the upper layer output values to account for the effects of input data transmission offset and increase.
Three branches are side by side ResBlock1D structures, and a residual structure is used in each ResBlock1D structure. Each branch has 3 levels, the convolution kernels in the branches are respectively 3,5 and 7, and different convolution kernel sizes are set for feature extraction on different scales. The residual structure is used to prevent the gradient vanishing problem after the network is deepened. And after finishing three residual processes, using a global average pooling layer globalafepal to perform average processing on the output elements, and reducing the dimension of feature output.
And the fusion structure fuses the branch characteristic data output by the three branch structures.
The model has four tasks, namely output of pulse rate, and as the pulse rate is a specific numerical value in units of sub-divisions, relu is applicable as an activation function of an output layer; a classification task of rhythms outputs a classification result of the alignment/non-alignment, and a sigmoid function is used as an activation function of an output layer; classification of pulse potential (fluency), outputting three classification results of slippery/astringent/no-two characteristics, and applying a softmax function as an activation function of an output layer; the output of pulse potential (tone) is a chord/plane classification result, and a sigmoid function is applied as an activation function of the output layer.
A loss weighting calculation layer connected to the multi-element result output structure, comprising: the first loss weighting calculation module is used for carrying out loss weighting calculation on the pulse rate by adopting an MSE function; and the second loss weighting calculation module is used for carrying out loss weighting calculation on the rhythm classification result, the pulse potential flow profit three classification result and the pulse potential tension classification result by adopting a weighted cross entropy loss function.
2. Resampling the signal;
since the pulse rate of the pulse signal is a specific value in units of sub-divisions, the value is between 40 and 200, and the pulse rate of the detected data is generally distributed in a middle position, the pulse rate results at two ends are less obtained, and the distribution is uneven. Therefore, the resampling is carried out on the signals, the resampling parameter is 1-1.3, the robustness of pulse rate output can be greatly improved, and the length of 1800samples is randomly intercepted for training after the resampling is finished. Meanwhile, the resampling mode does not influence the judgment result of rhythms and pulse potentials, so that the multitask model can obtain the output of four tasks simultaneously. If the combination of the loss weighted calculation and the signal resampling is used, the robustness of the model can be further improved, and the phenomenon of over fitting is prevented.
3. Training a network model;
and training the pulse condition signal recognition frame model by using the prepared data set, so that the computer automatically calculates parameters of the model to enable the output result to be similar to the marked result as much as possible. Pulse signals with the time length of 20 seconds are intercepted each time, and data in the middle of 18 seconds are randomly intercepted and taken as training data to prepare a training set. In the model training, an Adam optimizer is adopted as an optimizer for model training to train 700epoch. The initial set learning rate was 0.0001, shrinking to 0.1 fold after 300 epoch.
4. Pulse condition identification;
acquiring pulse condition data segments to be identified; based on the constructed multi-task pulse condition signal identification model, acquiring a multi-element pulse condition identification result corresponding to the pulse condition data segment;
the model of the embodiment can subdivide the results of the pulse signals from a single element, and not only output the pulse names, thereby being more in line with the theory of traditional Chinese medicine. By using the model obtained by data training, the average pulse rate error is lower than 1 time/minute, the classification accuracy of the rhythm is 99%, the classification accuracy of the pulse condition (fluency) is 93%, and the classification accuracy of the pulse condition (tension) is 92%.
Similar to the principles of the embodiments described above, the present invention provides a multi-tasking neural network pulse data processing system.
Specific embodiments are provided below with reference to the accompanying drawings:
FIG. 7 is a schematic diagram showing a structure of a system for processing pulse condition data of a multi-task neural network according to an embodiment of the present invention.
The system comprises:
a data acquisition module 71, configured to acquire a pulse condition data segment to be identified;
the recognition module 72 is connected with the data acquisition module 71 and is used for obtaining a multi-element pulse condition recognition result corresponding to the pulse condition data segment based on the constructed multi-task pulse condition signal recognition model; wherein the multi-element recognition result includes: pulse rate recognition results, rhythm recognition results, pulse potential fluency recognition results, and pulse potential tone recognition results;
the method for constructing the multi-task pulse condition signal identification model comprises the following steps:
constructing a pulse condition signal identification frame model;
training a pulse condition signal recognition frame model by using the training data set to obtain a multi-task pulse condition signal recognition model; and wherein the training data set comprises: a plurality of pulse condition data segment samples and a multi-element pulse condition recognition result corresponding to each pulse condition data segment sample.
It should be noted that, it should be understood that the division of the modules in the embodiment of the system of fig. 7 is merely a division of logic functions, and may be fully or partially integrated into a physical entity or may be physically separated. And these modules may all be implemented in software in the form of calls by the processing element; or can be realized in hardware; the method can also be realized in a mode that a part of modules are called by processing elements and software, and the part of modules are realized in a hardware mode;
for example, each module may be one or more integrated circuits configured to implement the above methods, e.g.: one or more application specific integrated circuits (Application Specific Integrated Circuit, abbreviated as ASIC), or one or more microprocessors (digital signal processor, abbreviated as DSP), or one or more field programmable gate arrays (Field Programmable Gate Array, abbreviated as FPGA), or the like. For another example, when a module above is implemented in the form of a processing element scheduler code, the processing element may be a general-purpose processor, such as a central processing unit (Central Processing Unit, CPU) or other processor that may invoke the program code. For another example, the modules may be integrated together and implemented in the form of a system-on-a-chip (SOC).
It should be noted that, since the structures and implementation principles of the multi-task neural network pulse condition data processing system and the multi-task pulse condition signal recognition model have been described in the foregoing embodiments, the description thereof is not repeated here.
Fig. 8 shows a schematic structural diagram of a multi-task neural network pulse condition data processing terminal 80 according to an embodiment of the present invention.
The multi-tasking neural network pulse condition data processing terminal 80 includes: a memory 81 and a processor 82, the memory 81 for storing a computer program; the processor 82 runs a computer program to implement the method for processing the data of the multi-tasking neural network pulse condition as described in fig. 1.
Alternatively, the number of the memories 81 may be one or more, and the number of the processors 82 may be one or more, and one is taken as an example in fig. 8.
Optionally, the processor 82 in the multi-task neural network pulse condition data processing terminal 80 loads one or more instructions corresponding to the process of the application program into the memory 81 according to the steps as shown in fig. 1, and the processor 82 runs the application program stored in the first memory 81, so as to implement various functions in the multi-task neural network pulse condition data processing method as shown in fig. 1.
Optionally, the memory 81 may include, but is not limited to, high speed random access memory, nonvolatile memory. Such as one or more disk storage devices, flash memory devices, or other non-volatile solid-state storage devices; the processor 82 may include, but is not limited to, a central processing unit (Central Processing Unit, CPU), a network processor (Network Processor, NP), etc.; but also digital signal processors (Digital Signal Processing, DSP for short), application specific integrated circuits (Application Specific Integrated Circuit, ASIC for short), field-programmable gate arrays (Field-Programmable Gate Array, FPGA for short) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components.
Alternatively, the processor 82 may be a general-purpose processor, including a central processing unit (Central Processing Unit, CPU for short), a network processor (Network Processor, NP for short), etc.; but also digital signal processors (Digital Signal Processing, DSP for short), application specific integrated circuits (Application Specific Integrated Circuit, ASIC for short), field-programmable gate arrays (Field-Programmable Gate Array, FPGA for short) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components.
The present invention also provides a computer readable storage medium storing a computer program which when run implements a method of processing data for a multi-tasking neural network as shown in fig. 1. The computer-readable storage medium may include, but is not limited to, floppy diskettes, optical disks, CD-ROMs (compact disk-read only memories), magneto-optical disks, ROMs (read-only memories), RAMs (random access memories), EPROMs (erasable programmable read only memories), EEPROMs (electrically erasable programmable read only memories), magnetic or optical cards, flash memory, or other type of media/machine-readable medium suitable for storing machine-executable instructions. The computer readable storage medium may be an article of manufacture that is not accessed by a computer device or may be a component used by an accessed computer device.
In summary, the method, the system and the terminal for processing the pulse condition data of the multi-task neural network acquire a pulse rate identification result, a rhythm identification result, a pulse potential fluency identification result and a pulse potential tension identification result corresponding to the pulse condition data segment through the constructed multi-task pulse condition signal identification model. The scheme of the invention can subdivide the results of the pulse condition signals from a single element, not just output the pulse condition names, accords with the theory of traditional Chinese medicine, and can obtain the pulse rate identification result, the rhythm identification result, the pulse potential fluency identification result and the pulse potential tension identification result with high accuracy.
The foregoing embodiments are merely illustrative of the principles of the present application and their effectiveness, and are not intended to limit the application. Modifications and variations may be made to the above-described embodiments by those of ordinary skill in the art without departing from the spirit and scope of the present application. Accordingly, it is intended that all equivalent modifications and variations which may be accomplished by persons skilled in the art without departing from the spirit and technical spirit of the disclosure be covered by the claims of this application.
Claims (4)
1. A method for processing pulse condition data of a multi-task neural network, comprising:
acquiring pulse condition data segments to be identified; the pulse condition data segments are obtained by resampling pulse condition signals;
based on the constructed multi-task pulse condition signal identification model, acquiring a multi-element pulse condition identification result corresponding to the pulse condition data segment; wherein the multi-element recognition result includes: simultaneously outputting pulse rate identification result, rhythm identification result, pulse potential fluency identification result and pulse potential tension identification result;
the method for constructing the multi-task pulse condition signal identification model comprises the following steps:
constructing a pulse condition signal identification frame model;
Training a pulse condition signal recognition frame model by using the training data set to obtain a multi-task pulse condition signal recognition model; and is also provided with
Wherein the training data set comprises: a plurality of pulse condition data segment samples and multi-element pulse condition recognition results corresponding to the pulse condition data segment samples;
wherein, the pulse condition signal identification frame model comprises:
the characteristic rough extraction structure is used for carrying out rough characteristic extraction on the input pulse condition data segments so as to output rough characteristic extraction data; wherein, the characteristic coarse extraction structure includes: two connected characteristic coarse sub extraction structures; wherein each feature coarse sub-extraction structure comprises: a convolution layer, a Dropout layer connected with the convolution layer, and a Batchnormal layer connected with the Dropout layer;
the characteristic fine extraction structure is connected with the characteristic coarse extraction structure and is used for carrying out fine characteristic extraction on the coarse characteristic extraction data so as to output fine characteristic extraction data; wherein, the feature fine extraction structure includes: a branch feature extraction structure comprising: three branch structures corresponding to different convolution kernels are used for respectively outputting corresponding branch feature data according to the input coarse feature extraction data; wherein each branching structure comprises: the ResBlock1D structure is connected in three layers and is used for carrying out three residual calculation on the input coarse feature extraction data; the global average pooling layer is used for carrying out average processing on the data subjected to three times of residual error processing; the fusion structure is connected with the branch feature extraction structure and is used for fusing the branch feature data output by the three branch structures and outputting fine feature extraction data;
The multi-element result output structure is connected with the characteristic fine extraction structure and is used for obtaining a multi-element identification result corresponding to the pulse condition data fragment based on the input fine characteristic extraction data; wherein the multi-element result output structure comprises: the pulse rate identification result output structure is used for extracting data based on the input fine features and outputting pulse rate identification results; a rhythm recognition result output structure for outputting a rhythm recognition result based on the input fine feature extraction data; the pulse potential flow utilization degree identification result output structure is used for extracting data based on the input fine characteristics and outputting pulse potential flow utilization degree identification results; the pulse tension recognition result output structure is used for extracting data based on the input fine features and outputting pulse tension recognition results; and wherein the activation function employed by the pulse rate identification result output structure is a relu function; the activation function adopted by the rhythm recognition result output structure is a sigmoid function; the activation function adopted by the pulse potential flow advantage degree identification result output structure is a softmax function; the activating function adopted by the pulse tension recognition result output structure is a sigmoid function;
a loss weighting calculation layer connected to the multi-element result output structure, comprising: the first loss weighting calculation module is used for carrying out loss weighting calculation on the pulse rate identification result by adopting an MSE function; the second loss weighting calculation module is used for carrying out loss weighting calculation on the rhythm recognition result, the pulse potential stream benefit recognition result and the pulse potential tension recognition result by adopting a weighted cross entropy loss function; wherein, control parameters are respectively set for the pulse rate recognition result, the rhythm recognition result, the pulse potential fluency recognition result and the pulse potential tone recognition result.
2. The method for processing the pulse condition data of the multi-tasking neural network according to claim 1, wherein the multi-element result output structure comprises:
a pulse rate recognition result output structure for extracting pulse rate values in units of sub-minute data output based on the input fine features;
a rhythm recognition result output structure for extracting data output rhythm classification results based on the input fine features; wherein the rhythm classification result comprises: corresponding to the results of rhythmic or dysrhythmic;
the pulse potential flow advantage degree identification result output structure is used for extracting data based on the input fine features and outputting pulse potential flow degree three-classification results; wherein, the three classification results of the pulse potential fluency comprise: a result corresponding to one of a slippery pulse, a astringent pulse, and a non-slippery pulse;
the pulse tension recognition result output structure is used for extracting data based on the input fine features and outputting pulse tension classification results; wherein the pulse tension classification result comprises: corresponding to a wiry pulse or a flat pulse.
3. A multitasking neural network pulse condition data processing system, comprising:
the data acquisition module is used for acquiring pulse condition data fragments to be identified; the pulse condition data segments are obtained by resampling pulse condition signals;
The identification module is connected with the data acquisition module and is used for acquiring a multi-element pulse condition identification result corresponding to the pulse condition data segment based on the constructed multi-task pulse condition signal identification model; wherein the multi-element recognition result includes: simultaneously outputting pulse rate identification result, rhythm identification result, pulse potential fluency identification result and pulse potential tension identification result;
the method for constructing the multi-task pulse condition signal identification model comprises the following steps:
constructing a pulse condition signal identification frame model;
training a pulse condition signal recognition frame model by using the training data set to obtain a multi-task pulse condition signal recognition model; and wherein the training data set comprises: a plurality of pulse condition data segment samples and multi-element pulse condition recognition results corresponding to the pulse condition data segment samples;
wherein, the pulse condition signal identification frame model comprises:
the characteristic rough extraction structure is used for carrying out rough characteristic extraction on the input pulse condition data segments so as to output rough characteristic extraction data; wherein, the characteristic coarse extraction structure includes: two connected characteristic coarse sub extraction structures; wherein each feature coarse sub-extraction structure comprises: a convolution layer, a Dropout layer connected with the convolution layer, and a Batchnormal layer connected with the Dropout layer;
The characteristic fine extraction structure is connected with the characteristic coarse extraction structure and is used for carrying out fine characteristic extraction on the coarse characteristic extraction data so as to output fine characteristic extraction data; wherein, the feature fine extraction structure includes: a branch feature extraction structure comprising: three branch structures corresponding to different convolution kernels are used for respectively outputting corresponding branch feature data according to the input coarse feature extraction data; wherein each branching structure comprises: the ResBlock1D structure is connected in three layers and is used for carrying out three residual calculation on the input coarse feature extraction data; the global average pooling layer is used for carrying out average processing on the data subjected to three times of residual error processing; the fusion structure is connected with the branch feature extraction structure and is used for fusing the branch feature data output by the three branch structures and outputting fine feature extraction data;
the multi-element result output structure is connected with the characteristic fine extraction structure and is used for obtaining a multi-element identification result corresponding to the pulse condition data fragment based on the input fine characteristic extraction data; wherein the multi-element result output structure comprises: the pulse rate identification result output structure is used for extracting data based on the input fine features and outputting pulse rate identification results; a rhythm recognition result output structure for outputting a rhythm recognition result based on the input fine feature extraction data; the pulse potential flow utilization degree identification result output structure is used for extracting data based on the input fine characteristics and outputting pulse potential flow utilization degree identification results; the pulse tension recognition result output structure is used for extracting data based on the input fine features and outputting pulse tension recognition results; and wherein the activation function employed by the pulse rate identification result output structure is a relu function; the activation function adopted by the rhythm recognition result output structure is a sigmoid function; the activation function adopted by the pulse potential flow advantage degree identification result output structure is a softmax function; the activating function adopted by the pulse tension recognition result output structure is a sigmoid function;
A loss weighting calculation layer connected to the multi-element result output structure, comprising: the first loss weighting calculation module is used for carrying out loss weighting calculation on the pulse rate identification result by adopting an MSE function; the second loss weighting calculation module is used for carrying out loss weighting calculation on the rhythm recognition result, the pulse potential stream benefit recognition result and the pulse potential tension recognition result by adopting a weighted cross entropy loss function; wherein, control parameters are respectively set for the pulse rate recognition result, the rhythm recognition result, the pulse potential fluency recognition result and the pulse potential tone recognition result.
4. A multi-tasking neural network pulse data processing terminal comprising: one or more memories and one or more processors;
the one or more memories are used for storing computer programs;
the one or more processors being connected to the memory for running the computer program to perform the method as claimed in claim 1 or 2.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210102125.XA CN114366047B (en) | 2022-01-27 | 2022-01-27 | Multi-task neural network pulse condition data processing method, system and terminal |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210102125.XA CN114366047B (en) | 2022-01-27 | 2022-01-27 | Multi-task neural network pulse condition data processing method, system and terminal |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN114366047A CN114366047A (en) | 2022-04-19 |
| CN114366047B true CN114366047B (en) | 2023-05-09 |
Family
ID=81146290
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210102125.XA Active CN114366047B (en) | 2022-01-27 | 2022-01-27 | Multi-task neural network pulse condition data processing method, system and terminal |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114366047B (en) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115374125B (en) * | 2022-09-01 | 2024-05-10 | 无锡市华焯光电科技有限公司 | Pulse diagnosis classification method, database construction method, device and storage medium |
Family Cites Families (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103034837B (en) * | 2011-09-30 | 2017-07-28 | Ge医疗系统环球技术有限公司 | Characteristic parameter is associated with pulse condition key element |
| CN110929697B (en) * | 2019-12-17 | 2021-04-13 | 中国人民解放军海军航空大学 | Neural network target identification method and system based on residual error structure |
| CN111202512A (en) * | 2020-03-05 | 2020-05-29 | 齐鲁工业大学 | Electrocardiogram classification method and device based on wavelet transformation and DCNN |
| CN112561863B (en) * | 2020-12-03 | 2022-06-10 | 吉林大学 | Medical image multi-classification recognition system based on improved ResNet |
| CN112613581B (en) * | 2020-12-31 | 2023-10-13 | 广州大学华软软件学院 | Image recognition method, system, computer equipment and storage medium |
| CN113397555A (en) * | 2021-07-21 | 2021-09-17 | 南通大学附属医院 | Arrhythmia classification algorithm of C-LSTM for physiological parameter monitoring |
-
2022
- 2022-01-27 CN CN202210102125.XA patent/CN114366047B/en active Active
Also Published As
| Publication number | Publication date |
|---|---|
| CN114366047A (en) | 2022-04-19 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Terekhov et al. | Knowledge transfer in deep block-modular neural networks | |
| Konstantopoulos et al. | Analyzing effect sizes: Fixed-effects models | |
| Ming et al. | Rapid reconstruction of 3D neuronal morphology from light microscopy images with augmented rayburst sampling | |
| CN110321872B (en) | Facial expression recognition method and device, computer equipment and readable storage medium | |
| CN114366047B (en) | Multi-task neural network pulse condition data processing method, system and terminal | |
| CN111967495A (en) | Classification recognition model construction method | |
| CN108122221A (en) | The dividing method and device of diffusion-weighted imaging image midbrain ischemic area | |
| CN114463605B (en) | Continuous learning image classification method and device based on deep learning | |
| CN111681233A (en) | US-CT image segmentation method, system and equipment based on deep neural network | |
| CN109102512A (en) | A kind of MRI brain tumor image partition method based on DBN neural network | |
| CN112149521B (en) | Palm print ROI extraction and enhancement method based on multitasking convolutional neural network | |
| CN111696119A (en) | Nuclear magnetic resonance image hippocampus segmentation method and device | |
| Korzhebin et al. | Comparison of combinations of data augmentation methods and transfer learning strategies in image classification used in convolution deep neural networks | |
| Monajemi et al. | Classification of HEp-2 cells using distributed dictionary learning | |
| Hashem et al. | Human identification using foot features | |
| CN112699907A (en) | Data fusion method, device and equipment | |
| CN111680132A (en) | Noise filtering and automatic classifying method for internet text information | |
| CN116385467A (en) | Cerebrovascular segmentation method based on self-supervision learning and related equipment | |
| Kumar et al. | White matter fiber analysis using kernel dictionary learning and sparsity priors | |
| CN110428012A (en) | Brain method for establishing network model, brain image classification method, device and electronic equipment | |
| CN113782121B (en) | Random grouping method, device, computer equipment and storage medium | |
| Yang et al. | Cascaded coarse-to-fine neural network for brain tumor segmentation | |
| US20230058500A1 (en) | Method and machine learning system to perform quantization of neural network | |
| CN109784406A (en) | A kind of user draws a portrait method, apparatus, readable storage medium storing program for executing and terminal device | |
| CN110634119A (en) | Method, device and computing equipment for segmenting vein blood vessel in magnetic sensitivity weighted image |
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 |