CN113647973B - Portable nondestructive testing device based on convolutional neural network - Google Patents

Portable nondestructive testing device based on convolutional neural network Download PDF

Info

Publication number
CN113647973B
CN113647973B CN202111204797.3A CN202111204797A CN113647973B CN 113647973 B CN113647973 B CN 113647973B CN 202111204797 A CN202111204797 A CN 202111204797A CN 113647973 B CN113647973 B CN 113647973B
Authority
CN
China
Prior art keywords
signal
data
prediction model
stool
shit
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
Application number
CN202111204797.3A
Other languages
Chinese (zh)
Other versions
CN113647973A (en
Inventor
张铁
黄泽铨
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
South China University of Technology SCUT
Original Assignee
South China University of Technology SCUT
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by South China University of Technology SCUT filed Critical South China University of Technology SCUT
Priority to CN202111204797.3A priority Critical patent/CN113647973B/en
Publication of CN113647973A publication Critical patent/CN113647973A/en
Application granted granted Critical
Publication of CN113647973B publication Critical patent/CN113647973B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B7/00—Instruments for auscultation
    • A61B7/02—Stethoscopes
    • A61B7/04—Electric stethoscopes
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/24—Detecting, measuring or recording bioelectric or biomagnetic signals of the body or parts thereof
    • A61B5/25—Bioelectric electrodes therefor
    • A61B5/251—Means for maintaining electrode contact with the body
    • A61B5/257—Means for maintaining electrode contact with the body using adhesive means, e.g. adhesive pads or tapes
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/24—Detecting, measuring or recording bioelectric or biomagnetic signals of the body or parts thereof
    • A61B5/316—Modalities, i.e. specific diagnostic methods
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/24—Detecting, measuring or recording bioelectric or biomagnetic signals of the body or parts thereof
    • A61B5/316—Modalities, i.e. specific diagnostic methods
    • A61B5/318—Heart-related electrical modalities, e.g. electrocardiography [ECG]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods

Landscapes

  • Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Public Health (AREA)
  • Animal Behavior & Ethology (AREA)
  • Theoretical Computer Science (AREA)
  • Veterinary Medicine (AREA)
  • Surgery (AREA)
  • Medical Informatics (AREA)
  • Heart & Thoracic Surgery (AREA)
  • Pathology (AREA)
  • Computational Linguistics (AREA)
  • Software Systems (AREA)
  • Mathematical Physics (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Computing Systems (AREA)
  • Evolutionary Computation (AREA)
  • Data Mining & Analysis (AREA)
  • Artificial Intelligence (AREA)
  • Acoustics & Sound (AREA)
  • Cardiology (AREA)
  • Measurement And Recording Of Electrical Phenomena And Electrical Characteristics Of The Living Body (AREA)

Abstract

The invention discloses a portable nondestructive testing device based on a convolutional neural network. The invention obtains real-time physiological data and stool labels of a plurality of subjects; preprocessing the acquired physiological data, forming a two-dimensional data set by the preprocessed physiological data, and dividing the two-dimensional data set into a training set and a testing set; inputting the training set and the shit labels corresponding to the physiological data in the training set into a convolutional neural network to train a shit early warning model, selecting the shit early warning model with the best prediction effect through the test set, and determining a final shit prediction model; and transplanting the data filtering algorithm and the final stool prediction model into an embeddable computer subsystem to predict whether the human body generates stool or not in real time. The invention collects the physiological data of the patient through the sensor, solves the problem that invasive monitoring is carried out when the physiological data is collected at present, and analyzes and processes the current signal in real time.

Description

Portable nondestructive testing device based on convolutional neural network
Technical Field
The invention relates to the field of signal detection and early warning, in particular to a portable nondestructive testing device based on a convolutional neural network.
Background
The disabled old people can have various complications after being bedridden for a long time, the life quality and the service life of the old people are seriously influenced, the intelligent excrement and urine monitoring technology and the processing function module based on individualized information are developed according to the long-term care and rehabilitation requirements of the disabled old people from clinical requirements, multi-scene application facing hospitals, nursing institutions, families and the like is established, the problem that the disabled old people are difficult to care for a long time can be solved, the social cost is saved, and the practical value and the social benefit are great.
On the other hand, the physiological signals for monitoring the excrement and urine are very weak, generally only about 50 microvolts, the amplitude range is 5 microvolts to 100 microvolts, and the frequency range is generally 0.5 Hz to 35 Hz. Because signals have characteristics of weak level, narrow band and the like, the signals are difficult to extract under strong electromyographic noise and power frequency interference, and complex processing on the signals is more difficult.
Among the urination detection and transmission devices (P2002-369810) is a urination detection and transmission device for constantly detecting the state of urine in the bladder from the inside of the bladder. The urination detecting and transmitting device is disposed inside the urinary bladder through the urethra, and if the detection result indicates that the stored urine amount reaches the urination level, the detection signal is launched to the side of the urination reporting device through the transmitting antenna. This method is invasive and highly invasive.
The existing excrement signal acquisition equipment has the following defects:
(1) various excrement signal acquisition equipment is high in price, large in size, complex to use and the like, so that the excrement signal acquisition equipment cannot be widely applied;
(2) the existing excrement and urine signal acquisition equipment adopts an invasive monitoring scheme, is difficult to correspond to various medical conditions and is difficult to observe in real time.
(3) The acquired data of the existing portable signal acquisition equipment cannot correspond to a server interface, and the electroencephalogram signals cannot be accurately acquired in real time.
Disclosure of Invention
The invention aims to overcome the defects of the prior art and provide a device for acquiring physiological signals of patients in real time, processing and analyzing the signals in real time and giving early warning to shit signals.
The purpose of the invention is realized by at least one of the following technical solutions.
A stool nondestructive testing device based on a convolutional neural network comprises a signal acquisition module, a signal preprocessing module, a stool prediction model training module and an embeddable computer subsystem;
the signal acquisition module acquires physiological signal data of a patient in real time and inputs the physiological signal data into the signal preprocessing module for preprocessing, the signal preprocessing module inputs the preprocessed physiological signal data into the binary prediction model training module, a final binary prediction model is obtained through training and is input into the embeddable computer subsystem, a data filtering algorithm and a final binary prediction model are transplanted on the embeddable computer subsystem, the physiological signal data sent by the signal acquisition module are received in real time, the physiological signal data are preprocessed through the data filtering algorithm and then input into the final binary prediction model for prediction, and a binary early warning signal is obtained.
Further, the signal acquisition module comprises:
the electronic stethoscope is used for measuring the bowel sounding signals of the patient; the electronic stethoscope comprises a stethoscope head, a sound guide tube and an ear hook;
a gastric electrical sensor for measuring a gastric electrical signal of the patient; the stomach electric sensor comprises a measuring electrode and a transmission interface;
the electrocardio sensor is used for measuring electrocardiosignals of a patient; the electrocardio sensor comprises a measuring electrode and a transmission interface;
the physiological signal data comprises intestinal tract bowel sounding signals, stomach electric signals and electrocardiosignals.
Furthermore, the real-time physiological signal data are acquired in the morning, the noon and the evening respectively in different situations, including fasting state and non-fasting state, and with and without defecation;
the real-time physiological signal data acquisition process is as follows:
intestinal borygmus signals, namely, tightly attaching an auscultation head to the skin, and respectively collecting the borygmus signals from the upper right quadrant, the upper left quadrant and the lower left quadrant of the abdomen of a subject;
the electrocardiosignal, the measuring electrode is clung to the chest part of the testee and collects the electrocardiosignal;
and (3) a stomach electric signal, namely tightly attaching a measuring electrode to the stomach of the subject and collecting the stomach electric signal.
Further, in the signal preprocessing module, the physiological signal data acquired by the signal acquisition module is preprocessed through a data filtering algorithm, including noise reduction, filtering and normalization processing, specifically as follows:
adopting wavelet to reduce noise of physiological signal data, selecting wavelet function to decompose physiological signal data, and setting the original signal expression of physiological signal data as:
Figure DEST_PATH_IMAGE001
wherein the content of the first and second substances,
Figure DEST_PATH_IMAGE002
representing signals
Figure DEST_PATH_IMAGE003
The time in (1) is (a),
Figure DEST_PATH_IMAGE004
by a set of functions
Figure DEST_PATH_IMAGE005
Expanding, namely representing a space formed by the piecewise constant functions of which all the discontinuous points are only in the integer set;
Figure DEST_PATH_IMAGE006
is a non-negative integer representing space
Figure 200390DEST_PATH_IMAGE004
The highest resolution of (c);
Figure DEST_PATH_IMAGE007
is a set of integers
Figure DEST_PATH_IMAGE008
Element(s) in (1), representing translation parameters;
Figure DEST_PATH_IMAGE009
is a signal
Figure 507743DEST_PATH_IMAGE003
Decomposed in space
Figure 490743DEST_PATH_IMAGE004
To (1) a
Figure DEST_PATH_IMAGE010
Wavelet packet coefficients;
Figure DEST_PATH_IMAGE011
representing signals
Figure 43209DEST_PATH_IMAGE003
In space
Figure 326423DEST_PATH_IMAGE004
To (3) is performed.
Figure DEST_PATH_IMAGE012
Is a scale function, defined as
Figure DEST_PATH_IMAGE013
Figure DEST_PATH_IMAGE014
Is that
Figure 539099DEST_PATH_IMAGE011
Can be decomposed into:
Figure DEST_PATH_IMAGE015
then there are:
Figure DEST_PATH_IMAGE016
then
Figure DEST_PATH_IMAGE017
Can be expressed as:
Figure DEST_PATH_IMAGE018
wherein the content of the first and second substances,
Figure DEST_PATH_IMAGE019
is a signal
Figure 635493DEST_PATH_IMAGE003
Decomposed in space
Figure 303235DEST_PATH_IMAGE004
Figure 303235DEST_PATH_IMAGE004
2 nd
Figure 690222DEST_PATH_IMAGE010
Wavelet packet coefficients;
Figure DEST_PATH_IMAGE020
is a signal
Figure 886849DEST_PATH_IMAGE003
Decomposed in space
Figure 110020DEST_PATH_IMAGE004
To (1) a
Figure DEST_PATH_IMAGE021
Wavelet packet coefficients;
Figure DEST_PATH_IMAGE022
is formed by
Figure DEST_PATH_IMAGE023
Stretching the linear combination of (1);
Figure DEST_PATH_IMAGE024
representing signals
Figure 191239DEST_PATH_IMAGE011
Is/are as follows
Figure 183466DEST_PATH_IMAGE022
A component;
Figure DEST_PATH_IMAGE025
representing signals
Figure 550993DEST_PATH_IMAGE011
Is/are as follows
Figure DEST_PATH_IMAGE026
A component;
Figure DEST_PATH_IMAGE027
is a wavelet function expressed as:
Figure DEST_PATH_IMAGE028
the decomposition process being followed by
Figure DEST_PATH_IMAGE029
Substitution
Figure DEST_PATH_IMAGE030
Continuing handle
Figure 605668DEST_PATH_IMAGE029
Is decomposed into
Figure DEST_PATH_IMAGE031
Obtaining:
Figure 818475DEST_PATH_IMAGE032
method for constructing FIR filter by adopting window function design, L-order FIR filter input
Figure DEST_PATH_IMAGE033
And
Figure DEST_PATH_IMAGE034
the output relational expression is as follows:
Figure DEST_PATH_IMAGE035
wherein the content of the first and second substances,
Figure DEST_PATH_IMAGE036
respectively representing an input signal time and an output signal time,
Figure DEST_PATH_IMAGE037
is the unit sample response of the finite impulse response filter;
the frequency response of the filter is:
Figure DEST_PATH_IMAGE038
Figure DEST_PATH_IMAGE039
representing the signal frequency;
after wavelet transformation and filtering processing, normalization processing is carried out on physiological signal data, and a calculation formula is as follows:
Figure DEST_PATH_IMAGE040
wherein the content of the first and second substances,
Figure DEST_PATH_IMAGE041
and
Figure DEST_PATH_IMAGE042
respectively representing signals
Figure 381654DEST_PATH_IMAGE003
Mean and standard deviation at all times;
and truncating the physiological signal data by using a time window to enable the dimensionality of each training signal data matrix to be two-dimensional data, and dividing the two-dimensional data set into a training set and a test set.
Furthermore, in the signal preprocessing module, a window function is selected according to the specification and parameter requirements of the FIR filter; calculating the size of the transition region according to the required passband edge frequency and stopband starting frequency, thereby calculating the length of the window function; and finally, calculating the unit impulse response of the required filter according to the window function and the unit impulse response of the ideal filter.
Further, in the stool prediction model training module, the training set and the stool labels corresponding to the physiological signal data in the training set are input into the convolutional neural network to train the stool prediction model, and the stool prediction model with the best prediction effect is selected through the test set to determine the final stool prediction model;
the convolutional neural network comprises an input layer, a plurality of convolutional layers, a plurality of pooling layers, a plurality of full-connection layers and an output layer which are connected in sequence;
wherein, in the input layer, the input value is the two-dimensional data generated after the data preprocessing in the step S2, and the two-dimensional data generated after the data preprocessing is transmitted to the convolutional layer, and each sample X has one dimension of
Figure DEST_PATH_IMAGE043
A tensor matrix of, wherein
Figure DEST_PATH_IMAGE044
Represents the number of channels of the electrical signal,
Figure DEST_PATH_IMAGE045
represents a time step;
the convolution layer receives data of an input layer and uses small-batch normalization to prevent gradient disappearance and over-fitting problems, the output after being filtered by each convolution filter is a two-dimensional tensor, and the two-dimensional tensor is transmitted to the pooling layer;
the down-sampling in the Pooling layer is carried out by Max-Pooling, the purpose is to carry out dimensionality reduction on the input tensor of the convolution layer, reduce the number of parameters under the condition of not losing characteristic information so as to reduce the calculated amount, flatten the multidimensional input tensor into a one-dimensional tensor form, and transmit the tensor to the full-connection layer so as to realize the transition from the convolution layer to the full-connection layer;
the fully connected layer plays a role of a 'classifier' in the whole convolutional neural network and is used for mapping the learned 'distributed feature representation' to a sample mark space; transport of full connecting layerIs output as a vector
Figure DEST_PATH_IMAGE046
And is and
Figure DEST_PATH_IMAGE047
；
Figure DEST_PATH_IMAGE048
for each neuron, i.e. predictive probability
Figure DEST_PATH_IMAGE049
And
Figure DEST_PATH_IMAGE050
respectively representing the probability of not detecting the stool signal and the probability of detecting the stool signal; therefore, the expression of the detection category of the binary signal is as follows:
Figure DEST_PATH_IMAGE051
wherein X represents an input sample, E represents a classifier obtained by learning, and the prediction probability of the full connection layer is used as the output of the binary prediction model.
Further, in a stool prediction model training module, obtaining a training set and stool labels corresponding to physiological signal data in the training set; the stool labels are divided into two types of stool labels with stool meaning and stool meaning;
using the shit labels corresponding to physiological signal data in a training set and a training set as the input of a shit prediction model, inputting the shit labels into the constructed shit prediction model for training, using shit signal prediction of the training set as the output of the shit prediction model to obtain the trained shit prediction model, and then selecting the shit prediction model with the best prediction effect through a test set;
and (3) taking the test set as test data, taking the excrement signal prediction of the test set as the output of the test set in an excrement prediction model, comparing the output with excrement signal labels corresponding to the physiological data of the actual test set to obtain the prediction accuracy of the test set, retraining by a method of setting a threshold value if the accuracy does not exceed the threshold value until the accuracy reaches the threshold value, and determining the final excrement prediction model.
Further, the embeddable computer subsystem adopts a Jetson TX2 computer subsystem.
Further, the final prediction model of the stool and urine obtained from the data filtering algorithm in the signal preprocessing module and the training module of the stool and urine prediction model is transplanted into the embeddable computer subsystem to predict whether the human body generates stool and urine in real time, which is specifically as follows:
the real-time physiological signal data of the patient are acquired through the signal acquisition module, preprocessed through a data filtering algorithm and finally input into a final stool prediction model for prediction, and a stool early warning signal is obtained.
Further, the detection process of the convolutional neural network-based two-dimensional nondestructive detection device comprises the following steps:
s1, acquiring real-time physiological signal data and shit labels of a plurality of subjects;
s2, preprocessing the acquired physiological signal data through a data filtering algorithm, forming a two-dimensional data set by the preprocessed physiological signal data, and dividing the two-dimensional data set into a training set and a test set;
s3, inputting the training set and the shit labels corresponding to the physiological data in the training set into a convolutional neural network to train a shit prediction model, selecting the shit prediction model with the best prediction effect through the test set, and determining a final shit prediction model;
and S4, transplanting the data filtering algorithm and the final stool prediction model into an embeddable computer subsystem to predict whether the human body generates stool intentions in real time.
Compared with the prior art, the invention has the advantages that:
the invention provides a measuring method for acquiring physiological signals of a patient in real time, and can process and analyze the signals in real time and obtain stool and urine signal early warning.
Drawings
Fig. 1 is a schematic diagram of a system structure of a convolutional neural network-based nondestructive testing apparatus according to an embodiment of the present invention.
FIG. 2a is a schematic diagram of an electronic stethoscope; figure 2b is a schematic diagram of an electrocardiograph sensor and a pyroelectric sensor.
Fig. 3 is a schematic diagram of a network structure of a convolutional neural network according to an embodiment of the present invention.
FIG. 4 is a diagram illustrating a data structure of a convolution binary signal according to an embodiment of the present invention.
Fig. 5 is a schematic diagram of CNN training data allocation in the embodiment of the present invention.
FIG. 6 is a schematic block diagram of a portable data filtering algorithm and a binary prediction model according to an embodiment of the present invention.
Detailed Description
The present invention will be described in further detail with reference to examples and drawings, but the present invention is not limited thereto.
Example (b):
a stool nondestructive testing device based on a convolutional neural network is shown in figure 1 and comprises a signal acquisition module, a signal preprocessing module, a stool prediction model training module and an embeddable computer subsystem;
the signal acquisition module acquires physiological signal data of a patient in real time and inputs the physiological signal data into the signal preprocessing module for preprocessing, the signal preprocessing module inputs the preprocessed physiological signal data into the binary prediction model training module, a final binary prediction model is obtained through training and is input into the embeddable computer subsystem, a data filtering algorithm and a final binary prediction model are transplanted on the embeddable computer subsystem, the physiological signal data sent by the signal acquisition module are received in real time, the physiological signal data are preprocessed through the data filtering algorithm and then input into the final binary prediction model for prediction, and a binary early warning signal is obtained.
Further, the signal acquisition module comprises:
the electronic stethoscope is used for measuring the bowel sounding signals of the patient; as shown in fig. 2a, the electronic stethoscope comprises a stethoscope head 1, a sound guide tube 2 and an ear hook 3; the intestinal bowel sounding signal is a gas-water passing sound generated between gas and liquid in an intestinal tube during intestinal peristalsis and can be used as an important index for detecting intestinal diseases; bowel sounds of different characteristics or different types, representing bowel movements in different pathological states;
a gastric electrical sensor for measuring a gastric electrical signal of the patient; as shown in fig. 2b, the gastric electrical sensor comprises a measuring electrode 4 and a transmission interface 5; the stomach electric signals are electric signals generated during gastrointestinal peristalsis, and the stomach electric signals with different characteristics reflect the movement state of the stomach and the intestine;
the electrocardio sensor is used for measuring electrocardiosignals of a patient; as shown in fig. 2b, the electrocardiograph sensor comprises a measuring electrode 4 and a transmission interface 5;
in this embodiment, the signal acquisition module comprises a stomach electric sensor, an electrocardio sensor and a 3M series Littmann3200 type electronic stethoscope configured by a Biosignals series 4-channel physiological recording monitor
The physiological signal data comprises intestinal tract bowel sounding signals, stomach electric signals and electrocardiosignals.
Furthermore, the real-time physiological signal data are acquired in the morning, the noon and the evening respectively in different situations, including fasting state and non-fasting state, and with and without defecation;
the subject needs to be a person with no history of intestinal tract diseases in nearly three months and normal gastrointestinal functions; then, arranging the tested volunteer to wear the gastric electricity sensor, the electrocardio sensor and the electronic stethoscope, and collecting data in a lying posture;
the real-time physiological signal data acquisition process is as follows:
intestinal bowel sounding signals are acquired from the right upper quadrant (ascending colon), the left upper quadrant (descending colon) and the left lower quadrant (sigmoid colon) of the abdomen of a subject by closely attaching an auscultatory head to the skin;
the electrocardiosignal, the measuring electrode is clung to the chest part of the testee and collects the electrocardiosignal;
and (3) a stomach electric signal, namely tightly attaching a measuring electrode to the stomach of the subject and collecting the stomach electric signal.
Further, in the signal preprocessing module, the physiological signal data acquired by the signal acquisition module is preprocessed through a data filtering algorithm, including noise reduction, filtering and normalization processing, specifically as follows:
adopting wavelet to reduce noise of physiological signal data, selecting wavelet function to decompose physiological signal data, and setting the original signal expression of physiological signal data as:
Figure 654503DEST_PATH_IMAGE001
wherein the content of the first and second substances,
Figure 852267DEST_PATH_IMAGE002
which represents the time in the signal or signals,
Figure 399923DEST_PATH_IMAGE004
by a set of functions
Figure 366742DEST_PATH_IMAGE005
Expanding, namely representing a space formed by the piecewise constant functions of which all the discontinuous points are only in the integer set;
Figure 357962DEST_PATH_IMAGE006
is a non-negative integer representing space
Figure 777443DEST_PATH_IMAGE004
The highest resolution of (c);
Figure 128789DEST_PATH_IMAGE007
is a set of integers
Figure 950115DEST_PATH_IMAGE008
Element(s) in (1), representing translation parameters;
Figure 610772DEST_PATH_IMAGE009
is a signal
Figure 783128DEST_PATH_IMAGE003
Decomposed in space
Figure 672586DEST_PATH_IMAGE004
To (1) a
Figure 613997DEST_PATH_IMAGE010
Wavelet packet coefficients;
Figure 681442DEST_PATH_IMAGE011
representing signals
Figure 75514DEST_PATH_IMAGE003
In space
Figure 768663DEST_PATH_IMAGE004
To (3) is performed.
Figure 299002DEST_PATH_IMAGE012
Is a scale function, defined as
Figure 301462DEST_PATH_IMAGE013
Figure 182830DEST_PATH_IMAGE014
Is that
Figure 679671DEST_PATH_IMAGE011
Can be decomposed into:
Figure 330095DEST_PATH_IMAGE015
then, there are:
Figure 739342DEST_PATH_IMAGE016
then
Figure 108006DEST_PATH_IMAGE017
Can be expressed as:
Figure 142958DEST_PATH_IMAGE018
wherein the content of the first and second substances,
Figure 913468DEST_PATH_IMAGE019
is a signal
Figure 992151DEST_PATH_IMAGE003
Decomposed in space
Figure 848112DEST_PATH_IMAGE004
Figure 848112DEST_PATH_IMAGE004
2 nd
Figure 686755DEST_PATH_IMAGE010
Wavelet packet coefficients;
Figure 46192DEST_PATH_IMAGE020
is a signal
Figure 62821DEST_PATH_IMAGE003
Decomposed in space
Figure 140498DEST_PATH_IMAGE004
To (1) a
Figure 782832DEST_PATH_IMAGE021
Wavelet packet coefficients;
Figure 527934DEST_PATH_IMAGE022
is formed by
Figure 964732DEST_PATH_IMAGE023
Stretching the linear combination of (1);
Figure 778973DEST_PATH_IMAGE024
representing signals
Figure 959419DEST_PATH_IMAGE011
Is/are as follows
Figure 293448DEST_PATH_IMAGE022
A component;
Figure 901147DEST_PATH_IMAGE025
representing signals
Figure 704149DEST_PATH_IMAGE011
Is/are as follows
Figure 422706DEST_PATH_IMAGE026
A component;
Figure 876821DEST_PATH_IMAGE027
is a wavelet function expressed as:
Figure 655421DEST_PATH_IMAGE028
the decomposition process being followed by
Figure 444255DEST_PATH_IMAGE029
Substitution
Figure 966503DEST_PATH_IMAGE030
Continuing handle
Figure 275124DEST_PATH_IMAGE029
Is decomposed into
Figure 490205DEST_PATH_IMAGE031
Obtaining:
Figure DEST_PATH_IMAGE052
；
method for constructing FIR filter by adopting window function design, L-order FIR filter input
Figure 459343DEST_PATH_IMAGE033
And
Figure 785282DEST_PATH_IMAGE034
the output relational expression is as follows:
Figure 948410DEST_PATH_IMAGE035
wherein the content of the first and second substances,
Figure 318081DEST_PATH_IMAGE036
respectively representing an input signal time and an output signal time,
Figure 832238DEST_PATH_IMAGE037
is the unit sample response of the finite impulse response filter;
the frequency response of the filter is:
Figure 696289DEST_PATH_IMAGE038
Figure 713924DEST_PATH_IMAGE039
representing the signal frequency.
After wavelet transformation and filtering processing, normalization processing is carried out on physiological signal data, and a calculation formula is as follows:
Figure 21540DEST_PATH_IMAGE040
wherein the content of the first and second substances,
Figure 757414DEST_PATH_IMAGE041
and
Figure 425156DEST_PATH_IMAGE042
respectively representing signals
Figure 297297DEST_PATH_IMAGE003
Mean and standard deviation at all times;
and truncating the physiological signal data by using a time window to enable the dimensionality of each training signal data matrix to be two-dimensional data, and dividing the two-dimensional data set into a training set and a test set.
In this embodiment, the physiological signal data is truncated with a time window of 0-600ms, such that each training signal data matrix has dimensions of
Figure DEST_PATH_IMAGE053
. Since the ratio of positive and negative samples of the training data set is 1:5, data imbalance is easily caused, and therefore all positive samples are copied by 4 times to achieve positive and negative sample balance.
In the signal preprocessing module, a window function is selected according to the specification and parameter requirements of the FIR filter; calculating the size of the transition region according to the required passband edge frequency and stopband starting frequency, thereby calculating the length of the window function; and finally, calculating the unit impulse response of the required filter according to the window function and the unit impulse response of the ideal filter.
In the stool prediction model training module, a training set and stool labels corresponding to physiological signal data in the training set are input into a convolutional neural network to train a stool prediction model, the stool prediction model with the best prediction effect is selected through a test set, and a final stool prediction model is determined;
as shown in fig. 3, in the present embodiment, an 11-layer convolutional neural network is adopted, which includes an input layer, a convolutional layer, a BN layer, a max pooling layer, a full link layer, and an output layer;
in this embodiment, as shown in fig. 4, the constructed convolutional neural network structure sequentially includes:
Figure DEST_PATH_IMAGE054
in the input layer, the input value is two-dimensional data generated by the physiological signal acquisition system after the data preprocessing in step S2, and each sample X has one dimension of
Figure 212032DEST_PATH_IMAGE043
The tensor matrix of (a), in this embodiment,
Figure 435203DEST_PATH_IMAGE044
3, the number of the electric signal channels is shown,
Figure 391789DEST_PATH_IMAGE045
at 651, a step of time is indicated,
and (3) constructing the 11-layer convolutional neural network model by using a machine learning library, then taking a two-dimensional tensor matrix corresponding to the bowel sound, the gastric electricity and the electrocardio time sequence physiological data input by the physiological signal acquisition system as the input of a shit prediction model, and transmitting the input to a final full-connection layer to output a result to show whether the result is convenient or not.
The fully-connected layer acts as a 'classifier' in the overall convolutional neural network for mapping the learned 'distributed feature representation' to the sample label space. The output of the fully-connected layer is a vector
Figure 384016DEST_PATH_IMAGE046
And is and
Figure 282702DEST_PATH_IMAGE047
；
Figure 462010DEST_PATH_IMAGE048
for each neuron, i.e. predictive probability
Figure 720822DEST_PATH_IMAGE049
And
Figure 833135DEST_PATH_IMAGE050
respectively representing the probability of not detecting the stool signal and the probability of detecting the stool signal; therefore, the expression of the detection category of the binary signal is as follows:
Figure 637143DEST_PATH_IMAGE051
wherein X represents an input sample, E represents a classifier obtained by learning, and the prediction probability of the full connection layer is used as the output of the binary prediction model.
In a stool prediction model training module, acquiring a training set and stool labels corresponding to physiological signal data in the training set; the stool labels are divided into two types of stool labels with stool meaning and stool meaning;
using the shit labels corresponding to physiological signal data in a training set and a training set as the input of a shit prediction model, inputting the shit labels into the constructed shit prediction model for training, using shit signal prediction of the training set as the output of the shit prediction model to obtain the trained shit prediction model, and then selecting the shit prediction model with the best prediction effect through a test set;
and taking the test set as test data, taking the shit signal prediction of the test set as the output of the test set in a shit prediction model, comparing the output with shit signal labels corresponding to the physiological data of the actual test set to obtain the prediction accuracy of the test set, and re-training until the accuracy reaches a threshold value by setting the threshold value to be 95% if the accuracy does not exceed the threshold value, so as to determine the final shit prediction model.
In this example, as shown in fig. 5, a total of 60 training and testing rounds were performed, and the total ecg beats for training and testing purposes were assigned as follows: early warning training data on stool and urine
Figure 834906DEST_PATH_IMAGE055
In use therein
Figure DEST_PATH_IMAGE056
To verify the algorithm.
In this embodiment, the embeddable computer subsystem adopts a Jetson TX2 computer subsystem, and the processor of the Nvidia Jetson TX2 computer subsystem takes 256 Nvidia cuda and 64-bit cpu as cores and is equipped with a data processing platform to process the acquired physiological electrical signals in real time. The processor functions include:
displaying: NVIDIA Jetson TX2 connects to the display through an HDMI interface.
Controlling: NVIDIA Jetson TX2 communicates with the peripheral subsystem through a uart serial port, usb interface.
Power supply: NVIDIA Jetson TX2 is connected to a power supply via a three-hole plug.
Starting up: the NVIDIA Jetson TX2 has four red buttons, wherein the lower face of the POWER Button is marked with POWER BTN font, and the following POWER Button are respectively a Force Recovery Button, a User Defined Button and a Reset Button.
As shown in fig. 6, the final stool prediction model obtained from the data filtering algorithm in the signal preprocessing module and the stool prediction model training module is transplanted into the embeddable computer subsystem to predict whether the human body generates stool intentions in real time, which is specifically as follows:
the real-time physiological signal data of the patient are acquired through the signal acquisition module, preprocessed through a data filtering algorithm and finally input into a final stool prediction model for prediction, and a stool early warning signal is obtained.
The convolutional neural network-based two-dimensional nondestructive testing process of the embodiment comprises the following steps:
s1, acquiring real-time physiological signal data and shit labels of a plurality of subjects;
s2, preprocessing the acquired physiological signal data through a data filtering algorithm, forming a two-dimensional data set by the preprocessed physiological signal data, and dividing the two-dimensional data set into a training set and a test set;
s3, inputting the training set and the shit labels corresponding to the physiological data in the training set into a convolutional neural network to train a shit prediction model, selecting the shit prediction model with the best prediction effect through the test set, and determining a final shit prediction model;
and S4, transplanting the data filtering algorithm and the final stool prediction model into an embeddable computer subsystem to predict whether the human body generates stool intentions in real time.
The above embodiments are preferred embodiments of the present invention, but the present invention is not limited to the above embodiments, and any other changes, modifications, substitutions, combinations, and simplifications which do not depart from the spirit and principle of the present invention should be construed as equivalents and all such changes, modifications, substitutions, combinations, and simplifications are intended to be included in the scope of the present invention.

Claims (5)

1. A stool nondestructive testing device based on a convolutional neural network is characterized by comprising a signal acquisition module, a signal preprocessing module, a stool prediction model training module and an embeddable computer subsystem;
the signal acquisition module acquires physiological signal data of a patient in real time and inputs the physiological signal data into the signal preprocessing module for preprocessing, the signal preprocessing module inputs the preprocessed physiological signal data into the binary prediction model training module, a final binary prediction model is obtained through training and is input into the embeddable computer subsystem, a data filtering algorithm and a final binary prediction model are transplanted on the embeddable computer subsystem, the physiological signal data sent by the signal acquisition module are received in real time, the physiological signal data are preprocessed through the data filtering algorithm and then input into the final binary prediction model for prediction, and a binary early warning signal is obtained;
the signal acquisition module includes:
the electronic stethoscope is used for measuring the bowel sounding signals of the patient; the electronic stethoscope comprises a stethoscope head, a sound guide tube and an ear hook;
a gastric electrical sensor for measuring a gastric electrical signal of the patient; the stomach electric sensor comprises a measuring electrode and a transmission interface;
the electrocardio sensor is used for measuring electrocardiosignals of a patient; the electrocardio sensor comprises a measuring electrode and a transmission interface;
the physiological signal data comprise intestinal tract bowel sounding signals, stomach electric signals and electrocardiosignals;
in the signal preprocessing module, the physiological signal data acquired by the signal acquisition module is preprocessed through a data filtering algorithm, wherein the preprocessing comprises noise reduction, filtering and normalization processing, and the preprocessing specifically comprises the following steps:
adopting wavelet to reduce noise of physiological signal data, selecting wavelet function to decompose physiological signal data, and setting the original signal expression of physiological signal data as:
Figure FDA0003504066920000011
wherein x represents the time in the signal f (x), VjBy a set of functions
Figure FDA0003504066920000012
Figure FDA0003504066920000013
Expanding, namely representing a space formed by the piecewise constant functions of which all the discontinuous points are only in the integer set; j is a non-negative integer representing the space VjThe highest resolution of (c); k is an element in the set of integers Z, representing a translation parameter; a isk jFor signal f (x) decomposition in space VjThe kth wavelet packet coefficient of (1); f. ofj(x) Representing the signal f (x) in space VjMapping of (2);
φ (x) is a scaling function defined as:
Figure FDA0003504066920000014
fjis fj(x) For short, decompose into: f. ofj＝ωj-1+fj-1Then, there are:
Figure FDA0003504066920000021
Figure FDA0003504066920000022
then m isk j-1、nk j-1Can be expressed as:
Figure FDA0003504066920000023
Figure FDA0003504066920000024
wherein, a2k jFor signal f (x) decomposition in space VjThe 2k th wavelet packet coefficient of (a); a is2k+1 jFor signal f (x) decomposition in space VjThe 2k +1 th wavelet packet coefficient; wj-1Is composed of { psi (2)j-1x-k), k ∈ Z }; omegaj-1Representing a signal fj(x) W of (2)j-1A component; f. ofj-1(x) Representing a signal fj(x) V ofj-1A component;
ψ (x) is a wavelet function expressed as:
ψ(x)＝φ(2x)-φ(2x-1)；
the decomposition process then replaces j with j-1 and continues to decompose j-1 to ωj-2+fj-2Obtaining:
fj＝ωj-1+ωj-2+…+ω0+f0；
the FIR filter is constructed by adopting a window function design method, and the relational expression of the input x (d) and the output y (n) of the L-order FIR filter is as follows:
Figure FDA0003504066920000025
wherein d and n respectively represent input signal time and output signal time, and h (d) is unit sampling response of the finite impulse response filter;
the frequency response of the filter is:
Figure FDA0003504066920000026
ω represents the signal frequency;
after wavelet transformation and filtering processing, normalization processing is carried out on physiological signal data, and a calculation formula is as follows:
Figure FDA0003504066920000027
wherein the content of the first and second substances,
Figure FDA0003504066920000028
and σ represents the mean and standard deviation of the signal f (x) at all times, respectively;
cutting off the physiological signal data by using a time window to enable the dimensionality of each training signal data matrix to be two-dimensional data, and dividing a two-dimensional data set into a training set and a test set;
in the stool prediction model training module, a training set and stool labels corresponding to physiological signal data in the training set are input into a convolutional neural network to train a stool prediction model, the stool prediction model with the best prediction effect is selected through a test set, and a final stool prediction model is determined;
the convolutional neural network comprises an input layer, a plurality of convolutional layers, a plurality of pooling layers, a plurality of full-connection layers and an output layer which are connected in sequence;
wherein, in the input layer, the input value is two-dimensional data generated after data preprocessing, and the two-dimensional data generated after data preprocessing is transmitted to the convolution layer, and each sample X has one dimension of [ N [ ]c×Nt]A tensor matrix of (2), wherein NcRepresenting the number of channels of the electrical signal, NtRepresents a time step;
the convolution layer receives data of an input layer and uses small-batch normalization to prevent gradient disappearance and over-fitting problems, the output after being filtered by each convolution filter is a two-dimensional tensor, and the two-dimensional tensor is transmitted to the pooling layer;
the down-sampling in the Pooling layer is carried out by Max-Pooling, the purpose is to carry out dimensionality reduction on the input tensor of the convolution layer, reduce the number of parameters under the condition of not losing characteristic information so as to reduce the calculated amount, flatten the multidimensional input tensor into a one-dimensional tensor form, and transmit the tensor to the full-connection layer so as to realize the transition from the convolution layer to the full-connection layer;
the fully-connected layer acts as a 'classifier' in the whole convolutional neural network and is used for learningThe 'distributed feature representation' of the object is mapped to the sample mark space; the output of the fully connected layer is vector a, and a ═ a0,a1]T；ai∈[0,1]For each neuron, i.e. a0And a1Respectively representing the probability of not detecting the stool signal and the probability of detecting the stool signal; therefore, the expression of the detection category of the binary signal is as follows:
Figure FDA0003504066920000031
wherein X represents an input sample, E represents a classifier obtained by learning, and the prediction probability of the full-connection layer is used as the output of the binary prediction model;
the real-time physiological signal data are acquired in the morning, at noon and at night respectively in different situations, including fasting state and non-fasting state, and defecation;
the real-time physiological signal data acquisition process is as follows:
intestinal borygmus signals, namely, tightly attaching an auscultation head to the skin, and respectively collecting the borygmus signals from the upper right quadrant, the upper left quadrant and the lower left quadrant of the abdomen of a subject;
the electrocardiosignal, the measuring electrode is clung to the chest part of the testee and collects the electrocardiosignal;
the stomach electric signal, the measuring electrode is tightly attached to the stomach of the subject and the stomach electric signal is collected;
in a stool prediction model training module, acquiring a training set and stool labels corresponding to physiological signal data in the training set; the stool labels are divided into two types of stool labels with stool meaning and stool meaning;
using the shit labels corresponding to physiological signal data in a training set and a training set as the input of a shit prediction model, inputting the shit labels into the constructed shit prediction model for training, using shit signal prediction of the training set as the output of the shit prediction model to obtain the trained shit prediction model, and then selecting the shit prediction model with the best prediction effect through a test set;
and (3) taking the test set as test data, taking the excrement signal prediction of the test set as the output of the test set in an excrement prediction model, comparing the output with excrement signal labels corresponding to the physiological data of the actual test set to obtain the prediction accuracy of the test set, retraining by a method of setting a threshold value if the accuracy does not exceed the threshold value until the accuracy reaches the threshold value, and determining the final excrement prediction model.
2. The convolutional neural network-based nondestructive testing apparatus according to claim 1, wherein in the signal preprocessing module, the window function is selected according to the specification and parameter requirements of the FIR filter; calculating the size of the transition region according to the required passband edge frequency and stopband starting frequency, thereby calculating the length of the window function; and finally, calculating the unit impulse response of the required filter according to the window function and the unit impulse response of the ideal filter.
3. The convolutional neural network-based nondestructive testing device as claimed in claim 1, wherein the embeddable computer subsystem is a Jetson TX2 computer subsystem.
4. The convolutional neural network-based binary nondestructive testing device of claim 1, wherein a final binary prediction model obtained from the data filtering algorithm in the signal preprocessing module and the binary prediction model training module is transplanted into an embeddable computer subsystem to predict whether a human body generates a convenient meaning in real time, specifically as follows:
the real-time physiological signal data of the patient are acquired through the signal acquisition module, preprocessed through a data filtering algorithm and finally input into a final stool prediction model for prediction, and a stool early warning signal is obtained.
5. The convolutional neural network-based binary nondestructive testing device as claimed in any one of claims 1 to 4, wherein the specific testing process comprises the following steps:
s1, acquiring real-time physiological signal data and shit labels of a plurality of subjects;
s2, preprocessing the acquired physiological signal data through a data filtering algorithm, forming a two-dimensional data set by the preprocessed physiological signal data, and dividing the two-dimensional data set into a training set and a test set;
s3, inputting the training set and the shit labels corresponding to the physiological data in the training set into a convolutional neural network to train a shit prediction model, selecting the shit prediction model with the best prediction effect through the test set, and determining a final shit prediction model;
and S4, transplanting the data filtering algorithm and the final stool prediction model into an embeddable computer subsystem to predict whether the human body generates stool intentions in real time.
CN202111204797.3A 2021-10-15 2021-10-15 Portable nondestructive testing device based on convolutional neural network Active CN113647973B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202111204797.3A CN113647973B (en) 2021-10-15 2021-10-15 Portable nondestructive testing device based on convolutional neural network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202111204797.3A CN113647973B (en) 2021-10-15 2021-10-15 Portable nondestructive testing device based on convolutional neural network

Publications (2)

Publication Number Publication Date
CN113647973A CN113647973A (en) 2021-11-16
CN113647973B true CN113647973B (en) 2022-04-22

Family

ID=78494525

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202111204797.3A Active CN113647973B (en) 2021-10-15 2021-10-15 Portable nondestructive testing device based on convolutional neural network

Country Status (1)

Country Link
CN (1) CN113647973B (en)

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113786205A (en) * 2021-11-17 2021-12-14 华南理工大学 Portable early warning system and method based on convolutional network and long-short term memory network

Family Cites Families (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2014039404A1 (en) * 2012-09-07 2014-03-13 The Regents Of The University Of California Multisensor wireless abdominal monitoring apparatus, systems, and methods
US11826170B2 (en) * 2012-10-10 2023-11-28 G-Tech Medical, Inc. Artificial intelligence models for wireless patch data acquisition for gastrointestinal electrodiagnostics
CN103654804B (en) * 2013-12-25 2015-04-01 上海交通大学 Defecation awareness signal detecting and defecation awareness reestablishing system

Also Published As

Publication number Publication date
CN113647973A (en) 2021-11-16

Similar Documents

Publication Publication Date Title
CN109009102B (en) Electroencephalogram deep learning-based auxiliary diagnosis method and system
CN111568446A (en) Portable electroencephalogram depression detection system combined with demographic attention mechanism
Azimi et al. Machine learning-based automatic detection of central sleep apnea events from a pressure sensitive mat
Zhao et al. An IoT-based wearable system using accelerometers and machine learning for fetal movement monitoring
Trivedy et al. Design and development of smartphone-enabled spirometer with a disease classification system using convolutional neural network
CN114999646B (en) Newborn exercise development assessment system, method, device and storage medium
CN113647973B (en) Portable nondestructive testing device based on convolutional neural network
Xu et al. Fetal movement detection by wearable accelerometer duo based on machine learning
Taşkıran et al. A deep learning based decision support system for diagnosis of Temporomandibular joint disorder
CN113593693A (en) Remote health management platform
CN113662578B (en) Human defecation prediction system based on residual error network
CN113768523B (en) Method and system for prewarning stool based on countermeasure generation network
CN110428893A (en) A kind of tele-medicine vital sign data acquisition device
CN114587310A (en) Method for realizing invasive blood pressure waveform estimation based on photoplethysmography
TWI748485B (en) Information processing system and method
CN115474897A (en) Wearable intelligent monitoring and identification system for audio and non-audio vibration signs
TWI756793B (en) A channel information processing system
CN219763346U (en) Cerebral apoplexy early warning and illness state real-time monitoring device
CN116965828A (en) Acoustic-electric dual-mode congenital heart disease prediction device based on incremental learning
Urdal Signal Processing for Newborn Survival: from labour to resuscitation
CN115979735A (en) Human urine smell detection device with disease detection function
Gopu et al. Digestive System Disorder Diagnosis using Electrogastrogram Signal Classification: Subtitle as needed (DSDD using EGG)
CN113786205A (en) Portable early warning system and method based on convolutional network and long-short term memory network
CN114118386A (en) Blood pressure estimation method based on attention mechanism and time domain convolution neural network
CN116548923A (en) Device for diagnosing cold and heat of internal organs based on human metabolism smell

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