CN108319928B - Deep learning method and system based on multi-target particle swarm optimization algorithm - Google Patents

Deep learning method and system based on multi-target particle swarm optimization algorithm Download PDF

Info

Publication number
CN108319928B
CN108319928B CN201810169310.4A CN201810169310A CN108319928B CN 108319928 B CN108319928 B CN 108319928B CN 201810169310 A CN201810169310 A CN 201810169310A CN 108319928 B CN108319928 B CN 108319928B
Authority
CN
China
Prior art keywords
target
parameter set
particle
particles
neural network
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
CN201810169310.4A
Other languages
Chinese (zh)
Other versions
CN108319928A (en
Inventor
高忠科
李彦里
杨宇轩
王新民
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.)
Tianjin University
Original Assignee
Tianjin University
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 Tianjin University filed Critical Tianjin University
Priority to CN201810169310.4A priority Critical patent/CN108319928B/en
Publication of CN108319928A publication Critical patent/CN108319928A/en
Application granted granted Critical
Publication of CN108319928B publication Critical patent/CN108319928B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/12—Classification; Matching
    • G06F2218/16—Classification; Matching by matching signal segments
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • 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

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Computation (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Evolutionary Biology (AREA)
  • Computational Linguistics (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Signal Processing (AREA)
  • Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Measurement And Recording Of Electrical Phenomena And Electrical Characteristics Of The Living Body (AREA)

Abstract

A deep learning model based on multi-objective particle swarm algorithm optimization is disclosed: acquiring an electroencephalogram signal, and preprocessing the acquired electroencephalogram signal to obtain a p-channel electroencephalogram signal data sample set with a p-dimension signal; normalizing the p-dimensional electroencephalogram data sample to be used as a deep learning model, namely, data input of a convolutional neural network, and using a corresponding imagined instruction class as the output of the last layer of the convolutional neural network; establishing a primary convolutional neural network; optimizing and adjusting the primary convolutional neural network by using a multi-objective particle swarm optimization algorithm to obtain a deep learning model; and realizing multi-target motion assistance by using a multi-target particle swarm optimization algorithm. The invention solves the problems of local optimization, low efficiency, need of priori knowledge and the like which may occur when the deep learning model is adjusted manually by utilizing a multi-target particle swarm optimization algorithm. The output of the constructed deep learning model can be used as a signal for controlling various devices such as a mechanical arm or an exoskeleton and the like.

Description

Deep learning method and system based on multi-target particle swarm optimization algorithm
Technical Field
The invention relates to a deep learning model. In particular to a deep learning method and a system based on multi-objective particle swarm optimization algorithm optimization.
Background
The brain electrical signal is a reflection mode of the physiological activity of neurons which can be detected by people in the cerebral cortex, and a large amount of physiological information can be extracted from different regional neuron potentials according to the specific condition of the brain electrical signal. Accurate detection, identification and classification of different human brain working states can provide a theoretical basis for certain special exercise assistance requirements in aspects such as a brain control system and the like. Therefore, the identification processing of the electroencephalogram signals has very important significance.
With the development of scientific technology, especially with the development of brain-computer interface (BCI) technology, in recent years, the research on the corresponding brain-computer interface technology becomes a new research hotspot, including SSVEP, motor imagery and other technologies. The research of motor imagery not only helps to improve understanding of related functions of the brain, but also obviously helps to research related fields of the brain-controlled motor-assisted system. The technology has the potential and prospect of providing a more convenient daily life means for the testee with the exercise assisting requirement, and under the ideal condition, the technology can enable the testee to control the degree of the basic or even complex life activities of the external equipment through the idea, and improve the mental and physical life experience of the testee by transmitting or realizing the idea to the outside through the idea.
In daily life, the movement of the limbs of people supports most living activities, and the movement function of both feet and the complex physiological function of both hands play a vital role in the normal life of people. Both the physical activities necessary for life and the basic creation of the substance in mental life require the corresponding limbs to support the physiology. However, for some subjects with special exercise or activity requirements, there may be environmental limitations due to the exercise or activity that needs to be performed, such as performing fault detection in a highly corrosive environment, in which case it is obviously not possible to perform these exercises and activities with the individual's body. Therefore, it is very important to use the brain-computer interface technology to help the examinee with the exercise assistance requirement to perform the activity.
As a new theoretical method emerging in recent years, deep learning has attracted wide attention from people in various fields, because of its unique and remarkable characteristics of learning, feeding back and realizing supervised classification on data features. The deep learning is used as a branch of machine learning, the algorithm of the deep learning is derived from simulation of a human brain, the deep learning belongs to further development of a neural network, the deep learning simulates the mechanism of the human brain to explain and analyze and learn data, the multi-layer representation of original data is automatically learned, and a plurality of hidden layers are used for forming a deep neural network structure. Typical network architectures are: convolutional Neural Networks (CNN), Generative Antagonistic Networks (GAN), long and short term memory networks (LSTM), and the like. With the gradual development of the deep learning method, solutions for various network overfitting problems are provided, so that a deep neural network with excellent effects can be trained and obtained, and compared with a conventional feature extraction and classification algorithm, the method has the advantages that the accuracy and other performances are remarkably improved, and therefore, the method is used for extracting the electroencephalogram motion feature signals of the testee and has remarkable effects.
Particle swarm optimization algorithm (PSO), an evolutionary computing technique developed by j.kennedy and r.c. eberhart, equal to 1995, was derived from a simplified social model, is a kind of clustering intelligence, and is incorporated into a multi-subject optimization system. PSO models bird flock predation behavior, in the context of which each solution to the optimization problem is equivalent to one bird in space, we call particles, all of which have a determined fitness value and their velocity of the optimization function. The PSO algorithm has the advantages of simplicity, easy implementation and no adjustment of many parameters, and is widely applied to the application fields of function optimization, neural network training, fuzzy system control and other genetic algorithms, and becomes an important branch of natural computation. The MOPSO algorithm (multi-target particle swarm optimization algorithm) which is subjected to the specialized improvement of multiple targets is adopted to optimize the deep learning network, so that the continuous manual try can be replaced, the network parameter optimization of deep learning is automatically carried out by using a computer, the efficiency is higher, and the effect is better.
Meanwhile, in the process of carrying out the motion assistance of the brain control system, the motion content of the electroencephalogram data of the testee is not limited to the target action which is attempted to be realized, but has more complicated thinking content, the electroencephalogram signal detection of the target action is carried out simply, other electroencephalogram signal actions are ignored, the situation that the actions are completed by mutually matching each limb organ in the physiological activities of normal people is not consistent, the effect of assisting the brain control action can be achieved, but the effect is probably superior to the manual operation mechanical execution and cannot have great breakthrough. If the MOPSO algorithm is adopted to carry out optimization assistance on the complex electroencephalogram signals, the single action assistance of a certain target is accurately replaced by the integral assistance of an action system, and the brain control action assistance with the best effect can be realized.
Disclosure of Invention
The invention aims to solve the technical problem of providing a deep learning method and system based on multi-target particle swarm optimization algorithm optimization.
The technical scheme adopted by the invention is as follows: a deep learning method based on multi-target particle swarm optimization algorithm optimization comprises the following steps:
1) acquiring an electroencephalogram signal, and preprocessing the acquired electroencephalogram signal to obtain a p-channel electroencephalogram signal data sample set x (t) with a two-dimensional signal;
2) normalizing the p-dimensional electroencephalogram data x (t) in the step 1) and then inputting the normalized p-dimensional electroencephalogram data x (t) as a data input of a deep learning model, namely a convolutional neural network, and taking a corresponding imagined instruction type as the output of the last layer of the convolutional neural network;
3) establishing a primary convolutional neural network;
4) optimizing and adjusting the primary convolutional neural network by using a multi-objective particle swarm optimization algorithm to obtain a deep learning model;
5) and realizing multi-target motion assistance by using a multi-target particle swarm optimization algorithm.
Step 1) acquiring an electroencephalogram signal by using a 32-channel electrode cap defined by an international 10-20 system, wherein the electroencephalogram signal is an electroencephalogram signal based on motor imagery, and the acquired content is a fixed limb motor imagery action.
The preprocessing of the acquired electroencephalogram signals in the step 1) comprises the following steps: band-pass filtering, sample normalization and elimination of the distorted electroencephalogram signals, wherein the range of the band-pass filtering is that corresponding electroencephalogram frequency bands are selected according to the characteristics of the motor imagery related tasks.
The step 3) comprises the following steps:
(1) selecting a sample from an electroencephalogram data sample set x (t) to enter a convolutional neural network;
(2) calculating actual output obtained by the sample entering the convolutional neural network, and at this stage, information is transmitted to an output layer from an input layer through gradual conversion, wherein the process is also a normal process after the network finishes training;
(3) calculating the difference between the actual output and the ideal output of the corresponding sample;
(4) and (4) adjusting the weight according to a method for minimizing errors to obtain a preliminary convolutional neural network.
The step 4) comprises the following steps:
(1) initializing a parameter group and a target parameter set of a preliminary convolutional neural network, wherein the parameter group of the preliminary convolutional neural network comprises a structure and corresponding hyper-parameters;
(2) carrying out particle group movement to obtain a new target parameter set;
(3) and inputting the obtained new target parameter set as a parameter of a new convolutional neural network, substituting the new target parameter set into an electroencephalogram data sample set x (t) for manual inspection, verifying the accuracy and efficiency of the new convolutional neural network, and using the verified convolutional neural network as a final convolutional neural network to form a deep learning model.
The step (1) includes randomly assigning an initial value to the parameter population, generating an initial convolutional neural network parameter population P1, and storing the best solution in the initial convolutional neural network parameter population P1 into the target parameter set as an initial best position archive set A1.
The step (2) includes that the current moving particle is j, and the following processes are carried out:
(2.1) calculating the dense information of the particles in the target parameter set, specifically, dividing the target parameter set into a plurality of regions by grids in space and equally, and taking the number of the particles contained in each region as the density information of the particles; the more the number of particles contained in the region where the particles are located, the larger the density value of the region, and conversely, the smaller the density value, the specific implementation is as follows:
calculating the boundary of the target parameter set space when the iteration number upper limit value t of the parameter group is calculated
Figure GDA0003523783800000031
Model of computational grid
Figure GDA0003523783800000032
Where M refers to the total number of regions into which the grid is divided, taking an integer,
Figure GDA0003523783800000033
and
Figure GDA0003523783800000034
is the objective function value; traversing the particles in the target parameter set and calculating the positions of the particles in the target parameter setNumbering of grids
Figure GDA0003523783800000035
Wherein Int is a rounding function; calculating grid information and a density estimation value of the particles;
(2.2) setting the parameters of the population of particles Pj，tIs the best particle G in the target parameter setj，tParticles Gj，tThe quality of the target particle swarm optimization algorithm determines the convergence performance and diversity of the multi-target particle swarm optimization algorithm, and the selection basis is the particle density information in the target parameter set; wherein j is the label of the current moving particle, and t is the value of the current iteration number; evaluating the searching potential of the target parameter set particles by the number of the target parameter set particles superior to the parameter population, wherein the more the target parameter set particles superior to the parameter population are, the greater the searching potential of the target parameter set is, and the specific algorithm is as follows:
calculating the preference among the set of target parameters over particle Pj，tParticle set A ofjFor integer k from 1 to AtIn the particle number range of (A)j＝Aj+{Ak，t|Ak，t＜Pj，t，Ak，t∈At}; then, the particle set A is calculatedjParticle set G with the lowest medium densityj，Gj＝min{Density(Ak)，k＝1,2，...，|Aj|，Ak∈Aj}; wherein A isjFor storing a target parameter set AtHas a medium to superior particle Pj，tSet of particles of (A)jThe particles with the lowest medium density are stored in the particle set GjPerforming the following steps; a. thek，tWhere t denotes the t-th iteration, the number of omitted iterations in the same iteration is Ak；Density(Ak) Is to calculate the particle Ak(ii) a density estimate of;
(2.3) updating the positions and the speeds of the particles in the parameter population, and searching for an optimal solution under the guidance of the global optimal particle G and the individual optimal particle P, namely updating a target parameter set;
(2.4) performing the partition operation of the target parameter set to avoid the number of particles exceeding the specified number;
and (2.5) outputting the particle information of the target parameter set to be a new target parameter set.
A system of a deep learning method based on multi-target particle swarm optimization algorithm optimization comprises the following steps:
1) initializing a preliminary multi-element limb action input parameter set and a target action parameter set, wherein the preliminary multi-element limb action input parameter set is the target action parameter set corresponding to classification information obtained after samples in an electroencephalogram data sample set x (t) are input into a deep learning model;
2) carrying out particle group movement to obtain a new target action parameter set;
3) and comparing the obtained new target action parameter set with a target action parameter set obtained by multiple actions of the testee, and verifying by adopting a threshold value screening method, wherein the new target action parameter set is used as a final target action parameter set to realize corresponding auxiliary actions if the new target action parameter set is the highest-frequency target action parameter set which is classified and judged for multiple times before, and otherwise, returning to the step 2).
Step 1) comprises randomly assigning an initial value to the multiple limb motion input parameter group, generating an initial multiple limb motion input parameter group P2, and storing the best solution in the initial multiple limb motion input parameter group P2 into a target motion parameter set as an initial best position archive set A2.
The step (2) includes that if the current moving particle is i, the following processes are carried out:
(2.1) calculating the dense information of the particles in the target motion parameter set, specifically, dividing the target motion parameter set into a plurality of regions by grids in space, and taking the number of the particles contained in each region as the density information of the particles; the more the number of particles contained in the region where the particles are located, the larger the density value of the region, and conversely, the smaller the density value, the specific implementation is as follows:
calculating the boundary of the target parameter set space when the iteration number upper limit value t of the parameter group is calculated
Figure GDA0003523783800000041
Model of computational grid
Figure GDA0003523783800000042
Where M refers to the total number of regions into which the grid is divided, taking an integer,
Figure GDA0003523783800000043
and
Figure GDA0003523783800000044
is the objective function value; traversing the particles in the target action parameter set, and calculating the number of the grid where the particles in the target action parameter set are positioned
Figure GDA0003523783800000045
Wherein Int is a rounding function; calculating grid information and a density estimation value of the particles;
(2.2) setting the parameters of the population of particles Pj，tIs the best particle G in the target motion parameter seti，tParticles Gi，tThe quality of the target particle swarm optimization algorithm determines the convergence performance and diversity of the multi-target particle swarm optimization algorithm, and the selection basis is the particle density information in the target action parameter set; wherein i is the label of the current moving particle, and t is the value of the current iteration number; evaluating the searching potential of the target action parameter set particles by the number of the target action parameter set particles superior to the parameter population, wherein the more the target action parameter set particles superior to the parameter population are, the greater the searching potential of the target action parameter set is, and the algorithm is specifically as follows:
computing target motion parameter set precedence over particle Pi，tParticle set A ofiFor integer k from 1 to AtIn the particle number range of (A)i＝Ai+{Ak，t|Ak，t＜Pi，t，Ak，t∈At}; then, the particle set A is calculatediParticle set G with the lowest medium densityi，Gi＝min{Density(Ak)，k＝1,2，...，|Ai|，Ak∈Ai}; wherein A isiFor storing a target motion parameter set AtHas a medium to superior particle Pi，tSet of particles of (A)iThe particles with the lowest medium density are stored in the particle set GiPerforming the following steps; a. thek，tWhere t denotes the t-th iteration, the number of omitted iterations in the same iteration is Ak；Density(Ak) Is to calculate the particle Ak(ii) a density estimate of;
(2.3) updating the positions and the speeds of the particles in the parameter population, and searching for an optimal solution under the guidance of the global optimal particle G and the individual optimal particle P, namely updating the target motion parameter set;
(2.4) performing the partition operation of the target action parameter set to avoid the number of particles exceeding the specified number;
and (2.5) outputting the particle information of the target action parameter set to be a new target action parameter set.
According to the multi-target particle swarm optimization algorithm-based deep learning method and system, the problems of local optimization, low efficiency, need of priori knowledge and the like which may occur when a deep learning model is adjusted manually are solved by using the multi-target particle swarm optimization algorithm. The output of the deep learning model constructed in the way can be used as signals for controlling various devices such as a mechanical arm or an exoskeleton, and the like, and compared with the traditional signal identification, the deep learning model has obvious advantages in excellent identification accuracy and extremely high identification efficiency. Meanwhile, the multi-target particle swarm algorithm is used for expanding and identifying, single action auxiliary identification is adopted as multi-target action auxiliary identification, the action auxiliary functionality is improved to a considerable extent, and the method has a wide application prospect.
Drawings
FIG. 1 is a flow chart of a deep learning method and system based on multi-objective particle swarm optimization according to the present invention.
Detailed Description
The following describes in detail a deep learning method and system based on multi-objective particle swarm optimization algorithm according to the present invention with reference to embodiments and drawings.
The invention discloses a deep learning method and a deep learning system based on multi-target particle swarm optimization algorithm optimization, and provides a method for establishing a deep learning network model based on electroencephalogram EEG signals, expanding PSO optimization algorithm to optimize the deep learning network and applying the deep learning network to brain control equipment, which is used for overcoming the defects of the prior art, carrying out efficient deep learning network construction by utilizing the advantages of MOPSO optimization algorithm, and simultaneously forming a complete action auxiliary system by utilizing MOPSO algorithm to realize single action auxiliary mode, optimization of main target and synchronous optimization of auxiliary target.
As shown in fig. 1, the deep learning method based on multi-objective particle swarm optimization of the present invention includes the following steps:
1) the method comprises the steps of collecting electroencephalogram signals, wherein the electroencephalogram signals are collected by using a 32-channel electrode cap defined by an international 10-20 system, are electroencephalogram signals based on motor imagery, and are collected by fixed limb motor imagery actions. The band-pass filtering range is specifically as follows: the motor imagery and the perceived control signal is the perceived motor rhythm. The perception of the motor rhythm is mainly characterized by the μ band (frequency range of 8-12Hz) and the band (frequency range of 18-26 Hz). The band-pass filtering range is therefore 1-30 Hz. Preprocessing the acquired electroencephalogram signals to obtain a p-channel electroencephalogram data sample set x (t) with two-dimensional signals, wherein the preprocessing of the acquired electroencephalogram signals is to manually remove the influence of an electrooculogram signal and the like by using corresponding methods such as ICA and the like, and determine the frequency range of a corresponding task through the perception rhythm of motor imagery to carry out band-pass filtering, and the method comprises the following steps: band-pass filtering, sample normalization and elimination of the distorted electroencephalogram signals, wherein the range of the band-pass filtering is that corresponding electroencephalogram frequency bands are selected according to the characteristics of the motor imagery related tasks.
2) Normalizing the p-dimensional electroencephalogram data x (t) sample in the step 1) to be used as a data input of a deep learning model, namely a Convolutional Neural Network (CNN), and using a corresponding imagined instruction class as an output of the last layer of the convolutional neural network;
3) establishing a primary convolutional neural network;
convolutional Neural Networks (CNN) are variations of the multi-layer perceptron (MLP), and a general neuro-cognitive machine includes two types of neurons, i.e., a sampling element for feature extraction and a convolution element for anti-variation, the sampling element relates to two parameter fields and a threshold, the field determines how many input connections are connected, and the threshold controls the degree of response to a feature sub-pattern. CNN essentially acts as a multi-layered perceptron, which is unique in that it not only reduces the number of weights making the network easy to optimize, but also reduces the risk of over-fitting. Compared with a common perceptron, the CNN network is closer to a biological neural network by the weight sharing characteristic structure, and the complexity of a network model is reduced. CNN is a multi-layered neural network, each layer consisting of multiple two-dimensional planes, each plane having multiple independent neurons. The network comprises simple elements and complex elements, the simple elements form a simple surface, the simple surfaces are aggregated to form a simple layer, and the complex elements, the complex surfaces and the complex layer are also defined in the same way. The middle part of the network is formed by connecting simple layers and complex parts in series, and the input layer is only one layer and directly receives the input two-dimensional electroencephalogram signals. Generally speaking, the simple layer is a feature extraction layer, the input of the neuron is connected with the local receptive field of the previous layer to extract the feature of the content of the receptive field, and if the feature is successfully extracted, the position relation between the feature and other features is determined; the complex layer is a feature mapping layer, each calculation layer of the network is composed of a plurality of feature mappings, each feature mapping is a plane, and the weights of all neurons on the plane are the same. Because the weight of the neuron on each mapping surface is shared, the number of free parameters of the network is reduced, and the complexity of network parameter selection is reduced. The CNN network comprises an input layer, an output layer, a middle convolutional layer, a sampling layer and a full-connection layer, wherein an original two-dimensional electroencephalogram signal is directly input into the input layer, the size of the original electroencephalogram data determines the size of an input vector, neurons extract local characteristics of the electroencephalogram data, each neuron is connected with a local receptive field of the previous layer, and the output of the network is given out on the output layer through the sampling layer (simple layer), the convolutional layer (complex layer) and the last full-connection layer which are alternately arranged. In CNN, network training and weight updating are usually based on the back propagation algorithm (BP algorithm),
the invention relates to a method for establishing a preliminary convolutional neural network, which comprises the following steps:
(1) selecting a sample from an electroencephalogram data sample set x (t) to enter a convolutional neural network;
(2) calculating actual output obtained by the sample entering the convolutional neural network, and at this stage, information is transmitted to an output layer from an input layer through gradual conversion, wherein the process is also a normal process after the network finishes training;
(3) calculating the difference between the actual output and the ideal output of the corresponding sample;
(4) and (4) adjusting the weight according to a method for minimizing errors to obtain a preliminary convolutional neural network.
4) Optimizing and adjusting the primary convolutional neural network by using a multi-objective particle swarm optimization algorithm (MOPSO) to obtain a deep learning model;
the multi-objective particle swarm optimization (MOPSO) algorithm is an improvement of a classical PSO optimization algorithm, and when the PSO algorithm is applied to deep learning, the problem that the deep learning involves many parameters and thus needs many optimized contents occurs, but the problem cannot be solved by a single-objective PSO optimization algorithm, so that the MOPSO algorithm is provided. When the individual optimal position P is selected in the single-target optimization, only comparison is needed to determine which is better, but for a deep learning network such as CNN, comparison of two particles cannot directly determine which is better, and if multiple targets are simultaneously best, the particles have the best effect, but in practice, different advantages are generally obtained. Similarly, for a population optimal location, there will be only one optimal individual in the population for a single-target PSO, but for multi-target optimization, there may be many optimal individuals. The solution of MOPSO for individual optimal positions is to randomly select a position as the historical optimal position under the condition that which position can not be strictly compared with better positions, and for group optimal positions, in an individual optimal set, an optimal position called as a leader is selected according to the distribution density degree of the individual optimal positions, and the leader usually selects a position particle which is not dense to play as the role. The method adopts an adaptive grid method to update the group optimal position and the individual optimal position, and the specific idea is to divide according to grids and assume the number n of particles in each gridB represents a second grid in which the probability that the particle is selected is
Figure GDA0003523783800000071
The more crowded the particles are, the lower the probability of selection. This is to ensure that unknown solution space regions can be explored, which is particularly important in the optimization of CNN networks.
The invention relates to a method for optimizing and adjusting a preliminary convolutional neural network by using a multi-objective particle swarm optimization algorithm (MOPSO) to obtain a deep learning model, which comprises the following steps of:
(1) initializing a parameter group of a preliminary convolutional neural network and a target parameter set (an archive set of optimal positions), wherein the parameter group of the preliminary convolutional neural network comprises a structure and corresponding super parameters;
the method comprises the steps of randomly giving an initial value to a parameter group, generating an initial convolutional neural network parameter group P1, and storing the best solution in a target parameter set in the initial convolutional neural network parameter group P1 as an initial best position archive set A1. In the standard for judging the quality, the accuracy of the brain electrical data acquisition and classification under the parameter group is occupied in the first judgment position, and the time required by operation, the parameter achievement requirement, the structural application universality and the like are adopted.
(2) Carrying out particle group movement to obtain a new target parameter set; the method comprises the following steps that if the current moving particle is j, the following processes are carried out:
(2.1) calculating the dense information of the particles in the target parameter set, specifically, dividing the target parameter set into a plurality of regions by grids in space and equally, and taking the number of the particles contained in each region as the density information of the particles; the more the number of particles contained in the region where the particles are located, the larger the density value of the region, and conversely, the smaller the density value, the specific implementation is as follows:
calculating the boundary of the target parameter set space when the iteration number upper limit value t of the parameter group is calculated
Figure GDA0003523783800000072
Model of computational grid
Figure GDA0003523783800000073
Where M refers to the total number of regions into which the grid is divided, taking an integer,
Figure GDA0003523783800000074
and
Figure GDA0003523783800000075
is the objective function value; traversing the particles in the target parameter set, and calculating the number of the grid where the particles in the target parameter set are located
Figure GDA0003523783800000076
Wherein Int is a rounding function; calculating grid information and a density estimation value of the particles;
(2.2) setting the parameters of the population of particles Pj，tIs the best particle G in the target parameter setj，tParticles Gj，tThe quality of the target particle group optimization algorithm determines the convergence performance and diversity of the multi-target particle group optimization algorithm, and the selection is based on the particle density information in the target parameter set; wherein j is the label of the current moving particle, and t is the value of the current iteration number; evaluating the searching potential of the target parameter set particles by the number of the target parameter set particles superior to the parameter population, wherein the more the target parameter set particles superior to the parameter population are, the greater the searching potential of the target parameter set is, and the specific algorithm is as follows:
calculating the preference among the set of target parameters over particle Pj，tParticle set A ofjFor integer k from 1 to AtIn the particle number range of (A)j＝Aj+{Ak，t|Ak，t＜Pj，t，Ak，t∈At}; then, the particle set A is calculatedjParticle set G with the lowest medium densityj，Gj＝min{Density(Ak)，k＝1,2，...，|Aj|，Ak∈Aj}; wherein A isjFor storing a target parameter set AtHas a medium to superior particle Pj，tSet of particles of (A)jThe particles with the lowest medium density are stored in the particle set GjPerforming the following steps; a. thek，tWhere t denotes the t-th iteration, with the iterations omitted in the same iterationThe generation number is Ak；Density(Ak) Is to calculate the particle Ak(ii) a density estimate of;
(2.3) updating the positions and the speeds of the particles in the parameter population, and searching for an optimal solution under the guidance of the global optimal particle G and the individual optimal particle P, namely updating a target parameter set;
(2.4) performing the partition operation of the target parameter set to avoid the number of particles exceeding the specified number;
and (2.5) outputting the particle information of the target parameter set to be a new target parameter set.
(3) And inputting the obtained new target parameter set as a parameter of a new convolutional neural network, substituting the new target parameter set into an electroencephalogram data sample set x (t) for manual inspection, verifying the accuracy and efficiency of the new convolutional neural network, and using the verified convolutional neural network as a final convolutional neural network to form a deep learning model.
5) And realizing multi-target motion assistance by using a multi-target particle swarm optimization algorithm.
In the invention, the MOPSO algorithm also plays a role in realizing the conversion of single action assistance into a multi-target action assistance system. The basic requirements for its implementation are: the method ensures that the action auxiliary function which is originally mainly operated is not influenced and interfered, and provides the auxiliary action auxiliary function which is realized as close to the action of the original normal person as possible, so as to ensure that the action auxiliary is close to the action of the normal person and the brain signal is coordinated.
The invention discloses a system of a deep learning method based on multi-target particle swarm optimization algorithm optimization, which comprises the following steps:
1) initializing a preliminary multi-element limb action input parameter set and a target action parameter set, wherein the preliminary multi-element limb action input parameter set is the target action parameter set corresponding to classification information obtained after samples in an electroencephalogram data sample set x (t) are input into a deep learning model; the initialization group is an input parameter set of another auxiliary motion device, and the optimization goal is a multi-element limb motion combination result consisting of one main motion auxiliary goal. Namely, the input values of the auxiliary devices at corresponding positions of all possible motion auxiliary devices needed by a brain control system user are used as an input parameter set, and when motion assistance is performed, the input of the auxiliary device which is most consistent with the current electroencephalogram signal triggering motion output and is nearby or related is quickly found through MOPSO optimization, so that a complete motion assistance system is formed, rather than a single joint or muscle, and the best motion combination configuration is used as the output value of the parameter. The judgment standard is derived from the correlation of the invoked action assisting system in the same or similar motor imagery of repeated times. The method specifically comprises the following steps:
randomly assigning an initial value to the multi-element limb movement input parameter group to generate an initial multi-element limb movement input parameter group P2, and storing the best solution in the initial multi-element limb movement input parameter group P2 into a target movement parameter set as an initial best position archive set A2.
2) Carrying out particle group movement to obtain a new target action parameter set; the method comprises the following steps of setting the current moving particle as i:
(2.1) calculating the dense information of the particles in the target motion parameter set, specifically, dividing the target motion parameter set into a plurality of regions by grids in space, and taking the number of the particles contained in each region as the density information of the particles; the more the number of particles contained in the region where the particles are located, the larger the density value of the region, and conversely, the smaller the density value, the specific implementation is as follows:
calculating the boundary of the target parameter set space when the iteration number upper limit value t of the parameter group is calculated
Figure GDA0003523783800000081
Model of computational grid
Figure GDA0003523783800000082
Where M refers to the total number of regions into which the grid is divided, taking an integer,
Figure GDA0003523783800000083
and
Figure GDA0003523783800000084
is the objective function value; traversing particles in a target action parameter setCalculating the number of the grid where the particles in the target action parameter set are located
Figure GDA0003523783800000091
Wherein Int is a rounding function; calculating grid information and a density estimation value of the particles;
(2.2) setting the parameters of the population of particles Pj，tIs the best particle G in the target motion parameter seti，tParticles Gi，tThe quality of the target particle swarm optimization algorithm determines the convergence performance and diversity of the multi-target particle swarm optimization algorithm, and the selection is based on the particle density information in the target action parameter set; wherein i is the label of the current moving particle, and t is the value of the current iteration number; evaluating the searching potential of the target action parameter set particles by the number of the target action parameter set particles superior to the parameter population, wherein the more the target action parameter set particles superior to the parameter population are, the greater the searching potential of the target action parameter set is, and the algorithm is specifically as follows:
computing target motion parameter set precedence over particle Pi，tParticle set A ofiFor integer k from 1 to AtIn the particle number range of (A)i＝Ai+{Ak，t|Ak，t＜Pi，t，Ak，t∈At}; then, the particle set A is calculatediParticle set G with the lowest medium densityi，Gi＝min{Density(Ak)，k＝1,2，...，|Ai|，Ak∈Ai}; wherein A isiFor storing a target motion parameter set AtHas a medium to superior particle Pi，tSet of particles of (A)iThe particles with the lowest medium density are stored in the particle set GiPerforming the following steps; a. thek，tWhere t denotes the t-th iteration, the number of omitted iterations in the same iteration is Ak；Density(Ak) Is to calculate the particle Ak(ii) a density estimate of;
(2.3) updating the positions and the speeds of the particles in the parameter population, and searching for an optimal solution under the guidance of the global optimal particle G and the individual optimal particle P, namely updating the target motion parameter set;
(2.4) performing the partition operation of the target action parameter set to avoid the number of particles exceeding the specified number;
and (2.5) outputting the particle information of the target action parameter set to be a new target action parameter set.
3) And comparing the obtained new target action parameter set with a target action parameter set obtained by multiple actions of the testee, and verifying by adopting a threshold value screening method, wherein the new target action parameter set is used as a final target action parameter set to realize corresponding auxiliary actions if the new target action parameter set is the highest-frequency target action parameter set which is classified and judged for multiple times before, and otherwise, returning to the step 2).
The brain electrical signals are subjected to deep learning model building learning processing, and the MOPSO optimization algorithm is used for optimizing and designing the deep learning model, so that the method can be finally applied to brain control equipment development of motor imagery.
The complete operation process is as follows: firstly, electroencephalogram signals of a target are collected by an electroencephalogram signal collecting device and are used as an input data set, a CNN is established, multi-target optimization is carried out on the network by an MOPSO algorithm, classification identification of electroencephalogram motor imagery signals is carried out by taking main action auxiliary accuracy as a first target and taking auxiliary action accuracy, reaction efficiency and the like as second targets; then, an MOPSO auxiliary action group optimizer which takes the prior accumulated information combination as a judgment standard is adopted to carry out systematic auxiliary parameter input optimization for the motor imagery action assistance; finally, inputting the corresponding parameter input optimization result into the auxiliary action device, and accordingly realizing the related functions of the brain control equipment.
The invention relates to a multi-target particle swarm optimization-based deep learning method and a system, which are characterized in that firstly, a multichannel signal measured by electroencephalogram data acquisition equipment is subjected to a preprocessing technology, the corresponding methods such as ICA (independent component analysis) and the like are used for manually removing the influence of an eye electrical signal and the like, and the frequency range of a corresponding task is determined through the perception rhythm of motor imagery; the method comprises the steps of inputting preprocessed data serving as training samples of a deep learning network into a CNN network, using brain instructions needing to be distinguished and identified in an actual task as output categories of the CNN, training by utilizing convolution and pooling operations of a plurality of training samples, adjusting and updating parameter groups of the network by utilizing an MOPSO algorithm, optimizing corresponding parameters of the network, further obtaining a deep learning model for motor imagery electroencephalogram signal identification, expanding single action assistance by adopting the MOPSO algorithm to form a complete action assistance system, and using the final identification and adjustment results of the deep learning network and a subsequent system as corresponding content of instruction classification of brain control equipment in the development of the brain control equipment to perform corresponding instruction operation of the brain control equipment.
The above description of the present invention and the embodiments is not limited thereto, and the description of the embodiments is only one of the implementation manners of the present invention, and any structure or embodiment similar to the technical solution without inventive design is within the protection scope of the present invention without departing from the inventive spirit of the present invention.

Claims (9)

1. A deep learning method based on multi-target particle swarm optimization algorithm is characterized by comprising the following steps:
1) acquiring an electroencephalogram signal, and preprocessing the acquired electroencephalogram signal to obtain a p-channel electroencephalogram signal data sample set x (t) with a two-dimensional signal;
2) normalizing the p-dimensional electroencephalogram data x (t) in the step 1) and then inputting the normalized p-dimensional electroencephalogram data x (t) as a data input of a deep learning model, namely a convolutional neural network, and taking a corresponding imagined instruction type as the output of the last layer of the convolutional neural network;
3) establishing a primary convolutional neural network;
4) optimizing and adjusting the primary convolutional neural network by using a multi-objective particle swarm optimization algorithm to obtain a deep learning model; the method comprises the following steps:
(1) initializing a parameter group and a target parameter set of a preliminary convolutional neural network, wherein the parameter group of the preliminary convolutional neural network comprises a structure and corresponding hyper-parameters;
(2) carrying out particle group movement to obtain a new target parameter set;
(3) inputting the obtained new target parameter set as the parameter of a new convolutional neural network, substituting the new target parameter set into an electroencephalogram data sample set x (t) for manual inspection, verifying the accuracy and efficiency of the new convolutional neural network, and using the verified convolutional neural network as a final convolutional neural network to form a deep learning model;
5) and realizing multi-target motion assistance by using a multi-target particle swarm optimization algorithm.
2. The deep learning method based on multi-target particle swarm optimization algorithm optimization according to claim 1, wherein step 1) is to use a 32-channel electrode cap defined by an international 10-20 system to collect electroencephalogram signals, the electroencephalogram signals are electroencephalogram signals based on motor imagery, and the collected content is a fixed limb motor imagery action.
3. The deep learning method based on multi-target particle swarm algorithm optimization according to claim 1, wherein the preprocessing of the acquired electroencephalogram signals in step 1) comprises: band-pass filtering, sample normalization and elimination of the distorted electroencephalogram signals, wherein the range of the band-pass filtering is that corresponding electroencephalogram frequency bands are selected according to the characteristics of the motor imagery related tasks.
4. The deep learning method based on multi-objective particle swarm algorithm optimization according to claim 1, wherein the step 3) comprises the following steps:
(1) selecting a sample from an electroencephalogram data sample set x (t) to enter a convolutional neural network;
(2) calculating actual output obtained by the sample entering the convolutional neural network, and at this stage, information is transmitted to an output layer from an input layer through gradual conversion, wherein the process is also a normal process after the network finishes training;
(3) calculating the difference between the actual output and the ideal output of the corresponding sample;
(4) and (4) adjusting the weight according to a method for minimizing errors to obtain a preliminary convolutional neural network.
5. The method of claim 1, wherein the step (1) comprises randomly assigning an initial value to the parameter population, generating an initial convolutional neural network parameter population P1, and storing the best solution in the initial convolutional neural network parameter population P1 as an initial best location archive set A1.
6. The deep learning method based on multi-objective particle swarm algorithm optimization according to claim 1, wherein the step (2) comprises the following steps of assuming that the currently moving particle is j:
(2.1) calculating the dense information of the particles in the target parameter set, specifically, dividing the target parameter set into a plurality of regions by grids in space and equally, and taking the number of the particles contained in each region as the density information of the particles; the more the number of particles contained in the region where the particles are located, the larger the density value of the region, and conversely, the smaller the density value, the specific implementation is as follows:
calculating the boundary of the target parameter set space when the iteration number upper limit value t of the parameter group is calculated
Figure FDA0003523783790000021
Model of computational grid
Figure FDA0003523783790000022
Wherein M refers to the total number of areas divided by the grid, taking an integer, F1 tAnd F2 tIs the objective function value; traversing the particles in the target parameter set, and calculating the number of the grid where the particles in the target parameter set are located
Figure FDA0003523783790000023
Wherein Int is a rounding function; calculating grid information and a density estimation value of the particles;
(2.2) setting the parameters of the population of particles Pj，tIs the best particle G in the target parameter setj，tParticles Gj，tThe quality of the multi-target particle swarm optimization algorithm determines the convergence performance and diversity of the multi-target particle swarm optimization algorithmThe selection basis is the particle density information in the target parameter set; wherein j is the label of the current moving particle, and t is the value of the current iteration number; evaluating the searching potential of the target parameter set particles by the number of the target parameter set particles superior to the parameter population, wherein the more the target parameter set particles superior to the parameter population are, the greater the searching potential of the target parameter set is, and the specific algorithm is as follows:
calculating the preference among the set of target parameters over particle Pj，tParticle set A ofjFor integer k from 1 to AtIn the particle number range of (A)j＝Aj+{Ak，t|Ak，t＜Pj，t，Ak，t∈At}; then, the particle set A is calculatedjParticle set G with the lowest medium densityj，Gj＝min{Density(Ak)，k＝1,2，...，|Aj|，Ak∈Aj}; wherein A isjFor storing a target parameter set AtHas a medium to superior particle Pj，tSet of particles of (A)jThe particles with the lowest medium density are stored in the particle set GjPerforming the following steps; a. thek，tWhere t denotes the t-th iteration, the number of omitted iterations in the same iteration is Ak；Density(Ak) Is to calculate the particle Ak(ii) a density estimate of;
(2.3) updating the positions and the speeds of the particles in the parameter population, and searching for an optimal solution under the guidance of the global optimal particle G and the individual optimal particle P, namely updating a target parameter set;
(2.4) performing the partition operation of the target parameter set to avoid the number of particles exceeding the specified number;
and (2.5) outputting the particle information of the target parameter set to be a new target parameter set.
7. The system of the multi-objective particle swarm algorithm optimization-based deep learning method of claim 1, characterized by comprising the following steps:
1) initializing a preliminary multi-element limb action input parameter set and a target action parameter set, wherein the preliminary multi-element limb action input parameter set is the target action parameter set corresponding to classification information obtained after samples in an electroencephalogram data sample set x (t) are input into a deep learning model;
2) carrying out particle group movement to obtain a new target action parameter set;
3) and comparing the obtained new target action parameter set with a target action parameter set obtained by multiple actions of the testee, and verifying by adopting a threshold value screening method, wherein the new target action parameter set is used as a final target action parameter set to realize corresponding auxiliary actions if the new target action parameter set is the highest-frequency target action parameter set which is classified and judged for multiple times before, and otherwise, returning to the step 2).
8. The system of claim 7, wherein step 1) comprises randomly assigning an initial value to the multiple limb motion input parameter group, generating an initial multiple limb motion input parameter group P2, and storing the best solution in the initial multiple limb motion input parameter group P2 as an initial best position archive set A2.
9. The system of multi-objective particle swarm algorithm optimization-based deep learning method according to claim 7, wherein the step (2) comprises the following steps of assuming that the currently moving particle is i:
(2.1) calculating the dense information of the particles in the target motion parameter set, specifically, dividing the target motion parameter set into a plurality of regions by grids in space, and taking the number of the particles contained in each region as the density information of the particles; the more the number of particles contained in the region where the particles are located, the larger the density value of the region, and conversely, the smaller the density value, the specific implementation is as follows:
calculating the boundary of the target parameter set space when the iteration number upper limit value t of the parameter group is calculated
Figure FDA0003523783790000031
Model of computational grid
Figure FDA0003523783790000032
Wherein M refers to the total number of areas divided by the grid, taking an integer, F1 tAnd F2 tIs the objective function value; traversing the particles in the target action parameter set, and calculating the number of the grid where the particles in the target action parameter set are positioned
Figure FDA0003523783790000033
Wherein Int is a rounding function; calculating grid information and a density estimation value of the particles;
(2.2) setting the parameters of the population of particles Pj，tIs the best particle G in the target motion parameter seti，tParticles Gi，tThe quality of the target particle swarm optimization algorithm determines the convergence performance and diversity of the multi-target particle swarm optimization algorithm, and the selection basis is the particle density information in the target action parameter set; wherein i is the label of the current moving particle, and t is the value of the current iteration number; evaluating the searching potential of the target action parameter set particles by the number of the target action parameter set particles superior to the parameter population, wherein the more the target action parameter set particles superior to the parameter population are, the greater the searching potential of the target action parameter set is, and the algorithm is specifically as follows:
computing target motion parameter set precedence over particle Pi，tParticle set A ofiFor integer k from 1 to AtIn the particle number range of (A)i＝Ai+{Ak，t|Ak，t＜Pi，t，Ak，t∈At}; then, the particle set A is calculatediParticle set G with the lowest medium densityi，Gi＝min{Density(Ak)，k＝1,2，...，|Ai|，Ak∈Ai}; wherein A isiFor storing a target motion parameter set AtHas a medium to superior particle Pi，tSet of particles of (A)iThe particles with the lowest medium density are stored in the particle set GiPerforming the following steps; a. thek，tWhere t denotes the t-th iteration, the number of omitted iterations in the same iteration is Ak；Density(Ak) Is to calculate the particle Ak(ii) a density estimate of;
(2.3) updating the positions and the speeds of the particles in the parameter population, and searching for an optimal solution under the guidance of the global optimal particle G and the individual optimal particle P, namely updating the target motion parameter set;
(2.4) performing the partition operation of the target action parameter set to avoid the number of particles exceeding the specified number;
and (2.5) outputting the particle information of the target action parameter set to be a new target action parameter set.
CN201810169310.4A 2018-02-28 2018-02-28 Deep learning method and system based on multi-target particle swarm optimization algorithm Active CN108319928B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201810169310.4A CN108319928B (en) 2018-02-28 2018-02-28 Deep learning method and system based on multi-target particle swarm optimization algorithm

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201810169310.4A CN108319928B (en) 2018-02-28 2018-02-28 Deep learning method and system based on multi-target particle swarm optimization algorithm

Publications (2)

Publication Number Publication Date
CN108319928A CN108319928A (en) 2018-07-24
CN108319928B true CN108319928B (en) 2022-04-19

Family

ID=62900991

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201810169310.4A Active CN108319928B (en) 2018-02-28 2018-02-28 Deep learning method and system based on multi-target particle swarm optimization algorithm

Country Status (1)

Country Link
CN (1) CN108319928B (en)

Families Citing this family (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109063319B (en) * 2018-07-27 2023-04-07 天津大学 Simulation method of biological ecosystem based on neural network
CN109003680B (en) * 2018-09-28 2021-04-09 四川大学 Epileptic data statistical method and device
CN109596326B (en) * 2018-11-30 2020-06-12 电子科技大学 Rotary machine fault diagnosis method based on convolution neural network with optimized structure
CN111436929B (en) * 2019-01-17 2021-06-01 复旦大学 Method for generating and identifying neurophysiological signals
CN110221684A (en) * 2019-03-01 2019-09-10 Oppo广东移动通信有限公司 Apparatus control method, system, electronic device and computer readable storage medium
CN110584597A (en) * 2019-07-15 2019-12-20 天津大学 Multi-channel electroencephalogram signal monitoring method based on time-space convolutional neural network and application
WO2021007812A1 (en) * 2019-07-17 2021-01-21 深圳大学 Deep neural network hyperparameter optimization method, electronic device and storage medium
CN110414732B (en) * 2019-07-23 2020-09-18 中国科学院地理科学与资源研究所 Travel future trajectory prediction method and device, storage medium and electronic equipment
CN112605983B (en) * 2020-12-01 2022-04-19 浙江工业大学 Mechanical arm pushing and grabbing system suitable for intensive environment
CN112990224B (en) * 2021-02-04 2023-07-11 郑州航空工业管理学院 Particle swarm method for identifying brain function division from fMRI data
CN116849942A (en) * 2023-07-28 2023-10-10 中国医学科学院生物医学工程研究所 Brain-control intelligent lifting and turning-over multifunctional medical care bed

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101711709A (en) * 2009-12-07 2010-05-26 杭州电子科技大学 Method for controlling electrically powered artificial hands by utilizing electro-coulogram and electroencephalogram information
CN104899436A (en) * 2015-05-29 2015-09-09 北京航空航天大学 Electroencephalogram signal time-frequency analysis method based on multi-scale radial basis function and improved particle swarm optimization algorithm
CN105447567A (en) * 2015-11-06 2016-03-30 重庆科技学院 BP neural network and MPSO algorithm-based aluminium electrolysis energy-saving and emission-reduction control method
CN105740887A (en) * 2016-01-26 2016-07-06 杭州电子科技大学 Electroencephalogram feature classification method based on PSO-SVM (Particle Swarm Optimization-Support Vector Machine)
WO2017014599A1 (en) * 2015-07-22 2017-01-26 주식회사 씨코헬스케어 Composition for stabilizing radiochemical purity of [18f]fluoro-dopa and method for preparing same
CN106821681A (en) * 2017-02-27 2017-06-13 浙江工业大学 A kind of upper limbs ectoskeleton control method and system based on Mental imagery

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101711709A (en) * 2009-12-07 2010-05-26 杭州电子科技大学 Method for controlling electrically powered artificial hands by utilizing electro-coulogram and electroencephalogram information
CN104899436A (en) * 2015-05-29 2015-09-09 北京航空航天大学 Electroencephalogram signal time-frequency analysis method based on multi-scale radial basis function and improved particle swarm optimization algorithm
WO2017014599A1 (en) * 2015-07-22 2017-01-26 주식회사 씨코헬스케어 Composition for stabilizing radiochemical purity of [18f]fluoro-dopa and method for preparing same
CN105447567A (en) * 2015-11-06 2016-03-30 重庆科技学院 BP neural network and MPSO algorithm-based aluminium electrolysis energy-saving and emission-reduction control method
CN105740887A (en) * 2016-01-26 2016-07-06 杭州电子科技大学 Electroencephalogram feature classification method based on PSO-SVM (Particle Swarm Optimization-Support Vector Machine)
CN106821681A (en) * 2017-02-27 2017-06-13 浙江工业大学 A kind of upper limbs ectoskeleton control method and system based on Mental imagery

Also Published As

Publication number Publication date
CN108319928A (en) 2018-07-24

Similar Documents

Publication Publication Date Title
CN108319928B (en) Deep learning method and system based on multi-target particle swarm optimization algorithm
CN113693613B (en) Electroencephalogram signal classification method, electroencephalogram signal classification device, computer equipment and storage medium
WO2021143353A1 (en) Gesture information processing method and apparatus, electronic device, and storage medium
Das et al. Unsupervised heart-rate estimation in wearables with Liquid states and a probabilistic readout
Du et al. A TSK-type convolutional recurrent fuzzy network for predicting driving fatigue
Adam et al. Feature selection using angle modulated simulated Kalman filter for peak classification of EEG signals
CN113693563B (en) Brain function network classification method based on hypergraph attention network
CN111544855A (en) Pure idea control intelligent rehabilitation method based on distillation learning and deep learning and application
Wang et al. An approach of one-vs-rest filter bank common spatial pattern and spiking neural networks for multiple motor imagery decoding
CN112043473A (en) Parallel nested and autonomous preferred classifier for brain-myoelectricity fusion perception of intelligent artificial limb
CN113111831A (en) Gesture recognition technology based on multi-mode information fusion
CN112528804A (en) Electromyographic signal noise reduction and classification method based on generation countermeasure network
Matsubara et al. Learning and adaptation of a stylistic myoelectric interface: EMG-based robotic control with individual user differences
Wu et al. Improving NeuCube spiking neural network for EEG-based pattern recognition using transfer learning
Li Hearing loss classification via AlexNet and extreme learning machine
Cai et al. Motor imagery decoding in the presence of distraction using graph sequence neural networks
Lv et al. Cognitive computing for brain–computer interface-based computational social digital twins systems
Zanghieri sEMG-based Hand gesture recognition with deep learning
Sinha et al. HSCAD: Heart sound classification for accurate diagnosis using machine learning and MATLAB
Bhalerao et al. Automatic detection of motor imagery EEG signals using swarm decomposition for robust BCI systems
CN110321856B (en) Time-frequency multi-scale divergence CSP brain-computer interface method and device
CN117322890A (en) Neuron group model-based generation type multichannel electroencephalogram modeling method
Venu et al. Optimized Deep Learning Model Using Modified Whale’s Optimization Algorithm for EEG Signal Classification
Zhang et al. Design on a wireless mechanomyography acquisition equipment and feature selection for lower limb motion recognition
Stuart et al. Deep learning shared bandpass filters for resource-constrained human activity recognition

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