CN112883914B - Multi-classifier combined mining robot idea sensing and decision making method - Google Patents
Multi-classifier combined mining robot idea sensing and decision making method Download PDFInfo
- Publication number
- CN112883914B CN112883914B CN202110296480.0A CN202110296480A CN112883914B CN 112883914 B CN112883914 B CN 112883914B CN 202110296480 A CN202110296480 A CN 202110296480A CN 112883914 B CN112883914 B CN 112883914B
- Authority
- CN
- China
- Prior art keywords
- classifier
- wavelet
- threshold
- algorithm
- 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
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/22—Matching criteria, e.g. proximity measures
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/25—Fusion techniques
- G06F18/253—Fusion techniques of extracted features
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; 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 OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/02—Preprocessing
- G06F2218/04—Denoising
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/08—Feature extraction
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/12—Classification; Matching
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
- Y02T10/00—Road transport of goods or passengers
- Y02T10/10—Internal combustion engine [ICE] based vehicles
- Y02T10/40—Engine management systems
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Health & Medical Sciences (AREA)
- Software Systems (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- Computational Linguistics (AREA)
- Health & Medical Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Measurement And Recording Of Electrical Phenomena And Electrical Characteristics Of The Living Body (AREA)
Abstract
The invention provides a multi-classifier combined mining robot idea sensing and decision-making method; step 1, data acquisition and preprocessing; step 2, changing a simple combination strategy of each similarity in template matching, and constructing a DCPM algorithm model; step 3, constructing a basic classifier, and constructing a linear discriminant classifier, an EEGNet neural network model and a DCPM algorithm; step 4, constructing a strong classifier, fusing characteristic information of the weak classifier, and outputting a control instruction of the mining robot; and 5, controlling the mining robot by using an unmanned aerial vehicle control instruction. The invention provides an improved wavelet threshold denoising method, which realizes the improvement of the filtering effect of electroencephalogram data. The invention uses the radial basis neural network to fuse the characteristic information extracted by the linear discriminant analysis, EEGNet neural network model and the improved DCPM algorithm, and outputs the control instruction of the mining robot to realize the perception and decision of the mining robot.
Description
Technical Field
The invention relates to the technical field of robots; in particular to a multi-classifier combined mining robot idea sensing and decision method.
Background
China is a large country for coal production and consumption, and the total amount of coal resources in China far exceeds other primary energy sources. According to the current situation of resources in China and the importance of coal in energy production and consumption structures, the energy structure taking coal as a main body in China can last for a quite long time. For a long time, coal is taken as a basic energy source and an important industrial raw material in China, so that strong power is provided for the development of national economy, and a foundation is laid for industries such as electric power, chemical industry, steel and the like. In view of the certain risk in the coal production process, china is highly concerned about the safe production of coal. At present, the safety production of coal is strictly regulated according to laws and regulations such as ' safety production method ', ' mine safety method ', ' safety production license regulations ', ' safety production safety seven regulations ' of mine ore length protection miners ' life safety in China. However, because of the large scale of coal production in China, coal mine safety accidents occur. The unmanned and the less-unmanned mine operation become trends for continuously advancing the safe production process of the coal mine industry;
in the aspect of extracting the characteristics based on the electroencephalogram data, several mature electroencephalogram characteristic extraction technologies have been developed at present, different original electroencephalogram characteristic information required by each method is obtained by each method, the data requirements among the methods are the same, but the difference of the extracted characteristic information is large, island of the characteristic information among different methods is caused, and further improvement of the performance of a characteristic extraction model on the basis of the existing research is not facilitated. Electroencephalogram data often contains a large amount of noise which masks weaker electroencephalogram signals, and denoising is a key step in electroencephalogram signal processing.
At present, wavelet transformation is a powerful tool for denoising brain electrical signals. The wavelet denoising method based on threshold processing is proposed by Donoho et al after deeply discussing wavelet transformation theory at the end of twentieth century, according to the theoretical basis that the difference of the coefficients of both the noise and useful signals exists after wavelet transformation, the part with the wavelet coefficient smaller than lambda is regarded as being caused by white noise by setting a proper critical value lambda, the part with the wavelet coefficient smaller than lambda is rejected, and the part with the wavelet coefficient larger than lambda is regarded as being caused by effective signals and is reserved, so that the removal of noise in the original signals can be completed. The original wavelet thresholding method includes a hard threshold and a soft threshold, and is widely applied in the field of denoising in view of its simplicity and effectiveness.
Assuming that the original signal is a compliant gaussian white noise, the noisy signal after adding noise can be expressed as:
f(t)＝s(t)+n(t)
performing discrete wavelet transformation on the noisy signal obtained by the formula to obtain a wavelet coefficient, wherein the wavelet coefficient comprises the following two parts:
wf(j,k)＝ws(j,k)+wn(j,k),j＝0,1,2,…,J；k＝0,1,2,…,N
wherein wf (j, k) is the wavelet coefficient value of the noisy signal f (t) at the j-th layer; ws (j, k) is the wavelet coefficient value of the original signal s (t) at the j-th layer; wn (j, k) is the wavelet coefficient value of the gaussian white noise signal n (t) at the j-th layer; j represents the maximum value of the wavelet decomposition layer number; n represents the signal length. Let wf (j, k) be wj, k, then select the appropriate threshold λ, process wj, k to obtain thresholded wavelet coefficient values that are as close as possible to ws (j, k). And finally, reconstructing to obtain a noise reduction signal.
The difference between the soft and hard thresholding functions in the traditional thresholding method proposed by Donoho is that the processing strategy for the wavelet coefficients is different. The expressions of the two functions are respectively:
the above equation is the hard threshold function and the soft threshold function, respectively. Lambda is a set threshold.
The hard threshold function has discontinuities at points-lambda and lambda, and poor continuity can cause the reconstructed signal to oscillate, i.e., pseudo-gibbs, thereby affecting the signal's smoothness. The signal processed by the soft threshold function has better overall continuity, but the wavelet coefficient obtained after thresholding has obvious difference from the actual wavelet coefficient, so that part of high-frequency information is lost, the signal distortion is easily caused by too large error, and the derivative of the signal is not continuous, so that mathematical analysis of the signal is very difficult.
The traditional threshold processing is really simple to realize, and the hard threshold denoising method can obtain the optimal estimation of the original signal, but the pseudo Gibbs phenomenon exists because of discontinuous functions; the soft threshold denoising method is characterized in that the obtained signal is smooth enough although the soft threshold denoising method is kept continuous at the threshold, the derivative of the soft threshold denoising method is discontinuous, so that constant deviation exists between the soft threshold denoising method and the original signal, errors in the reconstructed signal are unavoidable, and the denoising effect is affected.
In order to overcome the defects of the traditional threshold method, a plurality of experts study the optimization of the threshold function, so that the optimal compromise is achieved on the basis of the soft threshold and the hard threshold, the optimal estimation of the hard threshold function can be reserved, and a smooth signal can be obtained. However, most of the optimization functions have uncertain parameters, and a large number of experiments are needed to determine the values of the optimal parameters, so that the complexity only increases the difficulty of the experiments.
Disclosure of Invention
The invention aims to provide a multi-classifier combined mining robot idea sensing and deciding method.
The invention is realized by the following technical scheme:
the invention firstly acquires the multi-mode brain electrical signals of steady-state visual induction and motor imagery through an brain electrical signal acquisition experiment. Electroencephalogram data often contains a large amount of noise which masks weaker electroencephalogram signals, and denoising is a key step in electroencephalogram signal processing. After preprocessing of electroencephalogram data is completed, the invention builds a linear discriminant analysis, an EEGNet neural network model and an improved DCPM algorithm as basic classifiers.
The mining robot related by the invention mainly comprises five modules: the device comprises an electroencephalogram signal acquisition module, an electroencephalogram signal analysis module, a control instruction conversion module, a control signal transmission module and a control equipment module.
The invention relates to a multi-classifier combined mining robot idea sensing and deciding method, which comprises the following steps:
step 1, data acquisition and preprocessing, electroencephalogram acquisition experiments are carried out, multi-mode electroencephalogram data of steady-state vision and motor imagery are acquired, and the electroencephalogram data are subjected to filtering, artifact removal, baseline correction and 50Hz notch preprocessing;
step 2, changing a simple combination strategy of each similarity in template matching, and constructing a DCPM algorithm model;
step 3, constructing a basic classifier, namely constructing a linear discriminant classifier, an EEGNet neural network model and a DCPM algorithm by taking steady-state vision-induced and motor imagery multi-mode electroencephalogram signals as research objects;
step 4, constructing a strong classifier, constructing a neural network fusion model of a weak classifier based on a radial basis neural network, fusing characteristic information of the weak classifier, and outputting a control instruction of the mining robot;
and 5, controlling the mining robot by using an unmanned aerial vehicle control instruction.
Preferably, in step 1, the preprocessing of the electroencephalogram data for filtering, removing artifacts, correcting a baseline and carrying out 50Hz notch adopts an improved wavelet threshold denoising method.
The improved wavelet threshold denoising method adopts a new continuous and high-order conductive threshold function, and wavelet coefficients after threshold interception are as follows:
wherein j represents the number of wavelet decomposition layers, lambda j The given threshold value wj and k are the wavelet coefficient values on the j-th layer scale obtained by decomposition, and the wavelet coefficient values on the j-th layer scale after thresholding. The threshold function is free of uncertain parameters, and the threshold size can be adaptively adjusted along with the change of the decomposition order j. When one by oneGradually increasing, gradually tending towards wj, k; when gradually tending to lambda j At the time, gradually go to 0 or lambda j . Due to the threshold function being + -lambda j The process is continuous and the threshold function is higher-order conductive when the absolute value of the wavelet coefficients is satisfied.
The trend of the function is similar to that of the soft threshold function, the same continuity and smoothness characteristics as those of the soft threshold function are maintained, and no step phenomenon exists; and when |w j,k |→λ j In the time-course of which the first and second contact surfaces,the curve trend of the new function tends to the hard threshold function at the moment, so that the optimal estimation of the original signal can be achieved, and the defect that the soft threshold function has deviation from the original signal after being processed is avoided. Therefore, the threshold function can intuitively keep the signal smooth and continuous and can also keep effective information of the signal.
When the signal oscillates at the point of discontinuity, the denoising effect of wavelet threshold denoising is not ideal. The shift-invariant (TI) wavelet thresholding method changes the position of the singular point in the whole signal by changing the time domain order of the noisy signal, and is effective for eliminating the oscillations generated around the singular point of the signal at the time of wavelet transformation and thresholding, and the pseudo Gibbs phenomenon can be effectively suppressed.
Preferably, in step 2, the similarity difference combining strategies in the changing template matching are used to implement the improvement of the DCPM algorithm.
Preferably, in step 3, the multi-modal brain electrical signals induced by steady-state vision and imagined by motion are used to realize the perception and decision of the mine robot.
Preferably, in step 4, a multi-classifier based on a radial basis neural network is combined with a strong classifier to construct a strategy, a linear discriminant analysis, an EEGNet neural network and an improved DCPM algorithm are used as a basic network of a weak classifier, and the radial basis neural network is used as a combined network of the weak classifier to construct the strong classifier of the electroencephalogram signal.
The linear discriminant analysis is an algorithm for finding features that characterize or separate two or more classes of objects or events, and is commonly used in the fields of statistics, pattern recognition, and machine learning. The feature linear combination obtained by utilizing the linear discriminant analysis can be used for reducing the dimension of the data before classification or directly realizing the classification of the data. The linear discriminant analysis method has small dependence on the number of samples, does not need the problems of learning parameters, optimizing weights, selecting neuron activation functions and the like in the neural network, and is simple to realize and has better generalization capability.
The linear discriminant analysis is a supervised data dimension reduction and classification means, and the core idea of the algorithm is to spatially map the original data samples, so that the intra-class distance of the mapped data samples is as small as possible, and the inter-class distance is as large as possible. Linear discriminant analysis also seeks linear combinations of variables that best explain the data, as compared to principal component analysis algorithms, except that principal component analysis algorithms target processed data variances, and linear discriminant analysis algorithms supervise data labels, modeling with differences between data categories. Therefore, the linear discriminant analysis algorithm is more suitable for data processing with labels.
In the process of classifying electroencephalogram data by using linear discriminant analysis, firstly, frequency domain features of electroencephalogram signals are obtained by using wavelet transformation, and then the frequency domain features are used as analysis objects of the linear discriminant analysis. The wavelet transform can be considered an extension of the fourier transform, the algorithm working in a multi-scale manner to overcome the drawbacks of working on a single scale (time or frequency). Wavelet is a time-frequency analysis method that represents a special type of linear transformation of a signal, which is more efficient in terms of signal analysis than other transformation methods (e.g., fourier transform, short-time fourier transform). In wavelet analysis, instead of checking the entire signal through the same window, different portions of the wave are checked through windows of different sizes. The high frequency part of the signal uses a small window to provide good time resolution, while the low frequency part uses a large window to obtain good frequency information. Thus, the wavelet transform can provide both time and frequency information.
The processing procedure of the wavelet transformation to the signal in time-frequency transformation is specifically as follows:
let ψ (t) be the square integrable function, if the fourier transform of this function ψ (ω) satisfies the tolerability condition:
then ψ (t) can be defined as the wavelet mother function. The wavelet base function can be obtained by carrying out translation and expansion transformation on the wavelet mother function:
wherein a is a scale factor and b is a translation factor.
If the function f (t) is square, i.e. f (t) =l 2 (R), the continuous wavelet transform of function f (t) is defined as:
the reconstruction formula after successive wavelets of the function f (t) is as follows:
in the method, in the process of the invention,
the electroencephalogram data acquired by using the electroencephalogram acquisition equipment is generally discrete signals, so that continuous wavelet transformation is required to be subjected to discretization processing, and a discrete wavelet transformation formula is acquired. Scale factor a 0 And a translation factor b 0 Discretizing into a respectively 0 j And ka 0 j b 0 The discrete wavelet transform function is as follows:
the discrete wavelet coefficients are as follows:
the reconstruction formula of the discrete wavelet transform is as follows:
f(t)＝C∑ j∈N,k∈Z WT f (j,k)ψ j,k (t)
after obtaining frequency domain data of brain electrical data by wavelet transformation, assume that the frequency domain data is:
X＝{x 1 ,x 2 ,…,x N }
wherein x is i Represents the i-th sample, where the electroencephalogram data contains both fatigue and non-fatigue samples, so i e {0,1}, N represents the total number of samples, each sample being represented by M features, i.e., each sample is represented as a point (x i ∈R M )。
Let the sample mean of class i be μ i The total mean value of the samples is μ. Inter-class divergence S of class i Bi Mean μ representing class i i Distance from the total mean μ.
Inter-class divergence S of class i Bi The following is shown:
S Bi ＝(μ i -μ)(μ i -μ) T
assuming that the dimension after the sample is processed by the linear discriminant analysis is c, the divergence S of the processed sample B The following is shown:
wherein n is i Indicating the number of samples in class i.
After the samples are spatially mapped using linear discriminant analysis, the processed inter-class distances are as follows:
(m i -m) 2 ＝(W T μ i -W T μ) 2
＝W T (μ i -μ)(μ i -μ) T W
＝W T S Bi W
wherein m is i Representing the projection of the i-th class mean, m representing the projection of the total mean of all classes, and W representing the conversion matrix of the LDA algorithm.
The intra-class distance is represented by the sum of the average distance of each class of samples from the class, S is set Wi Intra-class divergence for the i-th class of samples is as follows:
S Wi ＝∑(x-μ i ) 2
intra-class divergence S of samples W The following is shown:
the linear discriminant analysis requires that the distance between classes of the processed data samples is as large as possible and the distance between classes is as small as possible, namely
S B W＝λS W W
S W -1 S B W＝λW
The optimal projection vector of the linear discriminant analysis is S W -1 S B A feature vector corresponding to the maximum feature value of (a).
The principle of the invention is as follows:
firstly, acquiring multi-mode brain electrical data of steady-state visual induction and motor imagery; the multi-mode electroencephalogram data of steady-state visual induction and motor imagery obtained by the invention uses NuAmps series electroencephalogram acquisition equipment of Neuroscan company in the United states as electroencephalogram acquisition equipment of the experiment. And recording the electroencephalogram signals of the tested person under two modes of steady-state visual induction and motor imagery, and obtaining the required basic electroencephalogram signal data.
Second, data preprocessing: the invention uses the data preprocessing modes of artifact elimination, filtering, data normalization and 50Hz notch for the multi-mode brain signal. In the aspect of preprocessing of the electroencephalogram data, the invention provides an improved wavelet threshold denoising method, and the improvement of the filtering effect of the electroencephalogram data is realized.
Thirdly, constructing a basic classifier: aiming at multi-mode electroencephalogram data classification of steady-state visual induction and motor imagery, the invention constructs a linear discriminant analysis, an EEGNet neural network model and an improved DCPM algorithm base classifier. The method changes a simple combination strategy of each similarity in the template matching according to the difference of each similarity judgment standard in the DCPM algorithm template matching, sets a difference weight for each similarity, and realizes the performance optimization of the DCPM algorithm.
Fourth, based on a multi-classifier model supporting radial basis: the invention uses the radial basis neural network to fuse the characteristic information extracted by the linear discriminant analysis, EEGNet neural network model and the improved DCPM algorithm, and outputs the control instruction of the mining robot to realize the perception and decision of the mining robot.
The invention has the following advantages: the invention uses the data preprocessing modes of artifact elimination, filtering, data normalization and 50Hz notch for the multi-mode brain signal. In the aspect of preprocessing of the electroencephalogram data, the invention provides an improved wavelet threshold denoising method, and the improvement of the filtering effect of the electroencephalogram data is realized. The invention uses the radial basis neural network to fuse the characteristic information extracted by the linear discriminant analysis, EEGNet neural network model and the improved DCPM algorithm, and outputs the control instruction of the mining robot to realize the perception and decision of the mining robot.
Drawings
FIG. 1 is a schematic diagram of a mining robot perception and decision making;
FIG. 2 is a diagram of the neural network architecture of EEGNet;
FIG. 3 is a block diagram of a DCPM algorithm;
FIG. 4 is a diagram of a multi-classifier fusion model based on RBF neural networks;
fig. 5 is a general design block diagram of a mining robot.
Detailed Description
The present invention will be described in detail with reference to specific examples. It should be noted that the following examples are only further illustrative of the present invention, but the scope of the present invention is not limited to the following examples.
Examples
The mining robot according to this embodiment is mainly composed of five modules: the device comprises an electroencephalogram signal acquisition module, an electroencephalogram signal analysis module, a control instruction conversion module, a control signal transmission module and a control equipment module. The overall design block diagram of the mining robot is shown in fig. 5.
The embodiment relates to a multi-classifier combined mining robot idea sensing and decision method, which is shown in fig. 1: the method comprises the following steps:
step 1, data acquisition and preprocessing, electroencephalogram acquisition experiments are carried out, multi-mode electroencephalogram data of steady-state vision and motor imagery are acquired, and the electroencephalogram data are subjected to filtering, artifact removal, baseline correction and 50Hz notch preprocessing;
step 2, changing a simple combination strategy of each similarity in template matching, and constructing a DCPM algorithm model;
step 3, constructing a basic classifier, namely constructing a linear discriminant classifier, an EEGNet neural network model and a DCPM algorithm by taking steady-state vision-induced and motor imagery multi-mode electroencephalogram signals as research objects;
step 4, constructing a strong classifier, constructing a neural network fusion model of a weak classifier based on a radial basis neural network, fusing characteristic information of the weak classifier, and outputting a control instruction of the mining robot;
and 5, controlling the mining robot by using an unmanned aerial vehicle control instruction.
Preferably, in step 1, the preprocessing of the electroencephalogram data for filtering, removing artifacts, correcting a baseline and carrying out 50Hz notch adopts an improved wavelet threshold denoising method.
The improved wavelet threshold denoising method adopts a new continuous and high-order conductive threshold function, and wavelet coefficients after threshold interception are as follows:
wherein j represents the number of wavelet decomposition layers, lambda j The given threshold value wj and k are the wavelet coefficient values on the j-th layer scale obtained by decomposition, and the wavelet coefficient values on the j-th layer scale after thresholding. The threshold function is free of uncertain parameters, and the threshold size can be adaptively adjusted along with the change of the decomposition order j. When gradually increased, gradually tends to wj, k; when gradually tending to lambda j At the time, gradually go to 0 or lambda j . Due to the threshold function being + -lambda j The process is continuous and the threshold function is higher-order conductive when the absolute value of the wavelet coefficients is satisfied.
The trend of the function is similar to that of the soft threshold function, the same continuity and smoothness characteristics as those of the soft threshold function are maintained, and no step phenomenon exists; and when |w j,k |→λ j In the time-course of which the first and second contact surfaces,the curve trend of the new function tends to the hard threshold function at the moment, so that the optimal estimation of the original signal can be achieved, and the defect that the soft threshold function has deviation from the original signal after being processed is avoided. Therefore, the threshold function can intuitively keep the signal smooth and continuous and can also keep effective information of the signal.
When the signal oscillates at the point of discontinuity, the denoising effect of wavelet threshold denoising is not ideal. The shift-invariant (TI) wavelet thresholding method changes the position of the singular point in the whole signal by changing the time domain order of the noisy signal, and is effective for eliminating the oscillations generated around the singular point of the signal at the time of wavelet transformation and thresholding, and the pseudo Gibbs phenomenon can be effectively suppressed.
Preferably, in step 2, the similarity difference combining strategies in the changing template matching are used to implement the improvement of the DCPM algorithm.
Preferably, in step 3, the multi-modal brain electrical signals induced by steady-state vision and imagined by motion are used to realize the perception and decision of the mine robot.
Preferably, in step 4, a multi-classifier based on a radial basis neural network is combined with a strong classifier to construct a strategy, a linear discriminant analysis, an EEGNet neural network and an improved DCPM algorithm are used as a basic network of a weak classifier, and the radial basis neural network is used as a combined network of the weak classifier to construct the strong classifier of the electroencephalogram signal.
The linear discriminant analysis is an algorithm for finding features that characterize or separate two or more classes of objects or events, and is commonly used in the fields of statistics, pattern recognition, and machine learning. The feature linear combination obtained by utilizing the linear discriminant analysis can be used for reducing the dimension of the data before classification or directly realizing the classification of the data. The linear discriminant analysis method has small dependence on the number of samples, does not need the problems of learning parameters, optimizing weights, selecting neuron activation functions and the like in the neural network, and is simple to realize and has better generalization capability.
The linear discriminant analysis is a supervised data dimension reduction and classification means, and the core idea of the algorithm is to spatially map the original data samples, so that the intra-class distance of the mapped data samples is as small as possible, and the inter-class distance is as large as possible. Linear discriminant analysis also seeks linear combinations of variables that best explain the data, as compared to principal component analysis algorithms, except that principal component analysis algorithms target processed data variances, and linear discriminant analysis algorithms supervise data labels, modeling with differences between data categories. Therefore, the linear discriminant analysis algorithm is more suitable for data processing with labels.
In the process of classifying electroencephalogram data by using linear discriminant analysis, firstly, frequency domain features of electroencephalogram signals are obtained by using wavelet transformation, and then the frequency domain features are used as analysis objects of the linear discriminant analysis. The wavelet transform can be considered an extension of the fourier transform, the algorithm working in a multi-scale manner to overcome the drawbacks of working on a single scale (time or frequency). Wavelet is a time-frequency analysis method that represents a special type of linear transformation of a signal, which is more efficient in terms of signal analysis than other transformation methods (e.g., fourier transform, short-time fourier transform). In wavelet analysis, instead of checking the entire signal through the same window, different portions of the wave are checked through windows of different sizes. The high frequency part of the signal uses a small window to provide good time resolution, while the low frequency part uses a large window to obtain good frequency information. Thus, the wavelet transform can provide both time and frequency information.
The processing procedure of the wavelet transformation to the signal in time-frequency transformation is specifically as follows:
let ψ (t) be the square integrable function, if the fourier transform of this function ψ (ω) satisfies the tolerability condition:
then ψ (t) can be defined as the wavelet mother function. The wavelet base function can be obtained by carrying out translation and expansion transformation on the wavelet mother function:
wherein a is a scale factor and b is a translation factor.
If the function f (t) is square, i.e. f (t) =l 2 (R), the continuous wavelet transform of function f (t) is defined as:
the reconstruction formula after successive wavelets of the function f (t) is as follows:
in the method, in the process of the invention,
the electroencephalogram data acquired by using the electroencephalogram acquisition equipment is generally discrete signals, so that continuous wavelet transformation is required to be subjected to discretization processing, and a discrete wavelet transformation formula is acquired. Scale factor a 0 And a translation factor b 0 Discretizing into a respectively 0 j And ka 0 j b 0 The discrete wavelet transform function is as follows:
the discrete wavelet coefficients are as follows:
the reconstruction formula of the discrete wavelet transform is as follows:
f(t)＝C∑ j∈N,k∈Z WT f (j,k)ψ j,k (t)
after obtaining frequency domain data of brain electrical data by wavelet transformation, assume that the frequency domain data is:
X＝{x 1 ,x 2 ,…,x N }
wherein x is i Represents the i-th sample, where the electroencephalogram data contains both fatigue and non-fatigue samples, so i e {0,1}, N represents the total number of samples, each sample represented by M features, i.e., each sample represented as one in M-dimensional spacePoint (x) i ∈R M )。
Let the sample mean of class i be μ i The total mean value of the samples is μ. Inter-class divergence S of class i Bi Mean μ representing class i i Distance from the total mean μ.
Inter-class divergence S of class i Bi The following is shown:
S Bi ＝(μ i -μ)(μ i -μ) T
assuming that the dimension after the sample is processed by the linear discriminant analysis is c, the divergence S of the processed sample B The following is shown:
wherein n is i Indicating the number of samples in class i.
After the samples are spatially mapped using linear discriminant analysis, the processed inter-class distances are as follows:
(m i -m) 2 ＝(W T μ i -W T μ) 2
＝W T (μ i -μ)(μ i -μ) T W
＝W T S Bi W
wherein m is i Representing the projection of the i-th class mean, m representing the projection of the total mean of all classes, and W representing the conversion matrix of the LDA algorithm.
The intra-class distance is represented by the sum of the average distance of each class of samples from the class, S is set Wi Intra-class divergence for the i-th class of samples is as follows:
S Wi ＝∑(x-μ i ) 2
intra-class divergence S of samples W The following is shown:
the linear discriminant analysis requires that the distance between classes of the processed data samples is as large as possible and the distance between classes is as small as possible, namely
S B W＝λS W W
S W -1 S B W＝λW
The optimal projection vector of the linear discriminant analysis is S W -1 S B A feature vector corresponding to the maximum feature value of (a).
In step 3, the EEGNet neural network model is used as another basic model of the classifier. EEGNet is a mature convolutional neural network model with excellent performance in a deep learning algorithm for extracting characteristics of an electroencephalogram signal. The model is a compact convolutional neural network specially designed for electroencephalogram characteristic extraction, and the model refers to a depth decomposable convolutional neural network model used in computer vision to construct a depth neural network special for electroencephalogram. The network encapsulates several common electroencephalogram feature extraction methods, such as an optimal spatial filter, and adapts to the task of electroencephalogram data processing while reducing training parameters. The neural network structure of EEGNet is shown in fig. 2.
The discriminant typical pattern matching algorithm (DCPM) is a novel electroencephalogram signal processing algorithm which appears in recent years. The DCPM algorithm uses a discriminant space mode algorithm (DSP) and a typical modal analysis algorithm (CCA) to construct weak classification, and the strong classifier effect of the DCPM algorithm is achieved through performance fusion of the weak classifier. The DCPM algorithm consists of three main components: and judging the construction of a spatial mode, the construction of a CCA template and the template matching. The structure of the DCPM algorithm is shown in fig. 3.
Let the parameter k e 1,2 denote the number of modes,for training samples, < >>To test a sample, N c Represents the number of channels, N t Representing time series, N s Is the experiment number>Is a template of category k, +.>And->The covariance matrix of (2) is:
x 1 and x 2 The variance of (2) is:
the projection matrix W constructed by the DSP algorithm realizes the distinction of two types of templates after DSP transformation.
S B ＝Σ 11 +Σ 22 -Σ 12 -Σ 21
Wherein lambda is i Is the eigenvalue of the projection matrix W.
After common mode noise in brain electricity is eliminated by using a DSP algorithm, a projection matrix U is constructed by using a CCA algorithm k And V k Realization ofAnd W is T Intrinsic correlation of Y.
Where ε represents the expected value.
In pattern matching, the similarity between the training template and the test signal is calculated as follows:
where corr represents the pearson correlation coefficient and dist represents the euclidean distance.
The more the test sample Y matches the template k, then itThe larger. The class determination for test sample Y is as follows:
considering that the similarity judgment standards in the template matching are different, the judgment of the template matching should provide information of the difference. Therefore, the invention changes the simple combination strategy of each similarity in the template matching, sets the difference weight value for each similarity, and realizes the performance optimization of the DCPM algorithm.
In the formula, h i For ρ ki Is used to determine the differential weight of the model.
Radial Basis Function (RBF) is a well-behaved forward network whose excitation function is typically a gaussian function. The RBF network has the performances of optimal approximation, simple training, fast learning convergence speed and overcoming the problem of local minimum values, and has the capability of approximating any continuous function with any precision, global approximation, fundamentally solves the problem of local optimum of the BP network, and has compact topological structure, separable learning of structural parameters and fast convergence speed. It has been widely used in the fields of pattern recognition, nonlinear control, image processing, and the like.
The invention uses radial basis functions as a multi-classification fusion neural network model, the multi-modal brain electrical characteristics extracted by a basic classifier are used as the input of the radial basis neural network, the model outputs the control instruction of the mining robot, and the multi-classifier fusion model based on the RBF neural network is shown in figure 4.
The foregoing describes specific embodiments of the present invention. It is to be understood that the invention is not limited to the particular embodiments described above, and that various changes and modifications may be made by one skilled in the art within the scope of the claims without affecting the spirit of the invention.
Claims (1)
1. The mining robot idea sensing and deciding method combining multiple classifiers is characterized by comprising the following steps:
step 1, data acquisition and preprocessing, electroencephalogram acquisition experiments are carried out, multi-mode electroencephalogram data of steady-state vision and motor imagery are acquired, and the electroencephalogram data are subjected to filtering, artifact removal, baseline correction and 50Hz notch preprocessing;
step 2, changing a simple combination strategy of each similarity in template matching, and constructing a DCPM algorithm model;
step 3, constructing a basic classifier, and constructing a linear discriminant classifier, an EEGNet neural network model and a DCPM algorithm;
step 4, constructing a strong classifier, fusing characteristic information of the weak classifier, and outputting a control instruction of the mining robot;
step 5, controlling the mining robot by using an unmanned aerial vehicle control instruction;
in the step 1, the preprocessing of the electroencephalogram data for filtering, removing artifacts, correcting a base line and carrying out 50Hz notch adopts an improved wavelet threshold denoising method;
the improved wavelet threshold denoising method adopts a new continuous and high-order conductive threshold function, and wavelet coefficients after threshold interception are as follows:
wherein j represents the number of wavelet decomposition layers, lambda j The given threshold value wj and k are wavelet coefficient values on the j-th layer scale obtained by decomposition, and the wavelet coefficient values on the j-th layer scale after threshold processing; the threshold function does not contain uncertain parameters, and the threshold size can be adaptively adjusted along with the change of the decomposition order j; when gradually increased, gradually tends to wj, k; when gradually tending to lambda j At the time, gradually go to 0 or lambda j The method comprises the steps of carrying out a first treatment on the surface of the Due to the threshold function being + -lambda j The process is continuous, and when the absolute value of the wavelet coefficient is satisfied, the threshold function is higher-order conductive;
the trend of the function is similar to that of the soft threshold function, the same continuity and smoothness characteristics as those of the soft threshold function are maintained, and no step phenomenon exists; and when w is j,k →λ j In the time-course of which the first and second contact surfaces,now new functionThe curve trend of the (2) tends to the hard threshold function, so that the optimal estimation of the original signal can be achieved, and the defect that the soft threshold function is deviated from the original signal after being processed is avoided; the threshold function can intuitively keep signals smooth and continuous and can also keep effective information of the signals;
when the signal oscillates at the discontinuous point, the denoising effect of wavelet threshold denoising is not ideal; the translation invariant wavelet threshold method changes the position of the singular point in the whole signal by changing the time domain sequence of the noise-containing signal, and the method is effective for eliminating the oscillation generated around the singular point of the signal during wavelet transformation and thresholding, and the pseudo Gibbs phenomenon can be effectively inhibited;
in the step 2, the similarity difference combination strategies in the changing template matching are used for realizing the improvement of the DCPM algorithm;
in the step 3, the multi-mode brain electrical signals induced by steady-state vision and imagined by the movement are used for realizing the perception and decision of the mine robot;
in step 4, a multi-classifier based on a radial basis neural network is combined with a strong classifier to construct a strategy, linear discriminant analysis, EEGNet neural network and improved DCPM algorithm are used as a basic network of a weak classifier, and the radial basis neural network is used as a combined network of the weak classification to realize the strong classifier construction of the electroencephalogram signals.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110296480.0A CN112883914B (en) | 2021-03-19 | 2021-03-19 | Multi-classifier combined mining robot idea sensing and decision making method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110296480.0A CN112883914B (en) | 2021-03-19 | 2021-03-19 | Multi-classifier combined mining robot idea sensing and decision making method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN112883914A CN112883914A (en) | 2021-06-01 |
| CN112883914B true CN112883914B (en) | 2024-03-19 |
Family
ID=76041432
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110296480.0A Active CN112883914B (en) | 2021-03-19 | 2021-03-19 | Multi-classifier combined mining robot idea sensing and decision making method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112883914B (en) |
Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2016110005A1 (en) * | 2015-01-07 | 2016-07-14 | 深圳市唯特视科技有限公司 | Gray level and depth information based multi-layer fusion multi-modal face recognition device and method |
| CN106292705A (en) * | 2016-09-14 | 2017-01-04 | 东南大学 | Many rotor wing unmanned aerial vehicles idea remote control system based on Bluetooth brain wave earphone and operational approach |
| WO2017084416A1 (en) * | 2015-11-17 | 2017-05-26 | 天津大学 | Feedback system based on motor imagery brain-computer interface |
| CN107616880A (en) * | 2017-08-01 | 2018-01-23 | 南京邮电大学 | A kind of intelligent electric wheelchair implementation method based on brain electricity idea and deep learning |
| CN109271975A (en) * | 2018-11-19 | 2019-01-25 | 燕山大学 | A kind of electrical energy power quality disturbance recognition methods based on big data multi-feature extraction synergetic classification |
| KR20190027026A (en) * | 2017-09-04 | 2019-03-14 | 주식회사 하이 | Method of Discriminating Sound Type from Living Noise and Home Social Robot Management System for Sharing Daily Activity Information Using It |
| CN109598222A (en) * | 2018-11-26 | 2019-04-09 | 南开大学 | Wavelet neural network Mental imagery brain electricity classification method based on the enhancing of EEMD data |
| WO2020042511A1 (en) * | 2018-08-28 | 2020-03-05 | 天津大学 | Motion potential brain-machine interface encoding and decoding method based on spatial filtering and template matching |
| CN112114670A (en) * | 2020-09-10 | 2020-12-22 | 季华实验室 | Man-machine co-driving system based on hybrid brain-computer interface and control method thereof |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US9358685B2 (en) * | 2014-02-03 | 2016-06-07 | Brain Corporation | Apparatus and methods for control of robot actions based on corrective user inputs |
-
2021
- 2021-03-19 CN CN202110296480.0A patent/CN112883914B/en active Active
Patent Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2016110005A1 (en) * | 2015-01-07 | 2016-07-14 | 深圳市唯特视科技有限公司 | Gray level and depth information based multi-layer fusion multi-modal face recognition device and method |
| WO2017084416A1 (en) * | 2015-11-17 | 2017-05-26 | 天津大学 | Feedback system based on motor imagery brain-computer interface |
| CN106292705A (en) * | 2016-09-14 | 2017-01-04 | 东南大学 | Many rotor wing unmanned aerial vehicles idea remote control system based on Bluetooth brain wave earphone and operational approach |
| CN107616880A (en) * | 2017-08-01 | 2018-01-23 | 南京邮电大学 | A kind of intelligent electric wheelchair implementation method based on brain electricity idea and deep learning |
| KR20190027026A (en) * | 2017-09-04 | 2019-03-14 | 주식회사 하이 | Method of Discriminating Sound Type from Living Noise and Home Social Robot Management System for Sharing Daily Activity Information Using It |
| WO2020042511A1 (en) * | 2018-08-28 | 2020-03-05 | 天津大学 | Motion potential brain-machine interface encoding and decoding method based on spatial filtering and template matching |
| CN109271975A (en) * | 2018-11-19 | 2019-01-25 | 燕山大学 | A kind of electrical energy power quality disturbance recognition methods based on big data multi-feature extraction synergetic classification |
| CN109598222A (en) * | 2018-11-26 | 2019-04-09 | 南开大学 | Wavelet neural network Mental imagery brain electricity classification method based on the enhancing of EEMD data |
| CN112114670A (en) * | 2020-09-10 | 2020-12-22 | 季华实验室 | Man-machine co-driving system based on hybrid brain-computer interface and control method thereof |
Non-Patent Citations (2)
| Title |
|---|
| 在线脑机接口中脑电信号的特征提取与分类方法;徐宝国;宋爱国;费树岷;;电子学报(05);全文 * |
| 基于样本熵和模式识别的脑电信号识别算法研究;沈晓燕;王雪梅;王燕;;计算机工程与科学(08);全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN112883914A (en) | 2021-06-01 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Shi et al. | Using multiple-feature-spaces-based deep learning for tool condition monitoring in ultraprecision manufacturing | |
| Gong et al. | A fast anomaly diagnosis approach based on modified CNN and multisensor data fusion | |
| CN118606796B (en) | Complex electromagnetic environment signal sorting method and system based on deep learning and time-frequency transformation | |
| CN108846323A (en) | A kind of convolutional neural networks optimization method towards Underwater Targets Recognition | |
| He et al. | Rotation invariant texture descriptor using local shearlet-based energy histograms | |
| Ma et al. | Multimodal convolutional neural network model with information fusion for intelligent fault diagnosis in rotating machinery | |
| Ye et al. | Overhead ground wire detection by fusion global and local features and supervised learning method for a cable inspection robot | |
| CN117036786B (en) | Image anomaly detection method integrating AVAE and SE modules | |
| Liu et al. | Feature entropy adaptive network for weak magnetic signal classification | |
| CN119646611A (en) | Partial discharge detection network system based on adaptive attention mechanism | |
| Wang et al. | An automatic identification framework for complex power quality disturbances based on ensemble CNN | |
| Ge et al. | Shuffle-RDSNet: a method for side-scan sonar image classification with residual dual-path shrinkage network | |
| CN112883914B (en) | Multi-classifier combined mining robot idea sensing and decision making method | |
| CN115687970A (en) | Method for improving electromyographic signal identification accuracy | |
| Assafo et al. | Unsupervised and semisupervised machine learning frameworks for multiclass tool wear recognition | |
| Abuzneid et al. | Improving human face recognition using deep learning based image registration and multi-classifier approaches | |
| CN119360891A (en) | An abnormal sound detection method for belt conveyor fault diagnosis | |
| Baranwal et al. | Indian sign language gesture recognition using discrete wavelet packet transform | |
| CN119302668A (en) | Physiological signal segment analysis method based on adaptive bidirectional selection state space | |
| CN119293588A (en) | A bearing fault diagnosis method based on crested porcupine optimized variable mode decomposition | |
| CN115984602B (en) | Method and system for improving accuracy of point cloud classification based on graph spectrum domain | |
| Trabelsi et al. | Enhancing Traffic Sign Recognition Through Daubechies Discret Wavelet Transform and Convolutional Neural Networks | |
| Ren et al. | Understanding Deep Face Representation via Attribute Recovery | |
| Liu et al. | Prediction of aluminum electrolysis superheat based on improved relative density noise filter SMO | |
| Gong | Expression recognition method of fusion Gabor filter and 2DPCA algorithm |
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 |