CN101840506B - Remote education student characteristic signal extraction and recognition method - Google Patents
Remote education student characteristic signal extraction and recognition method Download PDFInfo
- Publication number
- CN101840506B CN101840506B CN201010140911.6A CN201010140911A CN101840506B CN 101840506 B CN101840506 B CN 101840506B CN 201010140911 A CN201010140911 A CN 201010140911A CN 101840506 B CN101840506 B CN 101840506B
- Authority
- CN
- China
- Prior art keywords
- msub
- mrow
- physiological signal
- physiological
- students
- 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.)
- Expired - Fee Related
Links
- 238000000034 method Methods 0.000 title claims abstract description 43
- 238000000605 extraction Methods 0.000 title claims abstract description 13
- 238000012549 training Methods 0.000 claims abstract description 55
- 238000012706 support-vector machine Methods 0.000 claims abstract description 48
- 239000013598 vector Substances 0.000 claims description 30
- 230000008451 emotion Effects 0.000 claims description 13
- 238000001914 filtration Methods 0.000 claims description 12
- 230000002996 emotional effect Effects 0.000 claims description 11
- 238000005070 sampling Methods 0.000 claims description 11
- 239000008280 blood Substances 0.000 claims description 9
- 210000004369 blood Anatomy 0.000 claims description 9
- 238000002790 cross-validation Methods 0.000 claims description 8
- 238000004364 calculation method Methods 0.000 claims description 7
- 238000009499 grossing Methods 0.000 claims description 7
- 230000010349 pulsation Effects 0.000 claims description 7
- 230000029058 respiratory gaseous exchange Effects 0.000 claims description 5
- 238000010845 search algorithm Methods 0.000 claims description 5
- 230000009466 transformation Effects 0.000 claims description 5
- 230000007423 decrease Effects 0.000 claims description 4
- 239000011159 matrix material Substances 0.000 claims description 4
- 210000003205 muscle Anatomy 0.000 claims description 2
- 238000012545 processing Methods 0.000 abstract description 5
- 230000007613 environmental effect Effects 0.000 abstract description 3
- 238000007781 pre-processing Methods 0.000 abstract description 3
- 230000002349 favourable effect Effects 0.000 abstract description 2
- 230000010365 information processing Effects 0.000 abstract description 2
- 230000033764 rhythmic process Effects 0.000 abstract description 2
- 230000006870 function Effects 0.000 description 10
- 238000004422 calculation algorithm Methods 0.000 description 8
- 230000008569 process Effects 0.000 description 5
- 238000005457 optimization Methods 0.000 description 4
- 230000005672 electromagnetic field Effects 0.000 description 3
- 238000012360 testing method Methods 0.000 description 3
- 230000009286 beneficial effect Effects 0.000 description 2
- 230000008859 change Effects 0.000 description 2
- 230000001427 coherent effect Effects 0.000 description 2
- 230000007547 defect Effects 0.000 description 2
- 230000009977 dual effect Effects 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 238000002567 electromyography Methods 0.000 description 2
- 230000008909 emotion recognition Effects 0.000 description 2
- 230000008921 facial expression Effects 0.000 description 2
- 230000014509 gene expression Effects 0.000 description 2
- 230000003183 myoelectrical effect Effects 0.000 description 2
- 230000000241 respiratory effect Effects 0.000 description 2
- 230000036387 respiratory rate Effects 0.000 description 2
- 230000004044 response Effects 0.000 description 2
- 230000001131 transforming effect Effects 0.000 description 2
- 230000006854 communication Effects 0.000 description 1
- 230000003247 decreasing effect Effects 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 238000005259 measurement Methods 0.000 description 1
- 238000010606 normalization Methods 0.000 description 1
- 238000003909 pattern recognition Methods 0.000 description 1
- 230000035790 physiological processes and functions Effects 0.000 description 1
Images
Landscapes
- Image Analysis (AREA)
- Measurement Of The Respiration, Hearing Ability, Form, And Blood Characteristics Of Living Organisms (AREA)
Abstract
The invention relates to a remote education student characteristic signal extraction and recognition method in the technical field of information processing. The method comprises the following steps of: eliminating the signal biasing caused by individual difference and environmental factors by preprocessing physiological signals of students; processing the physiological signal characteristics by using a continuous limited Boltzmann machine to reduce the characteristic difference due to the individual difference; inputting the information into a support vector machine for training to obtain a sorter; and finally performing mode real-time recognition. Therefore, the feedback information of the students on teaching methods can be effectively acquired, the average teaching feedback information accuracy of the students on site reaches 82.6 percent, and the teaching level of the on-site remote education is improved. The remote education student characteristic signal extraction and recognition method can be widely applied to the remote education, can directly acquire the feedback information on the teaching acceptance degree of the students who take part in the remote education, timely adjusts and improves the teaching rhythm and level, is favorable for the remote education, and has obvious advantages on the sorting accuracy.
Description
Technical Field
The invention relates to a detection method in the technical field of information processing, in particular to a method for extracting and identifying characteristic signals of remote education students.
Background
With the gradual development of E-Learning, distance education techniques have been widely accepted by society. In the present distance education mode, because do not have face-to-face interchange process between mr and the student, mr can't in time know the student and go up the interest degree to the course content when class, also can't change according to the emotional state that the student shows in the course of going to class, suitably adjusts the mode of giving lessons, and this will influence distance education's teaching quality. Therefore, the method for identifying the emotional state of the student in real time through the computer can make up for the defects in the existing distance education mode and improve the teaching efficiency of distance education.
The document retrieval of the prior art finds that most of the existing emotion recognition methods use characteristics such as voice, facial expressions and the like, and then recognize the emotional state of the user through corresponding characteristic extraction methods. For example, chinese patent application No. 200610097301.6 discloses a speech emotion recognition method based on a support vector machine. There are many common problems with this technology, for example, in actual classroom teaching, students rarely have obvious facial expression changes, and there are not much voice communication processes. In such an environment lacking speech and expression changes, the existing methods are difficult to apply.
Disclosure of Invention
Aiming at the defects of the prior art, the invention provides a method for extracting and identifying characteristic signals of remote education students. The method can eliminate signal offset caused by individual difference and environmental factors by preprocessing physiological signals of remote education students, can effectively obtain feedback information of the students on teaching modes by utilizing a continuous limited wave Alzheimer machine and a support vector machine and finally carrying out mode real-time identification, and is favorable for improving the teaching level of on-site remote education.
The invention is realized by the following technical scheme, which comprises the following steps:
step one, filtering the noise of the physiological signal after receiving the signal of the physiological signal sensor.
Firstly, low-pass filtering is carried out on an input signal, and the interference of external electromagnetic fields such as alternating current and the like on the signal is weakened; the low pass filtered signal is then smoothed using a smoothing filter, which filters out gaussian random noise that may be introduced due to slight limb movements of the student.
The low-pass filtering means: the high-frequency component in the filtering signal is only allowed to pass through, and the function of the filtering signal is to filter the interference of the external electromagnetic field with high frequency to the input signal.
The smoothing filtering means: by processing the signals in the signal point time sequence field as the output signal points after filtering, the interference of random noise to physiological signals can be reduced.
And step two, extracting and determining a physiological signal reference signal.
The physiological signals of the students in the calm state are collected, and after the physiological signals are collected for a period of time continuously, the physiological signals of the students in the period of time are stored and extracted to be used as the current physiological signal reference.
The physiological signal reference refers to: under different environments or physiological states of students, the bias of physiological signals caused by external or internal non-emotional factors is accurately estimated, and the distortion of the physiological signals caused by the non-emotional factors can be effectively removed.
And step three, continuously collecting physiological signals of the students in class, segmenting the physiological signals according to a time window with fixed width, and taking each segment of physiological signals as a group of signal segments for the emotion stage.
The time window with fixed width refers to: a fixed length time interval is calculated from the physiological signal in the time interval to obtain a set of features that can be used for the emotional phase.
The signal segment refers to: the set of signals within the fixed time window described above. One signal segment is the data required for one emotional phase.
And step four, extracting the physiological signal characteristics.
The physiological signal feature extraction comprises time domain physiological signal feature extraction and frequency domain physiological signal feature extraction: such as: the time-domain physiological signal characteristics include: skin conductance, myoelectric response, blood volume pulsation, respiratory amplitude characteristics; the frequency domain physiological signal characteristics are obtained by transforming the physiological signal into a frequency domain through fast Fourier transform, and comprise: blood volume pulsation, heart rate variability high frequency energy to low frequency energy ratio, respiratory rate characteristics.
And step five, processing the physiological signal characteristics by using a continuous restricted wave Alzheimer machine.
The utilization of the continuous restricted wave Alzheimer's machine is as follows: training is carried out by using the physiological signal characteristics obtained through calculation to obtain a physiological signal characteristic and a generation model for physiological signal characteristic processing, and a group of corresponding physiological signal characteristic information can be obtained through sampling according to the input physiological signal characteristics.
The restricted wave alzheimer machine is generally used for unsupervised learning in pattern recognition. It contains two layers, one called the visible layer and the other called the hidden layer. Each layer consists of a series of nodes, commonly referred to as "experts," the node state of the restricted wave alzheimer machine being binary, i.e., 0 or 1. There is a bias node in each layer whose state does not change. Nodes between layers are all connected in a non-directional mode, and nodes in the layers do not have connection, so that an even graph is formed. Each connection is assigned a weight value describing the degree of association between nodes at both ends of the connection. In general, a vector v represents the visible layer state of the visible layer, a vector h represents the hidden layer state of the visible layer, a vector b is the connection weight of the hidden layer bias unit and the visible layer, a vector c is the connection weight of the visible layer bias and the hidden layer, and W represents a weight matrix connecting the hidden layer and the visible layer. It establishes the following distribution for the energy distribution of the data space:
E(v，h)＝exp(-Energy(v，h))
wherein,
Energy(v，h)＝-bTv-cTh-hTWv
normalizing the energy to obtain a joint probability distribution of visible and hidden states:
by building a P (v, h) generative model, it can be considered that, given v, h' sampled from the model is an implicit representation of v under the model, i.e., a set of codes. Generally, the implicit expression can learn some high-dimensional abstract features in the visible layer data, and has an auxiliary effect of improving the accuracy rate of subsequent classifier training and classification.
When the restricted wave Alzheimer's machine is used, firstly, a weight matrix W and bias vectors b and c are initialized randomly, and then the model is fitted to the actual data space energy distribution through a training algorithm. There is currently a faster algorithm called the coherent contrast Divergence (coherent contrast Divergence) method. The specific algorithm is as follows:
1. maintaining a set of negative samples v-initially in an all 0 state;
2. initializing W, b and c as random smaller numerical values;
3. for each group of training data, namely a positive sample v +, a group of h + is obtained by sampling P (h + | v +);
4. for negative samples, a set of h-is obtained by sampling P (h- | v-);
5. reconstructing a group of negative samples v-' by sampling P (v- | h-) to be used as negative samples of next sampling;
6. calculating gradient g ═ v + h +T-v-h-T；
7. Update W ═ W + rwG, rw is the learning rate;
8. repeating the steps 4-7 until the maximum training period is reached.
The continuous confined wave alzheimer machine is characterized in that: a generalization of a confined wave Zeeman machine having the same structure as the confined wave Zeeman machine except that each node of the confined wave Zeeman machine is a Bernoulli (Bernoulli) distributed random node and each node of successive confined wave Zeeman machines is a sigmoid node to which Gaussian random noise is added. The states of the hidden nodes are as follows:
Nj(0,1) is a Gaussian distribution with 0 as the mean and 1 as the variance, wherein,
θLand thetaHRespectively the minimum and maximum values, a, that the node may takejIs a random noise control parameter, for visible layer nodes, can be ajSet as a fixed value, a of hidden nodejAnd training by a training algorithm.
The continuous training method of the restricted wave alzheimer's machine is basically the same as the training method of the restricted wave alzheimer's machine:
1. maintaining a set of negative samples v-initially in an all 0 state;
2. initializing W, b, c and a to be random smaller numerical values;
3. for each set of training data, i.e., positive samples v +, by calculating hj＝φj(∑iwijvi+σ·Ni(0.1)), yielding a set of h +;
4. for negative samples, by calculating hj＝φj(∑iwijvi+σ·Nj(0, 1)), yielding a set of h-;
5. by calculating vi＝φi(∑iwijhj+σ·Ni(0, 1)) reconstructing a set of negative samples v-as negative samples for the next sampling;
6. calculating gradient g ═ v + h +T-v-h-T；
7. Update W ═ W + rwG, rw is the learning rate;
8. updating <math>
<mrow>
<msub>
<mi>a</mi>
<mi>j</mi>
</msub>
<mo>=</mo>
<msub>
<mi>r</mi>
<mi>a</mi>
</msub>
<mo>·</mo>
<mfrac>
<mn>1</mn>
<msubsup>
<mi>a</mi>
<mi>j</mi>
<mn>2</mn>
</msubsup>
</mfrac>
<mo>·</mo>
<mrow>
<mo>(</mo>
<msub>
<mi>h</mi>
<mi>j</mi>
</msub>
<msup>
<mo>+</mo>
<mn>2</mn>
</msup>
<mo>-</mo>
<msub>
<mi>h</mi>
<mi>j</mi>
</msub>
<msup>
<mo>-</mo>
<mn>2</mn>
</msup>
<mo>)</mo>
</mrow>
<mo>,</mo>
</mrow>
</math> ra is the learning rate;
9. repeating the steps 4-8 until the maximum training period is reached.
The generation model refers to: after training, P (v, h) which is continuously and firstly depicted by the Boltzmann machine can obtain v and h which accord with the data distribution probability from the model through sampling.
The physiological signal characteristic coding means: from the generated model, under the condition of giving a visible layer state vector v, a hidden layer state vector h obtained by sampling is a state vector of a hidden layer node.
And step six, training a support vector machine. After the physiological signal characteristic code is obtained, firstly, student emotion category marking is carried out on the code, and marked training data are obtained. And then training a support vector machine model by using training data through cross validation and a grid search algorithm to obtain a group of support vectors, weight vectors corresponding to the support vectors and a group of optimal hyper-parameters.
And then, processing the physiological signal characteristics by using a continuous restricted wave Alzheimer machine to reduce characteristic difference caused by individual difference, inputting the information into a support vector machine to train to obtain a classifier, and finally, carrying out mode real-time identification by using the support vector machine classifier to obtain feedback information of the teaching mode of the students.
The emotion category marking refers to: through experiments, a series of physiological signal characteristic codes and corresponding emotion classes are obtained.
The cross validation refers to: a common model performance estimation method. Dividing all training data into N parts, taking one part as test data each time, taking the rest N-1 parts as training data, training on given parameters to obtain an optimal model, and testing the performance of the model by using the test data. And taking the average performance of N times, namely the classification accuracy as an output result of the cross validation.
The grid search algorithm is as follows: a common searching method of hyper-parameters. The method includes traversing parameters of all the step points in a jointed hyperparameter space according to a fixed step length in sequence, selecting a best hyperparameter point, and then continuously reducing the step length in the neighborhood of the hyperparameter point to continue traversing until the set minimum step length. The resulting hyper-parameter point is an estimate of the optimal hyper-parameter point.
The support vector machine is as follows: a method is useful for pattern classification. The method reduces the VC dimension of the model and improves the generalization capability of the model by solving the maximum boundary margin of the characteristic vector. The two-class support vector machine can solve the two-class classification problem, and the core idea is to solve the following optimization problem:
to maximize LD
Is limited by
The dual problem is to maximize LP
Where K (×) is the kernel function. This is a quadratic optimization problem with a unique solution. X corresponding to α whose solution value is not 0 is called a support vector, and α itself is a weight vector.
The two-classification support vector machine can construct a multi-classification support vector machine by a One-VS-One method. For an N-class problem, N x (N-1)/2 two-class support vector machines can be constructed, and a corresponding two-class support vector machine is trained for each two classes of data. For a given input, inputting into the N x (N-1)/2 two-classification support vector machines respectively, and voting by all the two-classification support vector machines to determine the final classification.
The optimal hyper-parameter refers to: some parameters of the model that are not related to the training data. These hyper-parameters can be used to search for optimal solutions through grid search and cross-validation methods.
And step seven, supporting vector machine classification. And inputting the emotional feature codes into the trained support vector machine by using the trained support vector machine, and obtaining the classified output of the identification signals through the calculation of the multi-class support vector machine.
The multi-class support vector machine is characterized in that: and a multi-classification support vector machine constructed by using a two-classification support vector machine through an One-VS-One method. The decision function of the two-class support vector machine is:
the data may be classified according to the sign of the decision function. And finally voting by all the two-classification support vector machines to determine the final classification output.
The invention can eliminate signal bias caused by individual difference and environmental factors by preprocessing the physiological signals of students in the field teaching of remote education, then utilize continuous limited wave Alzheimer machines to process the characteristics of the physiological signals, reduce the characteristic difference caused by the individual difference, input the information into a support vector machine to train to obtain a classifier, and finally, by carrying out mode real-time identification, can effectively obtain the feedback information of the students on the teaching mode, and the accuracy rate of the field students on the feedback information of teaching can reach 82.6 percent on average, thus being beneficial to improving the teaching level of the field remote education. The invention can be widely applied to remote teaching, can directly acquire feedback information of students participating in the remote teaching on the teaching acceptance degree in the teaching, timely adjust and improve the teaching rhythm and level, is very beneficial to the remote teaching,
drawings
FIG. 1 is a schematic view of the working process of the present invention.
FIG. 2 is a schematic diagram of a graphical model of a continuous confined wave Alzheimer's machine.
FIG. 3 is a graph of a portion of the original physiological signal features and their corresponding features after reconstruction by a continuous confined wave Alzheimer's machine.
Detailed Description
The following detailed description of the embodiments of the present invention is provided with reference to the drawings, and the embodiments are implemented on the premise of the technical solution of the present invention, and the detailed implementation and the specific operation process are provided, but the protection scope of the present invention is not limited to the following embodiments.
As shown in fig. 1, the present embodiment includes the following steps:
step one, filtering the noise of the physiological signal. After receiving the signal of the physiological signal sensor, the low-pass filtering is firstly carried out on the input signal, and the interference of external electromagnetic fields such as alternating current and the like on the signal is weakened. The low pass filtered signal is then smoothed using a smoothing filter, which filters out gaussian random noise that may be introduced due to slight limb movements of the student.
The physiological signal sensor comprises: skin Conductance (Skin Conductance) sensor, blood volume Pulse (blood volume Pulse) sensor, Respiration (Respiration) sensor, and Electromyography (Electromyography) sensor. Fig. 2 is a typical waveform of an output signal of each sensor.
In this embodiment, the sampling frequency of the selected sensor is 256 Hz. The selected low-pass filter is a Butterworth filter, the Butterworth filter with the cutoff frequency of 50Hz and the gain of 1 is designed for skin conductance, blood volume pulsation and respiration signals, and the Butterworth filter with the cutoff frequency of 100Hz and the gain of 1 is designed for muscle electrical reaction signals. The selected smoothing filter is a mean smoothing filter, the window range of the mean smoothing filter is [ -32, 32], and the specific calculation method is as follows:
and step two, estimating the physiological signal reference. The physiological signals of the students in the calm state are collected, and after the physiological signals are collected for a period of time continuously, the physiological signals of the students in the period of time are stored, so that the current physiological signal reference is estimated.
The calm state refers to: the student is in a comfortable sitting position and the whole body of the student is relaxed.
In this example, the method adopted was to play a piece of soothing music, blue Danube, and tell the student to relax the body after the student was in a comfortable sitting position. The filtered physiological signal data is then continuously acquired for 40 seconds and stored in the computer as a file as the current physiological signal reference of the student.
And step three, continuously collecting physiological signals of students in class, and segmenting the physiological signals according to a time window with fixed width. Each segment of the physiological signal serves as a set of signal segments for emotion classification.
In the present embodiment, the fixed time window width is set to 25 seconds, i.e. each group of signal segments includes the physiological signals of the student within 25 seconds. In order to generate the training data required for the subsequent steps, a certain class listened by the student in this embodiment is specially designed, specifically a class of 45 minutes english. In order to simulate the actual situation in remote teaching, the course is pre-recorded and then played for students in a computer. In the course, a part of contents which can interest the students, a part of contents which need the students to think, a part of contents which can lead the students to be tired of frustration and a part of contents which can lead the students to regenerate the interest of learning from the frustration are intentionally arranged so as to generate the required data of four emotional states of the students.
And step four, extracting the physiological signal characteristics. The physiological signal feature extraction comprises two aspects of time domain physiological signal feature extraction and frequency domain physiological signal feature extraction, and the two aspects are as follows: the time-domain physiological signal characteristics include: skin conductance, myoelectric response, blood volume pulsation, respiratory amplitude characteristics; the frequency domain physiological signal characteristics are obtained by transforming the physiological signal into a frequency domain through fast Fourier transform, and comprise: blood volume pulsation, heart rate variability high frequency energy to low frequency energy ratio, respiratory rate characteristics. Specific measurements and signal extraction are performed for both aspects.
And step five, encoding the physiological signal characteristics. And training by using a continuous restricted wave Alzheimer machine and using the physiological signal characteristics obtained by calculation to obtain a generation model of the physiological signal characteristics and the physiological signal characteristic codes, and sampling to obtain a group of corresponding physiological signal characteristic codes according to the input physiological signal characteristics.
The continuous confined wave alzheimer machine is characterized in that: a generalization of a restricted wave Alzheimer's machine has the same structure as the restricted wave Alzheimer's machine except that each node of the restricted wave Alzheimer's machine is a Bernoulli (Bernoulli) distributed random node and each node of successive restricted wave Alzheimer's machines is a Sigmoid node to which Gaussian random noise is added. The graphical model is shown in fig. 2.
The states of hidden nodes of the continuous confined wave-zeeman machine are as follows:
Nj(0,1) is a Gaussian distribution with 0 as the mean and 1 as the variance, wherein,
θLand thetaHRespectively the minimum and maximum values, a, that the node may takejIs a random noise control parameter, for visible layer nodes, can be ajSet as a fixed value, a of hidden nodejAnd training by a training algorithm to continuously adjust.
The states of the visible layer nodes of a continuous confined wave alzheimer's machine are:
in this embodiment, the number of visible layer nodes is 11, and consists of 10 physiological signal characteristic inputs and 1 offset node. The number of hidden layer nodes is 51, and the hidden layer nodes are composed of 50 physiological signal coding nodes and 1 bias node. The value ranges of the visible layer unit and the hidden layer unit in the embodiment are both (0,1), so theta is selectedLAnd thetaHRespectively 0 and 1. The state of the bias node of the visible layer and the hidden layer is always 1.
For gaussian random noise cells, in this embodiment, the variance of the gaussian noise is fixed to 0.2.
For the noise control variable a, in the present embodiment, a of the visible layer is set to a fixed value, and each node is 0.01; the a of the hidden layer is set to be random decimal between [0, 0.01] during initialization and then is continuously adjusted through a training algorithm.
When the continuous restricted wave Alzheimer's machine is trained, the continuous restricted wave Alzheimer's machine gradually tends to a balanced state along with the increase of training time, so that a certain learning rate attenuation can ensure that a larger gradient decline is pursued at the initial training stage to accelerate the training speed, and a smaller training speed is guaranteed at the later training stage to ensure the stability of a learning algorithm.
Therefore, in this embodiment, rwIs 0.5, and decreases linearly by a factor of 0.3, as calculated as follows:
where t is the training period of this time (one training period is defined as one training process for completing all training data), and MaxTrainingEpoch is the largest training period, which is defined as 5000 in this embodiment.
raIs 0.2, and decreases linearly by a factor of 0.1, as calculated as follows:
in order to make the value range of the visible layer be (0,1), in this embodiment, the physiological signal features need to be normalized by a normalization factor, and the specific calculation method is to calculate the maximum and minimum values of each physiological signal feature on the training data, and then perform feature transformation according to the following formula:
wherein minyAnd maxyThe feature transformation formula is needed to be stored in the model, and after the model is trained, the given new features are all needed to be subjected to feature transformation by the transformation formula, so that the consistency of data space is ensured.
In order to improve the training speed, the training algorithm in the embodiment adopts a Mini-batch method to speed up, 100 training examples are taken as a training minimum unit, the average gradient of the 100 training examples is calculated, then model parameters are corrected, the accuracy of gradient estimation can be improved, and the operation speed can be improved due to the adoption of matrix operation.
The training effect is shown in fig. 3, where the green curve is the primary physiological signal characteristic and the blue curve is the visible layer node state reconstructed by a continuous confined wave alzheimer's machine. It can be seen that the visible layer node state reconstructed from the hidden layer node state of the continuous restricted wave alzheimer's machine is very similar to the original physical signal feature, that is, the hidden layer node state of the continuous restricted wave alzheimer's machine can well describe the original feature vector.
And step six, training a support vector machine. After the physiological signal characteristic code is obtained, firstly, student emotion category marking is carried out on the code, and marked training data are obtained. And then training a support vector machine model by using training data through cross validation and a grid search algorithm to obtain a group of support vectors, weight vectors corresponding to the support vectors and a group of optimal hyper-parameters.
In this embodiment, the emotion categories of the students include: "interesting", "confusing", "frustrated" and "full of hopes".
The grid search algorithm is as follows: a common searching method of hyper-parameters. The method comprises sequentially traversing parameters at all steps in a jointed hyperparametric space according to a fixed step length, selecting a best hyperparametric point, and then continuously reducing the step length in the neighborhood of the hyperparametric point to continue traversing until the set minimum stepLong. In this example, a common exponential step search is used, initially at [ 2]-7，27]Within the range, the search is performed in 2 steps. After estimating the rough optimized parameter range, decreasing the step size according to exponential decay step by step until the step size is 20.2Step size is multiplied.
The support vector machine is as follows: a method is useful for pattern classification. The two-class support vector machine can solve the two-class classification problem, and the core idea is to solve the following optimization problem:
to maximize LD
Is limited by
The dual problem is to maximize LP
Where K (×) is the kernel function. This is a quadratic optimization problem with a unique solution. X corresponding to α whose solution value is not 0 is called a support vector, and α itself is a weight vector.
The kernel function selected in this embodiment is a Radial Basis Function (RBF) expressed as
Wherein sigma is a hyper-parameter, and the optimal solution can be searched by a grid search and cross validation method.
The two-classification support vector machine can construct a multi-classification support vector machine by a One-VS-One method. For an N-class problem, N x (N-1)/2 two-class support vector machines can be constructed, and a corresponding two-class support vector machine is trained for each two classes of data. For a given input, inputting into the N x (N-1)/2 two-classification support vector machines respectively, and voting by all the two-classification support vector machines to determine the final classification.
The optimal hyper-parameter refers to: some parameters of the model that are not related to the training data include σ in the kernel function described above, and the like. These hyper-parameters can be used to search for optimal solutions through grid search and cross-validation methods.
And step seven, supporting vector machine classification. And inputting the emotion feature codes into the support vector machine obtained by training by using the support vector machine obtained by training, and obtaining emotion classification output through calculation of the multi-class support vector machine.
The multi-class support vector machine is characterized in that: and a multi-classification support vector machine constructed by using a two-classification support vector machine through an One-VS-One method. The decision function of the two-class support vector machine is:
the data may be classified according to the sign of the decision function. And finally voting by all the two-classification support vector machines to determine the final classification output.
The embodiment can be applied to a remote classroom environment, a student only needs to wear corresponding sensor equipment, after the current physiological signal reference signal of the student is acquired, the physiological signal of the student in class is recorded, the time domain and frequency domain physiological signal characteristic values of the student physiological signal of the time segment are calculated according to the time segment, the physiological signal characteristic vector of the student in the time segment is obtained, then the characteristic vector is input into a trained continuous limited wave Alzheimer's machine, a group of codes capable of representing the current student physiological signal characteristics is obtained by calculating the hidden layer state of the continuous limited wave Alzheimer's machine, and finally the codes are input into a pre-trained support vector machine to obtain the output for identifying the current emotion category of the student. The continuous limited wave Alzheimer's machine is matched with the support vector machine, so that the classification accuracy of 82.6% on average can be achieved, and the method has obvious advantages.
Claims (3)
1. A method for extracting and identifying characteristic signals of remote education students is characterized by comprising the following steps:
firstly, after receiving a signal of a physiological signal sensor, carrying out low-pass filtering on an input signal, and then carrying out smooth filtering on the input signal after the low-pass filtering by using a smoothing filter;
secondly, collecting physiological signals of students in a calm state, continuously collecting the physiological signals for a period of time, and storing the physiological signals of the students in the period of time as the current physiological signal reference;
continuously collecting physiological signals of students in class, and segmenting the physiological signals according to a time window with fixed width, wherein each segment of physiological signals is used as a group of signal segments for emotion classification; the emotion categories include four categories of "interesting", "confusing", "frustrated" and "full of hope";
extracting physiological signal characteristics, including time domain physiological signal characteristic extraction and frequency domain physiological signal characteristic extraction, wherein the time domain physiological signal characteristics comprise skin conductance, muscle electrical reaction, blood volume pulsation and respiration amplitude characteristics, and the frequency domain physiological signal characteristics are obtained after fast Fourier transformation to a frequency domain and comprise blood volume pulsation, heart rate variability high-frequency energy and low-frequency energy ratio and respiration frequency characteristics;
using a continuous restricted wave Alzheimer's machine to train according to the calculated physiological signal characteristics to obtain a generation model of the physiological signal characteristics and physiological signal characteristic codes, and sampling according to the input physiological signal characteristics to obtain a group of corresponding physiological signal characteristic codes; each node of the continuous restricted wave Alzheimer machine is a Sigmoid node added with Gaussian random noise, and the state of a hidden node is as follows:
Nj(0,1) is mean value of 0 and variance of 1Is a Gaussian distribution of
θLAnd thetaHRespectively the minimum and maximum values, a, that the node may takejIs a random noise control parameter, in which
W=W+rwG, where W is a weight matrix connecting the hidden layer and the visible layer, g is the gradient, rwFor the learning rate, its initial value is 0.5 and decreases linearly by a factor of 0.3:
t is the training period, and MaxTrainingepoch is the maximum training period;
after the physiological signal characteristic code is obtained, firstly carrying out student emotion category marking on the code to obtain marked training data, then using the training data to train a support vector machine model through cross validation and a grid search algorithm to obtain a group of support vectors, and a weight vector and a group of optimal hyper-parameters corresponding to the support vectors;
and step seven, inputting the emotional characteristic codes into the trained support vector machine by using the trained support vector machine, and obtaining the identification signal output through the calculation of the multi-class support vector machine.
2. The method of claim 1, wherein the physiological signal reference is an offset of a physiological signal of a student due to extrinsic or intrinsic non-emotional factors.
3. The method for extracting and identifying the characteristic signals of the remote education students as claimed in claim 1, wherein the physiological signal characteristic code is a state vector h of the hidden node sampled from a generative model described by a continuous training restricted wave alzheimer's machine under the condition that a visible layer state vector v, namely a physiological signal characteristic vector, is given.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201010140911.6A CN101840506B (en) | 2010-04-08 | 2010-04-08 | Remote education student characteristic signal extraction and recognition method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201010140911.6A CN101840506B (en) | 2010-04-08 | 2010-04-08 | Remote education student characteristic signal extraction and recognition method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN101840506A CN101840506A (en) | 2010-09-22 |
| CN101840506B true CN101840506B (en) | 2014-02-19 |
Family
ID=42743868
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201010140911.6A Expired - Fee Related CN101840506B (en) | 2010-04-08 | 2010-04-08 | Remote education student characteristic signal extraction and recognition method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN101840506B (en) |
Families Citing this family (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104408629B (en) * | 2014-11-10 | 2017-10-27 | 广东欧珀移动通信有限公司 | Shopping at network monitoring method, system and the terminal device comprising the system |
| CN107007291A (en) * | 2017-04-05 | 2017-08-04 | 天津大学 | Intense strain intensity identifying system and information processing method based on multi-physiological-parameter |
| CN107566471B (en) * | 2017-08-25 | 2021-01-15 | 维沃移动通信有限公司 | Remote control method and device and mobile terminal |
| CN109117731B (en) * | 2018-07-13 | 2022-02-18 | 华中师范大学 | Classroom teaching cognitive load measurement system |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN1952953A (en) * | 2006-11-16 | 2007-04-25 | 上海交通大学 | Posture recognition method of human's face based on limited Boltzmann machine neural network |
| CN1975856A (en) * | 2006-10-30 | 2007-06-06 | 邹采荣 | Speech emotion identifying method based on supporting vector machine |
| CN101599271A (en) * | 2009-07-07 | 2009-12-09 | 华中科技大学 | A kind of recognition methods of digital music emotion |
Family Cites Families (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US7472063B2 (en) * | 2002-12-19 | 2008-12-30 | Intel Corporation | Audio-visual feature fusion and support vector machine useful for continuous speech recognition |
| JP5217109B2 (en) * | 2006-04-17 | 2013-06-19 | 富士通株式会社 | Operator education support program, operator education support method, and operator education support apparatus |
| CN101587546B (en) * | 2009-06-12 | 2011-05-11 | 浙江大学 | Identification method of lower limb action pattern based on supporting vector multivariate classification |
-
2010
- 2010-04-08 CN CN201010140911.6A patent/CN101840506B/en not_active Expired - Fee Related
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN1975856A (en) * | 2006-10-30 | 2007-06-06 | 邹采荣 | Speech emotion identifying method based on supporting vector machine |
| CN1952953A (en) * | 2006-11-16 | 2007-04-25 | 上海交通大学 | Posture recognition method of human's face based on limited Boltzmann machine neural network |
| CN101599271A (en) * | 2009-07-07 | 2009-12-09 | 华中科技大学 | A kind of recognition methods of digital music emotion |
Non-Patent Citations (2)
| Title |
|---|
| JP特开2007-286302A 2007.11.01 |
| 付永钢 等.基于支持向量机的学习评价***.《计算机工程》.2004,第30卷(第8期), * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN101840506A (en) | 2010-09-22 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111461176B (en) | Multi-mode fusion method, device, medium and equipment based on normalized mutual information | |
| CN101317794B (en) | Myoelectric control ability detecting and training method for hand-prosthesis with multiple fingers and multiple degrees of freedom | |
| CN105488466B (en) | A kind of deep-neural-network and Acoustic Object vocal print feature extracting method | |
| CN112581979B (en) | Speech emotion recognition method based on spectrogram | |
| CN107293302A (en) | A kind of sparse spectrum signature extracting method being used in voice lie detection system | |
| CN108229659A (en) | Piano singly-bound voice recognition method based on deep learning | |
| Shao et al. | Single-channel SEMG using wavelet deep belief networks for upper limb motion recognition | |
| CN112820279B (en) | Parkinson detection model construction method based on voice context dynamic characteristics | |
| Hariharan et al. | Objective evaluation of speech dysfluencies using wavelet packet transform with sample entropy | |
| CN108122004A (en) | The brain electricity sorting technique of the sparse learning machine that transfinites is differentiated based on Fisher | |
| CN101840506B (en) | Remote education student characteristic signal extraction and recognition method | |
| Oweis et al. | ANN-based EMG classification for myoelectric control | |
| CN113616213A (en) | Electrocardiosignal denoising method, equipment and storage medium based on BP neural network and improved EMD method | |
| Salhi et al. | Voice disorders identification using hybrid approach: Wavelet analysis and multilayer neural networks | |
| Bernstein et al. | Using deep learning for alcohol consumption recognition | |
| CN114091529A (en) | Electroencephalogram emotion recognition method based on generation countermeasure network data enhancement | |
| Anjali et al. | Infant cry classification using transfer learning | |
| CN113343869A (en) | Electroencephalogram signal automatic classification and identification method based on NTFT and CNN | |
| Zhang et al. | Human movements classification using multi-channel surface EMG signals and deep learning technique | |
| Huang et al. | Classification of cough sounds using spectrogram methods and a parallel-stream one-dimensional deep convolutional neural network | |
| CN116570284A (en) | Depression recognition method and system based on voice characterization | |
| Fei | Application research of neural network technology in vocal music evaluation | |
| Yan et al. | Analyzing Wrist Pulse Signals Measured with Polyvinylidene Fluoride Film for Hypertension Identification. | |
| CN116013339A (en) | Single-channel voice enhancement method based on improved CRN | |
| CN114299925A (en) | Method and system for obtaining importance measurement index of dysphagia symptom of Parkinson disease patient based on voice |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| C14 | Grant of patent or utility model | ||
| GR01 | Patent grant | ||
| CF01 | Termination of patent right due to non-payment of annual fee | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20140219 Termination date: 20170408 |