[0065] In order to make the objects, technical solutions, and points of the invention, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.
[0066] Motion imaging classification method based on optimal narrowband feature fusion, including the following steps:
[0067] Step 1: Get a motion imagination eElectronic signal
[0068] Experiment Choose the BCI Competition IV 2A Quarterly Movement Imagination Data Set (hereinafter collectively referred to as BCI competition data set). These include left-handed, right hand, feet and tongue four types of motion imagination tasks, a total of EEG signals from 22 electrodes and an EOG signal from three electrodes, a signal sample rate of 250 Hz, and pass through a band pass filtering process of 0.5-100 Hz .
[0069] Step 1.1: The BCI competition data is concentrated, and 9 were tried to implement four types of motion imagination tasks, a total of two experiments. Each experiment includes 288 trials. Each test time is 3 seconds, and 288 test times obtained by the previous experiment as a training set, and the 288 test times obtained by the latter experiment as a test set. to evaluate.
[0070] Step 1.2: For the four-class motion imagination task, set four classifier to process the corresponding two-point tasks. Each classifier sets one of these motion imagination as the first class, all other motion imagination is a second type, constitutes a second classification task, and the first category of 4 classifiers is different. The training set and the test set are the same. When the training set uses a certain type as the first class, the test set also needs to use this class as a first class.
[0071] Step 2: Get the best narrowband
[0072] Step 2.1: Determine the wideband 0.1-32Hz as the root node, and divide it into 4 frequency bands, 0.1-8 Hz, 8-16 Hz, 16-24 Hz and 24-32 Hz, respectively.
[0073] Step 2.2: Belt-pass filtering by these four frequency bands, and obtains the filtered eElectronic signal.
[0074] Step 2.3: Extraction characteristics using the CSP algorithm for filtering EEG signals, specific:
[0075] The electromalid signal after the filter is set to X∈R C×N Where c represents the number of channels, n represents a sampling point; divide the training data of the electroelectric signal into two categories, respectively. 1 X 2 , Where x d ∈R C×N , D ∈ {1, 2} means category tag; x d = [X d，1 , ..., x d，N ] The mean is expressed as X d The time variance after using the airspace filter ω map is represented as:
[0076]
[0077] Where t is expressed;
[0078] The CSP algorithm is mapped to a new sub-space by building airspace filter ω, making the variance difference between the two categories after using the Ω mapping, so the airspace filter ω constraint condition is:
[0079]
[0080] Definition σ d A covariance matrix of an electromal signal indicating the category label D:
[0081]
[0082] The formula (2) is expressed as the following form using the formula (3):
[0083]
[0084] The expression (4) for solving the empty domain filter can be solved by the generalized feature value problem of the formula (5); 8 feature vectors corresponding to the smallest 4 and maximum four broad feature values to obtain the final nullial filter Ω∈R C×8;
[0085] Σ d Ω = λ (σ 1 + Σ 2 ω (5)
[0086] The calculated airspace filter ω handles the brain-filled electromal signal X after filtering, and the electrode signal X 'after the sub-spatial map is represented as:
[0087] X '= Ω T.X (6)
[0088] Extraction variance characteristics of EEG signal after sub-spatial mapping:
[0089] Y = σ 2 (X, ω) (7)
[0090] Where Y∈R 8;
[0091] Step 2.4: 80% of the training set as training data, 20% as test data. The support vector machine (SVM) model with a linear core is constructed, and the feature vector input is input as the ecpoC signal; the output is a category label; compared to the output category tag and the real class tag to obtain the classification accuracy under the current frequency band. The SVM model build method is:
[0092] The n dimensions are N dimensions n, which is shown as t = {(x 1 Y 1 ), (X 2 Y 2 ), ..., (x N Y N )}, Where X i ∈R n Y i ∈ {-1, 1}, i = 1, 2, ..., n. Select penalty parameters C> 0, construct and solve the convex secondary planning problem, find the ultra plane (W, b) that is mostly separated by another type of data from another type of data.
[0093]
[0094]
[0095] Where α indicates that the Lagrangian multiplier vector, k (•) represents the nuclear function, k (x, y) = φ (x) * φ (x), φ (•) represents the mapping function. Find the best solution Calculate Select α * One component Suitable condition Calculate B according to formula (9) *.
[0096]
[0097] Ask for separation super plane W * · X + b * = 0, classified decision function is f (x) = SIGN (W * · X + b * ).
[0098] During the training classifier, since the first type and the second type of sample ratio is 1: 3, it belongs to the number of categories distributed unbalanced sample, so the F1 Score is used to optimize the model evaluation index. Assuming the classification sample is a positive sample and negative sample, then predicting is that there may be two kinds, which are predicted to predicate the category (TP) and predict the negative prediction into a category (FP), and the accuracy is defined as:
[0099]
[0100] The recall rate refers to how many successful predictions have been successfully predicted relative to the original sample. There are also two possibilities, one is to predict the original category into a category (TP), and the other is to predict the original category into a negative (FN), the recall rate is defined as:
[0101]
[0102] F1 score is a measurement indicator of the classification problem, which is the average number of precision rates and recall rates. The maximum is 1, the minimum is 0, which is defined as:
[0103]
[0104] Step 2.5: Select the frequency band with the highest-class accuracy from the four frequency bands and the frequency bands in the frequency band adjacent to the category accuracy, the two frequency bands are common as the parent node, and this parent node is divided into 4 frequency bands, repeat steps 2.2-2.5 until the four frequency bands are 1 in the 4th frequency bands. At the end of each round search, the global classification accuracy maximum is updated, and the filter band corresponding to the maximum value is saved. The final output of the frequency band with the highest-class accuracy in the entire search process as the optimal narrow frequency band.
[0105] Step 3: Feature extraction phase
[0106] Step 3.1: Movement to step (1) by the optimal narrow frequency band obtained by step 2.5 to move the brain electrical signal, and then perform feature extraction of the filtered brain electrical signal:
[0107] 3.1.1 Time domain features, including maximum, minimum, mean, and standard deviation of electromal signals;
[0108] 3.1.2 Frequency domain features, including the frequency mean, frequency variance, and frequency entropy of the electromal signal;
[0109] The frequency mean is as follows:
[0110]
[0111] Wheref (k), k = 1, 2 ..., n indicates the frequency spectrum diagram after the fast Fourier transform, and n represents half of the highest frequency;
[0112] The frequency variance is represented as follows:
[0113]
[0114] The frequency entropy is represented as follows:
[0115]
[0116] Step 3.2: For 4 two-class classifier composed of quarter-category motion imaging tasks, obtain 4 8 × 7 feature matrices, and perform vertical cascading, generating dimensions 32 × 7 result matrices:
[0117]
[0118] among them The first feature in the first character vector of the optimal narrowband K, the range of K is 1 ≤ K ≤ 4, and the range of 1 ≤ i ≤ 8, J is 1 ≤ j ≤ 7, 1 7 respectively represent the maximum value, minimum value, mean, standard deviation, frequency mean, frequency variance, and frequency entropy, respectively, frequency.
[0119] Step 4: Construct a depth convolutional neural network model
[0120] The depth convolutional neural network model includes a first rolling layer, a second volume layer, a full connectivity layer; input to step (3) 32 × 7 result matrix A input , Output is a basic signal prediction category;
[0121] Step 4.1: The first volume layer contains two 2d convolutions, the core size is 1 × 3 and 8 × 3, and the features used in the fusion frequency band, and the size of the feature is reduced, and after the feature is fused through the frequency band, each The feature dimension of the output corresponding to the volume subscription is shown below.
[0122]
[0123] among them, The i-th feature of the fusion of the kth frequency band is shown in the range of 1 ≤ k ≤ 4, and the range is 1 ≤ i ≤ 7.
[0124] Step 4.2: Output matrix A of the first volume layer out1 As the input of the second volume layer. The second volume layer uses a 2D convolution of the core size of 4 × 1, and the characteristics of the fusion frequency band are used in the end of the second portion. The dimension of the feature is compressed to 1. After fusion of frequency band characteristics, the output feature dimension corresponding to each convolutionary claim is shown below.
[0125] A out1 ((A) 1 , A 2 , ..., A 7 )
[0126] Step 4.3: Since the input data has been pre-specifically, the pool processing is not performed through the entire network, and the mosction is made directly through the convolution layer to enhance the learning ability of the network. The normalized operation is performed and entered into the full connection layer.
[0127] Step 4.4: The full connection layer is used for the final classification, in order to avoid the fit, use Dropout, where Rate = 0.5. After experiment, RELU has a better effect relative to other activation functions, so the RELU activation function is set in the convolution layer and the full connect layer, and the activation function expression is as follows:
[0128] f (x) = max (0, x) (16)
[0129] Step 4.5: During the network training, use the cross entropy loss function is defined as follows:
[0130]
[0131] Where P is the target distribution, Q is the observed distribution. The model uses the ADAM optimizer to optimize, the learning rate is 1e-3, and the attenuation weight is 1e-7. The zero value and the normal distribution of the unit variance initialize the convolution layer weight, and the batch is initialized, the batch size is 16.
[0132] Step 5.6: Similar to MLP, each neuron in the fully attachment layer is fully attached to all neurons of the previous layer. The output value of the last layer is given an output, and the SoftMax logic regression is used. The classification result is compared with the correct result and calculate the loss function, and the modified parameters are modified by reverse transmission. After the model training is completed, it is applied to the test set and classified to obtain the final classification result.
