A kind of EMG Feature Extraction based on convolutional neural networksTechnical field
The present invention relates to physiological single processing and analytical technology more particularly to a kind of new sides for extracting electromyography signal featureMethod.
Background technique
The ultra-weak electronic signal generated when electromyography signal is contraction of muscle.The electric signal generated when contraction of muscle passes through internal groupThe conduction knitted forms potential change in skin surface.When this potential change passes through the processing of amplifying circuit, collected, storageGet off, referred to as surface electromyogram signal.Surface electromyogram signal mainly has both sides purposes: 1) clinical diagnosis and pathological analysis；2) human-computer interaction, as prosthetic hand or artificial limb leg control.Since electromyography signal is substantially closely related with the execution of user intention, lead toCrossing reasonable manner decoding electromyography signal can produce intuitive control command.Opposite EEG signals and nerve signal, myoelectricity letterNumber more stable and signal amplitude is larger, is acknowledged as most potential artificial limb end-of-pipe control signal source.
Nevertheless, electromyography signal still will receive the interference of various unfavorable factors, as muscular fatigue, dislocation of electrode, across withFamily otherness etc., so that the man-machine interface based on electromyography signal faces the test in terms of stability.Electromyography signal is substantially oneKind random signal.Feature extraction is to analyze a necessary means of electromyography signal.The feature of traditional electromyography signal can be divided into time domainFeature, frequency domain character and time and frequency domain characteristics.However, comprehensively utilizing these features, also there is no solve myoelectricity pattern classification precisionPractical problem low, stability is poor.It is made as example with myoelectric limb manual control, successful case in the market is only limitted to utilize binary channels fleshElectricity realizes the control and closure of prosthetic hand.From the point of view of current status, based on the human-computer interaction of electromyography signal still in the primary stage,Need further to be developed.A kind of myoelectricity feature with anti-noise ability is found, at the important opportunity for breaking through the bottleneck.
Convolutional neural networks are a kind of multilayer neural networks, are good to obtain from initial data and meet the reliable of target callInformation.The present invention proposes a kind of method for extracting invariant feature from original electromyography signal using convolutional neural networks.This methodA large amount of electromyography signals by multi-user, long time integration are input, train the convolutional neural networks of a high robust, and willCompletely new myoelectricity feature of the output of full articulamentum as similar traditional characteristic in network.Correlative data analysis shows that utilization is thisThe precision for the gesture identification every other day that the feature that method is extracted can be improved.
Therefore, those skilled in the art is dedicated to developing a kind of electromyography signal feature with compared with strong anti-interference ability,To be promoted based on electromyography signal as the stability of man-machine interface.
Summary of the invention
In view of the drawbacks described above of the prior art, the technical problem to be solved by the present invention is to traditional electromyography signal features cannotThe problem of meeting the stability requirement of man-machine interface control.
To achieve the above object, the present invention provides a kind of electromyography signal feature extraction side based on convolutional neural networksMethod,
Including the pre-training and essence training for training convolutional neural networks；With articulamentum output valve complete in convolutional neural networks workFor electromyography signal characteristic value, the convolutional neural networks are the multilayer neural networks comprising two convolutional layers, full articulamentum.
Further, the pre-training for training convolutional neural networks includes the following steps: with smart Training strategy
Step 1: carrying out pre-training to convolutional neural networks using the myoelectricity data of all subjects；
Step 2:, to above-mentioned pre-training neural network produced, being carried out further using the myoelectricity data of target subjectTraining obtains smart neural network model.
Further, the full articulamentum output valve includes the following steps: as the characteristic value of electromyography signal
Step 1: using the smart neural network model of acquisition as the network of feature extraction, and save network structure and parameter；
Step 2: the full articulamentum for above-mentioned network structure increases output interface.
Step 3: being input with real-time myoelectricity data, the output valve of full articulamentum is obtained, the feature as electromyography signal.
Detailed description of the invention
Fig. 1 is a kind of convolutional network knot of EMG Feature Extraction based on convolutional neural networks of the inventionStructure；
The wherein original electromyography signal of 1-, 2- convolutional layer, 3- convolutional layer, the full articulamentum of 4-, the full articulamentum of 5-, 6- output layer, 7- are completeThe output valve of articulamentum；
Fig. 2 is a kind of execution process of EMG Feature Extraction based on convolutional neural networks of the invention；
Fig. 3 is the test result that specific embodiment is obtained by pre-training and essence training；
Fig. 4 A- Fig. 4 C compared traditional characteristic space and the feature space based on convolutional neural networks when distinguishing myoelectricity sampleDifference row；
Fig. 5 illustrates influence of the feature and traditional characteristic of the invention extracted to gesture nicety of grading.
Specific embodiment
With reference to the accompanying drawings and detailed description, special to a kind of electromyography signal based on convolutional neural networks of the inventionSign extracting method is further described.
The embodiment is using a myoelectricity database as analysis object.The database contains the upper hand of 6 subjectsArm myoelectricity data, each subject acquire 10 days myoelectricity data.Myoelectricity data acquire under 13 different gesture motions.It should7 days electromyography signals are as training data before embodiment, and later 3 days electromyography signals are as test data.
The embodiment uses TensorFlow platform building network structure, and passes through the GeForce based on CUDA 8.0.44GTX 1080TI video card accelerating algorithm executes.
Network structure determined by the embodiment includes 1 input layer, 2 convolutional layers, 2 full articulamentums.The embodimentInput data be the dimension through over-segmentation be the original electromyography signal of 16*256.2 convolutional layers separately include 32 and 64 3*3Filter.The output of convolutional layer carries out Data Dimensionality Reduction by the MAX Pooling scheme of 2*2.2 full articulamentums wrap respectivelyContaining 128 and 13 concealed nodes.The output of 128 concealed nodes of trained network is the original flesh from 16*256The electromyography signal feature extracted in electric signal.The output of 13 concealed nodes can obtain 13 gestures by softmax functionThe classification of movement.In network training, second convolutional layer and first full articulamentum pass through Dropout processing respectively, keepProbability is respectively set to 0.8 and 0.5.
The pre-training of the embodiment uses the training data of all subjects, and obtains pre-training net by 500 iterationNetwork.The essence training of the embodiment uses the training data of target subject, and obtains essence training network by 500 iteration.The accuracy of identification variation of pre-training and essence training is as shown in Figure 3.
The feature of test data is extracted using the character network obtained above by pre-training with essence training, and 128 are tieed upAfter Feature Dimension Reduction, can get feature distribution as shown in Figure 4 as a result, wherein Fig. 4 A be sample traditional characteristic space distribution,It in the distribution of the feature space based on convolutional neural networks, Fig. 4 C is sample in two kinds of superimposed spaces of feature that Fig. 4 B, which is sample,In distribution.
Fig. 5 compared traditional characteristic and superposition state feature at LDA and SVM classifier to not as a result, i.e. traditional characteristicFeature of the superposition based on convolutional neural networks can obtain better classifying quality.