CN113688952A - Brain-computer interface decoding acceleration method and system based on self-adaptive electroencephalogram channel selection - Google Patents

Abstract

The present invention provides a brain-computer interface decoding acceleration method and system based on adaptive EEG channel selection, comprising: acquiring EEG data to be decoded; inputting the EEG data to be decoded into a decoding model, and outputting the EEG data to be decoded The decoding result of the intended decoding of the electrical data; wherein, the decoding model is used to perform feature extraction based on the minimum channel data compressed into the EEG data to be decoded to obtain a strategy feature, and select the optimal number of channels according to the strategy feature to After obtaining the optimal channel data, intentionally decode the EEG data to be decoded by using the optimal channel data. In order to solve the defect of low decoding efficiency caused by the use of multi-channel decoding of EEG data in the prior art, the conversion and selection of the decoding channel of the EEG data to be decoded through the decoding model is realized, and the decoding accuracy is not reduced or even improved. Improve decoding efficiency.

Description

Brain-computer interface decoding acceleration method and system based on adaptive EEG channel selection

technical field

The invention relates to the technical field of brain-computer interface decoding, in particular to a brain-computer interface decoding acceleration method and system based on adaptive EEG channel selection.

Background technique

A brain-computer interface is a direct connection pathway established between human or animal brains (or cultures of brain cells) and external devices, that is, directly decoding brain intentions from EEG signals to control external devices. Brain-computer interface rehabilitation training plays a very important role in the functional rehabilitation of patients with neurological injuries caused by stroke, spinal cord injury, etc. This method has been widely used in the fields of neurological rehabilitation and sports assistance. Specifically, by collecting the EEG information related to active movement intention of nerve injury patients, analyzing the EEG information, and controlling the rehabilitation training equipment based on the analysis results (related to active movement intention), the limb movement of nerve injury patients is carried out. functional training for rehabilitation. Compared with traditional rehabilitation methods and robot-assisted rehabilitation methods, rehabilitation training based on brain-computer interface technology is that patients with nerve injury actively participate in the control of rehabilitation training, and promote the occurrence of neuroplasticity by increasing the degree of neural participation, thereby effectively improving the effect of rehabilitation training.

One of the main challenges faced by brain-computer interface rehabilitation training is the low decoding efficiency of EEG signals. Specifically, the amount of input data is a key factor affecting the decoding efficiency of brain-computer interfaces. However, in order to ensure the decoding accuracy, a typical brain-computer interface system generally contains 32 to 128 EEG channels. Although the use of multiple channels can improve the decoding accuracy to a certain extent, the resulting high amount of EEG data and high computing resource requirements make the decoding efficiency of the brain-computer interface low, and limit the deployment platform of the brain-computer interface system. , that is, it is difficult to deploy on some devices that lack computing power.

SUMMARY OF THE INVENTION

The present invention provides a brain-computer interface decoding acceleration method and system based on adaptive EEG channel selection based on adaptive EEG channel selection, to solve the problem of low decoding efficiency caused by the use of multiple channels to decode EEG data in the prior art It realizes the conversion and selection of the decoding channel of the EEG data to be decoded through the decoding model, and improves the decoding efficiency without reducing or even improving the decoding accuracy.

The present invention provides a brain-computer interface decoding acceleration method based on adaptive EEG channel selection, comprising:

Obtain the EEG data to be decoded;

Inputting the EEG data to be decoded into a decoding model, and outputting a decoding result of intentionally decoding the EEG data to be decoded;

Wherein, the decoding model is used to perform feature extraction based on the minimum channel data compressed into the EEG data to be decoded to obtain strategy features, and select the optimal number of channels according to the strategy features to obtain the optimal channel data. The optimal channel data is intended to decode the EEG data to be decoded.

According to the brain-computer interface decoding acceleration method based on adaptive EEG channel selection, inputting the EEG data to be decoded into a decoding model, and outputting a decoding result of intentionally decoding the EEG data to be decoded, specifically include:

Based on the input EEG data to be decoded, compress the channel data to obtain the minimum channel data of the EEG data to be decoded;

Based on the minimum channel data, the strategy feature extraction related to the optimal channel number is performed to obtain the strategy feature related to the optimal channel number;

Based on the strategy feature, a decision probability calculation is performed on the to-be-decoded EEG data to obtain a decision-probability value in which each optimal number of channels in the to-be-decoded EEG data is selected;

Based on the decision probability value, obtain the optimal number of channels selected;

Based on the selected optimal number of channels, the EEG data to be decoded is converted into a data format corresponding to the selected optimal number of channels, and the data corresponding to the optimal channel number of the EEG data to be decoded is obtained Format decoding result of intent decoding.

The present invention also provides a brain-computer interface decoding acceleration system based on adaptive EEG channel selection, including:

The acquisition module is used to acquire the EEG data to be decoded;

an execution module, for inputting the EEG data to be decoded into a decoding model, and outputting a decoding result of intentionally decoding the EEG data to be decoded;

Wherein, the decoding model is used to perform feature extraction based on the minimum channel data compressed into the EEG data to be decoded to obtain strategy features, and select the optimal number of channels according to the strategy features to obtain the optimal channel data. The optimal channel data is intended to decode the EEG data to be decoded.

The present invention also provides a training method for the decoding model, which specifically includes:

Input the EEG data sample to be decoded;

Perform channel data compression on the EEG data samples to be decoded through multiple candidate transformation matrices constructed according to preset channel transformation rules to form multiple optimal channel data candidates;

Arranging the optimal channel data candidates in ascending order of the number of channels to construct an optimal channel data candidate library;

After obtaining the policy features related to the optimal channel data candidate with the least number of channels in the optimal channel data candidate database and the optimal channel number candidate, obtain a decision for a plurality of the optimal channel number candidates to be selected according to the policy features probability value;

Using the argmax function to select the optimal channel number of the EEG data samples to be decoded in the optimal channel number candidate based on the decision probability value;

After converting the EEG data samples to be decoded into a data format corresponding to the optimal number of channels, perform intention decoding;

Calculate the decoding loss of the EEG data sample to be decoded by using a pre-built loss function, and determine whether the decoding loss satisfies a preset loss standard;

If so, then use the candidate transformation matrix as the optimal transformation matrix, and obtain the decoding model that has been trained;

If not, the candidate transformation matrix is updated, and the EEG data sample to be decoded is returned to perform channel data compression through the updated candidate transformation matrix.

According to the training method of the decoding model, the preset channel conversion rules specifically include: a difference rule, an average rule, and a selective activation rule; wherein,

The difference rule realizes channel data compression by calculating the difference between the EEG data collected by the corresponding electrodes in the two brain regions;

The mean value rule realizes channel data compression by averaging the EEG data of adjacent channels;

The selective activation rule realizes channel data compression by removing the channel corresponding to the EEG data irrelevant to the task to be classified.

According to the training method of the decoding model, the loss function is the weighted sum of the decoding accuracy loss function and the decoding efficiency loss function; wherein,

The decoding accuracy loss function is a cross-entropy loss function constructed on the ShallowConvNet classification model based on the EEG data samples to be decoded and the real label;

The decoding efficiency loss function is a loss function constructed on the ShallowConvNet classification model by averaging the number of floating-point operations of the EEG data samples input to the decoding model to be decoded.

According to the training method of the decoding model, the updating of the candidate transformation matrix specifically includes:

The argmax function based on the decision probability value is approximated by Gumbel-Softmax, and the candidate transformation matrix is updated according to the approximation result.

The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored on the memory and running on the processor, characterized in that, when the processor executes the program, the above-mentioned program is implemented Steps of any one of the methods for accelerating the decoding of brain-computer interface based on adaptive EEG channel selection or the method for training the decoding model.

The present invention also provides a non-transitory computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, implements the brain-computer interface based on any one of the above-mentioned adaptive EEG channel selection The steps of the decoding acceleration method of the decoding model training method.

The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the decoding model of any of the above-mentioned methods for accelerating brain-computer interface decoding based on adaptive EEG channel selection. The steps of the training method.

The invention provides a brain-computer interface decoding acceleration method and system based on adaptive EEG channel selection based on adaptive EEG channel selection. The EEG data is converted into the minimum channel data, and the strategy features are extracted. After selecting the optimal channel number through the strategy features to obtain the optimal channel data, the EEG data to be decoded is intentionally decoded through the optimal channel data. , the optimal channel data obtained by the decoding model is the decoding channel data selected from the minimum channel data obtained by compressing the channel data of the EEG data to be decoded, which avoids the high amount of EEG data caused by multi-channel decoding. The disadvantage of high computing resource requirements improves the decoding efficiency.

Description of drawings

In order to explain the present invention or the technical solutions in the prior art more clearly, the following will briefly introduce the accompanying drawings that need to be used in the description of the embodiments or the prior art. Obviously, the accompanying drawings in the following description are the For some embodiments of the invention, for those of ordinary skill in the art, other drawings can also be obtained according to these drawings without any creative effort.

1 is a schematic flowchart of a brain-computer interface decoding acceleration method based on adaptive EEG channel selection provided by the present invention;

2 is a schematic structural diagram of a brain-computer interface decoding acceleration system based on adaptive EEG channel selection provided by the present invention;

3 is a schematic flowchart of a training process of a decoding model provided by the present invention;

Fig. 4 is the construction principle diagram of decoding model provided by the present invention;

Fig. 5 is the example diagram of the interpolation rule provided by the present invention;

Fig. 6 is the example diagram of the mean value rule provided by the present invention;

7 is an exemplary diagram of a selective activation rule provided by the present invention;

8 is a graph of brain decoding efficiency and decoding accuracy generated by an example of the decoding model of the present invention;

FIG. 9 is a performance comparison diagram of 10 subjects in the data source BCI competition IV dataset 2a (BCIC IV 2a) using the decoding model provided by the present invention and the existing ShallowConvNet model for decoding;

FIG. 10 is a schematic structural diagram of an electronic device provided by the present invention.

Detailed ways

In order to make the objectives, technical solutions and advantages of the present invention clearer, the technical solutions in the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention. , not all examples. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

Information obtained from high-density EEG channels is generally highly redundant (cross-correlated), and the key EEG channels involved in recognizing different types of brain intentions are also different. For example, for the BCI paradigm based on steady-state visual evokation, the information collected by the EEG channels distributed in the occipital lobe is crucial for intention decoding, while for the motor imagery BCI paradigm, only the information collected in the EEG channels distributed in the motor cortex may be required. CZ, C3 and C4 channels can achieve efficient and accurate decoding. It can be seen that EEG data decoded by multiple channels is not completely necessary. Based on this, the present invention proposes a brain computer based on adaptive EEG channel selection from the perspective of compressing the amount of input data. Interface decoding acceleration method. At the same time, for ease of understanding, the terms mentioned in the technical solution of the invention are explained here first, that is, "number of channels" refers to the number of channels; "number of optimal channels" refers to the number of optimal channels, that is, through the strategy feature The number of optimal channels selected for decoding the EEG data to be decoded, such as 6 channels, 9 channels, etc. mentioned in the following embodiments of the present invention; "channel data" refers to the specific data content of the channel; "Optimal channel data" refers to the optimal transformation matrix learned based on the policy network, which acts on the EEG data to be decoded and obtains the final channel data for intended decoding.

A brain-computer interface decoding acceleration method based on adaptive EEG channel selection of the present invention is described below in conjunction with FIG. 1. As shown in FIG. 1, the method includes the following steps:

101. Obtain the EEG data to be decoded;

102. Inputting the EEG data to be decoded into a decoding model, and outputting a decoding result of intentionally decoding the EEG data to be decoded;

It should be noted that the decoding model is obtained by training based on the EEG data samples to be decoded and the decoding results corresponding to the samples, and when using the decoding model for channel selection, the decoding model is used to convert the EEG data into the minimum channel data. For the strategy features obtained by feature extraction, the optimal channel data is selected from the EEG data to be decoded, so the EEG data to be decoded can be compressed first after being input into the decoding model, thereby removing redundant or other irrelevant channel data , reduce the amount of channel data, and then use the strategy features extracted from the compressed data to select the optimal number of channels to obtain the optimal channel data, which improves the decoding efficiency, that is, realizes the input of different EEG data to be decoded. After decoding the model, the required optimal channel data can be adaptively selected to improve the decoding efficiency.

In an embodiment of the present invention, inputting the EEG data to be decoded into a decoding model and outputting a decoding result of intentionally decoding the EEG data to be decoded specifically includes:

Based on the input EEG data to be decoded, compress the channel data to obtain the minimum channel data of the EEG data to be decoded;

Based on the minimum channel data, the strategy feature extraction related to the optimal channel number is performed to obtain the strategy feature related to the optimal channel number;

Based on the strategy feature, a decision probability calculation is performed on the to-be-decoded EEG data to obtain a decision-probability value in which each optimal number of channels in the to-be-decoded EEG data is selected;

Based on the decision probability value, obtain the optimal number of channels selected;

Based on the selected optimal number of channels, the EEG data to be decoded is converted into a data format corresponding to the selected optimal number of channels, and the data corresponding to the optimal channel number of the EEG data to be decoded is obtained Format decoding result of intent decoding.

It should be noted that when applying the decoding model to obtain the optimal channel data of the EEG data to be decoded, the decoding model first compresses the channel data of the EEG data to be decoded into the minimum channel data, that is, removes redundant, Useless equal-channel data to reduce the amount of initial input data, and then extract the strategy features related to the optimal number of channels for the minimum channel data, and according to the extracted strategy features, the EEG data to be decoded and the optimal number of channels are extracted. The relevant decision probability calculation is to obtain the decision probability value of each optimal number of channels in the EEG data to be decoded, and finally select the EEG data suitable for the EEG data to be decoded from the EEG data to be decoded according to the decision probability value. The number of optimal channels for decoding to obtain optimal channel data for decoding the EEG data to be decoded. Among them, by extracting strategy features from the least channel data, it is possible to use less channel data to obtain strategy features that comprehensively cover all channels in the EEG data to be decoded, so that in the process of strategy feature extraction, the amount of computation is significantly reduced. Effectively improve the decoding rate.

A brain-computer interface decoding acceleration system based on adaptive EEG channel selection provided by the present invention will be described below with reference to FIG. A brain-computer interface decoding acceleration method based on adaptive EEG channel selection can be referred to each other.

As shown in FIG. 2, a brain-computer interface decoding acceleration system based on adaptive EEG channel selection provided by the present invention includes an acquisition module 210 and an execution module 220; wherein,

The obtaining module 210 is used for obtaining the EEG data to be decoded;

The execution module 220 is configured to input the EEG data to be decoded into a decoding model to obtain the optimal channel for decoding the EEG data to be decoded;

Wherein, the decoding model is used to perform feature extraction based on the minimum channel data compressed into the EEG data to be decoded to obtain strategy features, and select the optimal number of channels according to the strategy features to obtain the optimal channel data. The optimal channel data is intended to decode the EEG data to be decoded.

It should be noted that a brain-computer interface decoding acceleration system based on adaptive EEG channel selection of the present invention uses a decoding model to select channel data, and specifically, performs channel data selection by converting EEG data to be decoded into the minimum channel data. For the strategy features obtained by feature extraction, the optimal channel data is selected from the EEG data to be decoded, so the EEG data to be decoded can be compressed first after being input into the decoding model, thereby removing redundant or other irrelevant channel data , reduces the amount of channel data, and then selects the optimal channel data from the compressed channel data, which improves the decoding efficiency, that is, after different EEG data to be decoded is input into the decoding model, it can be adaptively selected. The optimal channel data required to improve decoding efficiency.

In a preferred solution, the execution module 220 performs compression processing on the channel data based on the input EEG data to be decoded to obtain the minimum channel data of the EEG data to be decoded;

Based on the minimum channel data, the strategy feature extraction related to the optimal channel number is performed to obtain the strategy feature related to the optimal channel number;

Based on the strategy feature, a decision probability calculation is performed on the to-be-decoded EEG data to obtain a decision-probability value in which each optimal number of channels in the to-be-decoded EEG data is selected;

Based on the decision probability value, obtain the optimal number of channels selected;

Based on the selected optimal number of channels, the EEG data to be decoded is converted into a data format corresponding to the selected optimal number of channels, and the data corresponding to the optimal channel number of the EEG data to be decoded is obtained Format decoding result of intent decoding.

A brain-computer interface decoding acceleration system based on adaptive EEG channel selection provided by an embodiment of the present invention is used to sign the brain-computer interface decoding acceleration method based on adaptive EEG channel selection of various embodiments. For the specific methods and processes of the respective modules included in the brain-computer interface decoding acceleration system based on adaptive EEG channel selection, please refer to the above-mentioned embodiment of the brain-computer interface decoding acceleration method based on adaptive EEG channel selection, here No longer.

A brain-computer interface decoding acceleration system based on adaptive EEG channel selection of the present invention is used in the brain-computer interface decoding acceleration methods based on adaptive EEG channel selection in the foregoing embodiments. Therefore, the descriptions and definitions in the method for accelerating brain-computer interface decoding based on adaptive EEG channel selection in the foregoing embodiments can be used for the understanding of each execution module in the embodiments of the present invention.

A method for training a decoding model of the present invention will be described below with reference to FIGS. 3 to 9 . It can be understood that the decoding model trained by this method can be applied to the aforementioned acceleration of brain-computer interface decoding based on adaptive EEG channel selection. method or a brain-computer interface decoding acceleration system based on adaptive EEG channel selection; as shown in Figure 3, the method includes the following steps:

301. Input the EEG data sample to be decoded;

302. Perform channel data compression on the EEG data samples to be decoded through multiple candidate transformation matrices constructed according to preset channel transformation rules to form multiple optimal channel data candidates;

303. Arrange the optimal channel data candidates in ascending order of the number of channels to construct an optimal channel data candidate library;

304. After obtaining the policy features related to the optimal channel data candidate with the least number of channels in the optimal channel data candidate database and the optimal channel number candidate, obtain a plurality of the optimal channel number candidates to be selected according to the policy features. The decision probability value of ;

305. Use the argmax function to select the optimal channel number of the EEG data sample to be decoded in the optimal channel number candidate based on the decision probability value;

306. After converting the EEG data samples to be decoded into a data format corresponding to the optimal number of channels, perform intention decoding;

307. Calculate the decoding loss of the EEG data samples to be decoded using a pre-built loss function, and determine whether the decoding loss satisfies a preset loss standard; if so, go to 308; if not, jump to 309;

308. Use the candidate conversion matrix as the optimal conversion matrix, and obtain the decoding model that has been trained;

309 . Update the candidate transformation matrix, and return to compress the EEG data samples to be decoded by channel data compression through the updated candidate transformation matrix to form a new optimal channel data candidate, and then return to 303 .

It should be noted that, referring to the schematic diagram of the decoding model shown in FIG. 4 , the decoding model is obtained by continuously selecting the optimal decoding channel data of the EEG data samples to be decoded, wherein the EEG data samples to be decoded are all channels. EEG data to be decoded.

，即每个试次X包 含N个脑电通道，T个时序样本。通过上述公式1，候选转换矩阵将所述待解码脑电数据样本 的通道数据压缩，形成多个最优通道数据候选，最优通道数据候选库以通道数目升序排列 定义为: Specifically, the original input EEG data samples to be decoded are first defined as

, that is, each trial X contains N EEG channels and T time series samples. By above-mentioned formula 1, the candidate transformation matrix compresses the channel data of the EEG data samples to be decoded to form a plurality of optimal channel data candidates, and the optimal channel data candidate library is defined in ascending order of the number of channels as:

公式1；

Formula 1;

，

。 in,

,

.

被输入到策略网络层中，用于提取策略特征， 计算方法如下： Further, in order to reduce the calculation amount of the policy network layer, the channel data form of the optimal channel data candidate with the least number of channels in the optimal channel data candidate database is

is input into the policy network layer to extract policy features. The calculation method is as follows:

公式2；

formula 2;

Among them, PM represents the policy feature extraction model, which is mainly composed of two layers, the first layer is the time domain convolution layer, and the second layer is the spatial domain convolution layer.

The extracted policy features output different behavioral decision probabilities through the fully connected layer, that is, the probability P that each optimal channel number candidate in the optimal channel data candidate database is selected:

公式3；

formula 3;

In the training process of the conventional strategy network layer, it is necessary to obtain the discrete behavior A through the argmax operation, that is, the channel sample label corresponding to the optimal channel number candidate:

公式4；

formula 4;

代表Gumbel分布；

代表均匀分布。 Among them, i represents the optimal channel number candidate index;

Represents the Gumbel distribution;

represents a uniform distribution.

That is, the optimal number of channels is selected through the strategy network layer, the EEG data samples to be decoded are intentionally decoded in the data format corresponding to the optimal number of channels, and finally the decoding loss function is used to calculate the EEG data samples to be decoded. The decoding loss to continuously train and optimize the decoding model through the loss function.

Further, at the beginning of the training of the decoding model, a plurality of candidate transformation matrices need to be constructed according to the preset channel transformation rules, so as to facilitate the first channel data compression of the EEG data to be decoded. It can be understood that the first constructed candidate transformation matrix The more the matrix conforms to the EEG data decoding rules, the easier it is to obtain the optimal conversion matrix. Therefore, the preset channel conversion rules for compressing the EEG data samples to be decoded in this application are based on prior knowledge in the field of EEG signal decoding. That is, in another embodiment of the present invention, the preset channel conversion rule specifically includes: a difference rule, an average rule, and a selective activation rule; wherein,

The difference rule realizes channel data compression by calculating the difference between the EEG data collected by the corresponding electrodes in the two brain regions;

The mean value rule realizes channel data compression by averaging the EEG data of adjacent channels;

The selective activation rule realizes channel data compression by removing the channel corresponding to the EEG data irrelevant to the task to be classified.

It should be noted that brain activity, especially the brain signals stimulated by movement or motor imagery, generally shows the enhancement of the EEG signal amplitude in the ipsilateral brain area of the motor (motor imagery) limb and the inhibition of the EEG signal in the contralateral brain area. . Therefore, channel data compression can be achieved by calculating the difference between the EEG signals collected by the corresponding electrodes in the brain regions on both sides, that is, the difference rule; and considering the similarity and redundancy of adjacent EEG information, the The average principle of channel signals, that is, the mean value rule, is used to realize channel data compression; at the same time, the EEG signals of some brain areas may not be related to the task to be classified. In this case, the channel data can be directly removed. Therefore, the selective activation rule is used to realize the channel data. Compression of data.

Specifically, for example, as shown in FIG. 5-FIG. 7, an example of channel data compression using the channel conversion rule of the present invention is exemplified by taking 22-lead EEG as an example, wherein FIG. 5, FIG. 6 and FIG. 7 are respectively The 22 channels are compressed into 12 channels, 9 channels and 6 channels using difference rule, mean value rule and selective activation rule.

，以上述图7的 采用选择性激活规则将22条通道压缩为6通道为例，其对应的转化矩阵T6如公式5所示： Further, construct candidate conversion matrix with preset channel conversion rule

, taking the selective activation rule in Figure 7 to compress 22 channels into 6 channels as an example, the corresponding transformation matrix T6 is shown in formula 5:

公式5；

formula 5;

Through the following formula 6, the EEG data X to be decoded can be converted into a data format containing N ⁱ input channels:

公式6；

formula 6;

Among them, X* represents the EEG data to be decoded containing the * input channels.

It can be understood that the construction of the loss function can ensure that the optimal number of channels obtained by the trained decoding model can meet the decoding requirements, and the most important requirements for decoding are decoding accuracy and decoding efficiency. When the requirements for decoding accuracy and decoding efficiency are not the same, if the corresponding decoding model can be constructed according to the different needs of users, it will make the brain based on adaptive EEG channel selection of the present invention based on adaptive EEG channel selection. The machine interface decoding acceleration method is more flexible to use, so in another embodiment of the present invention, the loss function is the weighted sum of the decoding accuracy loss function and the decoding efficiency loss function; wherein,

The decoding accuracy loss function is a cross-entropy loss function constructed on the ShallowConvNet classification model based on the EEG data samples to be decoded and the real label;

The decoding efficiency loss function is a loss function constructed on the ShallowConvNet classification model by averaging the number of floating-point operations of the EEG data samples input to the decoding model to be decoded.

It should be noted that the ShallowConvNet classification model itself has an efficient decoding speed of EEG signals. Therefore, in the method of the present invention, ShallowConvNet is used to decode EEG data.

More specifically, both the decoding accuracy loss function and the decoding efficiency loss function are constructed on the ShallowConvNet classification model, and the accuracy loss function is:

公式7；

formula 7;

代表原始输入的脑电数据样本和决策网络层输出的行为标签；

代表ShallowConvNet分类模型，

代表模型参数。 in,

EEG data samples representing the original input and behavioral labels output by the decision network layer;

represents the ShallowConvNet classification model,

represents the model parameters.

For different inputs, based on the behavioral decision of the policy network layer, the number of channels of the input data of the classification model is different, resulting in different GFLOPs . In the present invention, the GFLOPs of the training data are averaged as the final decoding loss function.

公式8；

formula 8;

The final constructed loss function is the weighted sum of decoding accuracy loss and decoding efficiency loss:

公式9；

formula 9;

代表权重系数，用于权衡解码效率和解码精度，即

越大，解码效率越高。 in,

represents the weight coefficient, which is used to trade off decoding efficiency and decoding accuracy, namely

The larger the value, the higher the decoding efficiency.

It can be understood that since the argmax operation is non-differentiable, the traditional gradient backpropagation algorithm cannot be used for training, so in another embodiment of the present invention, the updating of the candidate transformation matrix specifically includes:

The argmax function based on the decision probability value is approximated by Gumbel-Softmax, and the candidate transformation matrix is updated according to the approximation result.

It should be noted that argmax is approximated using Gumbel-Softmax, as shown in Equation 10 below:

公式10。

Formula 10.

取不同值时，测试所述解码模型的性能，结果如表1所示： In the following, taking the data source BCI competition IV dataset 2a (BCIC IV 2a) as an example, combined with the 22-lead EEG provided above as an example, an example of channel compression using the channel conversion rule of the present invention uses 10 subjects, pair weight factor

When different values are taken, the performance of the decoding model is tested, and the results are shown in Table 1:

取不同值时解码模型性能的对照表 Table 1 Weight coefficient

Comparison table of decoding model performance when taking different values

取值下的计算 量，实线为解码精度和计算量的拟合曲线。 The performance of the decoding model corresponding to Table 1 is shown in Figure 8, where the asterisks are different

The calculation amount under the value, the solid line is the fitting curve of the decoding accuracy and the calculation amount.

Further, comparing the performance of 10 subjects in the data source BCI competition IV dataset 2a (BCIC IV 2a) using the decoding model provided by the present invention and using the existing ShallowConvNet classification model with the best decoding speed for decoding, the results are as follows: As shown in Figure 9, it can be seen that compared with ShallowConvNet, the decoding model provided by the present invention can reduce the calculation amount by 35% without reducing the accuracy (or even slightly improving). The BCI decoding acceleration method based on adaptive EEG channel selection has obvious advantages.

FIG. 10 illustrates a schematic diagram of the physical structure of an electronic device. As shown in FIG. 10 , the electronic device may include: a processor (processor) 810, a communication interface (Communications Interface) 820, a memory (memory) 830, and a communication bus 840, The processor 810 , the communication interface 820 , and the memory 830 communicate with each other through the communication bus 840 . The processor 810 can invoke the logic instructions in the memory 830 to execute the adaptive EEG channel selection-based brain-computer interface decoding acceleration method based on the adaptive EEG channel selection, the method comprising:

101. Obtain the EEG data to be decoded;

102. Inputting the EEG data to be decoded into a decoding model, and outputting a decoding result of intentionally decoding the EEG data to be decoded;

Wherein, the decoding model is used to perform feature extraction based on the minimum channel data compressed into the EEG data to be decoded to obtain strategy features, and select the optimal number of channels according to the strategy features to obtain the optimal channel data. The optimal channel data is intended to decode the EEG data to be decoded.

In addition, the above-mentioned logic instructions in the memory 830 may be implemented in the form of software functional units and may be stored in a computer-readable storage medium when sold or used as an independent product. Based on such understanding, the technical solution of the present invention can be embodied in the form of a software product in essence, or the part that contributes to the prior art or the part of the technical solution. The computer software product is stored in a storage medium, including Several instructions are used to cause a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, Read-Only Memory (ROM, Read-Only Memory), Random Access Memory (RAM, Random Access Memory), magnetic disk or optical disk and other media that can store program codes .

In another aspect, the present invention also provides a computer program product, the computer program product includes a computer program, the computer program can be stored on a non-transitory computer-readable storage medium, and when the computer program is executed by a processor, the computer can Execute the brain-computer interface decoding acceleration method based on adaptive EEG channel selection based on adaptive EEG channel selection provided by the above methods, and the method includes:

101. Obtain the EEG data to be decoded;

102. Inputting the EEG data to be decoded into a decoding model, and outputting a decoding result of intentionally decoding the EEG data to be decoded;

Wherein, the decoding model is used to perform feature extraction based on the minimum channel data compressed into the EEG data to be decoded to obtain strategy features, and select the optimal number of channels according to the strategy features to obtain the optimal channel data. The optimal channel data is intended to decode the EEG data to be decoded.

In yet another aspect, the present invention also provides a non-transitory computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, is implemented to perform the adaptive EEG channel selection-based algorithm provided by the above methods. A brain-computer interface decoding acceleration method based on adaptive EEG channel selection, the method includes:

101. Obtain the EEG data to be decoded;

102. Inputting the EEG data to be decoded into a decoding model, and outputting a decoding result of intentionally decoding the EEG data to be decoded;

Wherein, the decoding model is used to perform feature extraction based on the minimum channel data compressed into the EEG data to be decoded to obtain strategy features, and select the optimal number of channels according to the strategy features to obtain the optimal channel data. The optimal channel data is intended to decode the EEG data to be decoded.

The device embodiments described above are only illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in One place, or it can be distributed over multiple network elements. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution in this embodiment. Those of ordinary skill in the art can understand and implement it without creative effort.

From the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a necessary general hardware platform, and certainly can also be implemented by hardware. Based on this understanding, the above-mentioned technical solutions can be embodied in the form of software products in essence or the parts that make contributions to the prior art, and the computer software products can be stored in computer-readable storage media, such as ROM/RAM, magnetic A disc, an optical disc, etc., includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to perform the methods described in various embodiments or some parts of the embodiments.

Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, but not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that it can still be The technical solutions described in the foregoing embodiments are modified, or some technical features thereof are equivalently replaced; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims (10)

1. A brain-computer interface decoding acceleration method based on self-adaptive electroencephalogram channel selection is characterized by comprising the following steps:

acquiring electroencephalogram data to be decoded;

inputting the electroencephalogram data to be decoded into a decoding model, and outputting a decoding result of performing intended decoding on the electroencephalogram data to be decoded;

the decoding model is used for carrying out feature extraction on the minimum channel data compressed by the electroencephalogram data to be decoded to obtain strategy features, selecting the optimal channel number according to the strategy features to obtain the optimal channel data, and then carrying out intent decoding on the electroencephalogram data to be decoded through the optimal channel data.

2. The brain-computer interface decoding acceleration method based on adaptive electroencephalogram channel selection according to claim 1, wherein the inputting of the electroencephalogram data to be decoded into a decoding model and the outputting of a decoding result of the intended decoding of the electroencephalogram data to be decoded specifically comprises:

compressing channel data based on input electroencephalogram data to be decoded to obtain the minimum channel data of the electroencephalogram data to be decoded;

extracting the strategy features related to the optimal channel number based on the minimum channel data to obtain the strategy features related to the optimal channel number;

based on the strategy characteristics, performing decision probability calculation on the electroencephalogram data to be decoded to obtain decision probability values of the electroencephalogram data to be decoded, wherein the number of the optimal channels is selected;

obtaining the number of the selected optimal channels based on the decision probability value;

and converting the electroencephalogram data to be decoded into a data format corresponding to the selected optimal channel number based on the selected optimal channel number to obtain a decoding result of performing intended decoding on the electroencephalogram data to be decoded in the data format corresponding to the optimal channel number.

3. A brain-computer interface decoding acceleration system based on self-adaptive electroencephalogram channel selection is characterized by comprising:

the acquisition module is used for acquiring electroencephalogram data to be decoded;

the execution module is used for inputting the electroencephalogram data to be decoded into a decoding model and outputting a decoding result of the intended decoding of the electroencephalogram data to be decoded;

the decoding model is used for carrying out feature extraction on the minimum channel data compressed by the electroencephalogram data to be decoded to obtain strategy features, selecting the optimal channel number according to the strategy features to obtain the optimal channel data, and then carrying out intent decoding on the electroencephalogram data to be decoded through the optimal channel data.

4. A method for training a decoding model is characterized by specifically comprising the following steps:

inputting electroencephalogram data samples to be decoded;

performing channel data compression on the electroencephalogram data sample to be decoded through a plurality of candidate conversion matrixes constructed according to a preset channel conversion rule to form a plurality of optimal channel data candidates;

arranging the optimal channel data candidates in an ascending order of the number of channels to construct an optimal channel data candidate library;

obtaining a strategy characteristic related to an optimal channel data candidate with the least number of channels in the optimal channel data candidate library and the optimal channel number candidate, and then obtaining a decision probability value for selecting a plurality of optimal channel number candidates according to the strategy characteristic;

selecting the optimal channel number of the electroencephalogram data samples to be decoded in the optimal channel number candidates based on the decision probability value by utilizing an argmax function;

converting the electroencephalogram data sample to be decoded into a data format corresponding to the optimal channel number, and then performing intent decoding;

calculating the decoding loss of the electroencephalogram data sample to be decoded by using a pre-constructed loss function, and judging whether the decoding loss meets a preset loss standard;

if so, taking the candidate transformation matrix as an optimal transformation matrix, and obtaining a trained decoding model;

and if not, updating the candidate conversion matrix, and returning to perform channel data compression on the electroencephalogram data sample to be decoded through the updated candidate conversion matrix again.

5. The method for training a decoding model according to claim 4, wherein the preset channel conversion rule specifically includes: a difference rule, a mean rule, and a selective activation rule; wherein,

the difference rule realizes channel data compression by calculating the difference of the electroencephalogram data collected by the corresponding electrodes of the brain areas on the two sides;

the average rule is used for realizing channel data compression by averaging the electroencephalogram data of adjacent channels;

the selective activation rule removes channels corresponding to the electroencephalogram data irrelevant to the tasks to be classified, and channel data compression is achieved.

6. The method of claim 4, wherein the loss function is a weighted sum of a decoding precision loss function and a decoding efficiency loss function; wherein,

the decoding precision loss function is a cross entropy loss function which is constructed on a ShallowConvNet classification model based on the electroencephalogram data sample to be decoded and a real label;

the decoding efficiency loss function is a loss function constructed by averaging the floating point operation times of the electroencephalogram data samples to be decoded, which are input into the decoding model, on a ShallowConvNet classification model.

7. The method for training a decoding model according to claim 4, wherein the updating the candidate transformation matrix specifically includes:

and approximating the argmax function based on the decision probability value by using Gumbel-Softmax, and updating the candidate conversion matrix according to an approximation result.

8. An electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor when executing the program implements the steps of the method for accelerating decoding of a brain-computer interface based on adaptive brain electrical channel selection according to claim 1 or 2 or the method for training a decoding model according to any one of claims 4 to 7.

9. A non-transitory computer readable storage medium, on which a computer program is stored, wherein the computer program, when being executed by a processor, implements the steps of the method for accelerating decoding of a brain-computer interface based on adaptive brain electrical channel selection according to claim 1 or 2, or the method for training a decoding model according to any one of claims 4 to 7.

10. A computer program product comprising a computer program, characterized in that the computer program, when being executed by a processor, implements the steps of the method for accelerating the decoding of a brain-computer interface based on adaptive brain electrical channel selection according to claim 1 or 2 or the method for training a decoding model according to any one of claims 4 to 7.

Images