An ultra-lightweight epilepsy monitoring method and system based on multi-level multiplexing algorithm

Abstract

The present invention discloses an ultra-lightweight epilepsy monitoring method and system based on a multi-level multiplexing algorithm, belonging to the technical field of convolutional neural networks. The technical scheme thereof mainly comprises the following steps: obtaining brain wave sample data at the current time t, brain wave sample inference results at the time t-kn and graph convolution result data at the time t-k, wherein the brain wave sampling length at each time is m×k×n, k is the sampling step length, and n, m and k are all positive integers; extracting data overlapping with the sampling interval at the current time t from the brain wave sample inference results at the time t-kn to obtain first data, extracting data overlapping with the sampling interval at the current time t and not overlapping with the sampling interval at the time t-kn from the graph convolution result data at the time t-k to obtain second data; obtaining the brain wave inference result at the time t according to the first data and the second data, and reducing the amount of calculation of average inference by reusing the result data before the current time.

Description

Technical Field

[0001] The present invention relates to the technical field of convolutional neural networks, and more specifically to an ultra-lightweight epilepsy monitoring method and system based on a multi-level multiplexing algorithm. Background Art

[0002] In the scenario of real-time epilepsy monitoring, in order to ensure the long-term battery life of the device, there are extremely high requirements for the lightweight degree of the model. However, traditional network models often find it difficult to achieve a perfect balance between improving accuracy, reducing latency, and reducing the amount of calculation.

[0003] For example, the patent application with publication number CN113288173A provides an epilepsy detection device based on EEG signals. This technical solution combines low-level features through a convolutional neural network to form abstract high-level representation attribute categories or features, form a description and discrimination of the features, and then obtain epilepsy detection results.

[0004] However, this technical solution uses a neural network to calculate the directly collected data, which requires a large amount of calculation and will increase the calculation load of the computing system. Therefore, the existing technology has shortcomings. Summary of the invention

[0005] In view of the shortcomings of the prior art, the purpose of the present invention is to provide an ultra-lightweight epilepsy monitoring method and system based on a multi-level multiplexing algorithm, which can simultaneously have the advantages of improving accuracy, reducing latency and reducing calculation amount through multiple data multiplexing.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] The present invention provides an ultra-lightweight epilepsy monitoring method based on a multi-level multiplexing algorithm, comprising:

[0008] Get the EEG sample data at the current time t, the EEG sample inference results at time t-kn, and the graph convolution result data at time tk, where the EEG sampling length at each moment is m×k×n, k is the sampling step, and n, m, and k are all positive integers;

[0009] Extract the data overlapping with the sampling interval at the current time t from the EEG sample inference result at time t-kn to obtain the first data, and extract the data overlapping with the sampling interval at the current time t and not overlapping with the sampling interval at time t-kn from the graph convolution result data at time tk to obtain the second data;

[0010] The brain wave inference result at the current time t is obtained based on the first data and the second data.

[0011] As a further improvement of the present invention, obtaining the brain wave inference result at the current time t according to the first data and the second data includes:

[0012] Extract the last k frames of data from the brain wave sample data at the current time t to obtain group data;

[0013] Based on the group data, the first data and the second data, the brain wave inference result at the current time t is obtained.

[0014] As a further improvement of the present invention, obtaining the brain wave inference result at time t according to the first data and the second data further includes:

[0015] Inputting the group of data into a preset graph convolution network to obtain first graph convolution result data;

[0016] The step of obtaining the brain wave inference result at the current time t according to the group data, the first data and the second data includes:

[0017] According to the first graph convolution result data, the first data and the second data, the brain wave inference result at the current time t is obtained.

[0018] As a further improvement of the present invention, obtaining the brain wave inference result at the current time t according to the first graph convolution result data, the first data and the second data includes:

[0019] Combining the first graph convolution result data with the second data to obtain second graph convolution result data;

[0020] Obtaining convolution result data according to the second graph convolution result data and a preset convolution network;

[0021] The brain wave inference result at the current time t is obtained based on the convolution result data and the first data.

[0022] As a further improvement of the present invention, the step of obtaining the convolution result data according to the second graph convolution result data and a preset convolution network includes:

[0023] Inputting the second graph convolution result data into the convolution network to obtain first convolution result data;

[0024] The first convolution result data is multiplied by a weight vector in a positional manner to obtain the convolution result data, where the weight vector is obtained according to an attention channel mechanism.

[0025] As a further improvement of the present invention, obtaining the brain wave inference result at time t according to the convolution result data and the first data includes:

[0026] The convolution result data is combined with the first data to obtain the brain wave inference result at the current time t.

[0027] As a further improvement of the present invention, the above method also includes: voting on the EEG inference results at the current moment t to obtain the epilepsy monitoring results at the current moment t, the epilepsy monitoring results are epilepsy prediction results and epilepsy detection results, the epilepsy prediction results include the seizure period and the non-seizure period, and the epilepsy detection results include the pre-seizure period, the inter-seizure period and the post-seizure period.

[0028] The present invention provides an image monitoring method based on a multi-level multiplexing algorithm, comprising:

[0029] Get the sample data at the current time t, the sample inference results at time t-kn, and the graph convolution result data at time tk, where the sampling length at each time is m×k×n, k is the sampling step length, and n, m, and k are all positive integers;

[0030] Extract the data overlapping with the sampling interval at the current time t from the sample inference result at time t-kn to obtain the first data, and extract the data overlapping with the sampling interval at the current time t and not overlapping with the sampling interval at time t-kn from the graph convolution result data at time tk to obtain the second data;

[0031] An inference result at a current time t is obtained based on the first data and the second data.

[0032] The present invention provides an ultra-lightweight epilepsy monitoring system based on a multi-stage multiplexing algorithm, which is used to implement the above-mentioned ultra-lightweight epilepsy monitoring method based on a multi-stage multiplexing algorithm, specifically comprising:

[0033] Acquisition module: obtain the EEG sample data at the current time t, the EEG sample inference results at time t-kn, and the graph convolution result data at time tk, where the EEG sampling length at each moment is m×k×n, k is the sampling step, and n, m, and k are all positive integers;

[0034] Extraction module: extract the data overlapping with the sampling interval at the current time t from the EEG sample inference result at time t-kn to obtain the first data, and extract the data overlapping with the sampling interval at the current time t and not overlapping with the sampling interval at time t-kn from the graph convolution result data at time tk to obtain the second data;

[0035] Calculation module: obtain the brain wave inference result at the current time t according to the first data and the second data.

[0036] The present invention provides an image monitoring system based on a multi-level multiplexing algorithm, which is used to implement the above-mentioned image monitoring method based on a multi-level multiplexing algorithm, comprising:

[0037] Acquisition module: obtains the sample data at the current time t, the sample inference results at time t-kn, and the graph convolution result data at time tk, where the sampling length at each moment is m×k×n, k is the sampling step length, and n, m, and k are all positive integers;

[0038] Extraction module: extract the data overlapping with the sampling interval at the current time t from the sample inference result at time t-kn to obtain the first data, and extract the data overlapping with the sampling interval at the current time t and not overlapping with the sampling interval at time t-kn from the graph convolution result data at time tk to obtain the second data;

[0039] Calculation module: obtains the inference result at the current time t according to the first data and the second data.

[0040] The present invention only needs to perform inference calculations on the basis of group data each time, and reuse previous inference results to obtain the inference results at the current moment, and monitor epilepsy based on the inference results at the current moment, thereby reducing the average single inference calculation amount. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 A flowchart of the steps of an ultra-lightweight epilepsy monitoring method based on a multi-stage multiplexing algorithm of the present invention;

[0042] Figure 2 A flowchart of the steps of obtaining the brain wave reasoning result at the current time t in the present invention;

[0043] Figure 3 A schematic diagram of group multiplexing of the present invention;

[0044] Figure 4 It is a schematic diagram of convolution multiplexing of the present invention;

[0045] Figure 5 A schematic diagram of window multiplexing of the present invention;

[0046] Figure 6 It is the ACLPI variation trend diagram of different methods under different inference frequencies of the present invention;

[0047] Figure 7 It is a schematic diagram of the structure of the multi-layer neural network model of the present invention. DETAILED DESCRIPTION

[0048] The technical solution of the present invention is described in detail below through the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations of the technical solution of the present invention.

[0049] The term "and / or" in the following text is only a description of the association relationship between the associated objects, indicating that there can be three relationships. For example, A and / or B can mean: A exists alone, A and B exist at the same time, and B exists alone. In addition, the character " / " generally indicates that the associated objects before and after are in an "or" relationship.

[0050] The present application embodiment provides an image monitoring method based on a multi-level multiplexing algorithm, including:

[0051] Get the sample data at the current time t, the sample inference results at time t-kn, and the graph convolution result data at time tk, where the sampling length at each time is m×k×n, k is the sampling step length, and n, m, and k are all positive integers;

[0052] Extract the data overlapping with the sampling interval at the current time t from the sample inference result at time t-kn to obtain the first data, and extract the data overlapping with the sampling interval at the current time t and not overlapping with the sampling interval at time t-kn from the graph convolution result data at time tk to obtain the second data;

[0053] According to the first data and the second data, an inference result at time t is obtained.

[0054] This embodiment obtains the inference result at the current moment by reusing the data at the t-kn moment and the tk moment, which greatly reduces the amount of calculation compared to directly calculating based on the sample data at the current moment.

[0055] Exemplarily, this embodiment takes epilepsy detection as an example to specifically illustrate the above monitoring method.

[0056] In the prior art, there are usually two ways to monitor epilepsy in real time. In one way, the system first samples the EEG signal and then The samples are sent to the network model for inference. To avoid overlap between samples, the next sampling will start from the end of the previous sample. This method ensures that the number of inferences is reduced as much as possible while processing all data. However, this method has the problem of too large a sampling step. The system response delay is mainly affected by the sampling step and the network calculation and inference time. A sampling step that is too large seriously affects the system response speed.

[0057] Method 2 uses overlapping sampling to solve the problem of response delay, and defines the non-overlapping part of two adjacent samples as the step length , which is equivalent to the sampling delay of the system. The sampling delay can be reduced by reducing the step size, but this will increase the computational load of the system accordingly, and the amount of calculation will be ( / ) times.

[0058] Therefore, existing technologies often find it difficult to achieve a perfect balance among improving accuracy, reducing latency, and reducing the amount of computation.

[0059] In order to ensure the long-term endurance of the device in the scenario of real-time epilepsy monitoring on the end side, and to improve the lightweight degree of the detection system, and at the same time to shorten the response delay and respond to epileptic seizures in a timely manner, the embodiment of the present application provides an ultra-lightweight epilepsy monitoring method based on a multi-level multiplexing algorithm based on the above-mentioned image monitoring method, such as Figure 1 As shown, the method includes:

[0060] S1: Obtain the EEG sample data at the current time t, the EEG sample inference results at time t-kn, and the graph convolution result data at time tk, where the EEG sampling length at each moment is m×k×n, k is the sampling step, and n, m, and k are all positive integers;

[0061] S2: extract the data overlapping with the sampling interval at the current time t from the EEG sample inference results at time t-kn to obtain the first data, and extract the data overlapping with the sampling interval at the current time t and not overlapping with the sampling interval at time t-kn from the graph convolution result data at time tk to obtain the second data;

[0062] S3: Based on the first data and the second data, obtain the brain wave inference result at the current time t.

[0063] The end side can be understood as a small data processing unit that is closer to the data source, that is, closer to the patient, such as a local server set up in the ward. Compared with the cloud, it is closer to the site and can quickly receive data from terminal monitoring equipment, perform some real-time, relatively complex data processing and analysis, and respond quickly in the local network environment, such as issuing an alarm in time to remind medical staff that the patient may have an epileptic seizure, etc., reducing dependence on remote cloud transmission and processing, ensuring the timeliness and accuracy of real-time monitoring, and at the same time ensuring data security and privacy to a certain extent.

[0064] Among them, m is the number of window data contained in each sample data, and n is the number of group data contained in each window data. And each moment corresponds to a sampling length. For example, the data at time t corresponds to the data in the range of t-mkn frame to t frame, and the data at time t+nk corresponds to the data in the range of t-(m-1)nk to t+nk frame. The graph convolution result data is obtained according to the graph convolution network (GCN), which is a convolutional neural network that can directly process graph structure data and utilize its structural information. In the epilepsy monitoring task, the electroencephalogram (EEG) data collected by each electrode is regarded as a node, and the relationship between different electrodes is regarded as an edge, so as to construct a graph structure.

[0065] Further, such as Figure 2 As shown, this embodiment provides a step of obtaining the brain wave inference result at the current time t according to the first data and the second data, including:

[0066] S31: extract the last k frames of brain wave sample data at the current time t to obtain group data;

[0067] S32: input the group data into a preset graph convolution network to obtain first graph convolution result data;

[0068] S33: combining the first graph convolution result data with the second data to obtain the second graph convolution result data;

[0069] S34: Obtaining convolution result data according to the second graph convolution result data and a preset convolution network;

[0070] S35: Based on the convolution result data and the first data, the brain wave inference result at the current time t is obtained.

[0071] Specifically, assuming that the sampling step k is 8 frames and the sampling length at each moment is 2560 frames, the EEG sample data at each moment can be divided into m=10 window data, each window data has a length of 256 frames, and each window data can be divided into n=32 group data, each group data is 8 frames.

[0072] like Figure 3As shown, assuming that T+16 is the current time t, and T+8 is the time tk, there are 31 groups of data in the data corresponding to time T+8 and the data corresponding to time T+16 that are the same, and the 31 groups of data are represented by Group2-Group32. Therefore, the calculation results of Group2-Group32 in T+8 can be reused at time T+16, and the calculation results of Group2-Group32 are represented by Tensor2-Tensor32. Similarly, the reasoning results of Group1-Group31 in T+8 come from time T, and the reasoning results of Group1-Group31 are represented by Tensor1-Tensor31. Among them, Tensor2-Tensor32 is the second data, Group33 is the group data, and Tensor33 is the first graph convolution result data. The first graph convolution result data is combined with the second data to obtain the second graph convolution result data, that is, the graph convolution results Tensor2-Tensor33 at the current moment.

[0073] This embodiment divides each window data into multiple groups of data of equal length, so that only the first and last two groups of data in the sample data between two adjacent sampling steps are different. At this time, only the last group of data needs to be calculated, and the calculation results of the previous identical data are reused to obtain the graph convolution result at the current moment.

[0074] Furthermore, this embodiment provides a step of obtaining convolution result data according to the second graph convolution result data and a preset convolution network, including:

[0075] Inputting the second graph convolution result data into the convolution network to obtain the first convolution result data;

[0076] The first convolution result data is multiplied by the weight vector in position to obtain the convolution result data. The weight vector is obtained according to the attention channel mechanism.

[0077] The attention channel mechanism, as a feature detector, can increase the weight of important channels, so that the model can focus on the information of important channels in EEG data.

[0078] Exemplarily, the steps of obtaining the weight vector through the attention channel mechanism are:

[0079] Perform global maximum pooling and global average pooling in the spatial dimension on a W×C input feature map to obtain two 1×1×C feature maps. The purpose of pooling in the spatial dimension is to compress the spatial size to facilitate the subsequent learning of channel features, where C represents the number of channels contained in the feature map, and W represents the size-related quantity of the feature map in the spatial dimension, such as the width or a spatial dimension-related length after the feature map is expanded into a one-dimensional vector form. It and the number of channels C together define the shape and scale of the input feature map;

[0080] Then, the results of global maximum pooling and global average pooling are sent to a shared multi-layer perceptron (MLP) for learning, and two 1×1×C feature maps are obtained. The number of neurons in the first layer of MLP is C / r, and the activation function is Relu. The function of the first layer of neurons is to perform a certain degree of feature compression on the channel dimension, reduce the amount of calculation, and perform preliminary feature extraction. The number of neurons in the second layer is C. The function of the second layer of neurons is to restore the features after the first layer of compression learning to the dimension consistent with the original number of channels, so as to learn more abstract and critical features of the channel dimension and the importance of each channel. r is a reduction factor used to adjust the structure of the multi-layer perceptron;

[0081] The result of the MLP output is added (Add), and then mapped through the Sigmoid activation function to finally obtain the channel attention weight vector.

[0082] The convolutional network achieves gradual dimensionality reduction of data by setting up multiple convolutional layers. In the step of obtaining the convolution result data, the number of convolutions can be reduced in each convolution layer by reusing the previous result data, thereby improving the computational efficiency.

[0083] For example, Figure 4As shown in the figure, assuming that the sampling step is 1 frame, the convolution network contains only one convolution layer, that is, only one convolution calculation is required, T+1 is the current time t, T is the time tk, and the convolution result data of the second graph at time T and time T+1 can be represented as a matrix of 5 rows and 11 columns. Since the convolution kernel in the convolution network slides horizontally layer by layer, the 2nd to 10th columns of the first convolution result data at time T and the 1st to 9th columns of the first convolution result data at time T+1 are exactly the same, so when calculating the first convolution result data at time T+1, only the newly added 12 columns of convolution results need to be calculated, and then the 2nd to 10th columns of the result data at time T are reused to obtain the first convolution result data at the current time. When the convolution network contains multiple convolutional layers, each layer can use the above-mentioned example method to reuse the convolution results, that is, in two adjacent convolutional layers, when the next convolutional layer needs to convolve the calculation result of the previous convolutional layer, it is only necessary to perform convolution calculation on the convolution result of the newly added data in the previous convolutional layer, and reuse the convolution result of the next convolutional layer at the previous moment, so as to obtain the convolution result of the next convolutional layer. For example, Figure 3 As shown, at time T+16, when convolution layer 2 needs to perform convolution on convolution layer 1, it is only necessary to perform convolution calculation on the convolution result of Tensor33 in convolution layer 1, that is, it is only necessary to perform convolution calculation on the result obtained by convolution calculation 1, and reuse the result obtained by convolution layer 2 at time T+8 for the convolution result of Tensor2-Tensor32 in convolution layer 1, so as to obtain the convolution result of convolution layer 2 at time T+16, and then repeat the above steps at convolution layer 3 and convolution layer 4 to obtain the first convolution result data at time T+16.

[0084] This embodiment proposes a method for reusing convolution results in a convolution layer based on the above-mentioned group data reuse, so as to reduce the number of convolutions and the amount of calculation.

[0085] Furthermore, this embodiment provides a step of obtaining a brain wave inference result at the current time t according to the convolution result data and the first data, including:

[0086] The convolution result data is combined with the first data to obtain the brain wave inference result at the current time t.

[0087] For example, Figure 5 As shown in the figure, assuming that the sampling step length k is 8 frames and the sampling length at each moment is 2560 frames, the EEG sample data at each moment can be divided into m=10 window data, and each window data can be divided into n=32 group data, where T+256 is assumed to be the current moment t, and T is the moment t-kn. Figure 5It can be seen that the window data corresponding to time T is Window1-Window10, and its corresponding Figure 5 It is simplified to Win1-Win10, and the corresponding reasoning result is Result1-Result10. The window data corresponding to time T+256 is Window2-Window11, and the corresponding reasoning result is Result2-Result11. There are 9 window data in the samples at time T and time T+256 that are the same. The 9 window data are represented by Window2-Window10. Therefore, at time T+256, the reasoning results of Window2-Window10 at time T can be reused and only the data in Window11 need to be reasoned and calculated. The reasoning result of Window2-Window10 is represented by Result2-Result10. Then, the calculation result Result11 of Window11 is combined with the calculation result Result2-Result10 of Window2-Window10 to obtain the EEG reasoning result at the current time t. Then, the EEG reasoning result at the previous time t is voted to obtain the epilepsy monitoring result at the current time t, that is, Figure 5 The final result in , where Result11 is the convolution result data, and Result2-Result10 is the first data.

[0088] In combination with the above-mentioned group data-based reuse method, it can be seen that, except for the initial reasoning process at system startup, in each subsequent reasoning process, the reasoning results of the first 9 windows at the current time t can always reuse the results at time t-256, and in the last window, the calculation results of the first 31 group data (Group) at time t-8 can always be reused.

[0089] This embodiment reuses window data to divide the sample data at each moment into multiple window data of equal length, so that the sample data at time t and time t-kn are different only in the first and last two window data, while the middle window data are all the same. At this time, only the last window data needs to be calculated, and the calculation results of the same data are reused to obtain the current reasoning result, which greatly reduces the amount of calculation, improves the calculation efficiency and ensures the accuracy of reasoning.

[0090] Furthermore, the above method also includes: voting on the EEG inference results at the current time t to obtain the epilepsy monitoring results at the current time t. Whether the epilepsy monitoring results are finally output as binary or three-classification outputs needs to be determined according to the monitoring task. For example, in the seizure prediction task, the pre-seizure period, interictal period and post-seizure period are uniformly divided into the non-seizure period, and the model will perform binary classification on the seizure period and the non-seizure period; in the seizure detection task, the model will perform three-classification on the pre-seizure period, interictal period and post-seizure period.

[0091] The voting process is carried out in the fully connected layer. The data of each window needs to be input into the model for calculation, and the calculation results of each window data are voted by the majority, and the classification result is finally outputted based on the voting results.

[0092] An ultra-lightweight epilepsy monitoring method based on a multi-level multiplexing algorithm provided in an embodiment of the present application forms a multi-level multiplexing algorithm through window multiplexing, group multiplexing and convolution multiplexing. Inference calculations are performed only based on a small amount of changed data each time, thereby reducing the amount of calculation while ensuring calculation accuracy, and at the same time achieving the purpose of reducing latency by controlling the sampling step size.

[0093] Furthermore, the embodiment of the present application provides an ultra-lightweight epilepsy monitoring system based on a multi-level multiplexing algorithm, comprising:

[0094] Acquisition module: obtain the EEG sample data at the current time t, the EEG sample inference results at time t-kn, and the graph convolution result data at time tk, where the EEG sampling length at each moment is m×k×n, k is the sampling step, and n, m, and k are all positive integers;

[0095] Extraction module: extract the data overlapping with the sampling interval at the current time t from the EEG sample inference results at time t-kn to obtain the first data, and extract the data overlapping with the sampling interval at the current time t and not overlapping with the sampling interval at time t-kn from the graph convolution result data at time tk to obtain the second data;

[0096] Calculation module: obtain the brain wave inference result at the current time t according to the first data and the second data.

[0097] Furthermore, the embodiment of the present application provides an image monitoring system based on a multi-level multiplexing algorithm, comprising:

[0098] Acquisition module: obtains the sample data at the current time t, the sample inference results at time t-kn, and the graph convolution result data at time tk, where the sampling length at each moment is m×k×n, k is the sampling step length, and n, m, and k are all positive integers;

[0099] Extraction module: extract the data overlapping with the sampling interval at the current time t from the sample inference result at time t-kn to obtain the first data, and extract the data overlapping with the sampling interval at the current time t and not overlapping with the sampling interval at time t-kn from the graph convolution result data at time tk to obtain the second data;

[0100] Calculation module: according to the first data and the second data, obtain the inference result at the current time t. For example, in order to further illustrate the advantages of an ultra-lightweight epilepsy monitoring system based on a multi-level multiplexing algorithm provided by this embodiment, this embodiment uses the CHB-MIT data set for experiments. Since the sampling rate of the data set is 256 Hz (i.e., 256 frames of data are sampled per second), this embodiment uses the sampling step size Set to 8 frames (31.25 ms), sample length Set to 2560 frames (10 seconds), and set the inference frequency to 32 times per second. Each 10-second sample will be divided into 10 1-second window data (Window), and each window is further subdivided into 32 groups of 8 frames in length (Group).

[0101] The experimental results are shown in Table 1 and Figure 6 As shown in Table 1, D in the task column represents the seizure detection task, and P represents the seizure prediction task. As the inference frequency increases, the average computational load per inference (ACLPI) of the traditional inference model remains unchanged. However, the ACLPI of the ultra-lightweight epilepsy monitoring system based on the multi-level multiplexing algorithm proposed in this embodiment decreases significantly with the increase of the inference frequency. At an inference frequency of 32 times per second, the ACLPI of this system is reduced to only 0.051 million multiply-accumulate operations (MMACs), which is more than 97% less than that of previous research work, indicating that the ultra-lightweight epilepsy monitoring system based on the multi-level multiplexing algorithm provided in this embodiment has strong feasibility.

[0102] The ultra-lightweight epilepsy monitoring method and system based on the multi-level multiplexing algorithm provided in the embodiment of the present application uses the multi-level multiplexing algorithm and Figure 7 The multi-layer neural network model shown in Figure 1 can effectively detect epilepsy while improving accuracy, reducing latency and reducing the amount of calculation. Figure 7ReLU in the above is an activation function in the neural network layer, BN means batch normalization, Conv means convolution, S means step size, 5x1, 4x1, 3x1 means the size of convolution kernel, Max Pool means maximum pooling operation, 2x1 means the size of pooling window, GCN means graph convolution, FC means full connection, FC1 and FC2 means two different full connection layers, Input means the dimension of input data, Output means the dimension of output data, different Outputs correspond to different classification tasks, Dropout means random inactivation of neurons, Dropout(0.5) means that during the training process of neural network, the output of neurons is randomly set to 0 with a probability of 0.5, and Scale means weighting.

[0103] Table 1 ACLPI results of different methods

[0104]

[0105] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems) and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0106] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0107] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0108] The above is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should also be regarded as the protection scope of the present invention.

Claims

1. An ultra-lightweight epilepsy monitoring method based on a multi-level multiplexing algorithm, characterized in that: include: Get the EEG sample data at the current time t, the EEG sample inference results at time t-kn, and the graph convolution result data at time tk, where the EEG sampling length at each moment is m×k×n, k is the sampling step, and n, m, and k are all positive integers; Extract the data overlapping with the sampling interval at the current time t from the EEG sample inference result at time t-kn to obtain the first data, and extract the data overlapping with the sampling interval at the current time t and not overlapping with the sampling interval at time t-kn from the graph convolution result data at time tk to obtain the second data; Extract the last k frames of data from the brain wave sample data at the current time t to obtain group data; Inputting the group of data into a preset graph convolution network to obtain first graph convolution result data; Combining the first graph convolution result data with the second data to obtain second graph convolution result data; Obtaining convolution result data according to the second graph convolution result data and a preset convolution network; The brain wave inference result at the current time t is obtained based on the convolution result data and the first data.

2. The ultra-lightweight epilepsy monitoring method based on a multi-level multiplexing algorithm according to claim 1, characterized in that: The step of obtaining the convolution result data according to the second graph convolution result data and a preset convolution network includes: Inputting the second graph convolution result data into the convolution network to obtain first convolution result data; The first convolution result data is multiplied by a weight vector in a positional manner to obtain the convolution result data, where the weight vector is obtained according to an attention channel mechanism.

3. The ultra-lightweight epilepsy monitoring method based on a multi-level multiplexing algorithm according to claim 2, characterized in that: The step of obtaining the brain wave inference result at the current time t according to the convolution result data and the first data includes: The convolution result data is combined with the first data to obtain the brain wave inference result at the current time t.

4. The ultra-lightweight epilepsy monitoring method based on a multi-level multiplexing algorithm according to claim 3, characterized in that: The method also includes: voting on the EEG inference results at the current moment t to obtain the epilepsy monitoring results at the current moment t, wherein the epilepsy monitoring results are epilepsy prediction results and epilepsy detection results, wherein the epilepsy prediction results include the seizure period and the non-seizure period, and the epilepsy detection results include the pre-seizure period, the inter-seizure period, and the post-seizure period.

5. An image monitoring method based on a multi-level multiplexing algorithm, characterized in that: include: Get the sample data at the current time t, the sample inference results at time t-kn, and the graph convolution result data at time tk, where the sampling length at each time is m×k×n, k is the sampling step length, and n, m, and k are all positive integers; Extract the data overlapping with the sampling interval at the current time t from the sample inference result at time t-kn to obtain the first data, and extract the data overlapping with the sampling interval at the current time t and not overlapping with the sampling interval at time t-kn from the graph convolution result data at time tk to obtain the second data; Extract the last k frames of data from the sample data at the current time t to obtain group data; Inputting the group of data into a preset graph convolution network to obtain first graph convolution result data; Combining the first graph convolution result data with the second data to obtain second graph convolution result data; Obtaining convolution result data according to the second graph convolution result data and a preset convolution network; According to the convolution result data and the first data, an inference result at the current time t is obtained.

6. An ultra-lightweight epilepsy monitoring system based on a multi-stage multiplexing algorithm, which is used to implement an ultra-lightweight epilepsy monitoring method based on a multi-stage multiplexing algorithm as described in any one of claims 1 to 4, characterized in that: include: Acquisition module: obtain the EEG sample data at the current time t, the EEG sample inference results at time t-kn, and the graph convolution result data at time tk, where the EEG sampling length at each moment is m×k×n, k is the sampling step, and n, m, and k are all positive integers; Extraction module: extract the data overlapping with the sampling interval at the current time t from the EEG sample inference result at time t-kn to obtain the first data, and extract the data overlapping with the sampling interval at the current time t and not overlapping with the sampling interval at time t-kn from the graph convolution result data at time tk to obtain the second data; The calculation module extracts the last k frames of brain wave sample data at the current time t to obtain group data; inputs the group data into a preset graph convolution network to obtain first graph convolution result data; combines the first graph convolution result data with the second data to obtain second graph convolution result data; According to the second graph convolution result data and the preset convolution network, convolution result data is obtained; according to the convolution result data and the first data, the brain wave inference result at the current time t is obtained.

7. An image monitoring system based on a multi-level multiplexing algorithm, used to implement the image monitoring method based on a multi-level multiplexing algorithm as claimed in claim 5, characterized in that: include: Acquisition module: obtains the sample data at the current time t, the sample inference results at time t-kn, and the graph convolution result data at time tk, where the sampling length at each moment is m×k×n, k is the sampling step length, and n, m, and k are all positive integers; Extraction module: extract the data overlapping with the sampling interval at the current time t from the sample inference result at time t-kn to obtain the first data, and extract the data overlapping with the sampling interval at the current time t and not overlapping with the sampling interval at time t-kn from the graph convolution result data at time tk to obtain the second data; Computing module: extract the last k frames of data from the sample data at the current time t to obtain group data; input the group data into a preset graph convolution network to obtain first graph convolution result data; combine the first graph convolution result data with the second data to obtain second graph convolution result data; obtain convolution result data based on the second graph convolution result data and the preset convolution network; obtain inference result at the current time t based on the convolution result data and the first data.

Images