Electroencephalogram abnormal signal detection method, device and equipment based on convolutional neural network

The CNN-based EEG signal detection method addresses the challenges of automated EEG signal processing by enhancing noise reduction, segmentation, and multi-scale feature extraction, resulting in accurate and efficient abnormal signal detection.

CN119074015APending Publication Date: 2024-12-06COMP NETWORK INFORMATION CENT CHINESE ACADEMY OF SCI
View PDF 0 Cites 3 Cited by

Patent Information

Application Number
CN202411459668.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-10-18
Publication Date
2024-12-06

AI Technical Summary

Technical Problem

Current automated EEG signal detection algorithms face challenges in accurately locating abnormal areas and achieving sufficient recognition accuracy due to high time-variability, complex time-frequency characteristics, and interference from noise and artifacts, leading to unsatisfactory detection results.

Method used

A method using a convolutional neural network (CNN) for EEG signal processing, involving noise reduction, signal segmentation, transformation to two-dimensional frequency domain, and multi-scale feature extraction and fusion to detect and locate abnormal signals.

Benefits of technology

The method achieves accurate positioning and classification of abnormal EEG signals with improved computing efficiency, ensuring precise detection and localization of various signal types.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119074015A_ABST
    Figure CN119074015A_ABST
Patent Text Reader

Abstract

The invention provides an electroencephalogram abnormal signal detection method, device and equipment based on a convolutional neural network. Noise reduction processing and data segmentation are carried out on electroencephalogram abnormal signal data, one-dimensional time sequence information of the signal data is converted into two-dimensional frequency domain information, amplitude data and phase data of the signals are obtained, and then the two kinds of data are input into a preset target convolutional neural network prediction model. And outputting feature map information of the plurality of scale signals, determining a target range in the detected target signal data according to the feature information of the plurality of scale signals, and judging the type of the abnormal signal. According to the method, the abnormal information in the electroencephalogram time sequence data can be detected and positioned while certain calculation efficiency is guaranteed, and accurate positioning and correct classification of various types of information are realized through a multi-level detection algorithm for signals of different scales.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of bioinformatics, and in particular to a method, device and equipment for detecting abnormal EEG signals based on a convolutional neural network. Background Art

[0002] In the vast field of medical research and clinical practice, electroencephalogram (EEG) plays a vital role as an indispensable neurophysiological monitoring tool. This technology relies on an electrode array precisely arranged on the surface of the scalp to non-invasively capture and record the weak electrical signals generated by the activity of brain neuron clusters, thereby providing indirect and in-depth insights into the evaluation of brain functional status. EEG has shown unique value in the identification and evaluation of many neuropsychiatric diseases, especially in epilepsy, sleep disorders and cognitive disorders. Its non-invasiveness, high temporal resolution and versatility make it an important medical means to evaluate, monitor and study brain function. With the continuous advancement of technology and the deepening of its application, EEG will also play a more important role in future medical research and clinical diagnosis and treatment.

[0003] In the process of monitoring the patient's condition, the EEG signals of patients with neurological diseases are complex and diverse, and have different characteristics at different times. The existing detection algorithms may not be able to fully characterize these diverse signals. Therefore, in the existing medical field, the task of abnormal EEG signal detection is highly dependent on manual judgment by medical experts with deep professional knowledge and rich clinical experience.

[0004] Specifically, the EEG signals of patients with neurological diseases (such as epilepsy, Parkinson's disease, and Alzheimer's disease) often show high time-variability and complex time-frequency characteristics. These characteristics not only reflect the activities of different brain regions, but are also closely related to the progression of the disease and individual differences. Therefore, automated detection algorithms often face significant challenges when processing these signals, and it is difficult to achieve the same accuracy and sensitivity as human experts. In addition, current detection methods have many limitations, such as the influence of hardware noise, eye movement and heartbeat artifacts, and variable abnormal signal sizes, which are not conducive to detection. Current automated algorithms have problems such as the inability to accurately locate abnormal areas and insufficient recognition accuracy, resulting in unsatisfactory detection results. Summary of the invention

[0005] The present invention describes a method, device and equipment for detecting abnormal EEG signals based on convolutional neural network, which can solve the above technical problems.

[0006] According to a first aspect, a method for detecting abnormal EEG signals based on a convolutional neural network is provided, the method comprising:

[0007] Acquire signal data to be detected, perform noise reduction processing on the signal data, and segment the signal data after the noise reduction processing;

[0008] The segmented signal data is transformed from one-dimensional time series information to two-dimensional frequency domain information to obtain target signal data; wherein the target signal data includes amplitude data and phase data of the signal, the amplitude represents the strength of the signal at a certain scale and position, and the phase represents the relative time position and relative change trend of the signal;

[0009] Inputting the target signal data into a target convolutional neural network prediction model to obtain a model prediction result; wherein the target convolutional neural network prediction model includes an input end, a backbone network end, a feature fusion end, and an output end;

[0010] The backbone network end is composed of a CBL module and multiple Resc modules. The backbone network end performs preliminary feature extraction on the amplitude data and phase data of the signal through two channels, fuses the two extracted features, and then further extracts the fused features through three layers of Resc modules, wherein the two channels are composed of one layer of CBL module and two layers of Resc modules;

[0011] The feature fusion end is composed of a convolution layer, an upsampling layer, and a splicing operation layer, and is used to fuse the features extracted by the backbone network end to obtain feature maps of three scales; wherein the first scale feature map is obtained by convolution operation on the feature map extracted by the last layer Resc module of the backbone network end, the second scale feature map is obtained by fusion and convolution operation on the feature map extracted by the last layer and the middle layer Resc module of the backbone network end, and the third scale feature map is obtained by fusion and convolution operation on the feature map extracted by the second scale feature map and the first layer Resc module of the backbone network end;

[0012] Based on the model prediction results, the target range in the detected target signal data is determined, and the type of abnormal signal is judged.

[0013] In some embodiments, the method for performing noise reduction processing on signal data specifically includes:

[0014] After collecting signal data from multiple electrode channels, the average vector of all electrode lead potential values ​​is calculated at each time point;

[0015] For each electrode channel, the difference between its potential value and the average vector at that time point is calculated, and this is used as the reference adjusted electrode potential;

[0016] Remove low-frequency drift and high-frequency noise from your signals, and eliminate power line interference.

[0017] In some embodiments, the method for segmenting the signal data after noise reduction processing specifically includes:

[0018] The signal data after noise reduction is segmented using a certain window size to obtain data segments that can be quickly calculated by the model;

[0019] When the subsequent data is segmented, sliding segmentation is performed according to the set step size to capture the continuity characteristics in the data.

[0020] In some embodiments, the method for transforming the segmented signal data specifically includes:

[0021] The signal data is transformed using a continuous wavelet transform method to obtain a transformation result; wherein the result obtained by the continuous wavelet transform is a complex matrix composed of a real part and an imaginary part;

[0022] The conversion result is further converted into amplitude and phase; specifically, the amplitude calculation formula is:

[0023]

[0024] Where W(a,b) represents the result of continuous wavelet transform, |W(a,b)| represents the signal amplitude, Re{W(a,b)} and Im{W(a,b)} represent the real part and imaginary part of the complex matrix respectively;

[0025] The phase calculation formula is:

[0026]

[0027] Among them, θ(a,b) represents the phase, Represents the argument of a complex number;

[0028] The converted amplitude and phase are normalized and scaled to a certain value range.

[0029] In some embodiments, the Resc module in the backbone network end is composed of a CBL module and multiple ResU modules; wherein the ResU module is a convolution module composed of multiple branches, whose main trunk is composed of two CBL modules in series, and whose branch trunks are composed of a 1*1 convolution module and a ReLU activation function. When the two trunks are merged, the target signal data features will be accumulated element by element to achieve fusion. The branch trunks are used to alleviate the model degradation phenomenon caused by the model depth to ensure a good extraction effect of the target signal data features.

[0030] In some embodiments, the feature fusion end fuses the features extracted by the backbone network end to obtain a feature map of three scales, specifically including:

[0031] Extract the feature map from the last layer of the Resc module at the backbone network end, and perform convolution operation on it through the multi-layer CBL module to obtain the first scale feature map; wherein, the first scale feature map is used to detect larger abnormal signal targets;

[0032] The first scale feature map is extracted through the CBL module, and then deconvolved by the upsampling layer. The first scale feature map after deconvolution is fused with the feature map extracted from the middle layer Resc module at the backbone network end through the splicing operation layer. After further convolution operation, the second scale feature map is generated. The second scale feature map is used to detect medium-sized abnormal signal targets.

[0033] The second-scale feature map is extracted through the CBL module, and then deconvolved using the upsampling layer. The second-scale feature map after the deconvolution operation is fused with the feature map extracted from the first-layer Resc module at the backbone network end through the splicing operation layer. After further convolution operation, the third-scale feature map is generated; the third-scale feature map is used to detect small abnormal signal targets.

[0034] In some embodiments, the output end is responsible for outputting the prediction results of the model; wherein the output information includes: bounding box coordinates, confidence, and category probability. Specifically, the bounding box coordinates include the center coordinates and the offset of the edge from the center. The confidence indicates the confidence probability score of whether there is an abnormal signal in the bounding box, and the category probability indicates the probability that the abnormal signal in the bounding box belongs to each abnormal category.

[0035] In some embodiments, a method for determining a target range in detected target signal data and a type of abnormal signal according to a prediction result of a model specifically includes:

[0036] Adjust the anchor box coordinates and limit them between 0 and 1 through the sigmoid activation function;

[0037] The probability of an abnormal signal existing in the bounding box is calculated by the sigmoid function, and the probability of the abnormal signal belonging to each abnormal category is calculated by the softmax function;

[0038] The generated detection frames are subjected to non-maximum suppression through the NMS algorithm, the overlapping abnormal detection frames are removed, and the detection frames with the highest confidence are retained; the retained frames with the highest confidence represent the target range in the detected target signal data predicted by the model.

[0039] According to a second aspect, a device for detecting abnormal EEG signals based on a convolutional neural network is provided, comprising:

[0040] A data acquisition module, used to acquire abnormal EEG signal data;

[0041] A data processing module, used for performing noise reduction processing on the signal data, segmenting the signal data after the noise reduction processing, and transforming the segmented signal data;

[0042] A model prediction module is used to make predictions based on the processed signal data to obtain prediction results;

[0043] The prediction result determination module is used to determine the target range in the detected target signal data according to the prediction result and judge the type of abnormal signal.

[0044] According to a third aspect, an electronic device is provided, comprising a memory and a processor, wherein the memory stores a computer program executable on the processor, and wherein the processor implements the method described in the first aspect when executing the computer program.

[0045] The present invention provides a method, device and equipment for detecting abnormal EEG signals based on convolutional neural networks. By performing noise reduction processing and segmentation on abnormal EEG signal data, and converting the one-dimensional time series information of the signal data into two-dimensional frequency domain information, the amplitude data and phase data of the signal are obtained, and the two types of data are input into a preset target convolutional neural network prediction model to output feature map information of multiple scale signals, and determine the target range in the detected target signal data based on the feature information of the multiple scale signals, and judge the type of abnormal signal. The present invention can achieve the effect of detecting and locating abnormal information in EEG time series data while ensuring a certain computing efficiency, and realize accurate positioning and correct classification of various types of information through a multi-level detection algorithm for signals of different scales. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0047] Figure 1 A schematic diagram of a process for detecting abnormal EEG signals based on a convolutional neural network provided in an embodiment of this specification is shown;

[0048] Figure 2 A schematic diagram of a target convolutional neural network prediction model framework provided by an embodiment of this specification is shown;

[0049] Figure 3A schematic diagram of an abnormal EEG signal detection device based on a convolutional neural network provided in an embodiment of this specification is shown. DETAILED DESCRIPTION

[0050] The solution provided in this specification is described below in conjunction with the accompanying drawings.

[0051] In order to make the purpose, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described below in conjunction with the accompanying drawings.

[0052] In the description of the embodiments of the present application, words such as "exemplary", "for example" or "for example" are used to indicate examples, illustrations or descriptions. Any embodiment or design described as "exemplary", "for example" or "for example" in the embodiments of the present application should not be interpreted as being more preferred or more advantageous than other embodiments or designs. Specifically, the use of words such as "exemplary", "for example" or "for example" is intended to present related concepts in a concrete way.

[0053] In the description of the embodiments of the present application, the term "and / or" is merely a description of the association relationship of associated objects, indicating that three relationships may exist. For example, A and / or B may represent: A exists alone, B exists alone, and A and B exist at the same time. In addition, unless otherwise specified, the term "plurality" means two or more.

[0054] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features. The terms "include", "comprises", "has" and their variations all mean "including but not limited to", unless otherwise specifically emphasized.

[0055] In order to facilitate understanding of the solution provided by the embodiment of the present application, some terms involved in the solution are first briefly introduced.

[0056] Electrode lead potential value: In EEG (Electroencephalography) measurement, the electrode lead potential value refers to the strength of the brain electrical signal recorded by electrodes placed on the scalp. These electrodes are connected to amplifiers to amplify and record weak electrical signals for analysis.

[0057] Continuous Wavelet Transform: Continuous Wavelet Transform (CWT) is a time-frequency analysis tool used to analyze the frequency components of signals at different scales. Compared with the traditional Fourier transform, wavelet transform can provide localized information of signals on the time-frequency plane, which is particularly suitable for the analysis of non-stationary signals.

[0058] Figure 1 A schematic diagram of a method for detecting abnormal EEG signals based on a convolutional neural network is shown in the embodiment of this specification. Figure 1 As shown, the method comprises the following steps:

[0059] Step S101, obtaining signal data to be detected, performing noise reduction processing on the signal data, and segmenting the signal data after the noise reduction processing.

[0060] In one embodiment, the method for performing noise reduction processing on signal data specifically includes:

[0061] After collecting signal data from multiple electrode channels, the average vector of all electrode lead potential values ​​is calculated at each time point. For each electrode channel, the difference between its potential value and the average vector at that time point is calculated and used as the reference for the adjusted electrode potential. Low-frequency drift and high-frequency noise in the signal are removed, and power line interference is eliminated.

[0062] Exemplarily, an EEG acquisition device, such as an EEG cap or an EEG amplifier, can be used to collect EEG signals from multiple electrode channels, and the EEG signals are divided into a series of discrete time points according to the sampling period. For each time point, the average vector of all electrode lead potential values ​​is calculated. For each electrode channel, the difference between its potential value and the average vector at that time point is calculated at each time point, and this is used as a reference for the adjusted electrode potential. After completing the average lead processing, high-pass filtering (e.g., 1Hz) and low-pass filtering (e.g., 70Hz) are further applied to remove low-frequency drift and high-frequency noise, and a specific band-stop filter (e.g., 50Hz) is used to eliminate power line interference.

[0063] Through the above signal data noise reduction processing method, electrical noise, electromyographic activity, power line interference and other noises in the EEG signal can be effectively removed, ensuring that the final EEG signal is purer.

[0064] In one implementation, the method for segmenting the signal data after noise reduction processing specifically includes:

[0065] The signal data after noise reduction is segmented using a certain window size to obtain data segments that can be quickly calculated by the model. When segmenting the subsequent data, sliding segmentation is performed according to the set step size to capture the continuity characteristics in the data.

[0066] For example, first, a suitable window size needs to be determined according to the application scenario and signal characteristics, and the window is used to segment the signal data after noise reduction processing, and the subsequent data is slid and segmented by a set step size to capture the continuity characteristics in the data. It is worth noting that the size of the step size determines the degree of overlap between adjacent windows, and the size of the step size must be smaller than the value of the set window size. The specific step size can be set according to the actual situation. Preferably, the step size can be set to half of the window size.

[0067] Step S102, transform the segmented signal data from one-dimensional time series information to two-dimensional frequency domain information to obtain target signal data. The target signal data includes amplitude data and phase data of the signal, the amplitude represents the strength of the signal at a certain scale and position, and the phase represents the relative time position and relative change trend of the signal.

[0068] In one implementation, the method for transforming the segmented signal data specifically includes:

[0069] The signal data is transformed using a continuous wavelet transform method to obtain a transformation result. Specifically, the result obtained by the continuous wavelet transform is a complex matrix consisting of a real part and an imaginary part.

[0070] Exemplarily, an appropriate wavelet basis function, such as the Morlet wavelet, is selected, and the range of the scaling factor (scales) is determined, which determines the width variation range of the wavelet function. The signal is subjected to a continuous wavelet transform using the selected wavelet basis function and the scaling factor. The result of the continuous wavelet transform is a complex matrix containing a real part and an imaginary part.

[0071] It is worth noting that the continuous wavelet transform can provide time-frequency localized information. By using the scaled and translated wavelet function, the signal can be analyzed simultaneously in both the time and frequency domains. This localized characteristic makes it very effective in analyzing non-stationary signals such as brain waves, and can accurately capture instantaneous changes and mutations in the signal, providing the model with more refined and accurate features.

[0072] The above conversion results are further converted into amplitude and phase. Specifically, the amplitude calculation formula is:

[0073]

[0074] Where W(a,b) represents the result of continuous wavelet transform, |W(a,b)| represents the signal amplitude, Re{W(a,b)} and Im{W(a,b)} represent the real part and imaginary part of the complex matrix, respectively.

[0075] The phase calculation formula is:

[0076]

[0077] Among them, θ(a,b) represents the phase, Represents the argument of a complex number.

[0078] The converted amplitude and phase are normalized and scaled to a certain value range.

[0079] Exemplarily, the Min-Max normalization method may be used to normalize the amplitude and phase values. Preferably, the amplitude and phase data are scaled to a value range of [0, 255] and calculated using the following formula:

[0080]

[0081] Wherein, U is the upper limit of the target range (which can be 255 in this implementation), L is the lower limit of the target range (which can be 0 in this implementation), and x norm is the normalized value, x max is the maximum value in the data, x min is the minimum value in the data.

[0082] By calculating the amplitude of the signal data, we can obtain the energy distribution of the signal at different scales and positions, thereby revealing the local characteristics and global change trends of the signal. Phase information provides the relative time position and relative change trend of the signal, which can help better understand the structure and dynamic behavior of the signal. By converting the complex results into amplitude and phase, we can more comprehensively analyze the characteristics of the signal at different scales and positions, help reveal the deep structure and internal laws of the signal, and thus provide rich information support for the feature extraction of the model.

[0083] Step S103, input the target signal data into the target convolutional neural network prediction model to obtain the model prediction result. The target convolutional neural network prediction model includes an input end, a backbone network end, a feature fusion end and an output end. Figure 2 FIG. 2 shows a schematic diagram of a target convolutional neural network prediction model framework provided by this embodiment. Figure 2 As shown:

[0084] The backbone network consists of a CBL module and multiple Resc modules. The backbone network first performs preliminary feature extraction on the amplitude data and phase data of the signal through two channels, then fuses the two extracted features, and then further extracts the fused features through three layers of Resc modules. Specifically, both channels are composed of one layer of CBL module and two layers of Resc modules.

[0085] The feature fusion end is composed of a convolution layer, an upsampling layer, and a splicing operation layer, which is used to fuse the features extracted by the backbone network end to obtain feature maps of three scales. Specifically, the first-scale feature map is obtained by convolution operation on the feature map extracted by the last layer Resc module of the backbone network end, the second-scale feature map is obtained by fusion and convolution operation on the feature map extracted by the last layer and the middle layer Resc module of the backbone network end, and the third-scale feature map is obtained by fusion and convolution operation on the feature map extracted by the second-scale feature map and the first layer Resc module of the backbone network end.

[0086] Before processing the signal data at the backbone network end, the input end first receives the processed signal data, that is, the target signal data. The target data is divided into two groups, one group is amplitude data and the other group is phase data. Figure 2 The input end sets two ports to receive amplitude data and phase data respectively, and then the two sets of data enter the backbone network end and enter the feature extraction process. Figure 2 , the amplitude data and phase data pass through the same two channels respectively, extract the preliminary features of the two sets of data, and generate the corresponding feature maps. Specifically, both channels are composed of one layer of CBL module and two layers of Resc module. After extracting the feature maps of the two sets of data, the two sets of feature maps are fused so that the amplitude data features and the phase data features are superimposed. Then, the fused features are further extracted using three layers of Resc modules connected in series. As the number of layers increases, the features are gradually extracted from low-level features to high-level features.

[0087] In one embodiment, the Resc module in the backbone network is composed of a CBL module and multiple ResU modules. Specifically, the ResU module is a convolution module composed of multiple branches, whose main trunk is composed of two CBL modules in series, and whose branch trunk is composed of a 1*1 convolution module and a ReLU activation function. When the two trunks are merged, the target signal data features will be accumulated element by element to achieve fusion. The branch trunk is used to alleviate the model degradation phenomenon caused by the model depth to ensure a good extraction effect of the target signal data features.

[0088] In one implementation, the feature fusion end fuses the features extracted by the backbone network end to obtain a feature map of three scales, specifically including:

[0089] Continue to refer Figure 2 , extract the feature map from the last layer of the Resc module at the backbone network end, and perform convolution operation on it through the multi-layer CBL module to obtain the first scale feature map. Larger abnormal signal targets can be detected based on the first scale feature map. Preferably, the CBL module can be 5 layers.

[0090] Continue to refer Figure 2 , the first scale feature map is extracted through the CBL module, and then the upsampling layer is used to perform a deconvolution operation on it. The first scale feature map after the deconvolution operation is fused with the feature map extracted from the middle layer Resc module of the backbone network through the splicing operation layer, and a second scale feature map is generated after further convolution operation. The second scale feature map can be used to detect medium-sized abnormal signal targets.

[0091] Continue to refer Figure 2 , the second-scale feature map is extracted through the CBL module, and then the upsampling layer is used to perform a deconvolution operation on it. The second-scale feature map after the deconvolution operation is fused with the feature map extracted from the first-layer Resc module at the backbone network end through the splicing operation layer, and a third-scale feature map is generated after further convolution operation. The third-scale feature map can be used to detect small abnormal signal targets.

[0092] As the signal type changes, the scale of abnormal signals also varies greatly (for example, the length of sharp_wave signals is usually less than 40 sampling points, while the length of slow signals is usually more than 100 sampling points), so multi-scale detection is very important for this solution. By extracting feature maps from different layers of the backbone network, using these feature maps for multi-scale signal detection, and through this multi-scale feature fusion, better detection of signals of different sizes, this module processes and enhances the feature maps through a series of convolutional layers, deconvolutional layers and fusion operations, thereby improving the expressiveness of features and the accuracy and robustness of target detection.

[0093] In one embodiment, the output end of the target convolutional neural network prediction model is responsible for outputting the prediction results of the model. The output information includes: bounding box coordinates, confidence, and category probability. Specifically, the bounding box coordinates include the center coordinates and the offset of the edge from the center. The confidence indicates the confidence probability score of whether there is an abnormal signal in the bounding box, and the category probability indicates the probability that the abnormal signal in the bounding box belongs to each category.

[0094] In another embodiment, returning to Figure 2 , three output ports are set at the output end to output the information of the corresponding three scale feature maps.

[0095] Since the feature fusion end generates feature maps of three scales, an independent feature output port is set for each scale feature map. The output layer of each scale feature map predicts 3 anchor boxes on each grid unit. Therefore, the output feature map of each grid unit has (Bx(3+C)) channels, where B is the number of bounding boxes predicted by each grid unit, 3 is the 2 coordinate values ​​and 1 confidence value of each bounding box, and C is the number of categories.

[0096] Step S104, determining the target range in the detected target signal data based on the model prediction result, and judging the type of the abnormal signal.

[0097] In one embodiment, a method for determining a target range in detected target signal data and a type of abnormal signal according to a prediction result of a model specifically includes:

[0098] First, adjust the coordinates of the anchor box, and limit the coordinates to (0,1) through the sigmoid activation function. Then, calculate the probability (i.e., confidence) of the existence of an abnormal target in the bounding box through the sigmoid function, and calculate the probability that the abnormal target belongs to each abnormal category through the softmax function. According to the calculation results of the above abnormal category probabilities, identify the type of abnormal signal, for example, it can be eyes_blink, seiz, sharp_slow, sharp_wave, spike_wave, slow, spike_slow, spindle and other abnormal waveform types. Finally, the generated detection box is non-maximum suppressed by the NMS (Non-Maximum Suppression) algorithm, and the overlapping abnormal detection boxes are removed. The detection box information with the highest confidence is retained. The retained box represents the abnormal signal position predicted by the model, that is, the target range in the detected target signal data.

[0099] Corresponding to the above method provided by the present invention, the present invention also provides a device. Figure 3 FIG. 2 is a schematic diagram showing a structure of an abnormal EEG signal detection device based on a convolutional neural network provided in an embodiment of the present specification. Figure 3 As shown, the device comprises:

[0100] The data acquisition module S301 is used to acquire abnormal EEG signal data;

[0101] The data processing module S302 is used to perform noise reduction processing on the signal data, segment the signal data after the noise reduction processing, and transform the segmented signal data;

[0102] The model prediction module S303 is used to make predictions based on the processed signal data to obtain prediction results;

[0103] The prediction result determination module S304 is used to determine the target range in the detected target signal data according to the prediction result, and to determine the type of the abnormal signal.

[0104] It is worth mentioning that Figure 3 The device shown is Figure 1 The method embodiments shown correspond to and can be applied to Figure 1 The acquisition of abnormal EEG signal data, the processing of signal data, the input of the processed data into the prediction model for prediction, the determination of the target range in the detected target signal data according to the model prediction results, and the determination of the type of abnormal signal in the method embodiment shown will not be repeated here. Those skilled in the art should easily realize that, in combination with the units and algorithm steps of each example described in the embodiments disclosed herein, the present application can be implemented in the form of hardware or a combination of hardware and computer software. Whether a function is executed in the form of hardware or computer software driving hardware depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0105] According to another embodiment of the present invention, a computing device is provided, including a memory and a processor, wherein the memory stores executable code, and when the processor executes the executable code, the Figure 1 The method described.

[0106] According to another embodiment, there is also provided a computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to execute a combination of Figure 1 The method described.

[0107] Those skilled in the art should be aware that in one or more of the above examples, the functions described in the present invention can be implemented by hardware, software, firmware or any combination thereof. When implemented by software, these functions can be stored in a computer-readable medium or transmitted as one or more instructions or codes on a computer-readable medium.

[0108] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made on the basis of the technical solution of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for detecting abnormal EEG signals based on convolutional neural network, characterized in that: The method comprises: Acquire signal data to be detected, perform noise reduction processing on the signal data, and segment the signal data after the noise reduction processing; The segmented signal data is transformed from one-dimensional time series information to two-dimensional frequency domain information to obtain target signal data; wherein the target signal data includes amplitude data and phase data of the signal, the amplitude represents the strength of the signal at a certain scale and position, and the phase represents the relative time position and relative change trend of the signal; Inputting the target signal data into a target convolutional neural network prediction model to obtain a model prediction result; wherein the target convolutional neural network prediction model includes an input end, a backbone network end, a feature fusion end, and an output end; The backbone network end is composed of a CBL module and multiple Resc modules. The backbone network end performs preliminary feature extraction on the amplitude data and phase data of the signal through two channels, fuses the two extracted features, and then further extracts the fused features through three layers of Resc modules, wherein the two channels are composed of one layer of CBL module and two layers of Resc modules; The feature fusion end is composed of a convolution layer, an upsampling layer and a splicing operation layer, which is used to fuse the features extracted by the backbone network end to obtain feature maps of three scales; wherein, the first scale feature map is formed by the feature map extracted by the last layer Resc module of the backbone network end through a convolution operation, the second scale feature map is formed by the feature map extracted by the last layer and the middle layer Resc modules in the backbone network end through fusion and convolution operations, and the third scale feature map is formed by the second scale feature map and the feature map extracted by the first layer Resc module in the backbone network end through fusion and convolution operations; based on the model prediction results, the target range in the detected target signal data is determined, and the type of abnormal signal is judged.

2. The method according to claim 1, characterized in that: The method for performing noise reduction processing on signal data specifically includes: After collecting signal data from multiple electrode channels, the average vector of all electrode lead potential values ​​is calculated at each time point; For each electrode channel, the difference between its potential value and the average vector at that time point is calculated, and this is used as the reference adjusted electrode potential; Remove low-frequency drift and high-frequency noise from your signals, and eliminate power line interference.

3. The method according to claim 1, characterized in that The method for segmenting the signal data after noise reduction processing specifically includes: The signal data after noise reduction is segmented using a certain window size to obtain data segments that can be quickly calculated by the model; When the subsequent data is segmented, sliding segmentation is performed according to the set step size to capture the continuity characteristics in the data.

4. The method according to claim 1, characterized in that The method for transforming the segmented signal data specifically includes: The signal data is transformed using a continuous wavelet transform method to obtain a transformation result; wherein the result obtained by the continuous wavelet transform is a complex matrix composed of a real part and an imaginary part; The conversion result is further converted into amplitude and phase; specifically, the amplitude calculation formula is: Where W(a,b) represents the result of continuous wavelet transform, |W(a,b)| represents the signal amplitude, Re{W(a,b)} and Im{W(a,b)} represent the real part and imaginary part of the complex matrix respectively; The phase calculation formula is: Among them, θ(a,b) represents the phase, Represents the argument of a complex number; The converted amplitude and phase are normalized and scaled to a certain value range.

5. The method according to claim 1, characterized in that The Resc module in the backbone network end is composed of a CBL module and multiple ResU modules; wherein the ResU module is a convolution module composed of multiple branches, whose main trunk is composed of two CBL modules in series, and whose branch trunk is composed of a 1*1 convolution module and a ReLU activation function. When the two trunks are merged, the target signal data features will be accumulated element by element to achieve fusion. The branch trunk is used to alleviate the model degradation phenomenon caused by the model depth to ensure a good extraction effect of the target signal data features.

6. The method according to claim 1, characterized in that The feature fusion end fuses the features extracted by the backbone network end to obtain a method for feature maps of three scales, specifically comprising: Extract the feature map from the last layer of the Resc module at the backbone network end, and perform convolution operation on it through the multi-layer CBL module to obtain the first scale feature map; wherein, the first scale feature map is used to detect larger abnormal signal targets; The first scale feature map is extracted through the CBL module, and then deconvolved by the upsampling layer. The first scale feature map after deconvolution is fused with the feature map extracted from the middle layer Resc module at the backbone network end through the splicing operation layer. After further convolution operation, the second scale feature map is generated. The second scale feature map is used to detect medium-sized abnormal signal targets. The second-scale feature map is extracted through the CBL module, and then deconvolved using the upsampling layer. The second-scale feature map after the deconvolution operation is fused with the feature map extracted from the first-layer Resc module at the backbone network end through the splicing operation layer. After further convolution operation, the third-scale feature map is generated; the third-scale feature map is used to detect small abnormal signal targets.

7. The method according to claim 1, characterized in that The output end is responsible for outputting the prediction results of the model; wherein the output information includes: bounding box coordinates, confidence, and category probability. Specifically, the bounding box coordinates include the center coordinates and the offset of the edge from the center. The confidence indicates the confidence probability score of whether there is an abnormal signal in the bounding box, and the category probability indicates the probability that the abnormal signal in the bounding box belongs to each abnormal category.

8. The method according to claim 1, characterized in that The method of determining the target range in the detected target signal data and the type of the abnormal signal according to the prediction result of the model specifically includes: Adjust the anchor box coordinates and limit them between 0 and 1 through the sigmoid activation function; The probability of an abnormal signal existing in the bounding box is calculated by the sigmoid function, and the probability of the abnormal signal belonging to each abnormal category is calculated by the softmax function; The generated detection frames are subjected to non-maximum suppression through the NMS algorithm, the overlapping abnormal detection frames are removed, and the detection frames with the highest confidence are retained; the retained frames with the highest confidence represent the target range in the detected target signal data predicted by the model.

9. A device for detecting abnormal EEG signals based on convolutional neural network, characterized in that: Abnormal EEG signal detection is performed by the method according to any one of claims 1 to 8, the device comprising: A data acquisition module, used to acquire abnormal EEG signal data; A data processing module, used for performing noise reduction processing on the signal data, segmenting the signal data after the noise reduction processing, and transforming the segmented signal data; A model prediction module is used to make predictions based on the processed signal data to obtain prediction results; The prediction result determination module is used to determine the target range in the detected target signal data according to the prediction result and judge the type of abnormal signal.

10. An electronic device comprising a memory and a processor, wherein the memory stores a computer program that can be run on the processor, wherein: When the processor executes the computer program, the method according to any one of claims 1 to 8 is implemented.

Citation Information

Cited By

  • Network information security intrusion detection method based on artificial intelligence

    CN120692061A

  • Abnormal signal detection and signal transmission protocol identification method and device

    CN121665244A

  • Sleep spindle wave detection method and device and medium

    CN122220691A