Application of signal detection methods in physiological signal classification

Abstract

This invention discloses an application of a signal detection method in physiological signal classification, which uses a convolutional network weighted model to filter feature information from different frequency bands. The convolutional network weighted model includes: processing the input signal using a multi-scale convolutional network to obtain multiple feature map attention signals; and cascading the multiple feature map attention signals. This invention, by using a multi-scale convolutional network, can capture features across a wider frequency range, which is beneficial for improving detection accuracy.

Description

[0001] This application is a divisional application of Chinese patent application filed on April 6, 2023, with application number 2023103544389, entitled "Signal Detection Method, Detection Model, Detection Equipment and Application". Technical Field

[0002] This invention relates to the field of signal detection technology, and in particular to a signal detection method, detection model, detection equipment, and application. Background Technology

[0003] Signal detection technology is widely used in various fields. By detecting signals in real time, the changing trends of signal characteristics can be discovered, thereby obtaining the desired information. In existing technologies, signal detection mainly involves averaging data collected from different channels or manually setting a weight. This approach ignores the weight changes when the subject is in different states and the weight changes of different channels at different frequencies, resulting in low accuracy of detection results and poor generalization ability and robustness of the detection algorithm. Summary of the Invention

[0004] The technical problem to be solved by this invention is: how to improve the accuracy and robustness of signal detection methods.

[0005] This invention provides a signal detection method, detection model, detection equipment, and application, which can improve the accuracy and robustness of detection results.

[0006] The technical solution adopted by this invention to solve its technical problem is: a signal detection method, comprising:

[0007] S1: Preprocess the original signal to obtain the signal to be processed;

[0008] S2: Obtain dynamically weighted signals using a weighting mechanism;

[0009] S3: Output a discrete array to obtain the classification results;

[0010] The weighting mechanism includes at least one of a channel weighting model and a convolutional network weighting model; wherein, the weighting mechanism is suitable for adjusting the weights of channels in the signal to be processed; and the convolutional network weighting model is suitable for filtering signals of interest frequency bands through multi-scale convolution processing.

[0011] When both the channel-weighted model and the convolutional network weighted model exist, the output of the channel-weighted model is used as the input signal of the convolutional network weighted model to adjust the weights of the corresponding frequency bands of each channel in the input signal.

[0012] Furthermore, the channel weighting model includes:

[0013] The signal to be processed is sequentially subjected to time pooling, channel convolution, and activation processing to obtain a channel-weighted signal.

[0014] Furthermore, the weighted model of the convolutional network includes:

[0015] A multi-scale convolutional network is used to process the input signal to obtain multiple feature map attention signals;

[0016] The attention signals of the multiple feature maps are cascaded.

[0017] Furthermore, the multi-scale convolutional network includes at least two convolutional branches, each of which contains a temporal convolutional layer and a channel convolutional layer;

[0018] The kernel sizes of the multiple temporal convolutional layers are different;

[0019] The convolution kernel of the temporal convolutional layer is equal to [sampling rate / minimum value of the frequency band of interest] + 1, where [] indicates rounding.

[0020] The convolution kernel of the temporal convolutional layer is used to obtain features of the frequency band of interest, and the kernel parameters are constrained by limiting the weights in order to filter out the frequency band of interest.

[0021] Furthermore, between steps S2 and S3, step S2-1 is also included: fine-tuning the dynamic weighted signal;

[0022] The fine-tuning process includes:

[0023] Acquire the raw signal of the target object and the raw signals of other objects;

[0024] K1-fold training and validation were performed using the original signals from other objects to obtain a pre-trained network;

[0025] A portion of the original signal of the target object is taken as fine-tuning training data, and the pre-trained network is trained and verified using weight-limited loss to obtain the fine-tuned network.

[0026] The fine-tuning network is tested using the remaining portion of the original signal of the target object as fine-tuning test data.

[0027] Furthermore, the preprocessing of the original signal includes at least one of: frequency domain filtering, exponential moving normalization, and multiple data enhancement combined; wherein,

[0028] The multi-data enhancement process includes: performing random flipping, Gaussian noise processing, time-domain swapping, and time-domain combination processing on the original signal to expand the data volume of the signal.

[0029] Furthermore, the output discrete array comprises: obtaining a discrete array by sequentially performing activation processing, batch normalization processing, and linear processing on the dynamically weighted signal; wherein,

[0030] The activation process includes: squared activation, average pooling, and LOG activation;

[0031] The signal to be processed was not subjected to sliding window processing.

[0032] The present invention also provides a detection model for the signal detection method described above, comprising:

[0033] The preprocessing module is used to preprocess the raw signal;

[0034] The weighting mechanism module is used to run the weighting mechanism to obtain dynamic weighting signals;

[0035] The output module is used to output discrete arrays.

[0036] The present invention also provides a detection device, comprising:

[0037] The lower-level machine is used to acquire raw signals, store signal data, and transmit data to the upper-level machine for processing.

[0038] The host computer is used to execute the signal detection method described above;

[0039] A display, used to show the detection results output by the host computer;

[0040] The execution component is used to execute corresponding instructions based on the detection results.

[0041] The present invention also provides an application of the signal detection method described above in physiological signal classification, which uses a convolutional network weighted model to filter out feature information of different frequency bands.

[0042] The beneficial effects of this invention are that it utilizes a weighting mechanism to process the signal to be processed, obtaining a dynamically weighted signal. This dynamically weighted signal can better reflect changes in physiological state, thereby improving the accuracy of subsequent detection results. The weighting mechanism can assign dynamic weights to different channels and capture a wider frequency band, thus possessing high application value. Attached Figure Description

[0043] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0044] Figure 1 This is a flowchart of the signal detection method of the present invention.

[0045] Figure 2 This is a flowchart illustrating the specific implementation of the signal detection method of the present invention.

[0046] Figure 3 This is a schematic diagram of the channel weighting model of the present invention.

[0047] Figure 4 This is a schematic diagram of the weighted model of the convolutional network in this invention.

[0048] Figure 5 This is a schematic diagram of the fine-tuning process of the present invention.

[0049] Figure 6 This is a schematic diagram of the multi-data enhancement combined in this invention.

[0050] Figure 7 This is a flowchart of the multi-data enhancement combined in this invention.

[0051] Figure 8 This is the confusion matrix of Embodiment 1 of the present invention.

[0052] Figure 9 This is the confusion matrix of Comparative Example 1.

[0053] Figure 10 This is a comparison chart of the detection accuracy of Example 1 and Comparative Example 1, showing a pair of different subjects.

[0054] Figure 11 The confusion matrix of Embodiment 2 of the present invention.

[0055] Figure 12 This is the confusion matrix of Comparative Example 2.

[0056] Figure 13 This is a comparison chart of the detection accuracy of Example 2 and Comparative Example 2 for different subjects. Detailed Implementation

[0057] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0058] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," "counterclockwise," "axial," "radial," and "circumferential," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, features defined with "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, unless otherwise stated, "a plurality of" means two or more.

[0059] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0060] like Figures 1 to 2 As shown, the signal detection method of the present invention includes: S1: preprocessing the original signal to obtain a signal to be processed; S2: obtaining a dynamically weighted signal using a weighting mechanism; S3: outputting a discrete array to obtain a classification result. It should be noted that the present invention uses a weighting mechanism to process the signal to be processed to obtain a dynamically weighted signal. The dynamically weighted signal can better reflect changes in physiological state, thereby improving the accuracy of subsequent classification results. After processing the dynamically weighted signal, a discrete array can be obtained, applicable to different fields and products. The number of discrete arrays can be used to represent the number of classification results.

[0061] The weighting mechanism of this invention includes at least one of a channel-weighted model and a convolutional network weighting model. The weighting mechanism is suitable for adjusting the weights of channels in the signal to be processed; the convolutional network weighting model is suitable for filtering signals in frequency bands of interest through multi-scale convolution processing. When both the channel-weighted model and the convolutional network weighting model exist simultaneously, the output of the channel-weighted model serves as the input signal for the convolutional network weighting model to adjust the weights of the corresponding frequency bands of each channel in the input signal. That is, in specific applications, this invention can choose either the channel-weighted model or the convolutional network weighting model, or a combination of both. It should be noted that the original signal in this invention is, for example, a multi-channel EEG signal or other physiological signal, with each channel acquiring signals through electrodes placed at different target locations. Therefore, when the subject's physical or mental state changes, the characteristics of the EEG signals acquired from different channels will differ. To address this, this invention utilizes the channel-weighted model to assign different weights to each channel and uses the convolutional network weighting model to filter signals in frequency bands of interest, such as 8Hz to 13Hz. When a channel-weighted model and a convolutional network-weighted model are used together, the output of the channel-weighted model is input into the convolutional network-weighted model to adjust the weights of the corresponding frequency bands for each channel, thereby improving the accuracy and specificity of the detection results. The corresponding frequency bands for each channel are obtained through the convolutional network-weighted model.

[0062] like Figure 3 As shown, the channel-weighted model includes: sequentially performing temporal pooling, channel convolution, and activation processing on the signal to be processed to obtain a channel-weighted signal. It should be noted that the signal to be processed includes a time parameter, the number of channels, and the feature map dimension. First, the time parameter is pooled, and global pooling is performed on the time dimension to obtain the global receptive field. After pooling, the feature dimension changes from "feature map × channel × time" to "feature map × channel × 1". Then, channel convolution is performed on the pooling result to assign weights to different channels. Finally, the result is mapped using activation functions (sequentially performing squared activation, overlap pooling, and logarithmic activation). Thus, a channel-weighted signal is obtained, meaning each channel has a corresponding weight. For example, if there are n channels, then there will be n weights w1 to w2. n .

[0063] like Figure 4As shown, the weighted convolutional network model includes: processing the input signal using a multi-scale convolutional network to obtain multiple feature map attention signals; concatenating these multiple feature map attention signals, i.e., after feature map attention, the two branches output a matrix of feature map × channel × time respectively; the concatenation operation merges the matrices in the time dimension, resulting in a matrix of feature map × channel × (2 × time). The multi-scale convolutional network includes at least two convolutional branches, each containing a temporal convolutional layer and a channel convolutional layer. The kernel sizes of the multiple temporal convolutional layers are different. The kernels of the temporal convolutional layers are used to acquire features within the range of interest, and the kernel parameters are constrained by limiting the weights to select the interest band.

[0064] In other words, the weighted convolutional network model processes the same input signal using multiple convolutional branches, and then concatenates the results from different convolutional branches to obtain the final output. The difference between each convolutional branch lies in the size of the temporal convolutional kernel. This allows for the extraction of local features at different time scales. Based on the duality of time and frequency, different convolutional branches can extract local features at different frequencies, with the extracted frequency range being [sampling rate / kernel size, +∞). Then, the outputs from different convolutional branches are concatenated and fused to form a deep network with multi-band selectivity.

[0065] For example, the kernel size of a temporal convolutional layer is calculated as [sampling rate / minimum value of the band of interest] + 1, where [] represents the rounding down sign. The kernel size of a channel convolutional layer is equal to the number of channels. Both the temporal and channel kernel sizes are constrained by the square norm. For example, the constraint formula is:

[0066] For example, taking two convolutional branches, the temporal convolution kernel of the first convolutional branch = [sampling rate / x] + 1, corresponding to the frequency band range greater than x Hz in the original signal; the temporal convolution kernel of the second convolutional branch = [sampling rate / y] + 1, corresponding to the frequency band range greater than y Hz in the original signal, where x ≠ y. The channel convolution kernels of the first and second convolutional branches are the same size, and the output of the channel convolution is the input of the channel convolution. Therefore, the outputs of the two convolutional branches are not the same, allowing the extraction of feature information from different frequency bands. It should be noted that although multi-scale convolutional networks can extract information from different frequency bands, the temporal input of multiple convolutional branches is the same. Setting too many convolutional branches can lead to temporal redundancy, increasing the number of parameters in subsequent processing and causing overfitting in the weighting mechanism, thus increasing the complexity of the algorithm. Therefore, to balance the multi-frequency band extraction capability and algorithm complexity of multi-scale convolutional networks, this invention preferably uses two convolutional branches. Compared with conventional methods of filtering frequency bands, conventional methods have drawbacks such as filter transition limitations and lack of channel weighting. However, this invention uses a multi-scale convolutional network, which avoids transition bands and can obtain feature weighting for different channels. By using overlapping convolution to extract features in the time direction, shallow features can be obtained, thus obtaining more features. This is a significant difference from conventional physical filtering for selecting frequency bands.

[0067] It should be noted that this method includes step S2-1 between steps S2 and S3: fine-tuning the dynamically weighted signal. The fine-tuning process includes: acquiring the original signal of the target object and the original signals of other objects; performing K1-fold training and verification using the original signals of other objects to obtain a pre-trained network; taking a portion of the original signal of the target object as fine-tuning training data, and performing K2-fold training and verification on the pre-trained network with weight-constrained loss as a constraint to obtain a fine-tuned network; and testing the fine-tuned network using the remaining portion of the original signal of the target object as fine-tuning test data.

[0068] It should be noted that steps S1-S3 of this invention can be understood as a detection model. This model is trained with a large amount of data before being applied to improve its detection accuracy. In practical applications, the detection model can be used in single-subject and cross-subject scenarios. Single-subject refers to training data consisting of the EEG data of a single subject (subject A), and the trained detection model is applied to that subject (subject A). Cross-subject refers to training data consisting of the EEG data of other subjects (subject B), and the trained detection model is applied to subject A. Therefore, cross-subject scenarios place higher demands on the generalization performance of the detection model. To improve the generalization performance of the detection model across subjects, this invention performs fine-tuning of the dynamically weighted signal. For example… Figure 5As shown, multiple subjects B can be tested, denoted as Subject B1 to Subject B2. The EEG data of all subjects B are shuffled and divided into training data (e.g., percentage a%) and validation data (e.g., percentage (100-a)%) according to a certain ratio. The pre-trained network ω is obtained using a K1-fold training and validation method. 0 That is, the training data is divided into K1 parts, where K1 is, for example, 5 to 10. Each part serves as a test set, and the remaining parts serve as the training set (for example, when K1 = 5, the training data is divided into 5 equal parts, with 1 part serving as the test set and the remaining 4 parts as the training set). After training, validation data is used for verification. Obtain the pre-trained network ω. 0 Then, a portion of the EEG data from subject A is selected as fine-tuning training data (e.g., b%), and the remainder is selected as fine-tuning test data (e.g., (100-b)%). K2-fold training is performed with weight-constrained loss as a constraint to obtain the fine-tuned network ω. The fine-tuning training data is divided into K2 parts, where K2 is, for example, 5-10. Each part serves as a test set, and the remaining parts serve as the training set (e.g., when K2 = 5, the fine-tuning training data is divided into 5 equal parts, with 1 part as the test set and the remaining 4 parts as the training set). The formula for the weight-constrained loss is: Where α is a constant, and f(ω) represents the additional loss function in the fine-tuning process.

[0069] Therefore, the dynamically weighted signal obtained after fine-tuning has better generalization performance and can be used in application scenarios across subjects. It should be noted that this invention uses a multi-scale convolutional network to filter feature information in different frequency bands, which can be applied to scenarios with EEG activity in specific frequency bands, such as motor imagery (8Hz~13Hz) and emotion recognition (4Hz~28Hz). This method allows the model to select frequency bands according to the application scenario, thereby improving the model's recognition accuracy and robustness.

[0070] The preprocessing of the original signal in this invention includes at least one of the following: frequency domain filtering, exponential sliding normalization, and multiple data augmentation. The multiple data augmentation includes random flipping, Gaussian noise reduction, time-domain swapping, and time-domain combination processing to expand the signal's data volume. It should be noted that frequency domain filtering includes high-pass filtering (removing DC components from the original signal), low-pass filtering (removing electromyographic signals from the original signal and reducing aliasing frequencies), and notch filtering (removing power frequency interference such as 50Hz / 60Hz). Exponential sliding normalization is used to remove high-frequency noise from the original signal. The data volume is expanded to M times the original amount through multiple augmentation, increasing the amount of data used for training, avoiding model overfitting, and improving the robustness of the detection model. For example... Figures 6 to 7As shown, the original signal is divided into multiple segments (Trials), and each segment (Trials) is further divided into multiple slices (Crops). A segment is selected with a probability of p%, and its slice is flipped along the time domain to obtain augmented data one. Another segment is selected with a probability of p%, and Gaussian noise with a mean of μ and a standard deviation of σ is added to it to obtain augmented data two. Two segments Ti and Tj with the same label are randomly selected, and their corresponding slices Ck are swapped to obtain augmented data three. Augmented data three is then superimposed onto the original signal, and this data augmentation process is repeated until the data volume is increased to M times its original size. After the data volume is increased to M times, 1 / q of the data from segment Ti is randomly selected to replace the data in segment Tj, and time-domain combination is performed. Thus, through preprocessing, noise and interference in the original signal can be removed, and the data volume can be increased to improve the accuracy of the detection model training and enhance the model's robustness.

[0071] In this invention, the output discrete array is obtained by sequentially performing activation processing, batch normalization processing, and linear processing on the dynamically weighted signal. The activation processing includes squared activation, average pooling, and LOG activation. It should be noted that the signal to be processed in this invention does not undergo sliding window processing. This is because sliding window processing conflicts with average pooling (e.g., squared average pooling). This invention uses average pooling, which can significantly reduce the number of parameters and extract signal features. If sliding window processing is applied to the signal, it will cause secondary overlap in average pooling, introducing excessive redundant information and leading to overfitting of the detection model. In this invention, squared average pooling and LOG activation need to be used together. The squaring operation will cause the activation values ​​to tend to be larger and smaller. Overlapping average pooling will further reduce the dimensionality of the expanded data and enhance the influence of the expansion. For example, if the activation values ​​are 0.1 and 100, before squaring, average pooling yields (0.1+100) / 2 = 50.05. After squared average pooling, it yields (0.12+1002) / 2 = 5000.005 > 50.052. Then, the LOG activation function is used for secondary scaling to suppress excessively large activation values ​​and amplify excessively small activation values ​​(log5000.005 = 3.7, log50.05 = 1.7), making it easier to fit and activate the data during subsequent linear processing.

[0072] For example, the output discrete array is [p1, p2, p3, ..., pn], where p1 to pn represent probability values, each corresponding to a different classification result. It should be noted that the detection method of this invention can be applied to scenarios such as motor imagery, emotion recognition, epilepsy detection, sleep staging, and fatigue detection. For example, taking motor imagery as an example, motor imagery (MI) refers to the activation of specific brain regions when a person imagines their limbs (or muscles) moving without actual movement output. By analyzing EEG signals and detecting and identifying the activation effects of different brain regions, the user's intention can be determined, thereby achieving direct communication and control between the human brain and external devices. For example, common motor imagery locations are: left hand, right hand, foot, and tongue. Therefore, the output discrete array can be set to [p1, p2, p3, p4], corresponding to the left hand, right hand, foot, and tongue, respectively. The values ​​of p1, p2, p3, and p4 are compared, and the location with the highest probability value is taken as the final detection result. For example, if p2 has the highest value, then the final output detection result is: the motor imagery location is "right hand".

[0073] The technical effects of the present invention will be illustrated below through specific examples.

[0074] Example 1: With 22 channels and a single-subject scenario, the signal detection method of this invention was used to process the raw signals of 9 subjects, outputting a four-class discrete array. The multi-scale convolutional network was configured with two convolutional branches. The temporal convolutional kernel size of the first convolutional branch was set to 1*30, i.e., 30 = [250 / x] + 1, where 250 represents the sampling rate, and the minimum value of the corresponding frequency band is 8.5Hz. That is, the frequency band corresponding to the first convolutional branch is a frequency band greater than 8.5Hz. The temporal convolutional kernel size of the second convolutional branch was set to 1*50, i.e., 50 = [250 / y] + 1, where 250 represents the sampling rate, and the minimum value of the corresponding frequency band is 5.1Hz. That is, the frequency band corresponding to the second convolutional branch is a frequency band greater than 5.1Hz.

[0075] Comparative Example 1: The difference between Comparative Example 1 and Example 1 is that the M-Shallow ConvNet method (shallow convolutional network) is used to process and output a four-class discrete array. The processing of the M-Shallow ConvNet method is to take the original motor imagery EEG signal as input and obtain the probability of the input original motor imagery signal belonging to each category through operations such as temporal convolution (convolution kernel size is 1*30), spatial convolution, and average pooling.

[0076] Table 1 shows the model parameters for Example 1, and Table 2 shows the model parameters for Comparative Example 1. Comparing Table 1 and Table 2, the number of parameters involved in this method is greater than that in Comparative Example 1. This method can improve the accuracy of the final classification results by increasing the number of parameters. Figure 8This is the confusion matrix of Example 1. Figure 9 This is the confusion matrix for Comparison Example 1. Figure 8 and Figure 9 As can be seen, the classification accuracy of this method is generally higher than that of Comparative Example 1. Please continue to refer to [the relevant documentation / reference]. Figure 10 , Figure 10 The table shows a comparison of the detection accuracy of this method and the method in Comparative Example 1 for 9 subjects. For different subjects, the overall detection accuracy of this method is higher than that of the method in Comparative Example 1. Referring to Table 3, the average accuracy of this method is 0.83, and the average kappa coefficient is 0.773. The average accuracy of Comparative Example 1 is 0.816, and the average kappa coefficient is 0.755. The kappa coefficient is used for consistency testing, with values ​​ranging from -1 to 1, and measures the consistency between the classification result and the actual result. A higher kappa coefficient indicates better consistency. Compared with Shallow ConvNet, the method of this invention has a 1.4% higher recognition accuracy and a 0.018 higher kappa coefficient, indicating a significant improvement in the classification effect of motor imagery EEG signals (single subject).

[0077] Example 2: The difference between Example 2 and Example 1 is that Example 2 is applied to a cross-subject scenario.

[0078] Comparative Example 2: The difference between Comparative Example 2 and Example 2 is that the PSTS-CNN (Motion Imagery Classification Network Based on Parallel Spatiotemporal Self-Attention) method is used.

[0079] Figure 11 This is the confusion matrix of Example 2. Figure 12 This is the confusion matrix for example two, in comparison. Figure 11 and Figure 12 As can be seen, the classification accuracy of this method is generally higher than that of Comparative Example 1. Please continue to refer to [the relevant documentation / reference]. Figure 13 , Figure 13 The table shows a comparison of the detection accuracy of this method and the method in Comparative Example 1 across 9 subjects. For different subjects, the overall detection accuracy of this method is higher than that of the method in Comparative Example 1. Referring to Table 4, the average accuracy of this method is 0.756, and the average kappa coefficient is 0.675, while the average accuracy of Comparative Example 1 is 0.74, and the average kappa coefficient is 0.655. Compared with PSTS-CNN, the method of this invention has a 1.6% higher recognition accuracy and a 0.02 higher kappa coefficient, indicating that this method also significantly improves the classification performance of motor imagery EEG signals (across subjects).

[0080] Table 1

[0081] name Number of channels nuclear size Input size Output size Number of parameters (k) Channel attention 1 3 1*22*751 1*22*751 1.5 Temporal Convolution 1 24 1*30 1*22*751 24*22*751 0.7 Temporal Convolution 2 24 1*50 1*22*751 24*22*751 1.2 Channel convolution 1 24 22*1 24*22*751 24*1*751 12.7 Channel convolution 2 24 22*1 24*22*751 24*1*751 12.7 Feature Map Attention 1 24 3 24*1*751 24*1*751 0.003 Feature Map Attention 2 24 3 24*1*751 24*1*751 0.003 Average pooling 24 1*75 24*1*1502 24*1*65 0 linear layer - - 1560 4 6.2

[0082] Table 2

[0083] name Number of channels nuclear size Input size Output size Number of parameters (k) Temporal convolution 24 1*30 1*22*751 24*22*751 0.7 Channel convolution 24 22*1 24*22*751 24*1*751 12.7 Average pooling 24 1*75 24*1*751 24*1*31 0 linear layer - - 751 4 3.0

[0084] Table 3

[0085]

[0086] Table 4

[0087] Methods used Average accuracy Average kappa coefficient PSTS-CNN 0.74 0.655 This method 0.756 0.675

[0088] This invention also provides a detection model for a signal detection method, comprising: a preprocessing module for preprocessing the original signal; a weighting mechanism module for acquiring a dynamically weighted signal; and an output module for outputting a discrete array. The preprocessing module performs frequency domain filtering, exponential moving average normalization, and multi-data enhancement processing on the original data. The output module performs activation processing, batch normalization, and linear processing on the dynamically weighted signal before outputting a discrete array. The preprocessing module, weighting mechanism module, and output module can all be implemented using a computer (e.g., a PC).

[0089] The present invention also provides a detection device, comprising: a lower-level computer for acquiring raw signals; a higher-level computer for executing a signal detection method; a display for displaying the detection results output by the higher-level computer; and an execution unit for executing corresponding instructions based on the detection results.

[0090] This invention also provides an application of a signal detection method in physiological signal classification. Specifically, it uses a weighted convolutional network model to filter feature information from different frequency bands. This method can be applied to scenarios involving electroencephalogram (EEG) activity with specific frequency bands, particularly in the computation of multiple application scenarios. It eliminates the need to reconstruct the detection model or repeat preprocessing and channel weighting steps; only the parameters of the multi-scale convolutional network need to be adjusted. Physiological signals include, but are not limited to, EEG signals, ECG signals, electrooculogram (EOG) signals, and electromyogram (EMG) signals. For example, the application scenarios and frequency bands of EEG signals are generally based on clinical experience, including but not limited to: emotion recognition (frequency band of interest: 4Hz–28Hz), epilepsy detection (frequency band of interest is related to the type of epilepsy and can cover the entire frequency band, or a specific frequency band of interest can be set as needed), motor imagery recognition (frequency band of interest: 8Hz–13Hz), sleep stage recognition (frequency band of interest: 0.5–12Hz), and fatigue detection (frequency band of interest: 4Hz–30Hz). These can be divided into two scenarios: 1) Physical fatigue leads to drowsiness, resulting in a decrease in alpha band energy and an increase in theta band energy; 2) Brain tension and fatigue lead to a decrease in alpha band energy and an increase in beta band energy. For N-class classification, the category with the largest value from the 1×N discrete array output by the signal detection method is selected as the detection result. The classification steps are the same for different application scenarios. For example, taking the five stages of sleep as an example, sleep is mainly divided into two major states: non-rapid eye movement sleep (NREM sleep) and rapid eye movement sleep (REM sleep). NREM can be further divided into stages 1 to 3, denoted as NREM1, NREM2, and NREM3 respectively. Therefore, the five stages of sleep include: Wake / NREM1 / NREM2 / NREM3 / REM stages, that is, there are 5 types to be identified, i.e., N=5. The signal detection method outputs a 1×5 discrete array, such as [p1,p2,p3,p4,p5], which represent the identification probability values ​​of Wake / NREM1 / NREM2 / NREM3 / REM stages respectively. The stage corresponding to the largest value among p1 to p5 is selected as the final identification result.

[0091] In summary, the signal detection method, detection model, detection device, and application of this invention improve the accuracy and robustness of the detection results by dynamically weighting the signal to be processed. This allows the weights of different channels to change dynamically with the subject's physical or mental activities, and the weights of channels at different frequencies also change dynamically. Using multi-scale convolutional networks can capture features across a wider frequency band, which is beneficial for improving detection accuracy. Preprocessing the original signal removes interference and increases the data volume, which is beneficial for improving the robustness of this method. Squared average pooling and LOG activation can simulate the relative changes in the rhythmic energy of EEG signals, and scaling the relative scale using the LOG function can reduce the impact of jumps.

[0092] Based on the above-described preferred embodiments of the present invention, and through the foregoing description, those skilled in the art can make various changes and modifications without departing from the inventive concept. The technical scope of this invention is not limited to the contents of the specification, but must be determined by the scope of the claims.

Claims

1. An application of a signal detection method in physiological signal classification, characterized in that, Feature information of different frequency bands is filtered out using a weighted model of convolutional networks; The weighted model of the convolutional network includes: A multi-scale convolutional network is used to process the input signal to obtain multiple feature map attention signals; The multi-scale convolutional network includes at least two convolutional branches, each of which contains a temporal convolutional layer and a channel convolutional layer; The convolution kernels of the temporal convolutional layer are used to obtain features of the frequency band of interest, and the convolution kernel parameters are constrained by limiting the weights in order to filter out the frequency band of interest. The temporal convolution kernel size is different for each convolutional branch. The attention signals of the multiple feature maps are cascaded.

2. The application as described in claim 1, characterized in that, The kernel sizes of the multiple temporal convolutional layers are different; The convolution kernel of the temporal convolutional layer is equal to [sampling rate / minimum value of the frequency band of interest] + 1, where [] represents rounding.

3. The application as described in claim 1, characterized in that, The kernel size of the channel convolutional layer is equal to the number of channels.

4. The application as described in claim 1, characterized in that, The kernel size of both the temporal convolutional layer and the channel convolutional layer is constrained by the square norm.

5. The application as described in claim 2, characterized in that, The multi-scale convolutional network includes a first convolutional branch and a second convolutional branch. The temporal convolution kernel of the first convolution branch = [sampling rate / x] + 1, corresponding to the frequency band range greater than x Hz in the input signal; The temporal convolution kernel of the second convolution branch = [sampling rate / y] + 1, corresponding to the frequency band range greater than yHz in the input signal; Where, x y.

6. The application as described in claim 1, characterized in that, Applications of physiological signals include: emotion recognition, epilepsy detection, motor imagery recognition, sleep stage recognition, and fatigue detection.

Images