A frequency band attention method for olfactory EEG recognition

Abstract

The present invention discloses a frequency band attention method for olfactory EEG recognition, comprising the following steps: using an EEG signal acquisition system to respectively obtain the olfactory EEG data induced by a subject under various odor stimulations, after using the data to establish an EEG sample, noise reduction is performed by low-pass and band-pass filtering methods in sequence, downsampling is used to reduce the amount of data, and then the data is divided into frequency bands according to the EEG rhythm, and the data of each frequency band of the sample is cut and segmented according to time using a time window, and the EEG data of each frequency band in different time periods before and after cutting are reconstructed, and a model with a frequency band attention mechanism is constructed based on a residual module and a convolutional block attention module, and the reconstructed EEG data of different time periods are input in parallel into the model with a frequency band attention mechanism for odor identification. The present invention can realize rapid recognition of odors based on olfactory EEG, and the recognition effect is good, and its accuracy and F1 score are both above 97%.

Description

Technical Field

[0001] The present invention relates to electroencephalogram (EEG) analysis and deep learning technology, and in particular to a frequency band attention method for olfactory EEG recognition. Background Art

[0002] Humans can perceive the external environment through hearing, vision, smell, taste and touch. At present, with the development of optical, acoustic and somatic sensory systems, the research on vision, hearing and touch has been relatively mature. However, there are relatively few studies on smell and taste. On the one hand, smell and taste are based on complex chemical reactions and neural transmission mechanisms. On the other hand, sensory stimuli such as sound and light are quantified by loudness and intensity, but smell cannot be expressed and quantified by corresponding physical variables. At present, the sensory analysis of food structure, aroma and flavor mainly relies on human senses and machine perception. However, both methods have their limitations. In terms of artificial sensory evaluation, there are problems such as descriptive differences, lack of units and difficulty in forming a unified evaluation index. In terms of machine perception, there are problems such as limitations on the number and type of sensors and neglect of consumer psychological factors. Therefore, a method that combines human perception with quantitative measurement is urgently needed in olfactory sensory analysis. Odor-evoked EEG (olfactory EEG) has unique advantages in olfactory sensory analysis. It realizes the combination of human perception and quantitative measurement and solves the above technical problems well. However, when using olfactory EEG to identify odorous substances, we often face many influencing factors and uncertainties in parameter selection, such as determining the amount of sample used to form the odor (such as liquid volume, solid volume, size of volumetric flask, etc.), determining the standing time, setting the odor inducing parameters (such as gas path pressure, time and interval of odor stimulation, etc.), selecting olfactory EEG signals (such as time length, window width, signal period, signal frequency band, etc.), determining multiple parameters in the recognition algorithm, and many other issues. There are countless possibilities for the permutations and combinations of these key parameters, and poor parameter selection will seriously affect the recognition results of the olfactory EEG. Summary of the invention

[0003] The main purpose of the present invention is to provide a frequency band attention method for olfactory EEG recognition, determine a series of key parameters affecting olfactory EEG recognition, and construct an effective recognition algorithm.

[0004] The technical solution adopted by the present invention is: a frequency band attention method for olfactory EEG recognition, comprising:

[0005] S1. Several types of samples for olfactory EEG induction are placed in sealed sampling bottles to produce stimulating odors, and the EEG signal acquisition system is used to obtain the olfactory EEG data induced by the subjects under various odor stimulations;

[0006] S2. After the collected EEG data is used to establish an EEG sample, the noise is reduced by low-pass and band-pass filtering methods in turn, and the amount of data is reduced by downsampling. Then, the data is divided into frequency bands according to the EEG rhythm, and the data of each frequency band of the sample is cut and segmented according to time using a time window, and the EEG data of each frequency band at different time periods before and after cutting are reconstructed;

[0007] S3. Build a model with frequency band attention mechanism based on residual module and convolutional block attention module;

[0008] S4. The reconstructed EEG data of different time periods are input in parallel into a model with frequency band attention mechanism for odor identification.

[0009] Furthermore, in step S1, there are 8 kinds of irritating odors. When the odors are generated, 50 ml of each liquid sample and 25 g of each solid sample are sealed in 250 ml sampling bottles respectively, and left to stand for ten minutes to allow the odors to evaporate completely.

[0010] Furthermore, in step S1, an EEG signal acquisition system with a sampling frequency of 256 Hz is used to obtain olfactory EEG data of 21 channels of 15 subjects according to the international 10-20 system; when conducting experiments on each subject, ensure that the pressure in the gas delivery tube with an inner diameter of 6 mm is maintained at 0.3±0.05 MPa, and the 8 odor stimuli are randomly arranged, each odor stimulation includes 20 parallel experiments, wherein the stimulation of each parallel experiment lasts for 10 seconds, and the interval between two parallel experiments is 5 seconds; after the experiment, 2-9 seconds of the EEG data obtained in each parallel experiment are extracted.

[0011] Furthermore, in step S2, when establishing the EEG sample, the 2400 EEG data with a length of 8 seconds are segmented into 2-second segments to establish 9600 samples, and each sample is subjected to a 50 Hz low-pass filter and a 49 Hz-51 Hz band-pass filter in turn to achieve noise reduction of the EEG sample, and the sampling frequency of the sample is reduced to 128 Hz by downsampling to reduce the amount of data, and then the downsampled data is divided into δ , θ , α、β and γ , there are 5 frequency bands: 0-4Hz, 4-8Hz, 8-13Hz, 13-30Hz and 30-50Hz.

[0012] Furthermore, in step S2, when the data of each frequency band of the sample is cut and segmented by time, the EEG data of 5 frequency bands with a length of 2s in each sample is segmented by a window with a length of 1s and a step size of 0.5s, so that the sample is divided into 3 time periods of data, and each time period of data contains 5 EEG data of different frequency bands with a length of 1s.

[0013] Furthermore, in step S2, when reconstructing the EEG data of each frequency band in different time periods before and after cutting, the data time period before cutting is 0-2 seconds, and the data time period after cutting is 0-1 second, 0.5-1.5 seconds and 1-2 seconds, and each time period before and after cutting contains EEG data of 5 frequency bands, wherein the data length before cutting is 2 seconds, and the data format of each frequency band is 21×256, the data length after cutting is 1 second, and the data format of each frequency band is 21×128, so a 21-channel EEG sample can generate 4 data segments, which are in the form of 5×21×256, 5×21×128, 5×21×128, and 5×21×128, respectively, and then the third dimension of these 4 data segments is expanded to the second dimension and converted into data in the form of 5×84×64, 5×42×64, 5×42×64, and 5×42×64, respectively.

[0014] Furthermore, in the step S3, the purpose of constructing a model with a frequency band attention mechanism is to fully explore the characteristic information of the olfactory EEG in each frequency band, and to fully explore the characteristic information of its key frequency bands without the need to manually select effective frequency bands; when constructing a model with a frequency band attention mechanism based on the residual module and the convolutional block attention module, the convolutional block attention module has channel attention and spatial attention mechanisms, which can be embedded in the residual module to constitute a residual module with an attention mechanism. Multiple residual modules with attention mechanisms are combined to form a residual network with an attention mechanism. The model with a frequency band attention mechanism consists of 4 parallel residual networks with attention mechanisms at the front end, and a splicing layer, a fully connected layer and a softmax layer at the back end.

[0015] Furthermore, in step S4, when the reconstructed EEG data of different time periods are input in parallel to the model with frequency band attention mechanism, the EEG data of the four time periods are respectively input into the four parallel residual networks with attention mechanism at the front end of the model with frequency band attention mechanism, wherein the number of input channels of each residual network with attention mechanism is 5, corresponding to the five frequency bands of EEG data respectively. Therefore, the attention mechanism of the model can be used to effectively realize automatic attention to the key frequency bands in the olfactory EEG, and then its spatial attention mechanism can realize the effective mining of key spatiotemporal features in the EEG data in the form of 42×64 or 84×64 in the key frequency bands.

[0016] Furthermore, in step S4, when the reconstructed EEG data in the form of 5×84×64, 5×42×64, 5×42×64, and 5×42×64 are input in parallel to the model with the frequency band attention mechanism, the reconstructed EEG data are subjected to a series of attention, convolution, and pooling operations of its four parallel residual networks with the attention mechanism to obtain data in the form of 1×4×512, 1×2×512, 1×2×512, and 1×2×512, respectively, and then the data is flattened The four segments of data are in the form of 1×2048, 1×1024, 1×1024 and 1×1024, so as to realize the mining of olfactory EEG data in four time periods. Finally, the flattened feature data is integrated with the feature data in the key frequency bands of the four time periods through the splicing layer, so that the model can automatically make full use of the key feature information in the EEG data according to the classification task. The spliced ​​data in the form of 1×5120 is then passed through the fully connected layer and the softmax layer in turn to obtain the predicted output of 8 odors.

[0017] Advantages of the present invention:

[0018] The present invention can realize rapid recognition of odors based on olfactory EEG, and has good recognition effect, with the accuracy and F1 score both being above 97%.

[0019] In addition to the above-described purposes, features and advantages, the present invention has other purposes, features and advantages. The present invention will be further described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The drawings constituting a part of this application are used to provide a further understanding of the present invention. The illustrative embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0021] Figure 1 This is a flow chart of a method for quickly identifying odors based on olfactory EEG according to an embodiment of the present invention;

[0022] Figure 2 Schematic diagram of a model with a frequency band attention mechanism in an embodiment of the present invention;

[0023] Figure 3 A convolutional block attention module in an embodiment of the present invention;

[0024] Figure 4 is a residual module with an attention mechanism in an embodiment of the present invention;

[0025] Figure 5 This is a result diagram of odor recognition in an embodiment of the present invention. DETAILED DESCRIPTION

[0026] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0027] A frequency band attention method for olfactory EEG recognition comprises the following steps:

[0028] S1. Several types of samples for olfactory EEG induction are placed in sealed sampling bottles to produce stimulating odors, and the EEG signal acquisition system is used to obtain the olfactory EEG data induced by the subjects under various odor stimulations;

[0029] S2. After the collected EEG data is used to establish an EEG sample, the noise is reduced by low-pass and band-pass filtering methods in turn, and the amount of data is reduced by downsampling. Then, the data is divided into frequency bands according to the EEG rhythm, and the data of each frequency band of the sample is cut and segmented according to time using a time window, and the EEG data of each frequency band at different time periods before and after cutting are reconstructed;

[0030] S3. Build a model with frequency band attention mechanism based on residual module and convolutional block attention module;

[0031] S4. The reconstructed EEG data of different time periods are input in parallel into a model with frequency band attention mechanism for odor identification.

[0032] In step S1, there are 8 kinds of irritating odors. When the odors are generated, 50 ml of each liquid sample and 25 g of each solid sample are sealed in 250 ml sampling bottles respectively and left to stand for ten minutes to allow the odors to evaporate completely.

[0033] In step S1, an EEG signal acquisition system with a sampling frequency of 256 Hz is used to obtain olfactory EEG data of 21 channels of 15 subjects according to the international 10-20 system. When conducting experiments on each subject, ensure that the pressure in the gas delivery tube with an inner diameter of 6 mm is maintained at 0.3±0.05Mpa, and the 8 odor stimuli are randomly ordered. Each odor stimulation contains 20 parallel experiments, where the stimulation of each parallel experiment lasts for 10 seconds, and the interval between two parallel experiments is 5 seconds. After the experiment, extract 2-9 seconds of the EEG data obtained in each parallel experiment.

[0034] In step S2, when establishing the EEG sample, the 2400 (15×8×20) EEG data with a length of 8 seconds are segmented into 2-second segments to establish 9600 samples, and each sample is subjected to a 50 Hz low-pass filter and a 49 Hz-51 Hz band-pass filter in turn to achieve noise reduction of the EEG sample. The sampling frequency of the sample is reduced to 128 Hz by downsampling to reduce the amount of data, and then the downsampled data is divided into δ (0-4Hz), θ (4-8Hz), α (8-13Hz), β (13-30Hz) and γ (30-50Hz)5 frequency bands.

[0035] In step S2, when the data of each frequency band of the sample is cut and segmented by time, the EEG data of 5 frequency bands with a length of 2s in each sample is segmented by a window with a length of 1s and a step size of 0.5s. Therefore, the sample is divided into 3 time periods of data, and each time period of data contains 5 EEG data of different frequency bands with a length of 1s.

[0036] In step S2, when reconstructing the EEG data of each frequency band in different time periods before and after cutting, the data time period before cutting is 0-2 seconds, and the data time period after cutting is 0-1 second, 0.5-1.5 seconds and 1-2 seconds, and each time period before and after cutting contains EEG data of 5 frequency bands, wherein the data length before cutting is 2 seconds, and the data format of each frequency band is 21×256, the data length after cutting is 1 second, and the data format of each frequency band is 21×128, so a 21-channel EEG sample can generate 4 data segments, which are in the form of 5×21×256, 5×21×128, 5×21×128, and 5×21×128, respectively, and then the third dimension of these 4 data segments is expanded to the second dimension and converted into data in the form of 5×84×64, 5×42×64, 5×42×64, and 5×42×64, respectively.

[0037] In step S3, the purpose of constructing a model with a frequency band attention mechanism is to fully explore the characteristic information of the olfactory EEG in each frequency band, and to fully explore the characteristic information of its key frequency band without the need to manually select effective frequency bands. When constructing a model with a frequency band attention mechanism based on a residual module and a convolutional block attention module, the convolutional block attention module has a channel attention and a spatial attention mechanism, which can be embedded in the residual module to form a residual module with an attention mechanism. Multiple residual modules with an attention mechanism are combined to form a residual network with an attention mechanism. The model with a frequency band attention mechanism consists of 4 parallel residual networks with an attention mechanism at the front end, and a splicing layer, a fully connected layer, and a softmax layer at the back end.

[0038] In the step S4, when the reconstructed EEG data of different time periods are input in parallel to the model with the frequency band attention mechanism, the EEG data of the four time periods are respectively input into the four parallel residual networks with the attention mechanism at the front end of the model with the frequency band attention mechanism, wherein the number of input channels of each residual network with the attention mechanism is 5, corresponding to the five frequency bands of the EEG data, so the attention mechanism of the model can be used to effectively realize automatic attention to the key frequency bands in the olfactory EEG, and then its spatial attention mechanism can realize the effective mining of the key spatiotemporal features in the EEG data in the form of 42×64 or 84×64 in the key frequency bands.

[0039] In step S4, when the reconstructed EEG data in the form of 5×84×64, 5×42×64, 5×42×64, and 5×42×64 are input in parallel to the model with the frequency band attention mechanism, the reconstructed EEG data are subjected to a series of attention, convolution, and pooling operations of the four parallel residual networks with the attention mechanism to obtain data in the form of 1×4×512, 1×2×512, 1×2×512, and 1×2×512, respectively, and then the data is flattened into the form The four segments of data are 1×2048, 1×1024, 1×1024 and 1×1024, so as to realize the mining of olfactory EEG data in four time periods. Finally, the flattened feature data is integrated with the feature data in the key frequency bands of the four time periods through the splicing layer, so that the model can automatically make full use of the key feature information in the EEG data according to the classification task. The data in the form of 1×5120 after splicing is then passed through the fully connected layer and the softmax layer in turn to obtain the predicted output of 8 odors.

[0040] like Figure 1 As shown, the method for quickly identifying degradable and non-degradable plastics according to the embodiment of the present invention comprises the following steps:

[0041] S1. Obtain olfactory EEG data, prepare stimulating odors, take 50 ml of mint syrup, rose syrup, orange juice, and strawberry juice, take 25 g of garlic sauce, shrimp paste, sour bamboo shoots, and durian sauce, seal them in 250 ml sampling bottles and let them stand for ten minutes to allow the odor to completely evaporate. Use the NCERP-P EEG signal acquisition system with a sampling frequency of 256 Hz, and obtain olfactory EEG data of 21 channels of 15 subjects according to the international 10-20 system. When conducting experiments on each subject, ensure that the pressure in the gas delivery tube with an inner diameter of 6 mm is maintained at 0.3 ± 0.05 MPa. The 8 odor stimuli are randomly ordered, and each odor stimulus contains 20 parallel experiments, in which the stimulation of each parallel experiment lasts for 10 seconds, and the interval between two parallel experiments is 5 seconds. After the experiment, extract 2-9 seconds of the EEG data obtained from each parallel experiment.

[0042] S2. Create EEG samples. Divide the acquired 2400 (15×8×20) EEG data with a length of 8 seconds into segments of 2 seconds to create 9600 samples.

[0043] S3, filtering, denoising and downsampling. The finite impulse response in EEGLAB is used to perform a 50Hz low-pass filter on the EEG samples to filter out high-frequency noise. Then a notch filter with a lower passband edge of 49Hz and an upper passband edge of 51Hz is used to eliminate power frequency noise. The sampling frequency of the data is then reduced to 128Hz to reduce the amount of data.

[0044] S4, frequency band division, divide the downsampled data into δ (0-4Hz), θ (4-8Hz), α (8-13Hz), β (13-30Hz) and γ (30-50Hz)5 frequency bands.

[0045] S5. Cut the data of each frequency band. The EEG data of 5 frequency bands with a length of 2s in each sample are segmented by a window with a length of 1s and a step size of 0.5s. Therefore, the sample is divided into 3 time periods, and each time period contains 5 EEG data of different frequency bands with a length of 1s.

[0046] S6, data reconstruction, the data time period before cutting is 0-2 seconds, the data time period after cutting is 0-1 second, 0.5-1.5 seconds and 1-2 seconds, each time period before and after cutting contains EEG data of 5 frequency bands, wherein the data length before cutting is 2 seconds, and the data format of each frequency band is 21×256, the data length after cutting is 1 second, and the data format of each frequency band is 21×128, so a 21-channel EEG sample can generate 4 data segments, which are in the form of 5×21×256, 5×21×128, 5×21×128, and 5×21×128 respectively, and then the third dimension of these 4 data segments is expanded to the second dimension and converted into data in the form of 5×84×64, 5×42×64, 5×42×64, and 5×42×64 respectively.

[0047] S7. According to Figure 2 A model with a band attention mechanism is constructed. The network parameters of the residual network 0 with the attention mechanism are shown in Table 1. The network parameters of the residual network 0 with the attention mechanism differ from those of the other three parallel networks only in the last flattening layer. Figure 3 Construct a convolutional block attention module, which has channel attention and spatial attention mechanisms, and can be embedded in the residual module to form a residual module with attention mechanism. Therefore, according to Figure 4 A residual module with an attention mechanism is constructed. The model with a band attention mechanism consists of four parallel residual networks with an attention mechanism at the front end, and a concatenation layer, a fully connected layer, and a softmax layer at the back end.

[0048] S8, odor identification, when the reconstructed EEG data of different time periods are input in parallel to the model with frequency band attention mechanism, the EEG data of the four time periods are respectively input into the four parallel residual networks with attention mechanism at the front end of the model with frequency band attention mechanism, wherein the number of input channels of each residual network with attention mechanism is 5, corresponding to the five frequency bands of EEG data respectively, and when the reconstructed EEG data in the form of 5×84×64, 5×42×64, 5×42×64, and 5×42×64 are input in parallel to the model with frequency band attention mechanism, the reconstructed EEG data are subjected to a series of attention, convolution and After operations such as pooling, data in the form of 1×4×512, 1×2×512, 1×2×512 and 1×2×512 are obtained, and then the data is flattened into 4 segments of data in the form of 1×2048, 1×1024, 1×1024 and 1×1024, so as to realize the mining of olfactory EEG data in 4 time periods. Finally, the flattened feature data is integrated through the splicing layer to realize the integration of feature data in the key frequency bands of the 4 time periods, so that the model can automatically make full use of the key feature information in the EEG data according to the classification task. The spliced ​​data in the form of 1×5120 is then passed through the fully connected layer and the softmax layer in turn to obtain the predicted output of 8 odors. When training and testing the model, 2 / 3 samples were randomly selected from the data set as the training set, and 1 / 3 samples were randomly selected as the prediction set. The number of iterations of model training was 100 times, and the learning rate was set to 0.03. Finally, the Adam optimizer and cross entropy loss were used to update the network parameters. The results obtained in the test set after repeated training for 5 times were averaged. The final odor recognition results are shown in the figure. Figure 5 As shown in the figure, both the accuracy and F1 score values ​​are over 97%.

[0049] Table 1

[0050]

[0051] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A frequency band attention method for olfactory EEG recognition, characterized in that: include: S1. Several types of samples for olfactory EEG induction are placed in sealed sampling bottles to produce stimulating odors, and the EEG signal acquisition system is used to obtain the olfactory EEG data induced by the subjects under various odor stimulations; S2. After the collected EEG data is used to establish an EEG sample, the noise is reduced by low-pass and band-pass filtering methods in turn, and the amount of data is reduced by downsampling. Then, the data is divided into frequency bands according to the EEG rhythm, and the data of each frequency band of the sample is cut and segmented according to time using a time window, and the EEG data of each frequency band at different time periods before and after cutting are reconstructed; S3. Build a model with frequency band attention mechanism based on residual module and convolutional block attention module; S4, inputting the reconstructed EEG data of different time periods in parallel into a model with frequency band attention mechanism for odor identification; In the step S3, the purpose of constructing a model with a frequency band attention mechanism is to fully explore the characteristic information of the olfactory EEG in each frequency band, and to fully explore the characteristic information of its key frequency band without manually selecting the effective frequency band; when constructing a model with a frequency band attention mechanism based on the residual module and the convolutional block attention module, the convolutional block attention module has a channel attention and a spatial attention mechanism, which is embedded in the residual module to form a residual module with an attention mechanism, and multiple residual modules with an attention mechanism are combined to form a residual network with an attention mechanism. The model with a frequency band attention mechanism consists of 4 parallel residual networks with an attention mechanism at the front end, and a splicing layer, a fully connected layer and a softmax layer at the back end; In the step S4, when the reconstructed EEG data of different time periods are input in parallel to the model with the frequency band attention mechanism, the EEG data of the four time periods are respectively input into the four parallel residual networks with the attention mechanism at the front end of the model with the frequency band attention mechanism, wherein the number of input channels of each residual network with the attention mechanism is 5, corresponding to the five frequency bands of the EEG data, so that the attention mechanism of the model is used to effectively realize the automatic attention to the key frequency bands in the olfactory EEG, and then the spatial attention mechanism realizes the effective mining of the key spatiotemporal features in the EEG data in the form of 42×64 or 84×64 in the key frequency bands; In step S4, when the reconstructed EEG data in the form of 5×84×64, 5×42×64, 5×42×64, and 5×42×64 are input in parallel to the model with the frequency band attention mechanism, the reconstructed EEG data are subjected to a series of attention, convolution, and pooling operations of the four parallel residual networks with the attention mechanism to obtain data in the form of 1×4×512, 1×2×512, 1×2×512, and 1×2×512, respectively, and then the data is flattened into Four data segments in the form of 1×2048, 1×1024, 1×1024 and 1×1024 are used to realize the mining of olfactory EEG data in four time periods. Finally, the flattened feature data is integrated through the splicing layer to realize the integration of feature data in the key frequency bands of the four time periods, so that the model can automatically make full use of the key feature information in the EEG data according to the classification task. The spliced ​​data in the form of 1×5120 is then passed through the fully connected layer and the softmax layer in turn to obtain the predicted output of 8 odors.

2. The frequency band attention method for olfactory EEG recognition according to claim 1, It is characterized in that In step S1, there are 8 kinds of irritating odors. When the odors are generated, 50 ml of each liquid sample and 25 g of each solid sample are sealed in 250 ml sampling bottles respectively and left to stand for ten minutes to allow the odors to evaporate completely.

3. The frequency band attention method for olfactory EEG recognition according to claim 1, characterized in that: In step S1, an EEG signal acquisition system with a sampling frequency of 256 Hz is used to obtain olfactory EEG data of 21 channels of 15 subjects according to the international 10-20 system; when conducting experiments on each subject, it is ensured that the pressure in the entire gas delivery tube with an inner diameter of 6 mm is maintained at 0.3±0.05 MPa, and the 8 odor stimuli are randomly arranged, each odor stimulation includes 20 parallel experiments, wherein the stimulation of each parallel experiment lasts for 10 seconds, and the interval between two parallel experiments is 5 seconds; after the experiment, 2-9 seconds of the EEG data obtained in each parallel experiment are extracted.

4. The frequency band attention method for olfactory EEG recognition according to claim 1, characterized in that: In step S2, when establishing EEG samples, the 2400 EEG data with a length of 8 seconds are segmented into 2-second segments to establish 9600 samples, and each sample is subjected to 50 Hz low-pass filtering and 49 Hz-51 Hz band-pass filtering in turn to achieve noise reduction of the EEG samples. The sampling frequency of the samples is reduced to 128 Hz by downsampling to reduce the amount of data, and then the downsampled data is divided into δ , θ , α、β and γ , there are 5 frequency bands: 0-4Hz, 4-8Hz, 8-13Hz, 13-30Hz and 30-50Hz.

5. The frequency band attention method for olfactory EEG recognition according to claim 1, characterized in that: In step S2, when the data of each frequency band of the sample is cut and segmented by time, the EEG data of 5 frequency bands with a length of 2s in each sample is segmented by a window with a length of 1s and a step size of 0.5s. Therefore, the sample is divided into 3 time periods of data, and each time period of data contains 5 EEG data of different frequency bands with a length of 1s.

6. The frequency band attention method for olfactory EEG recognition according to claim 1, characterized in that: In step S2, when reconstructing the EEG data of each frequency band in different time periods before and after cutting, the data time period before cutting is 0-2 seconds, and the data time period after cutting is 0-1 second, 0.5-1.5 seconds and 1-2 seconds. Each time period before and after cutting contains EEG data of 5 frequency bands, wherein the data length before cutting is 2 seconds, and the data format of each frequency band is 21×256, the data length after cutting is 1 second, and the data format of each frequency band is 21×128, so a 21-channel EEG sample generates 4 data segments, which are in the form of 5×21×256, 5×21×128, 5×21×128, and 5×21×128, respectively, and then the third dimension of these 4 data segments is expanded to the second dimension and converted into data in the form of 5×84×64, 5×42×64, 5×42×64, and 5×42×64, respectively.

Images