A method for underwater target recognition based on brain network

Abstract

The invention discloses an underwater target recognition method based on brain network, comprising the following steps: S1, recruiting multiple sonar operators as subjects for experiment, and obtaining multiple groups of EEG signals corresponding to the multiple subjects; S2, preprocessing the multiple groups of EEG signals obtained in step S1; S3, based on an improved brain network acquisition algorithm, calculating and solving the EEG signals preprocessed in step S2, and obtaining multiple groups of PSI adjacency matrices; S4, selecting connecting nodes of the PSI adjacency matrix as features to form a feature set, and sorting the features in each feature set, and then using a 6-fold cross-validation method to divide the sorted feature set into a training set and a test set, the training set is used to complete the training of the classification model, and the test set is used to obtain the final classification accuracy of the subject; S5, when actually performing underwater target recognition, the sonar operator selects the corresponding classification model to complete the real-time collection and classification of EEG data, so as to realize automatic recognition of underwater targets.

Description

Technical Field

[0001] The present invention relates to the technical field of electroencephalogram (EEG) signal processing, and in particular to an underwater target recognition method based on brain network. Background Art

[0002] Sonar is an important technology that is widely used in underwater detection, underwater target identification, positioning, etc. Compared with traditional underwater cameras, this technology can effectively avoid the limitations of low visibility and weak lighting in underwater environments. In view of this, acoustic methods have been widely used in underwater species identification, fish community structure, biomass and population dynamics research.

[0003] In recent years, as people's exploration of the ocean has become more and more in-depth, the development of fishery technology has become more and more mature. In order to improve fishing efficiency and output, people have begun to pay attention to underwater fish identification technology based on sonar technology to quickly identify and locate fish schools. However, relying solely on machine learning algorithms to identify sonar images of target fish schools often fails to achieve the desired results because the classification model cannot cover all samples for training. When the marine environment is complex, the accuracy of this method cannot meet the current fishing needs.

[0004] In the fishery, relying on experienced sonar operators, sonar can help fishermen quickly locate the position and quantity of fish. Due to the brain's memory storage mechanism, it can quickly identify targets subconsciously. If the relevant areas of the sonar personnel's brain can be analyzed, it will help people understand the brain's processing mechanism in cognitive judgment tasks. At present, the research on EEG-based visual-auditory dual-mode underwater target recognition is limited, and it is difficult to realize the in-depth application of brain-computer interface technology in underwater target recognition. Summary of the invention

[0005] In order to solve the problems raised in the above background technology, the technical solution adopted by the present invention is:

[0006] A method for underwater target recognition based on brain network includes the following steps:

[0007] S1. Recruit multiple sonar operators as subjects for the experiment. During the experiment, provide stimulation signals of multiple types of underwater targets. When multiple subjects judge the underwater targets, obtain multiple sets of EEG signals corresponding to the multiple subjects.

[0008] S2, preprocessing the multiple groups of EEG signals obtained in step S1;

[0009] S3, based on the improved brain network extraction algorithm, the EEG signal preprocessed in step S2 is calculated and solved to obtain multiple groups of PSI adjacency matrices;

[0010] S4, selecting the connection nodes of the PSI adjacency matrix as features to form a feature set, obtaining the feature set corresponding to each subject, and sorting the features in each feature set, and then using a 6-fold cross-validation method to divide the sorted feature set into a training set and a test set. The training set is used to complete the training of the classification model, and the test set is used to obtain the final classification accuracy of the subject;

[0011] S5. When actually identifying underwater targets, the sonar operator selects the corresponding classification model to complete the real-time collection and classification of EEG data, thereby realizing automatic identification of underwater targets.

[0012] Furthermore, in step S1, the following steps are specifically included:

[0013] Several sonar operators were recruited as subjects. Before the experiment, each subject was trained to be familiar with the stimulation signals of ocean white noise and various underwater targets. The stimulation signals were video signals composed of audio and waveforms, ensuring that each subject could quickly and accurately distinguish ocean white noise from various underwater targets and form long-term memory.

[0014] During the experiment, each subject wore an acoustic signal providing device and an EEG signal collecting device, and viewed the sonar waveform through a display screen. An operating keyboard was also provided. Ocean white noise and stimulation signals of various types of underwater targets were used as stimulation signals and randomly provided to the subjects. After receiving the stimulation signals, the subjects made judgments on the stimulation signals and selected their judgment results through the operating keyboard. At the same time, the EEG signal collecting device collected EEG signals. The above process was repeated multiple times for each subject to obtain multiple groups of EEG signals corresponding to multiple subjects.

[0015] Further, in step S1, during the experiment, the EEG signal acquisition device includes 64 Ag / AgCl EEG channels, as well as a CPz reference electrode, an EOG electrode, and an AFz ground electrode.

[0016] Furthermore, in step S2, preprocessing the EEG signal specifically includes:

[0017] Data inspection to remove obvious interference fragments from the data;

[0018] Bandpass filtering the data using a filter pair;

[0019] The mastoid electrodes M1 and M2 were used instead of the reference electrodes;

[0020] Remove artifacts from EEG signals through independent component analysis;

[0021] Perform baseline correction;

[0022] Event-related potential analysis was performed to determine the start and end time range of the event-related potential for subsequent data analysis.

[0023] Furthermore, in step S3, the following steps are specifically included:

[0024] Welch evaluation method is used to solve the cross spectrum. Given a multi-channel time series T = {t 1 ,t 2 ,t 3 ,…,t N}, the solution of complex coherence is defined as:

[0025]

[0026] Where S is the cross spectrum; i,j represents the time series of channels i and j in T (1 <i,j<N,i≠j);

[0027] The improved brain network extraction algorithm is as follows:

[0028]

[0029] where α ij (f)=|C ij (f)| is the frequency-dependent weight value, and:

[0030] H(Φ(f))=P(Φ(f+δf))-P(Φ(f))

[0031] P represents the new sequence obtained after polynomial fitting of the Φ(f), f∈F sequence;

[0032] The improved brain network extraction algorithm is used to solve the brain network for each trial and obtain the corresponding PSI adjacency matrix.

[0033] Furthermore, in step S4, the F_score method is specifically used to sort the features, and the calculation method of F_score is as follows:

[0034]

[0035] Among them, i represents the i-th feature in the feature set, j is the j-th category in the category, and F i is the F_score value of the i-th feature, l is the number of categories, n j is the number of samples in the jth class, k is the kth feature instance under the jth class, is the total average of the i-th feature, is the average of the i-th feature under the j-th category, F i What is actually measured is the ratio of the inter-class distance to the intra-class distance under the i-th feature;

[0036] After the feature set is sorted, the data set is divided using the 6-fold cross-validation method. For each fold validation, features are added to the feature set in descending order according to the F_score to complete the training of the classification model. The classification model uses SVM, and the kernel function is a linear kernel.

[0037] The divided test set is used to obtain the classification accuracy of the subject. For subjects whose classification accuracy is lower than the preset threshold, the model parameters are adjusted and re-trained until the classification accuracy reaches the preset threshold, and finally the classification model corresponding to each subject is obtained.

[0038] Compared with the prior art, the present invention has the following beneficial effects:

[0039] The brain network-based underwater target recognition method provided by the present invention combines the brain's memory and experience to further process the sonar signal, and then uses a machine learning algorithm to classify the brain signal, which is equivalent to allowing the human brain to complete the feature extraction work, and then being learned by the classification model. The brain can classify the target at an extremely fast speed in the subconscious. Therefore, combined with the brain-computer interface technology, the present invention not only greatly improves the generalization ability of the classification model, but also can achieve rapid response of the entire system. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 A schematic flow chart of the brain network-based underwater target recognition method provided by the present invention. DETAILED DESCRIPTION

[0041] In order to make the technical means, creative features, objectives and effects achieved by the present invention easy to understand, the following further describes how the present invention is implemented in conjunction with the accompanying drawings and specific implementation methods.

[0042] Reference Figure 1 As shown, the present invention provides an underwater target recognition method based on brain network, comprising the following steps:

[0043] S1. Recruit multiple sonar operators as subjects for the experiment. During the experiment, provide stimulation signals of multiple types of underwater targets. When multiple subjects judge the underwater targets, obtain multiple sets of EEG signals corresponding to the multiple subjects.

[0044] S2. Preprocess the multiple groups of EEG signals acquired in step S1.

[0045] S3. Based on the improved brain network extraction algorithm, the EEG signal preprocessed in step S2 is calculated and solved to obtain multiple groups of PSI adjacency matrices.

[0046] S4. Select the connecting nodes of the PSI adjacency matrix as features to form a feature set, obtain the feature set corresponding to each subject, and sort the features in each feature set. Then, use the 6-fold cross-validation method to divide the sorted feature set into a training set and a test set. The training set is used to complete the training of the classification model, and the test set is used to obtain the final classification accuracy of the subject.

[0047] S5. When actually identifying underwater targets, the sonar operator selects the corresponding classification model to complete the real-time collection and classification of EEG data, thereby realizing automatic identification of underwater targets.

[0048] Furthermore, in step S1, the following steps are specifically included:

[0049] Usually, the sonar signals received by sonar operators are mostly ocean white noise. When the target signal suddenly appears, the sonar operator is required to detect, judge, locate and track it in time. Therefore, a number of sonar operators were recruited as subjects. Before the experiment, each subject was trained and familiar with the stimulation signals of ocean white noise and various types of underwater targets. The stimulation signal is a video signal composed of audio and waveform, which enables the subject to receive visual and auditory stimulation at the same time, ensuring that each subject can quickly and accurately distinguish between ocean white noise and various types of underwater targets and form long-term memory;

[0050] During the experiment, each subject wore an acoustic signal providing device and an EEG signal collecting device, and viewed the sonar waveform through a display screen. An operating keyboard was also provided. Ocean white noise and stimulation signals of various types of underwater targets were used as stimulation signals and randomly provided to the subjects. After receiving the stimulation signals, the subjects made judgments on the stimulation signals and selected their judgment results through the operating keyboard. At the same time, the EEG signal collecting device collected EEG signals. The above process was repeated multiple times for each subject to obtain multiple groups of EEG signals corresponding to multiple subjects.

[0051] In this embodiment, the biased stimulus signal comes from fish signals collected by passive sonar, including lobsters, dolphins and manta rays, and the standard stimulus is ocean white noise. All stimulus signals will be presented to the subjects in the form of video signals consisting of waveforms and audio to ensure that they can receive visual and auditory stimulation at the same time. The experimental paradigm consists of 100 trials, including 70 standard stimuli and 30 biased stimuli. In order to ensure the authenticity of the experimental simulation and the effectiveness of the stimulus, the signals of the three types of fish in the biased stimulus will be evenly distributed, and the 10 biased stimuli of the same type will come from sonar signals collected at different time points. All audio signals used for stimulation will be adjusted to the same level. In addition to the standard stimulus, all other stimulus signals are mixed with ship noise. Among them, the mixing weight of ship noise is 0.1~0.2, and each stimulus signal will randomly mix 1-3 types of ship noise.

[0052] A total of 45 skilled sonar operators were recruited as subjects in this example, with the following requirements: 1) working in marine fisheries for at least 1 year, mainly as sonar operators; 2) normal vision and hearing; 3) aged between 26 and 30 years old; 4) no history of mental illness; 5) all subjects are right-handed. Two weeks before the formal experiment, the study provided the subjects with a pure signal without any ship noise. This signal is consistent with the stimulus signal used in the formal experiment. Each subject is required to perform memory training on all signals at least three times a day to ensure that they can quickly distinguish four types of signals (ocean white noise, lobsters, dolphins, and manta rays) and form long-term memory. Before the experiment officially begins, the subjects will be asked to perform a 10-minute training task, which is consistent with the formal experimental paradigm used.

[0053] In the experiment, a "+" sign will first appear on the screen for a duration of 300ms. After that, the screen will turn black and the duration will be randomly between 500 and 1500 milliseconds. The stimulus signal will appear later and last for 1000ms. During this period, the subject must respond quickly and press the corresponding key according to the different stimulus types. At the same time, the subject needs to complete the classification of the stimulus signal in the brain. Finally, the mode will provide a short rest period of 1000 milliseconds. The above is a complete experiment. The whole experiment consists of 3 stages, and for each subject, each stage contains 100 separate trials, a total of 300 trials.

[0054] During the experiment, the EEG signal acquisition equipment was based on the international 10-10 system, including 64 Ag / AgCl EEG channels, as well as the CPz reference electrode, EOG electrode, and AFz ground electrode.

[0055] Furthermore, in step S2, preprocessing the EEG signal specifically includes:

[0056] Data inspection: remove obvious interference fragments from the data; perform 0.3-30Hz bandpass filtering on the data through a filter; use mastoid electrodes M1 and M2 to replace the reference electrode; use independent component analysis (ICA) to remove artifacts in the EEG signal, such as electroretinogram interference; extract each epoch type and select the 200ms before the baseline for baseline correction; perform event-related potential analysis to determine the start and end time range of the event-related potential (ERP) for subsequent data analysis.

[0057] Furthermore, in step S3, the following steps are specifically included:

[0058] In the calculation of the phase slope index (PSI), the solution of the coherent imaginary part is the basis of the entire estimation process. Before this, it is necessary to first calculate the cross spectrum between time series. The present invention adopts the Welch evaluation method to solve the cross spectrum. Given a multi-channel time series T = {t 1 ,t 2 ,t 3 ,…,t N}, the solution of complex coherence is defined as:

[0059]

[0060] Where S is the cross spectrum; i,j represents the time series of channels i and j in T (1 <i,j<N,i≠j)。

[0061] In the traditional definition of PSI, it can be expressed as:

[0062]

[0063] in Represents the operation of taking the imaginary part of a complex value. For this formula, there is a more intuitive rewrite, the rewritten formula is:

[0064]

[0065] where α ij (f)=|C ij (f)| is the frequency-dependent weight value.

[0066] It can be seen that the traditional PSI uses coherence to weight the phase difference, and then accumulates it to obtain the final evaluation, and finally obtains the PSI adjacent feature matrix. However, this method will cause the evaluation performance of the PSI algorithm to deteriorate under the condition of large noise. In the present invention, it is improved and the improved brain network algorithm is used as follows:

[0067]

[0068] in:

[0069] H(Φ(f))=P(Φ(f+δf))-P(Φ(f))

[0070] P represents the new sequence obtained after polynomial fitting of the Φ(f), f∈F sequence. The main purpose is to remove discrete values ​​to ensure the stability of the sequence.

[0071] The improved brain network extraction algorithm is used to solve the brain network for each trial and obtain the corresponding PSI adjacency matrix.

[0072] Further, in step S4, as mentioned above, in this embodiment, for each subject, a total of 300 trials were performed, and 300 PSI adjacency matrices were obtained through brain network calculation. For the PSI adjacency matrix in each trial, since the absolute value of the matrix is ​​a symmetric matrix, its upper triangular matrix and lower triangular matrix are redundant. In this embodiment, the upper triangular matrix of each PSI adjacency matrix is ​​selected and stretched into a one-dimensional matrix. In this embodiment, the original matrix is ​​61*61 (number of channels*number of channels, the original number of channels is 64, and the electrooculogram (EOG) channel and two mastoid electrode channels are removed), and the diagonal is 1, so the connection nodes finally retained are the upper triangular matrices that do not contain diagonal data in the PSI adjacency matrix. After being stretched into a one-dimensional matrix, its size is 1*1830. Therefore, 1*1830 connection nodes can be obtained for each trial and used as features to establish a feature set, which is used for the final classification.

[0073] When classifying, the F_score method is used to sort the features, and the features are put into the feature set in descending order according to the F_score value for classification. The calculation method of F_score is as follows:

[0074]

[0075] Among them, i represents the i-th feature in the feature set, j is the j-th category in the category, and F i is the F_score value of the i-th feature, l is the number of categories, n j is the number of samples in the jth class, k is the kth feature instance under the jth class, is the total average of the i-th feature, is the average of the i-th feature under the j-th category, F i What is actually measured is the ratio of the inter-class distance to the intra-class distance under the i-th feature.

[0076] After the feature set is sorted, the data set is divided using the 6-fold cross-validation method. For each fold of validation, features are added to the feature set one by one according to the F_score sorting from large to small to complete the training of the classification model. The classification model uses SVM, and the kernel function is a linear kernel; each subject will get a corresponding model, and 45 subjects will eventually get 45 models.

[0077] The divided test set is used to obtain the classification accuracy of the subject. For subjects whose classification accuracy is lower than a preset threshold (such as 80%), the model parameters are adjusted and retrained until the classification accuracy reaches the preset threshold, and finally the classification model corresponding to each subject is obtained. In this embodiment, the classification results show that the highest average classification accuracy of 45 subjects obtained by this method is 82.43%. It can be understood that in the present invention, because the selected subjects are experienced sonar testers who have undergone a lot of training in advance, it can be considered that the judgment results of the sonar testers on underwater targets are accurate; the above-mentioned "classification accuracy" refers to the classification accuracy of the classification model corresponding to the subject, rather than the accuracy of the subject's own judgment of underwater targets.

[0078] After the model training is completed, further, in step S5, the sonar operator can select the corresponding classification model to complete the real-time collection and classification of EEG data to achieve automatic recognition of underwater targets.

[0079] In summary, the brain network-based underwater target recognition method provided by the present invention combines the brain's memory and experience to further process the sonar signals, and then uses a machine learning algorithm to classify the brain signals, which is equivalent to allowing the human brain to complete the feature extraction work, and then being learned by the classification model. The brain can subconsciously classify the target at an extremely fast speed. Therefore, combined with the brain-computer interface technology, the present invention not only greatly improves the generalization ability of the classification model, but also can achieve rapid response of the entire system.

[0080] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution of the present invention, which should be included in the scope of the claims of the present invention.

Claims

1. A method for underwater target recognition based on brain network, characterized in that: The steps include: S1. Recruit multiple sonar operators as subjects for the experiment. During the experiment, provide stimulation signals of multiple types of underwater targets. When multiple subjects judge the underwater targets, obtain multiple sets of EEG signals corresponding to the multiple subjects. S2, preprocessing the multiple groups of EEG signals obtained in step S1; S3, based on the improved brain network extraction algorithm, the EEG signal preprocessed in step S2 is calculated and solved to obtain multiple groups of PSI adjacency matrices; S4, selecting the connection nodes of the PSI adjacency matrix as features to form a feature set, obtaining the feature set corresponding to each subject, and sorting the features in each feature set, and then using a 6-fold cross-validation method to divide the sorted feature set into a training set and a test set. The training set is used to complete the training of the classification model, and the test set is used to obtain the final classification accuracy of the subject; S5. When actually identifying underwater targets, the sonar operator selects the corresponding classification model to complete the real-time collection and classification of EEG data, thus realizing automatic identification of underwater targets; Wherein, in step S3, the following steps are specifically included: The Welch evaluation method is used to solve the cross spectrum. Given a multi-channel time series T = {t1, t2, t3, …, t N }, the solution of complex coherence is defined as: Where S is the cross spectrum; i,j represents the time series of channels i and j in T (1 <i,j<N,i≠j); The improved brain network algorithm is as follows: where α ij (f)=|C ij (f)| is the frequency-dependent weight value, and: H(Φ(f))=P(Φ(f+δf))-P(Φ(f)) P represents the new sequence obtained after polynomial fitting of the Φ(f), f∈F sequence; The improved brain network extraction algorithm is used to solve the brain network for each trial and obtain the corresponding PSI adjacency matrix.

2. The method for underwater target recognition based on brain network according to claim 1, characterized in that: In step S1, the following steps are specifically included: Several sonar operators were recruited as subjects. Before the experiment, each subject was trained to be familiar with the stimulation signals of ocean white noise and various underwater targets. The stimulation signals were video signals composed of audio and waveforms, ensuring that each subject could quickly and accurately distinguish ocean white noise from various underwater targets and form long-term memory. During the experiment, each subject wore an acoustic signal providing device and an EEG signal collecting device, and viewed the sonar waveform through a display screen. An operating keyboard was also provided. Ocean white noise and stimulation signals of various types of underwater targets were used as stimulation signals and randomly provided to the subjects. After receiving the stimulation signals, the subjects made judgments on the stimulation signals and selected their judgment results through the operating keyboard. At the same time, the EEG signal collecting device collected EEG signals. The above process was repeated multiple times for each subject to obtain multiple groups of EEG signals corresponding to multiple subjects.

3. The underwater target recognition method based on brain network according to claim 1 is characterized in that: In step S1, during the experiment, the EEG signal acquisition equipment includes 64 Ag / AgCl EEG channels, as well as a CPz reference electrode, an EOG electrode, and an AFz ground electrode.

4. The method for underwater target recognition based on brain network according to claim 1, characterized in that: In step S2, the EEG signal is preprocessed, specifically including: Data inspection to remove obvious interference fragments from the data; Bandpass filtering the data using a filter pair; The mastoid electrodes M1 and M2 were used instead of the reference electrodes; Remove artifacts from EEG signals through independent component analysis; Perform baseline correction; Event-related potential analysis was performed to determine the start and end time range of the event-related potential for subsequent data analysis.

5. The method for underwater target recognition based on brain network according to claim 4, characterized in that: In step S4, the F_score method is specifically used to sort the features. The calculation method of F_score is as follows: Among them, i represents the i-th feature in the feature set, j is the j-th category in the category, and F i is the F_score value of the i-th feature, l is the number of categories, n j is the number of samples in the jth class, k is the kth feature instance under the jth class, is the total average of the i-th feature, is the average of the i-th feature under the j-th category, F i What is actually measured is the ratio of the inter-class distance to the intra-class distance under the i-th feature; After the feature set is sorted, the data set is divided using the 6-fold cross-validation method. For each fold validation, features are added to the feature set in descending order according to the F_score to complete the training of the classification model. The classification model uses SVM, and the kernel function is a linear kernel. The divided test set is used to obtain the classification accuracy of the subject. For subjects whose classification accuracy is lower than the preset threshold, the model parameters are adjusted and re-trained until the classification accuracy reaches the preset threshold, and finally the classification model corresponding to each subject is obtained.

Images