An auditory neural feature extraction method for ship radiated noise identification

Abstract

The present application belongs to the field of signal processing technology. The present application provides an auditory nerve feature extraction method for ship radiation noise identification. The disclosed embodiment collects EEG data induced by ship radiation noise stimulation, pre-processes the EEG signals induced by different ship radiation noise stimulations to obtain EEG segmented data; performs EEG feature statistical analysis on the EEG segmented data to obtain an EEG data set; performs feature extraction and dimensionality reduction operations on the EEG data set to obtain EEG features; trains and tests the SVM model using the EEG features to obtain a trained SVM model; inputs the test data set into the trained SVM model to obtain auditory nerve features. According to the discrimination strategies adopted by different types of ships, a basis is provided for designing a more effective recognition mechanism, especially when classifying ships, a personalized recognition method can be adopted, thereby improving the adaptability and accuracy of the system.

Description

Technical Field

[0001] The embodiments of the present disclosure relate to the field of signal processing technology, and in particular to an auditory nerve feature extraction method for identifying ship radiated noise. Background Art

[0002] Ship radiated noise, generated by the mechanical operation and motion of a ship and radiated into the water, is a major source of marine environmental noise. Using ship radiated noise to identify and track different types of vessels has important applications in areas such as marine traffic safety and marine environmental monitoring. However, the ocean acoustic environment is complex and dynamic, with multiple natural and human factors interacting to influence the propagation of sound waves through the water, posing significant challenges to the identification of ship radiated noise.

[0003] Early ship radiated noise identification relied primarily on traditional signal processing techniques, analyzing the spectral and temporal characteristics of the radiated noise to identify the ship type. With the rapid development of machine learning and pattern recognition technologies, researchers have begun exploring the use of feature extraction and classification algorithms to improve the accuracy of ship radiated noise identification. In recent years, the application of deep learning technology has provided new solutions in this field. Deep learning models such as convolutional neural networks (CNNs) and recurrent neural networks (RNNs) have been widely used to improve recognition efficiency and accuracy. Researchers have also constructed a public underwater acoustic target recognition dataset, Deepship, and based on this dataset, have conducted extensive research on the identification of different ship types. Despite these achievements, existing deep learning recognition models cannot yet directly replace human operators in real-world scenarios. This is because a good deep learning recognition model requires extensive dataset training, models trained on small datasets tend to have poor generalization performance, and deep learning models suffer from poor interpretability.

[0004] In contrast, humans possess advantages over machines in certain areas. For example, they can rely on experience and intuition to make accurate decisions even with small sample sizes. Therefore, exploring human recognition mechanisms is of great research value in informing machine recognition algorithms. Electroencephalogram (EEG) signals reflect the human brain's activity during the perception, understanding, and cognition of acoustic signals. By extracting and analyzing EEG features, human cognitive abilities can be incorporated into machine detection systems to improve the performance of ship radiated noise detection. In the field of computer vision, researchers have decoded EEG signals during image detection and applied them to machine detection systems. In the field of human-computer voice interaction, a team has proposed an intent recognition method for speech interaction that integrates EEG, text, and speech features. In the field of acoustic target detection, another team has proposed a method for detecting sound targets based on EEG signals. These studies have achieved promising results.

[0005] However, there is no relevant research on the key neural response feature extraction methods for ship radiated noise type identification.

[0006] Therefore, it is necessary to improve one or more problems existing in the above-mentioned related technical solutions.

[0007] It should be noted that this section is intended to provide background or context for the technical solutions of the present disclosure stated in the claims. The description herein is not admitted to be prior art by virtue of being included in this section. Summary of the Invention

[0008] The purpose of the embodiments of the present disclosure is to provide an auditory nerve feature extraction method for ship radiated noise identification, thereby overcoming one or more problems caused by the limitations and defects of related technologies, at least to a certain extent.

[0009] According to an embodiment of the present disclosure, a method for extracting auditory nerve features for identifying ship radiated noise is provided, the method comprising:

[0010] Acquiring electroencephalogram (EEG) data induced by ship radiated noise stimulation, and preprocessing the EEG data to obtain EEG segmented data;

[0011] Performing EEG feature statistical analysis on the EEG segmented data to obtain an EEG dataset;

[0012] Performing feature extraction and dimensionality reduction operations on the EEG dataset to obtain EEG features;

[0013] Using the EEG features to train and test the SVM model to obtain the trained SVM model;

[0014] The data set to be tested is input into the trained SVM model to obtain auditory nerve features.

[0015] Furthermore, the step of preprocessing the EEG data to obtain EEG segmented data includes:

[0016] The EEG data are sequentially subjected to downsampling, low-pass filtering and high-pass filtering, mains power elimination, artifact elimination, re-reference, independent component analysis, independent component selection, dipfit source localization and segmentation operations to obtain the EEG segmented data; wherein the EEG segmented data includes S ship short group, Y ship short group, S ship long group, Y ship long group, correct short group, incorrect short group, correct long group and incorrect long group.

[0017] Furthermore, the downsampling operation is: reducing the sampling rate from 1000 Hz to 256 Hz; the low-pass filtering and high-pass filtering are 60 Hz low-pass filtering and 1 Hz high-pass filtering;

[0018] The mains frequency of the mains power elimination is 50Hz;

[0019] The segmentation operation is: according to the ship type discrimination task, to obtain the EEG segmentation data.

[0020] Furthermore, the step of performing EEG feature statistical analysis on the EEG segmented data to obtain an EEG dataset includes:

[0021] The EEG segmented data are grouped to obtain an S-boat short group-Y-boat short group, an S-boat long group-Y-boat long group, a correct short group-an incorrect short group, and a correct long group-an incorrect long group;

[0022] Pre-calculating EEG features for statistical analysis based on the S-boat-short group-Y-boat-short group, the S-boat-long group-Y-boat-long group, the correct short group-incorrect short group, and the correct long group-incorrect long group, and clustering independent components using ERSP and dipole position as clustering features;

[0023] Set statistical parameters for permutation testing and FDR correction;

[0024] Drawing parameters are set and statistics are performed to obtain the EEG dataset.

[0025] Furthermore, the EEG features include Spectrum features and ERSP features.

[0026] Furthermore, the step of performing feature extraction and dimensionality reduction operations on the EEG dataset to obtain EEG features includes:

[0027] Extracting time-frequency features of the EEG dataset;

[0028] The PCA algorithm is used to reduce the dimension of the time-frequency features to obtain the auditory nerve features, and the number of feature dimensions with principal component cumulative contribution rates of 0.5, 0.6, 0.7, 0.8, and 0.9 are selected as the final number of feature vectors.

[0029] Furthermore, the steps of training and testing the SVM model using the EEG features to obtain the trained SVM model include:

[0030] The EEG features were divided into five parts using 5-fold cross-validation, with four of the parts used as training sets and one as a test set in turn; the ratio of S-ship and Y-ship samples in the EEG features was 3:1, and stratified sampling was used when dividing the training set and the test set to ensure that the class ratio in each fold was the same as the class ratio in the original dataset;

[0031] The SVM model is trained and tested using the training set and the test set to obtain the trained SVM model; wherein, during the training process of the SVM model, the misclassification cost of the S ship and the Y ship is adjusted to 1:3.

[0032] Furthermore, the SVM model is implemented by the fitcsvm() function in MATLAB, KernelFunction is the Gaussian kernel rbf, BoxConstraint is set to 5, KernelScale is set to auto, the misclassification cost Cost is set to cost_matrix = [0 3; 1 0], and Standardize is set to true.

[0033] The technical solutions provided by the embodiments of the present disclosure may have the following beneficial effects:

[0034] In the embodiments disclosed herein, the auditory neural feature extraction method for ship radiation noise identification uses the following methods: on the one hand, EEG data induced by ship radiation noise stimulation is collected, and EEG signals induced by different ship radiation noise stimulations are preprocessed to obtain EEG segmented data; EEG feature statistical analysis is performed on the EEG segmented data to obtain an EEG dataset; feature extraction and dimensionality reduction operations are performed on the EEG dataset to obtain EEG features; an SVM model is trained and tested using the EEG features to obtain a trained SVM model; and the test dataset is input into the trained SVM model to obtain auditory neural features. On the other hand, the discrimination strategies adopted by different types of ships provide a basis for designing more effective recognition mechanisms, especially when classifying ships, which allows for personalized recognition methods to improve the adaptability and accuracy of the system. This method is used to explore the key neural representations of human auditory perception when distinguishing ship radiation noise types, in order to provide a foundation for subsequent human-machine fusion ship radiation noise type recognition and improvement of ship radiation noise recognition methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] The accompanying drawings are incorporated into and constitute a part of the specification, illustrate embodiments consistent with the present disclosure, and together with the specification, are used to explain the principles of the present disclosure. Obviously, the drawings described below are only some embodiments of the present disclosure, and those skilled in the art can derive other drawings based on these drawings without inventive effort.

[0036] Figure 1 A diagram showing the steps of an auditory nerve feature extraction method for identifying ship radiated noise in an exemplary embodiment of the present disclosure;

[0037] Figure 2Showing the types of errors in hypothesis testing in an exemplary embodiment of the present disclosure;

[0038] Figure 3 A schematic diagram illustrating the principle of an SVM linear separable model in an exemplary embodiment of the present disclosure is shown;

[0039] Figure 4 FIG. 1 shows a schematic diagram of EEG data processing grouping in an exemplary embodiment of the present disclosure;

[0040] FIG5( a ) shows the average reaction time under different ship noise stimuli in an exemplary embodiment of the present disclosure;

[0041] FIG5( b ) shows the accuracy rates under different ship noise stimuli in an exemplary embodiment of the present disclosure;

[0042] FIG5( c ) shows the average reaction time under correct decision and incorrect decision in an exemplary embodiment of the present disclosure;

[0043] Figure 6 FIG. 1 shows a short group dipole cluster diagram without adding noise in an exemplary embodiment of the present disclosure;

[0044] FIG7 (a) shows a cluster diagram showing significant differences in the spectra of the Cls4 Spectrum under the stimulation of two types of ships in the short group without adding noise in an exemplary embodiment of the present disclosure;

[0045] FIG7( b ) shows a cluster diagram of Cls7 Spectrum in an exemplary embodiment of the present disclosure, in which the spectra of the short group without noise are significantly different under the stimulation of two types of ships;

[0046] FIG7( c ) shows a cluster diagram showing significant differences in the spectra of the Cls9 Spectrum group under the stimulation of two types of ships without adding noise in an exemplary embodiment of the present disclosure;

[0047] FIG7( d ) shows a cluster diagram showing significant differences in spectra under two types of ship stimulation in the Cls10 Spectrum group without adding noise in an exemplary embodiment of the present disclosure;

[0048] FIG7( e ) shows a cluster diagram showing significant differences in spectra under two types of ship stimulation in the Cls11 Spectrum group without adding noise in an exemplary embodiment of the present disclosure;

[0049] FIG8( a ) shows a cluster diagram of the ERSPs of the short group without noise showing significant differences under the stimulation of two types of ships in Cls4 ERSP in an exemplary embodiment of the present disclosure;

[0050] FIG8( b ) shows a cluster diagram showing significant differences in ERSP under two types of ship stimulation in the no-noise short group at Cls9 ERSP in an exemplary embodiment of the present disclosure;

[0051] FIG8( c ) shows a cluster diagram showing significant differences in ERSP under two types of ship stimulation in the no-noise short group at Cls10 ERSP in an exemplary embodiment of the present disclosure;

[0052] FIG8( d ) shows a cluster diagram of the ERSP of the short group without noise when Cls11 ERSP under two types of ship stimulation in an exemplary embodiment of the present disclosure.

[0053] Figure 9 FIG. 1 shows a long group dipole cluster diagram without adding noise in an exemplary embodiment of the present disclosure;

[0054] Figure 10 It shows that the Spectrum of the two types of ships in the no-noise group have significant differences in clustering in the exemplary embodiment of the present disclosure;

[0055] FIG11( a ) shows the clustering of ERSPs with significant differences under the stimulation of two types of ships in the long group without adding noise at Cls4 ERSP in an exemplary embodiment of the present disclosure;

[0056] FIG11( b ) shows the clustering of ERSPs with significant differences under the stimulation of two types of ships in the long group without adding noise at Cls6 ERSP in an exemplary embodiment of the present disclosure;

[0057] Figure 12 FIG. 1 shows a noisy short group dipole cluster diagram in an exemplary embodiment of the present disclosure;

[0058] FIG13( a ) shows the clustering of spectra with significant differences under the stimulation of two types of ships in the short noise group when using Cls4 Spectrum in an exemplary embodiment of the present disclosure;

[0059] FIG13( b ) shows the clustering of spectra with significant differences under the stimulation of two types of ships in the Cls7 Spectrum group with added noise in an exemplary embodiment of the present disclosure;

[0060] Figure 14 FIG. 1 shows a short group dipole cluster diagram without adding noise in an exemplary embodiment of the present disclosure;

[0061] FIG15( a ) shows the clustering of the Cls3 Spectrum in the correct and incorrect decision cases without adding noise in the short group in an exemplary embodiment of the present disclosure;

[0062] FIG15( b ) shows the clustering of Cls6 Spectrum in the correct and incorrect decision cases without adding noise in the short group, where the Spectrum has significant differences in the exemplary embodiment of the present disclosure;

[0063] FIG15( c ) shows the clustering of Cls7 Spectrum in the exemplary embodiment of the present disclosure, in which the Spectrum is significantly different between the correct and incorrect decisions of the short group without adding noise;

[0064] FIG15( d ) shows the clustering of the spectra with significant differences between the correct and incorrect decisions of the short group without adding noise when the Cls10 spectra are used in an exemplary embodiment of the present disclosure;

[0065] FIG15( e ) shows the clustering of the Cls12 Spectrum in the correct and incorrect decision of the short group without adding noise in the exemplary embodiment of the present disclosure;

[0066] FIG16( a ) shows the clustering of ERSPs with significant differences between correct and incorrect decisions in the unnoised short group when using Cls7 ERSP in an exemplary embodiment of the present disclosure;

[0067] FIG16( b ) shows the clustering of ERSPs with significant differences between correct and incorrect decisions in the short group without adding noise when ERSP is Cls12 in an exemplary embodiment of the present disclosure;

[0068] Figure 17 FIG. 1 shows a long group dipole cluster diagram without adding noise in an exemplary embodiment of the present disclosure;

[0069] Figure 18 The figure shows that the Spectrum of the correct and incorrect decisions of the long group without adding noise in the exemplary embodiment of the present disclosure has significant differences;

[0070] Figure 19 The clustering of ERSP under correct and incorrect decisions without adding noise in the exemplary embodiment of the present disclosure is shown;

[0071] Figure 20 FIG. 1 shows a noisy short group dipole cluster diagram in an exemplary embodiment of the present disclosure;

[0072] Figure 21 The diagram shows clusters with significant differences in Spectrum between correct and incorrect decisions of the noisy short group in an exemplary embodiment of the present disclosure.

[0073] FIG22( a ) shows the clusters with significant differences in ERSP between correct and incorrect decisions in the noisy short group at Cls4 ERSP in an exemplary embodiment of the present disclosure;

[0074] FIG22( b ) shows the clusters with significant differences in ERSP between correct and incorrect decisions in the noisy short group at Cls7 ERSP in an exemplary embodiment of the present disclosure. DETAILED DESCRIPTION

[0075] Example embodiments will now be described more fully with reference to the accompanying drawings. However, example embodiments can be embodied in many forms and should not be construed as limited to the examples set forth herein; rather, these embodiments are provided so that this disclosure will be thorough and complete and will fully convey the concepts of the example embodiments to those skilled in the art. The described features, structures, or characteristics may be combined in any suitable manner in one or more embodiments.

[0076] In addition, the accompanying drawings are merely schematic illustrations of embodiments of the present disclosure and are not necessarily drawn to scale. Like reference numerals in the figures represent like or similar parts, and thus repeated descriptions thereof will be omitted. Some of the blocks shown in the accompanying drawings are functional entities and do not necessarily correspond to physically or logically separate entities.

[0077] This example embodiment provides an auditory neural feature extraction method for ship radiated noise identification. Figure 1 As shown in , the auditory nerve feature extraction method for ship radiated noise identification may include: steps S101 to S105.

[0078] Step S101: acquiring EEG data induced by ship radiated noise stimulation, and preprocessing the EEG data to obtain EEG segmented data;

[0079] Step S102: performing EEG feature statistical analysis on the EEG segmented data to obtain an EEG dataset;

[0080] Step S103: performing feature extraction and dimensionality reduction operations on the EEG dataset to obtain EEG features;

[0081] Step S104: using the EEG features to train and test the SVM model to obtain the trained SVM model;

[0082] Step S105: inputting the data set to be tested into the trained SVM model to obtain auditory nerve features.

[0083] The above-described auditory neural feature extraction method for ship radiated noise identification involves, on the one hand, collecting EEG data induced by ship radiated noise stimulation and preprocessing the EEG signals induced by different ship radiated noise stimulation to obtain EEG segmented data. EEG feature statistical analysis is performed on the segmented EEG data to obtain an EEG dataset. Feature extraction and dimensionality reduction are performed on the EEG dataset to obtain EEG features. A support vector machine (SVM) model is trained and tested using the EEG features to obtain a trained SVM model. The test dataset is then input into the trained SVM model to obtain auditory neural features. Furthermore, the discrimination strategies adopted by different types of ships provide a basis for designing more effective recognition mechanisms. In particular, personalized recognition methods can be used for ship classification, thereby improving the system's adaptability and accuracy. This method is used to explore the key neural representations of human auditory perception when distinguishing ship radiated noise types, hoping to provide a foundation for subsequent human-machine fusion ship radiated noise type recognition and improved ship radiated noise recognition methods.

[0084] Below, we will refer to Figures 1 to 2 2(b) provides a more detailed description of each step of the above-mentioned auditory neural feature extraction method for ship radiated noise identification in this exemplary embodiment.

[0085] In step S101, EEG data induced by ship radiation noise stimulation is acquired, and the EEG data is preprocessed to obtain EEG segmented data.

[0086] Specifically, EEG signals are very weak, typically only tens of microvolts, and are easily contaminated by irrelevant noise, resulting in various artifacts. Therefore, preprocessing and noise reduction are necessary to minimize or eliminate the impact of these artifacts and preserve the original, authentic EEG information. Specific steps include downsampling, filtering, re-referencing, segmentation, baseline correction, independent component analysis (ICA), and artifact subspace reconstruction (ASR).

[0087] In step S102 and step S103, EEG feature statistical analysis is performed on the EEG segmented data to obtain an EEG dataset; and feature extraction and dimensionality reduction operations are performed on the EEG dataset to obtain EEG features.

[0088] Specifically, a spectrum is a graph of a signal's distribution over frequency, describing the energy distribution of a signal at different frequencies. Spectral analysis transforms a signal from the time domain to the frequency domain, enabling a more intuitive analysis of the signal's frequency distribution characteristics. The foundation of spectrum analysis is the Fourier transform. The Fourier transform converts a time-domain signal into a frequency-domain signal by representing it as the sum of a series of sine and cosine functions.

[0089] In practice, EEG signals collected by EEG devices must be processed by analog-to-digital converters. Therefore, the EEG signals collected and analyzed are sampled at a series of discrete time points. Furthermore, because computers can only store and process data using a limited number of bits, the time, frequency, and amplitude domains of EEG signals are all discretized.

[0090] EEG signals are highly non-stationary, meaning their statistical properties and spectral density vary over time, generating time-varying spectral features. Spectral analysis cannot identify the time-varying spectral features of non-stationary EEG signals, leading to the neglect of important information conveyed by these features. Therefore, time-frequency analysis techniques are required to provide the distribution of signal power at each time point and frequency point, allowing observation of how the spectrum changes over time.

[0091] Common time-frequency analysis methods are based on a general sliding window analysis approach. This method assumes that non-stationary signals can be divided into a series of internally stationary short segments. That is, while the overall signal spectrum is time-varying, the spectrum of any short segment is fixed. Conventional spectral analysis methods are then performed on each short segment. Finally, the spectral estimates for all short segments are stacked together to form a spectral power distribution plot in the joint time-frequency domain.

[0092] Time-frequency analysis is commonly used to detect and analyze ERS (Event-Related Synchronization) and ERD (Event-Related Desynchronization) in EEG signals. ERS and ERD can be detected from ERSP (Event-Related Spectral Perturbation). ERSP describes the temporal changes in EEG signal power at different frequencies when an event or stimulus occurs, i.e., event-related EEG spectral changes. These changes are relative to the baseline spectrum of spontaneous EEG. An increase in spectral power in a specific frequency band is called ERS, while a decrease is called ERD.

[0093] Permutation testing is a nonparametric method used in hypothesis testing. Unlike traditional parametric tests (such as t-tests and analysis of variance), permutation testing does not rely on specific distributional assumptions about the data. It remains effective even when the sample size is small or the data do not meet the normal distribution assumption. Therefore, permutation testing is particularly ideal for statistical analysis of EEG data. The principle of permutation testing is to construct a permutation distribution of the test statistic by repeatedly randomly rearranging the data labels to generate new sample combinations and calculating the test statistic for each combination. Ultimately, by comparing the position of the original test statistic in the permuted distribution, the probability of its occurrence in this distribution, p-value, is calculated and compared with the preset significance level α to assess whether the observed differences in the original data are significant.

[0094] Significance check:

[0095] The false positive problem is affected by the significance level Control. Figure 2 The error types in hypothesis testing are shown in Figure 1. If a set of hypothesis tests are performed simultaneously, the probability of one or more false positives is called the overall type I error rate. If only one comparison is performed, the probability of a false positive is If m independent comparisons are made, the probability of at least one false positive, that is, the overall type I error rate, is When m takes a larger value, even if there is no significant difference between the samples under the two conditions, the probability of detecting one or more false positives is very high, and it cannot be guaranteed that the difference is significant at the significance level. This is the so-called multiple comparisons problem, so a multiple comparisons correction is needed to control the overall Type I error rate.

[0096] False Discovery Rate (FDR) correction is a multiple comparison correction method that attempts to balance false positives and false negatives, keeping the ratio of false positives to true positives within a certain range, thus providing strong statistical power. The FDR is defined as the expected value of the ratio of the number of false positives to the number of rejected null hypotheses. FDR correction corrects for errors caused by multiple comparisons by controlling the FDR. When rejecting multiple null hypotheses (H0), the FDR correction method can control the probability of false positives and find an appropriate combination of results.

[0097] Feature extraction

[0098] EEG signal features are primarily classified into three categories: time domain features, frequency domain features, and time-frequency domain features. Time domain features reflect EEG signal information in the time domain, frequency domain features reflect EEG signal information in the frequency domain, and time-frequency domain features are a combination of time and frequency domain features. EEG signals are typically divided into the delta band (1-4Hz), theta band (4-8Hz), alpha band (8-13Hz), beta band (13-30Hz), and gamma band (30-50Hz) based on frequency. This study extracted frequency domain features and time-frequency domain features of these five frequency bands for subsequent learning and classification.

[0099] Power spectral density (PSD) is a commonly used frequency domain feature used to describe the distribution of the power of a random signal along the frequency. It is the result of square normalization of the spectrum.

[0100] PCA Dimensionality Reduction

[0101] Extracted EEG features typically have extremely high data dimensions, but the number of samples is very limited. Directly feeding these features into machine learning algorithms for training often leads to overfitting and reduces training and prediction speeds. Dimensionality reduction algorithms can reduce high-dimensional features to lower dimensions while retaining most of the original data information, reducing data complexity and computational costs. This, in turn, reduces the risk of overfitting in machine learning and accelerates training and prediction speeds.

[0102] Principal Component Analysis (PCA) is a mainstream, simple, unsupervised linear dimensionality reduction algorithm. It aims to minimize reconstruction error or maximize scatter. It uses an orthogonal transformation to linearly transform the observed values ​​of a series of potentially correlated variables, projecting them into a series of linearly uncorrelated variables, called principal components.

[0103] Assume there are m n-dimensional data, , where each x is an n-dimensional column vector. The main calculation process of the PCA algorithm is as follows:

[0104] (1) Decentralize the original data:

[0105] (1)

[0106] (2) Calculate the covariance matrix:

[0107] (2)

[0108] (3) Perform eigenvalue decomposition on the covariance matrix to obtain the characteristic matrix (arranged from large to small according to the eigenvalue), and take the first k columns to form the matrix , P is equivalent to a new coordinate system, each column of which is a coordinate axis.

[0109] (4) Projecting the original data into the P coordinate system yields the reduced-dimensional data:

[0110] (3)

[0111] In step S104 and step S105, the SVM model is trained and tested using the EEG features to obtain the trained SVM model; the data set to be tested is input into the trained SVM model to obtain auditory nerve features.

[0112] Specifically, support vector machines (SVMs) and their related learning algorithms and recognition patterns are commonly used as supervised learning models in classification and regression tasks. Their basic principle is to find an optimal hyperplane to partition the data space, separating samples of different categories by as large a margin as possible. This hyperplane is determined by the nearest data points, which are called support vectors.

[0113] The most basic support vector machine is a non-probabilistic binary linear classifier. Given a set of linearly separable training instances, each labeled as belonging to one or the other of two categories, the SVM training algorithm builds a model that assigns new instances to one of the two categories. The SVM model represents instances as points in space, mapping them so that instances of the individual categories are separated by as wide a clear gap as possible. New instances are then mapped into the same space and their category is predicted based on which side of the gap they fall. Figure 3 This is a schematic diagram of the SVM classifier when it is linearly separable.

[0114] In practical applications, most problems are linearly inseparable. In such cases, a linear classifier alone cannot achieve correct classification. When faced with such nonlinear classification problems, Support Vector Machines (SVMs) use a nonlinear kernel function to map the original low-dimensional nonlinear classification problem into a higher-dimensional space, thereby transforming it into a linear classification problem.

[0115] In a specific embodiment, the present application proposes a method for constructing a neural representation model of the human brain's auditory perception of ship radiated noise. The specific process is as follows: Figure 1 The collected data set is shown in Table 1. This application uses two methods to group and analyze EEG data. The specific methods and groups are as follows: Figure 4 shown.

[0116] Table 1 Data composition of each subject

[0117]

[0118] like Figure 1 As shown, the technical solution of this application mainly involves three aspects: EEG data preprocessing, EEG feature statistical analysis and machine learning classification.

[0119] First, the collected EEG data induced by ship radiated noise stimulation need to be preprocessed, including: reducing the sampling rate from 1000Hz to 256Hz, 60Hz low-pass filtering and 1Hz high-pass filtering, eliminating the mains (mains frequency is 50Hz), ASR, re-reference (averaging), independent component analysis, dipfit source localization and segmentation according to grouping criteria (grouping such as Figure 4 As shown, the duration of the short group without noise is 5s, the duration of the long group without noise is 25s, the duration of the short group with noise is 8s, and the duration of the long group with noise is 35s).

[0120] Statistical analysis was then performed on the preprocessed EEG segmented data. The specific steps were: selecting and importing paired EEG segmented data using the same grouping criteria (S-ship short group vs. Y-ship short group, S-ship long group vs. Y-ship long group, correct short group vs. incorrect short group, correct long group vs. incorrect long group), precalculating the EEG features (spectrum and ERSP) for statistical analysis, clustering independent components using ERSP and dipole position as clustering features (using the k-means method, with the number of clusters set to 10 for the unnoised group and 6 for the noisy group), setting statistical parameters (permutation test and FDR correction) and plotting parameters (time domain range = grouping duration, frequency band range = 1-60 Hz), and obtaining statistical results. The ships were divided into S-ship and Y-ship based on their differences in radiated noise.

[0121] In addition, behavioral analysis was performed on the raw data. Based on the results of these behavioral and EEG feature statistical analyses, as well as considering sample size constraints, we selected EEG data from ship type discrimination without ambient noise for the machine learning dataset. The behavioral analysis revealed that the short-term data for the three subjects had higher confidence levels, and the EEG feature statistical analysis also showed more significant differences in the short-term data. Therefore, we segmented the preprocessed EEG signals into a data segment spanning [-1s, 5s] relative to the onset of the target ship noise stimulus. Data analysis revealed that only the first 122 channels of EEG data were valid. Because the 128-channel data already contains significant overlap and redundancy, removing six of these channels did not significantly impact the final machine learning results.

[0122] This application adopts two methods to analyze the EEG data of the classification task. One is to divide the EEG data into S ship noise stimulation group and Y ship noise stimulation group according to the ship type to which the current ship radiation noise stimulus belongs. The other is to divide the EEG data into correct decision group and incorrect decision group according to whether the subject's decision is correct.

[0123] The first classification method further divides each data set into the S-boat short reaction time group, the Y-boat short reaction time group, the S-boat long reaction time group, and the Y-boat long reaction time group based on the subject's key reaction time. The average reaction time and accuracy rate for each data group are calculated and the results are plotted as a bar graph. Then, a paired-sample t-test was performed on the average reaction time and accuracy rate under S-ship noise stimulation and Y-ship noise stimulation in the short and long groups, respectively. (This was performed in Excel. For the paired-sample t-test in the short group, because the sample sizes of S-ship and Y-ship were different, the sample with the smaller sample size was used as the benchmark to sample the sample with the larger sample size. For example, the sample size of the S-ship short group was 147, and the sample size of the Y-ship short group was 32. In this case, when performing the paired-sample t-test, all the samples of the Y-ship short group were included. Then, 32 samples were extracted from the S-ship short group using the sampling method in Excel's Data-Data Analysis. Then, a paired-sample t-test was performed using the "t-Test: Paired Two-Sample Analysis of Means" method in Excel's Data-Data Analysis. The paired-sample t-test was performed in the long group using the same operation.) The significance level α was set to 0.05, and the p-value was calculated. α and p-value were compared. If p < α, it means that there is a significant difference in the average reaction time and accuracy rate under S-ship noise stimulation and Y-ship noise stimulation in the short and long groups.

[0124] The second classification method further divides each data set into a correct decision short group, an incorrect decision short group, a correct decision long group, and an incorrect decision long group based on the subject's key-pressing reaction time. The average reaction time for each data group is calculated and the results are plotted as a bar graph. A paired-sample t-test (same test steps as above) is then performed on the average reaction times for correct and incorrect decisions within the short and long groups, setting a significance level of α = 0.05. The resulting p-value is then compared with the α and p-values. A p < α indicates a significant difference in the average reaction times for correct and incorrect decisions within the short and long groups.

[0125] The goal of machine learning is to distinguish ship types based on EEG features. The EEG data for distinguishing ship types without environmental noise not only has a considerable sample size, but also is selected to construct a machine learning dataset based on the results of behavioral analysis and statistical analysis of EEG features.

[0126] After processing, we finally obtained 681 EEG data samples under S-ship noise stimulation and 223 EEG data samples under Y-ship noise stimulation. The size of each data sample is [122, 1536], where 122 is the number of effective channels and 1536 is the sampling point, which is equal to the product of the sampling rate (256 Hz) and the time length of the sample data segment (6 s).

[0127] For all samples, the Welch method and short-time Fourier transform are used to extract frequency domain features (PSD) and time-frequency features, respectively. The frequency domain information obtained from each channel is a one-dimensional vector containing all frequency domain feature points, while the time-frequency domain information is a matrix with F rows and T columns (F represents the frequency point and T represents the time point). The frequency domain information (size [F, 122]) and time-frequency information (size [F, T, 122]) of the 122 channels are then rearranged into a one-dimensional vector, which serves as the original frequency domain feature vector and time-frequency domain feature vector for the sample.

[0128] Specifically, the Welch method is used to extract the frequency domain features (PSD). This method is implemented using the pwelch() function in MATLAB. A Hamming window is used as the window function, with a window size of 256 (1 second) and a window step size of 128. The Welch method is applied to each channel of each trial to obtain the spectral density (pxx), a one-dimensional vector of size [F, 1], where F is the number of frequency bins. All pxx are then stored at the corresponding frequency positions of the corresponding channels in a three-dimensional matrix of shape [F, channels, trials], obtaining the frequency domain features for all trials and channels. The reshape method in MATLAB is then used to convert the three-dimensional matrix into a two-dimensional matrix of shape [F*channels, trials]. This flattens the frequency domain features of all channels for each trial into a one-dimensional vector. Finally, the zscore method is used to normalize the eigenvectors of the two-dimensional matrix, and PCA is then used to calculate the principal components of the normalized data.

[0129] STFT is implemented using the Spectrogram() function in MATLAB, using a Hamming window as the window function, with the window size set to 256 (1 second) and the window moving step set to 10.

[0130] The above-mentioned Welch method and short-time Fourier transform are used to extract the frequency domain features PSD and time-frequency features of the five frequency bands, and after PCA dimensionality reduction, they are saved as .mat files.

[0131] The statistical analysis results of EEG features show that the delta frequency band (1-4 Hz), theta frequency band (4-8 Hz), alpha frequency band (8-13 Hz), beta frequency band (13-30 Hz) and gamma frequency band (30-50 Hz) all contribute to the discrimination of ship types, so features are extracted for these frequency bands respectively.

[0132] Finally, the z-score method was used to normalize the eigenvectors of the two-dimensional matrix. The principal components of the normalized data were then calculated using the PCA method. The eigenvectors were then dimensionality reduced using the PCA algorithm (implemented using the MATLAB pca() function). The final number of eigenvectors was selected as the number of feature dimensions with cumulative principal component contributions of 0.5, 0.6, 0.7, 0.8, and 0.9.

[0133] A support vector machine (SVM) was selected as the sample classification algorithm. The previously extracted EEG features were used as input to train a ship type discrimination model. The model's performance was evaluated using a 5-fold cross-validation method. The 5-fold cross-validation method split the dataset into 5 parts, with 4 parts used as training data and 1 part used as test data for training and testing. The distribution of S- and Y-ship classes in the dataset was uneven (the ratio of S- and Y-ship samples was 3:1). To ensure stability and representativeness of the cross-validation, stratified sampling was used when dividing the training and test sets to ensure that the class ratio in each fold was consistent with that in the original dataset. Furthermore, during SVM training, the misclassification cost for S- and Y-ships was adjusted to 1:3 to prevent bias in favoring one class during training. The SVM was implemented using the fitcsvm() function in MATLAB. The Gaussian kernel rbf was selected for KernelFunction, BoxConstraint was set to 5, KernelScale was set to auto, the misclassification cost was set to cost_matrix = [0 3; 1 0], and Standardize was set to true.

[0134] Specifically, in the SVM stage, based on the PCA method's return value (score: data after PCA dimensionality reduction; latent: variance of each principal component, indicating the importance of each principal component), feature dimensions with cumulative principal component contributions (calculated based on the latent variance) of 0.5, 0.6, 0.7, 0.8, and 0.9 were selected as the final number of feature vectors. The data was then randomly shuffled and saved. These processed EEG features were then used as input to train an SVM model for ship type discrimination, and model performance was evaluated using 5-fold cross-validation. This 5-fold cross-validation method split the dataset into five parts, alternating between using four parts as training data and one part as testing data for training and testing. The distribution of S- and Y-ship classes in the dataset was uneven (the ratio of S- and Y-ship samples was 3:1). To ensure stability and representativeness of the cross-validation, stratified sampling was used when dividing the training and test sets to ensure that the class ratio in each fold was consistent with that in the original dataset. During SVM training, the misclassification cost for S and Y ships was adjusted to 1:3 to prevent the model from favoring one class over the other during training. The SVM was implemented using the fitcsvm() function in MATLAB. The Gaussian kernel rbf was selected for KernelFunction, BoxConstraint was set to 5, KernelScale was set to auto, the misclassification cost Cost was set to cost_matrix = [0 3; 1 0], and Standardize was set to true.

[0135] In a specific embodiment, 1. Behavioral data analysis:

[0136] Table 2 t-test results of behavioral data

[0137] (* indicates p < 0.05, and the numbers in the first row correspond to the group numbers in Figure 5)

[0138]

[0139] Figures 5(a) through 5(c), along with Table 2, show the results of behavioral data statistical analysis. Figures 5(a) through 5(c) show the behavioral data analysis results for the different groups of subjects (* indicates groups with significant differences; see Table 2 for details). Figure 5(a) shows the average reaction time under different ship noise stimuli; Figure 5(b) shows the accuracy rate under different ship noise stimuli; and Figure 5(c) shows the average reaction time for correct and incorrect decisions.

[0140] The results showed that, behaviorally, some samples of the same type of ship required longer reaction times, while others had shorter reaction times, indicating significant intra-sample variability. Comparing the average reaction times and accuracy rates for different ship types revealed that, regardless of whether or not ocean noise was added, shorter reaction times for S-ships correlated with higher accuracy rates, while longer reaction times for Y-ships correlated with higher accuracy rates. This suggests that different discrimination strategies are required for S-ships and Y-ships. Comparing the behavioral responses of the two types of ships also revealed that Y-ships had a lower accuracy rate, were more challenging to discriminate, and required a longer time to make a comprehensive decision based on stimulus information and prior knowledge and experience.

[0141] 2. Statistical Analysis of EEG Characteristics

[0142] 2.1 EEG characteristics under different ship noise stimulations

[0143] Here are the results of a statistical analysis of EEG characteristics induced by ship noise stimulation. By comparing the statistical characteristics of EEG signals induced by S-ship noise stimulation and Y-ship noise stimulation, the differences in brain responses to different ship noise stimuli were investigated. The EEG data under different ship noise stimuli were divided into a short group without noise, a long group without noise, a short group with noise, and a long group with noise. Within each group, the EEG characteristics under S-ship noise stimulation and Y-ship noise stimulation were statistically analyzed to determine the dipole clustering locations, spectra, and ERSPs of different clusters within each group. The results showed that the spectra and ERSPs of the EEG signals induced by the two types of ship stimuli in the short group without noise, the long group without noise, and the short group with noise showed varying degrees of difference, while the spectra and ERSPs in the long group with noise did not differ significantly.

[0144] (1) Short group without noise

[0145] Table 3 shows the brain region information of the short group clustering without noise.

[0146] like Figure 6 As shown, it is the short group dipole clustering diagram without noise.

[0147] As shown in Figures 7(a), 7(b), 7(c), 7(d) and 7(e), there are significant differences in the Spectrum of the two types of ship stimulation in the short group without noise.

[0148] As shown in Figures 8(a), 8(b), 8(c), and 8(d), there are clusters with significant differences in ERSP under the stimulation of two types of ships in the short group without noise.

[0149] Table 3 Brain region information of short group clustering without noise

[0150] (* indicates brain regions with significant differences in Spectrum, # indicates brain regions with significant differences in ERSP)

[0151]

[0152] (2) No noise long group

[0153] Table 4 shows the brain region information of the long group clustering without noise.

[0154] like Figure 9 As shown, this is the long group dipole clustering diagram without noise.

[0155] like Figure 10 As shown in the figure, there are significant clusters of Spectra under the stimulation of two types of ships in the no-noise group.

[0156] As shown in Figure 11 (a) and Figure 11 (b), there are significant differences in the ERSP clusters under the stimulation of the two types of ships in the no-noise group.

[0157] Table 4 Brain region information of clustered long group without noise

[0158] (* indicates brain regions with significant differences in Spectrum, # indicates brain regions with significant differences in ERSP)

[0159]

[0160] (3) Noise-added short group

[0161] Table 5 shows the brain region information of the short group clustering with noise.

[0162] like Figure 12 As shown, it is a noisy short group dipole cluster diagram.

[0163] As shown in Figure 13 (a) and Figure 13 (b), the Spectrum of the noise-added short group under the stimulation of two types of ships has significant differences in clustering.

[0164] Table 5 Brain region information of the short noise group clustering

[0165] (* indicates brain regions with significant differences in Spectrum)

[0166]

[0167] Figure 6-Figure 1Results from Figure 3(b) and Tables 3-5 show that, for the ship type discrimination task, the spectral and time-frequency characteristics of the occipital, frontal, parietal, and temporal lobes in the short-noise group showed significant differences in both spectral and temporal characteristics across the two ship types. These differences were observed within 1000ms of stimulus onset, with the most pronounced EEG features in the theta (4-8Hz), alpha (8-13Hz), and beta (13-30Hz) bands. In contrast, the long-noise group showed significant differences only in the occipital and frontal EEG features, with more pronounced time-frequency characteristics. In the short-noise group, only minor spectral differences were observed in the occipital, parietal, and temporal lobes. Combining the results from the short-noise and long-noise groups, we find that the occipital and frontal lobes are highly stable and important in ship type discrimination. We hypothesize that the frontal lobe is primarily responsible for higher-level cognitive functions such as attention control and decision-making in this task, while the occipital lobe may be involved in associating and remembering ship-radiated noise. Results from both the no-noise and noise-added short-duration groups indicate that the addition of ocean noise increases the difficulty of auditory perception and cognitive discrimination, significantly disrupting the differences in EEG signatures. However, significant differences in spectral characteristics were still observed in the occipital, parietal, and temporal lobes, suggesting that these three brain regions play a key role in responding to noise interference. Across all conditions, the occipital lobe exhibited significant EEG differences, suggesting its crucial role in distinguishing ship types.

[0168] 2.2 EEG characteristics of correct and incorrect decisions

[0169] Here, we compared the EEG signal statistical characteristics when participants made correct and incorrect decisions, hoping to uncover differences in brain activity during these decisions. The EEG data were again divided into the unnoised short group, the unnoised long group, the noised short group, and the noised long group. The EEG signal spectra and ERSPs for the unnoised short group, the unnoised long group, and the noised short group showed varying degrees of difference under different decision conditions.

[0170] (1) Short group without noise

[0171] like Figure 14 As shown, it is the short group dipole clustering diagram without noise.

[0172] Table 6 shows the brain region information of the short group clustering without noise.

[0173] Table 6 Brain region information of short group clustering without noise

[0174] (* indicates brain regions with significant differences in Spectrum, # indicates brain regions with significant differences in ERSP)

[0175]

[0176] As shown in Figures 15(a), 15(b), 15(c), 15(d), and 15(e), the Spectrum clusters under correct and incorrect decisions of the unnoised short group have significant differences.

[0177] As shown in Figure 16 (a) and Figure 16 (b), there are clusters with significant differences in ERSP under correct and incorrect decisions in the short group without noise.

[0178] (2) No noise long group

[0179] like Figure 17 As shown, this is the long group dipole clustering diagram without noise.

[0180] Table 7 shows the brain region information of the long group clustering without noise.

[0181] Table 7 Brain region information of clustering of the long group without noise

[0182] (* indicates brain regions with significant differences in Spectrum, # indicates brain regions with significant differences in ERSP)

[0183]

[0184] like Figure 18 As shown in Figure 3, there are significant differences in the Spectrum of the correct and incorrect decisions of the long group without noise.

[0185] like Figure 19 As shown in Figure 3, there are significant differences in ERSP between correct and incorrect decisions in the noisy group.

[0186] (3) Noise-added short group

[0187] like Figure 20 As shown, it is a noisy short group dipole cluster diagram.

[0188] Table 8 shows the brain region information of the short group clustering with noise.

[0189] Table 8 Brain region information of the short noise group clustering

[0190] (* indicates brain regions with significant differences in Spectrum, # indicates brain regions with significant differences in ERSP)

[0191]

[0192] like Figure 21 As shown in Figure 3, there are significant differences in the Spectrum of the correct and incorrect decisions in the noisy short group.

[0193] As shown in Figure 22 (a) and Figure 22 (b), there are clusters with significant differences in ERSP under correct and incorrect decisions in the noisy short group.

[0194] Figure 14-Figure 2The results in Figure 2(b) and Tables 6-8 show that the frontal lobe plays a key role in the decision-making process in the ship type discrimination task. The frontal lobe has high-level cognitive functions that regulate decision-making. Therefore, in the ship type discrimination task, the activity of the frontal lobe may contribute to more accurate judgment and classification.

[0195] 3. Machine Learning Classification Results

[0196] Table 9 SVM classification results of frequency domain features

[0197] (The best average classification results under the current contribution rate are marked in bold)

[0198]

[0199] Table 10 SVM classification results of time-frequency domain features

[0200] (The best average classification results under the current contribution rate are marked in bold)

[0201]

[0202] Tables 9 and 10 show the SVM classification accuracy for frequency-domain features and time-frequency-domain features at different principal component cumulative contribution rates for different frequency bands. Analysis of the machine learning classification results shows that, among frequency-domain features, the beta band performs better, while among time-frequency features, the delta, theta, and beta bands perform better. This is consistent with the statistical analysis results of EEG features for the ship type discrimination task. Furthermore, the time-frequency features perform better than the frequency-domain features. All results are significantly above the chance level (50%), demonstrating the effectiveness of the proposed EEG feature extraction method for ship type discrimination, as well as the feasibility and application potential of combining EEG features for ship radiated noise identification.

[0203] The above details the principles, processes, and results of constructing the neural representation model of the human brain's auditory perception of ship radiated noise proposed in this application. It can be seen that this application has the following beneficial effects:

[0204] (1) Behavioral analysis found that different types of ships require different discrimination strategies, which can provide a basis for designing more effective identification mechanisms, especially when classifying ships, personalized identification methods can be used to improve the adaptability and accuracy of the system.

[0205] (2) The EEG characteristics of the theta (4-8 Hz), alpha (8-13 Hz), and beta (13-30 Hz) frequency bands are particularly different and can be quickly detected within 1000 ms after stimulation, indicating that these low-frequency bands play a key role in ship type discrimination. This provides theoretical support for the development of a ship noise recognition model based on these specific frequency bands.

[0206] (3) The frontal lobe plays an important role in the decision-making process of the ship type discrimination task. This finding provides a direction for further research on the function of this brain region and how to enhance its performance in the corresponding task.

[0207] (4) The results of machine learning analysis show that the classification of ship types based on EEG features is excellent in the delta, theta, and beta frequency bands. In particular, the classification effect of time-frequency features is significantly better than that of frequency domain features. This demonstrates the discriminative power of EEG signals in complex tasks and their practical application potential in ship noise recognition.

[0208] The above-described auditory neural feature extraction method for ship radiated noise identification involves, on the one hand, collecting EEG data induced by ship radiated noise stimulation and preprocessing the EEG signals induced by different ship radiated noise stimulation to obtain EEG segmented data. EEG feature statistical analysis is performed on the segmented EEG data to obtain an EEG dataset. Feature extraction and dimensionality reduction are performed on the EEG dataset to obtain EEG features. A support vector machine (SVM) model is trained and tested using the EEG features to obtain a trained SVM model. The test dataset is then input into the trained SVM model to obtain auditory neural features. Furthermore, the discrimination strategies adopted by different types of ships provide a basis for designing more effective recognition mechanisms. In particular, personalized recognition methods can be used for ship classification, thereby improving the system's adaptability and accuracy. This method is used to explore the key neural representations of human auditory perception when distinguishing ship radiated noise types, hoping to provide a foundation for subsequent human-machine fusion ship radiated noise type recognition and improved ship radiated noise recognition methods.

[0209] In the description of this specification, the reference terms "one embodiment", "some embodiments", "example", "specific example" or "some examples" mean that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present disclosure. In this specification, the schematic expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art can combine and combine different embodiments or examples described in this specification.

[0210] Those skilled in the art will readily appreciate other embodiments of the present disclosure after considering the specification and practicing the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of the present disclosure that follow the general principles of the present disclosure and include common knowledge or customary techniques in the art not disclosed herein. The description and examples are to be considered as exemplary only, with the true scope and spirit of the present disclosure being indicated by the appended claims.

Claims

1. An auditory nerve feature extraction method for ship radiated noise identification, characterized in that: The method includes: Acquiring electroencephalogram (EEG) data induced by ship radiation noise stimulation, and preprocessing the EEG data to obtain EEG segmented data; preprocessing the EEG data specifically includes: sequentially performing downsampling, low-pass filtering and high-pass filtering, mains power removal, artifact removal, re-referencing, independent component analysis, independent component selection, dipfit source localization, and segmentation operations on the EEG data to obtain the EEG segmented data; wherein the ship radiation noise includes S ship noise and Y ship noise, and the EEG segmented data includes an S ship short group, a Y ship short group, an S ship long group, a Y ship long group, a correct short group, an incorrect short group, a correct long group, and an incorrect long group; The EEG segmented data are grouped to obtain an S-short group-Y-short group, an S-long group-Y-long group, a correct short group-incorrect short group, and a correct long group-incorrect long group; based on the S-short group-Y-short group, the S-long group-Y-long group, the correct short group-incorrect short group, and the correct long group-incorrect long group, EEG features for statistical analysis are pre-calculated, and independent components are clustered using ERSP and dipole position as clustering features; statistical parameters are set to perform a permutation test and FDR correction; drawing parameters are set, and statistics are performed to obtain statistical results; Performing a paired sample t-test on the average reaction time and accuracy rate under the stimulation of the S-ship noise and the Y-ship noise in the S-ship short group, the Y-ship short group, the S-ship long group, and the Y-ship long group, respectively, and combining the significance level values ​​to obtain a first analysis result; Performing a paired sample t-test on the average reaction times under correct decisions and incorrect decisions in the correct short group, the incorrect short group, the correct long group, and the incorrect long group, respectively, and combining the significance level values ​​to obtain a second analysis result; Constructing a machine learning data set based on the statistical results, the first analysis results, and the second analysis results, and performing feature extraction and dimensionality reduction operations on the machine learning data set to obtain EEG features; wherein the EEG features include Spectrum features and ERSP features; The SVM model is trained and tested using the EEG features to obtain the trained SVM model; wherein the SVM model is implemented using the fitcsvm() function in MATLAB, KernelFunction is a Gaussian kernel rbf, BoxConstraint is set to 5, KernelScale is set to auto, the misclassification cost Cost is set to cost_matrix = [0 3; 1 0], and Standardize is set to true; The data set to be tested is input into the trained SVM model to obtain auditory nerve features.

2. The auditory nerve feature extraction method for ship radiated noise identification according to claim 1 is characterized in that: The downsampling operation is to reduce the sampling rate from 1000 Hz to 256 Hz; the low-pass filtering and high-pass filtering are 60 Hz low-pass filtering and 1 Hz high-pass filtering; The mains frequency of the mains power elimination is 50Hz; The segmentation operation is: according to the ship type discrimination task, to obtain the EEG segmentation data.

3. The auditory nerve feature extraction method for ship radiated noise identification according to claim 2 is characterized in that: The steps of extracting features and reducing dimensionality of machine learning datasets to obtain EEG features include: Extracting time-frequency features of the machine learning dataset; The PCA algorithm is used to reduce the dimension of the time-frequency features to obtain the auditory nerve features, and the number of feature dimensions with principal component cumulative contribution rates of 0.5, 0.6, 0.7, 0.8, and 0.9 are selected as the final number of feature vectors.

4. The auditory nerve feature extraction method for ship radiated noise identification according to claim 3 is characterized in that: The steps of training and testing the SVM model using the EEG features to obtain the trained SVM model include: The EEG features were divided into five parts using 5-fold cross-validation, with four of the parts used as training sets and one as a test set in turn; the ratio of S-ship and Y-ship samples in the EEG features was 3:1, and stratified sampling was used when dividing the training set and the test set to ensure that the class ratio in each fold was the same as the class ratio in the original dataset; The SVM model is trained and tested using the training set and the test set to obtain the trained SVM model; wherein, during the training process of the SVM model, the misclassification cost of the S ship and the Y ship is adjusted to 1:3.

Images