Multi-dimensional feature selection method and system for underwater target recognition

Abstract

The invention provides a multi-dimensional feature selection method and system for underwater target recognition, and the method comprises the steps: carrying out the normalization of a feature vector, and obtaining a feature space; filtering multiple collinear features based on a variance expansion factor; the working characteristic curve of the subject is used for retaining the characteristics with high classification capability; calculating feature importance scores to obtain a feature importance sequence; and adding features into the feature space according to the feature importance sequence to obtain an optimal feature subset. According to the invention, by filtering out multiple collinear features and underwater target features with poor classification capability on scale, the redundancy of the features is reduced, and the distinction degree of the features is improved; the defects of a filtering method are made up through feature sorting, and compared with a wrapping method, the time complexity is low, and compared with an embedding method, the interpretability is better; through a refined feature selection method, the feature dimension is reduced, the operation efficiency and the generalization performance of the model are improved, and the path is simple and efficient.

Description

Technical Field

[0001] The present invention belongs to the technical field of underwater acoustic signal processing, and in particular, relates to a multi-dimensional feature selection method and system for underwater target recognition. Background Art

[0002] With the rapid development of artificial intelligence, the automatic recognition capability of anti-frogman sonar has attracted more and more attention and has become one of the main indicators for measuring sonar performance. The underwater environment is complex, especially in a high signal-to-noise ratio environment. The detection range will detect more moving objects and input them to the recognition end. Some false alarm targets, such as fish schools, have some features similar to frogmen or UUVs at certain times and are easily identified as threat targets, resulting in frequent false alarms. Therefore, finding features with strong target discrimination ability is the key to accurately identifying threat targets.

[0003] The main features of underwater targets include signal echo features and motion behavior features. Through time domain, time-frequency domain, pattern decomposition and other means, physical quantities or statistics that characterize these features can be extracted. The recognition accuracy of underwater targets can be improved by fusing features from multiple information sources. However, some redundant or irrelevant features can cause overfitting of the training model and increase the computational cost. Therefore, feature selection can select the most discriminative features from the original feature set, reduce the feature dimension, and improve the prediction accuracy and generalization ability of the model.

[0004] Common feature selection methods can usually be summarized into three categories: filtering, wrapping, and embedding. Among them, the filtering method mainly calculates statistical properties such as variance and frequency to determine whether to filter out the feature. This method relies on data characteristics and is not closely integrated with the classifier. The wrapping method uses the performance of the classifier to be used as the evaluation criterion for the feature subset to select the best feature subset, such as recursive feature elimination and forward selection. This method has high computational complexity. The embedding method embeds feature selection into the training of the classifier, such as LASSO regression and random forest. Although some embedding methods have superior performance, the model itself is relatively complex and the interpretability of the selected features is poor.

[0005] The patent document "Feature Selection Method, Device, Electronic Device and Computer Readable Storage Medium" (CN114912628A) discloses that by grouping the features to be selected according to the reference column and then randomly shuffling the real labels of the features to be selected within the group, overfitting features can be screened out more quickly and effectively to improve the accuracy of the model. The focus is on reducing the false importance features caused by feature noise, and the variance inflation factor and receiver operating characteristic curve used in the present invention filter out multicollinear features and underwater target features with poor classification ability on scale, reducing feature redundancy and tending to describe how to select important features from a large number of features.

[0006] Taking into account the shortcomings of existing methods, in order to select features that can stably express the inherent mechanism of the target and are easy to train and interpret, and to solve the problems of high false alarm frequency and low recognition accuracy of existing anti-frogman sonar, the present invention proposes a multidimensional feature selection method for underwater target recognition. Summary of the invention

[0007] In view of the defects in the prior art, the object of the present invention is to provide a multi-dimensional feature selection method and system for underwater target recognition.

[0008] The multi-dimensional feature selection method for underwater target recognition provided by the present invention comprises:

[0009] Step S1: normalize the feature vector to obtain the feature space;

[0010] Step S2: filter out multiple collinear features and retain features with strong classification capabilities;

[0011] Step S3: Calculate the feature importance score and obtain the feature importance ranking;

[0012] Step S4: Add features to the feature space according to the order of feature importance to obtain the best feature subset.

[0013] Preferably, the targets include three types: UUV, frogman and other types.

[0014] The features are signal echo features and motion behavior features, including tracking scale calculation, azimuth change, azimuth fluctuation, curvature, curvature fluctuation, speed, speed change, speed fluctuation, trajectory fluctuation, trajectory linearity, aspect ratio, area, rectangularity, center of mass ratio, peak-to-peak width ratio, first distribution probability and second distribution probability.

[0015] In the step S1, the signal echo and the motion behavior feature vector extracted from the underwater target are normalized in sequence to obtain a normalized feature space.

[0016] The filtering of multiple collinear features includes establishing a multivariate linear regression model for each underwater target feature, solving the regression coefficient, and calculating the corresponding variance inflation factor.

[0017] The retaining of features with strong classification capabilities includes defining multiple thresholds along the feature scale of the underwater target, calibrating the true positive rate and the false positive rate under different thresholds, and screening the classification capability of each underwater target feature.

[0018] In step S3, the F value between each feature and the category is calculated in turn, and the importance ranking of the underwater target features is established according to the correlation between the features and the categories.

[0019] In step S4, an underwater target classification model is constructed, and the underwater target features are sorted according to their importance, and the features are added to the feature space in sequence. The number of features with the best model performance is retained according to the model training results to obtain the best underwater target feature subset.

[0020] Preferably, the normalization calculation is

[0021] Among them, X represents the selected eigenvalue;

[0022] X max Indicates the maximum value corresponding to the selected eigenvalue;

[0023] X min Indicates the minimum value corresponding to the selected eigenvalue;

[0024] X * Represents the normalized result.

[0025] In the multilayer perceptron model in step S4, 5 hidden layers are used, each hidden layer contains 100 neurons, a RELU activation function is used, batch_size is set to 64, an Adam optimization algorithm is used, a learning rate is set to 0.001, and it is iterated 100 times.

[0026] According to the model training results, when the accuracy reaches 100%, the features with the best model performance are retained to obtain the best feature subset.

[0027] Preferably, filtering multiple collinear features comprises:

[0028] For each underwater target feature X i , and use them as dependent variables in turn, and the remaining features as independent variables to establish a multiple linear regression model X1 = α1 + α2X2 + ... + α N X N +ε;

[0029] Among them, X i represents the feature of the i-th underwater target;

[0030] N indicates that there are N underwater target features, i∈N;

[0031] a i represents the i-th regression coefficient;

[0032] ε represents the error term.

[0033] Use the least squares method to solve the regression coefficient and calculate the coefficient of determination in turn Calculate the variance inflation factor for each underwater feature

[0034] in, represents the predicted value of the jth underwater target sample of the i-th feature;

[0035] x i,j represents the true value of the jth sample of the i-th underwater target feature;

[0036] represents the average value of the feature of the i-th underwater target;

[0037] VIF i represents the variance expansion factor corresponding to the i-th underwater target feature;

[0038] n represents the sample size.

[0039] The features that retain strong classification capabilities include:

[0040] The targets were divided into three categories according to the two-by-two categories: "frogmen and UUVs", "UUVs and other types", and "frogmen and other types", and the true probability was calculated. False Positive Rate

[0041] Take FPR as the horizontal coordinate of the ROC curve, TPR as the vertical coordinate of the ROC curve, and the ROC curve above the diagonal line represents the feature classification ability of the feature. Calculate the area AUC below the ROC curve to judge the performance of each target feature in turn.

[0042] If the ROC curve of a feature has an AUC less than 0.65 for both pairwise classifications, the feature is deleted. If the AUC for both pairwise classifications is not less than 0.65, the feature is retained according to the inclusion principle.

[0043] Among them, FP represents the number of negative samples mistakenly identified as positive samples;

[0044] TN represents the number of correctly identified negative samples;

[0045] TP represents the number of correctly identified positive samples;

[0046] FN represents the number of positive samples that are incorrectly identified as negative samples.

[0047] Preferably, in step S3, the inter-group difference of each feature is calculated and intra-group differences Calculate the F value between each feature and the class

[0048] The P value is calculated based on the F value and degrees of freedom. When P < 0.05, the difference between the features is considered significant, and the feature importance is ranked from large to small based on the F value.

[0049] Where n represents the number of samples;

[0050] n i Represents the number of samples in the i-th category;

[0051] k represents the number of categories;

[0052] represents the sample mean of the i-th category;

[0053] represents the average value of all samples;

[0054] x i,j represents the jth sample of the i-th category;

[0055] SSA represents the mean difference between different groups;

[0056] SSE represents the difference between samples in the same group.

[0057] A multi-dimensional feature selection system for underwater target recognition provided by the present invention comprises:

[0058] Module M1: normalize the feature vector to obtain the feature space;

[0059] Module M2: Filter multiple collinear features and retain features with strong classification ability;

[0060] Module M3: Calculate feature importance scores and get feature importance rankings;

[0061] Module M4: Add features to the feature space according to the order of feature importance to obtain the best feature subset.

[0062] Preferably, the targets include three types: UUV, frogman and other types.

[0063] The features are signal echo features and motion behavior features, including tracking scale calculation, azimuth change, azimuth fluctuation, curvature, curvature fluctuation, speed, speed change, speed fluctuation, trajectory fluctuation, trajectory linearity, aspect ratio, area, rectangularity, center of mass ratio, peak-to-peak width ratio, first distribution probability and second distribution probability.

[0064] The module M1 normalizes the extracted signal echo and motion behavior feature vector of the underwater target in turn to obtain a normalized feature space.

[0065] The filtering of multiple collinear features includes establishing a multivariate linear regression model for each underwater target feature, solving the regression coefficient, and calculating the corresponding variance inflation factor.

[0066] The retaining of features with strong classification capabilities includes defining multiple thresholds along the feature scale of the underwater target, calibrating the true positive rate and the false positive rate under different thresholds, and screening the classification capability of each underwater target feature.

[0067] The module M3 calculates the F value between each feature and the category in turn, and establishes the importance ranking of the underwater target features according to the correlation between the features and the categories.

[0068] In the module M4, an underwater target classification model is constructed, and the underwater target features are sorted according to their importance, and the features are added to the feature space in sequence. The number of features with the best model performance is retained according to the model training results to obtain the best underwater target feature subset.

[0069] Preferably, the normalization calculation is

[0070] Among them, X represents the selected eigenvalue;

[0071] X max Indicates the maximum value corresponding to the selected eigenvalue;

[0072] X min Indicates the minimum value corresponding to the selected eigenvalue;

[0073] X * Represents the normalized result.

[0074] The multilayer perceptron model in the module M4 uses 5 hidden layers, each hidden layer contains 100 neurons, uses RELU activation function, batch_size is set to 64, uses Adam optimization algorithm, the learning rate is set to 0.001, and iterates 100 times.

[0075] According to the model training results, when the accuracy reaches 100%, the features with the best model performance are retained to obtain the best feature subset.

[0076] Preferably, filtering multiple collinear features comprises:

[0077] For each underwater target feature X i , and use them as dependent variables in turn, and the remaining features as independent variables to establish a multiple linear regression model X1 = α1 + a2X2 + ... + a N X N +ε;

[0078] Among them, X i represents the feature of the i-th underwater target;

[0079] N indicates that there are N underwater target features, i∈N;

[0080] a irepresents the i-th regression coefficient;

[0081] ε represents the error term.

[0082] Use the least squares method to solve the regression coefficient and calculate X j ,i=1,2,...,k,i≠j, coefficient of determination for regression Calculate the variance inflation factor for each underwater feature

[0083] in, represents the predicted value of the jth underwater target sample of the i-th feature;

[0084] x i,j represents the true value of the jth sample of the i-th underwater target feature;

[0085] represents the average value of the feature of the i-th underwater target;

[0086] VIF i represents the variance expansion factor corresponding to the i-th underwater target feature;

[0087] n represents the sample size.

[0088] The features that retain strong classification capabilities include:

[0089] The targets were divided into three categories according to the two-by-two categories: "frogmen and UUVs", "UUVs and other types", and "frogmen and other types", and the true probability was calculated. False Positive Rate

[0090] Take FPR as the horizontal coordinate of the ROC curve, TPR as the vertical coordinate of the ROC curve, and the ROC curve above the diagonal line represents the feature classification ability of the feature. Calculate the area AUC below the ROC curve to judge the performance of each target feature in turn.

[0091] If the ROC curve of a feature has an AUC less than 0.65 for both pairwise classifications, the feature is deleted. If the AUC for both pairwise classifications is not less than 0.65, the feature is retained according to the inclusion principle.

[0092] Among them, FP represents the number of negative samples mistakenly identified as positive samples;

[0093] TN represents the number of correctly identified negative samples;

[0094] TP represents the number of correctly identified positive samples;

[0095] FN represents the number of positive samples that are incorrectly identified as negative samples.

[0096] Preferably, the module M3 calculates the inter-group difference of each feature and intra-group differences Calculate the F value between each feature and the class

[0097] The P value is calculated based on the F value and degrees of freedom. When P < 0.05, the difference between the features is considered significant, and the feature importance is ranked from large to small based on the F value.

[0098] Where n represents the number of samples;

[0099] n i Represents the number of samples in the i-th category;

[0100] k represents the number of categories;

[0101] represents the sample mean of the i-th category;

[0102] represents the average value of all samples;

[0103] x i,j represents the jth sample of the i-th category;

[0104] SSA represents the mean difference between different groups;

[0105] SSE represents the difference between samples in the same group.

[0106] Compared with the prior art, the present invention has the following beneficial effects:

[0107] 1. The present invention filters out multiple collinear features and underwater target features with poor classification ability on scale through variance inflation factor and receiver operating characteristic curve, thereby reducing feature redundancy and improving feature discrimination.

[0108] 2. The present invention establishes an importance ranking of underwater target features, and screens features according to the performance of the classification model to obtain the best feature subset, which makes up for the shortcomings of the filtering method and can be closely integrated with the classifier performance. It has lower time complexity than the wrapping method and better interpretability than the embedding method.

[0109] 3. The present invention reduces feature dimensions, improves computational efficiency and the generalization performance of the model through a refined feature selection method, and has a simple and efficient path.

[0110] 4. The present invention has a mathematical theoretical basis and is combined with the actual classification effect of the classifier, and is more suitable for the needs of underwater target classification. BRIEF DESCRIPTION OF THE DRAWINGS

[0111] Other features, objects and advantages of the present invention will become more apparent from the detailed description of non-limiting embodiments made with reference to the following drawings:

[0112] Figure 1 Schematic diagram of the process of multi-dimensional feature selection method for underwater target recognition;

[0113] Figure 2 This is a schematic diagram of the characteristic VIF calculation results;

[0114] Figure 3 Schematic diagram of ROC curve for some features;

[0115] Figure 4 Schematic diagram of feature importance ranking;

[0116] Figure 5 This is a schematic diagram of the recognition accuracy results. DETAILED DESCRIPTION

[0117] The present invention is described in detail below in conjunction with specific embodiments. The following embodiments will help those skilled in the art to further understand the present invention, but are not intended to limit the present invention in any form. It should be noted that, for those of ordinary skill in the art, several changes and improvements can also be made without departing from the concept of the present invention. These all belong to the protection scope of the present invention.

[0118] According to a multidimensional feature selection method for underwater target recognition provided by the present invention, Figure 1 For example, it includes:

[0119] Through feature selection, redundant and irrelevant features are removed, features with strong classification capabilities are screened out, the ability to fine-tune the identification of underwater multiple targets is improved, and the frequency of false alarms is reduced. While being closely integrated with the classifier, it has good interpretability and low complexity.

[0120] In more preferred examples, taking the measured sonar data of a certain sea area as an example, after detection and tracking, the coordinate information of the moving target is obtained, and 17 signal echo features and motion behavior features of three types of targets including unmanned submarines (UUVs), frogmen, and other types are extracted. The extracted features are specifically divided into tracking scale calculation, azimuth change, azimuth fluctuation, curvature, curvature fluctuation, speed, speed change, speed fluctuation, trajectory fluctuation, trajectory linearity, aspect ratio, area, rectangularity, centroid ratio, peak-to-peak width ratio, distribution probability 1, and distribution probability 2. The data set has a total of 3133 samples, which are divided into training set and test set in a ratio of 8:2.

[0121] Step S1: normalization of feature vector;

[0122] The extracted feature vectors such as the signal echo and motion behavior of the underwater target are normalized in turn to obtain a normalized feature space. The normalization method is as follows.

[0123]

[0124] Among them, X is a certain characteristic value;

[0125] X max is the maximum value corresponding to the eigenvalue;

[0126] X min is the minimum value corresponding to this eigenvalue;

[0127] X * is the normalized result.

[0128] Step S2: filtering multicollinear features based on variance inflation factor (VIF);

[0129] Multicollinearity means that in the feature space, some features can be approximately equal to the linear combination of other features, that is, the features are correlated with each other. When extracting the signal echo features and motion behavior features of underwater targets, there may be features that are physically related. Multicollinearity will lead to unstable result analysis and feature redundancy. In the underwater target classification model, it is expected that each feature is independent of other features, that is, there is no collinearity between them.

[0130] For each underwater target feature X i , and use them as dependent variables and the remaining features as independent variables to establish a multiple linear regression model. Multiple linear regression model:

[0131] X1=α1+α2X2+...+α N X N +ε;

[0132] Among them, X i represents the feature of the i-th underwater target;

[0133] X1, ..., k are independent variables;

[0134] a1, ..., a k is the regression coefficient;

[0135] ε is the error term;

[0136] N indicates that there are N underwater target features in total, i∈N.

[0137] The least squares method is used to solve the regression coefficients to minimize the error between the model's predicted values ​​and the actual observed values. The calculation formula is:

[0138]

[0139] in, Represents the predicted value of the jth sample of the i-th feature;

[0140] x i,j Represents the true value of the jth sample of the i-th feature;

[0141] Represents the average value of the i-th feature

[0142] n represents the sample size.

[0143] Determines the explanatory power of the explanatory variable on the dependent variable. The larger the value, the higher the degree of explanation of the dependent variable by the explanatory variable, and the higher the degree of linear correlation. The variance inflation factor (VIF) corresponding to each underwater feature is calculated using the following formula:

[0144]

[0145] Among them, VIF i Represents the variance expansion factor corresponding to the i-th underwater target feature.

[0146] VIF can measure the severity of multicollinearity in a multiple linear regression model. The larger the VIF, the higher the correlation between the feature and other features. It is generally believed that when VIF>10, multicollinearity is more serious.

[0147] by Figure 2 For example, when there are underwater target features with VIF>10, such as "azimuth change" and "azimuth fluctuation", the feature with the largest VIF, such as "azimuth change", is deleted, and then the regression analysis is performed again and the VIF value of each feature is calculated until no feature with VIF>10 is found, and the feature filtering is completed. When deleting features, it is necessary to consider the theoretical significance of the target feature. If a feature is highly correlated with other features but has different information sources and has the significance of explaining the model, choosing to retain the feature can better understand the data. On the contrary, for example, the feature "azimuth change" is highly correlated with "azimuth fluctuation" and has the same information source, so it is reasonable to delete the feature "azimuth change". By filtering multiple collinear features, underwater target features with better independence can be obtained, reducing redundancy in features.

[0148] Step S3: Use the receiver operating characteristic curve (ROC) to retain features with strong classification ability;

[0149] The performance of underwater target characteristics such as target strength and speed in scale distribution can reveal the physical characteristics and behavior patterns of the target. Different types of underwater targets have different scale distributions due to differences in their structure, motion state, etc. ROC curves are mainly used to analyze binary classification models. Here, the ROC curve is extended from binary classification problems to multi-classification problems, and the ROC curves of two target categories are drawn.

[0150] In more preferred examples, because there are three categories, they can be divided into three categories according to the two categories: "frogmen and UUVs", "UUVs and other types", and "frogmen and other types". Multiple thresholds are defined along the characteristic scale of the underwater target, and the true positive rate and false positive rate under different thresholds are calibrated, and the calculation is as follows:

[0151]

[0152] Among them, FP represents the number of negative samples mistakenly identified as positive samples;

[0153] TN represents the number of correctly identified negative samples;

[0154] TP represents the number of correctly identified positive samples;

[0155] FN represents the number of positive samples that are incorrectly identified as negative samples.

[0156] FPR is the false positive rate, that is, the proportion of negative samples predicted as positive samples, which is the horizontal coordinate of the ROC curve. TPR is the true positive rate, that is, the proportion of positive samples predicted as positive samples, which is the vertical coordinate of the ROC curve.

[0157] The ROC method is used to screen the classification ability of each underwater target feature. The ROC curves of some features are shown in Figure 3 For example, the ROC curve above the diagonal line indicates that the feature has feature classification ability. The area under the ROC curve (AUC) is calculated. The larger the AUC, the stronger the feature classification ability on the scale. The performance of individual target features is analyzed in turn. If the ROC curve of the feature has a pairwise classification AUC of less than 0.65, the feature is deleted. If not all of them are less than 0.65, the inclusion principle is adopted to retain the feature.

[0158] In more preferred examples, the ROC curves of the features "azimuth fluctuation", "trajectory fluctuation", "rectangularity", "area", "aspect ratio" and "centroid rate" have a pairwise classification AUC of < 0.65, and these features are deleted.

[0159] Before establishing the classification model, the VIF and ROC methods were used to filter out multiple collinear features and underwater target features with poor classification ability in scale, which reduced the redundancy of features and improved the discrimination of features.

[0160] After the classification model is established, the importance ranking of underwater target features is established according to the correlation between features and categories, and the features are filtered according to the performance of the classification model to obtain the best feature subset. This makes up for the shortcomings of the filtering method, can be closely combined with the classifier performance, has lower time complexity than the wrapping method, and has better interpretability than the embedding method.

[0161] Since the decision tree has limitations in calculating feature importance, some effective features may be filtered out, such as features with average classification effect for a single feature but good classification effect when multiple features are associated. Considering the feature selection in the field of underwater acoustics (because it is applied to equipment), the stability and interpretability of the feature selection method need to be considered, so a combination of multiple selection methods is used to have a certain degree of inclusiveness.

[0162] Step S4: Calculate the feature importance score and obtain the feature importance ranking;

[0163] The F value of variance analysis can reflect whether there is a significant difference in a certain feature of different underwater target categories. The F value between each feature and category is calculated in turn to obtain the feature importance score, which can quantify the significance of the impact of different underwater target features on the target. The calculation formula is:

[0164]

[0165] Where n is the number of samples;

[0166] k is the number of categories;

[0167] n i Represents the number of samples in the i-th category;

[0168] SSA is the difference between groups, indicating the mean difference between different groups;

[0169] SSE is the intra-group difference, which indicates the difference between samples in the same group.

[0170] is the sample mean of the i-th category;

[0171] is the average value of all samples;

[0172] x i,j is the jth sample of the i-th category.

[0173] The P value can be calculated based on the F value and the degree of freedom. When P < 0.05, the difference between the features is considered significant. Figure 4 For example, the feature importance is sorted from large to small according to the F value.

[0174] Step S5: Evaluate the model performance to obtain the best feature subset.

[0175] An underwater target classification model is constructed, and the underwater target features are sorted according to their importance. The features are added to the feature space in sequence. The number of features with the best model performance is retained according to the model training results to obtain the optimal underwater target feature subset.

[0176] Through refined feature selection methods, the feature dimension can be reduced, the computational efficiency and the generalization performance of the model can be improved, and the path is simple and efficient.

[0177] In more preferred examples, taking the multi-layer perceptron (MLP) as an example, the model uses 5 hidden layers, each hidden layer contains 100 neurons, uses the RELU activation function, the batch_size is set to 64, uses the Adam optimization algorithm, the learning rate is set to 0.001, and iterates 100 times. The recognition accuracy is Figure 5 For example, according to the model training results, the number of features with the best model performance is retained. When the number of features is 5, the accuracy reaches 100%, and these 5 features are retained as the best feature subset. From the results, it can be seen that feature selection can reduce the feature dimension while having better recognition performance.

[0178] The present invention also provides a multidimensional feature selection system for underwater target identification. The multidimensional feature selection system for underwater target identification can be implemented by executing the process steps of the multidimensional feature selection method for underwater target identification, that is, those skilled in the art can understand the multidimensional feature selection method for underwater target identification as a preferred implementation of the multidimensional feature selection system for underwater target identification.

[0179] A multi-dimensional feature selection system for underwater target recognition provided by the present invention comprises:

[0180] Module M1: normalize the feature vector to obtain the feature space;

[0181] Module M2: Filter multiple collinear features and retain features with strong classification ability;

[0182] Module M3: Calculate feature importance scores and get feature importance rankings;

[0183] Module M4: Add features to the feature space according to the order of feature importance to obtain the best feature subset.

[0184] In more preferred examples, the targets include three types: UUVs, frogmen and other types.

[0185] The features are signal echo features and motion behavior features, including tracking scale calculation, azimuth change, azimuth fluctuation, curvature, curvature fluctuation, speed, speed change, speed fluctuation, trajectory fluctuation, trajectory linearity, aspect ratio, area, rectangularity, center of mass ratio, peak-to-peak width ratio, first distribution probability and second distribution probability.

[0186] The module M1 normalizes the extracted signal echo and motion behavior feature vector of the underwater target in turn to obtain a normalized feature space.

[0187] The filtering of multiple collinear features includes establishing a multivariate linear regression model for each underwater target feature, solving the regression coefficient, and calculating the corresponding variance inflation factor.

[0188] The retaining of features with strong classification capabilities includes defining multiple thresholds along the feature scale of the underwater target, calibrating the true positive rate and the false positive rate under different thresholds, and screening the classification capability of each underwater target feature.

[0189] The module M3 calculates the F value between each feature and the category in turn, and establishes the importance ranking of the underwater target features according to the correlation between the features and the categories.

[0190] In the module M4, an underwater target classification model is constructed, and the underwater target features are sorted according to their importance, and the features are added to the feature space in sequence. The number of features with the best model performance is retained according to the model training results to obtain the best underwater target feature subset.

[0191] In more preferred embodiments, the normalized calculation is

[0192] Among them, X represents the selected eigenvalue;

[0193] X max Indicates the maximum value corresponding to the selected eigenvalue;

[0194] X min Indicates the minimum value corresponding to the selected eigenvalue;

[0195] X * Represents the normalized result.

[0196] The multilayer perceptron model in the module M4 uses 5 hidden layers, each hidden layer contains 100 neurons, uses RELU activation function, batch_size is set to 64, uses Adam optimization algorithm, the learning rate is set to 0.001, and iterates 100 times.

[0197] According to the model training results, when the accuracy reaches 100%, the features with the best model performance are retained to obtain the best feature subset.

[0198] In more preferred embodiments, the filtering of multiple collinear features includes:

[0199] For each underwater target feature X i , and use them as dependent variables in turn, and the remaining features as independent variables to establish a multiple linear regression model X1 = a1 + a2X2 + ... + a N X N +ε;

[0200] Among them, X i represents the feature of the i-th underwater target;

[0201] N indicates that there are N underwater target features, i∈N;

[0202] a i represents the i-th regression coefficient;

[0203] ε represents the error term.

[0204] Use the least squares method to solve the regression coefficient and calculate the coefficient of determination in turn Calculate the variance inflation factor for each underwater feature

[0205] in, represents the predicted value of the jth underwater target sample of the i-th feature;

[0206] x i,j represents the true value of the jth sample of the i-th underwater target feature;

[0207] represents the average value of the feature of the i-th underwater target;

[0208] VIF i represents the variance expansion factor corresponding to the i-th underwater target feature;

[0209] n represents the sample size.

[0210] The features that retain strong classification capabilities include:

[0211] The targets were divided into three categories according to the two-by-two categories: "frogmen and UUVs", "UUVs and other types", and "frogmen and other types", and the true probability was calculated. False Positive Rate

[0212] Take FPR as the horizontal coordinate of the ROC curve, TPR as the vertical coordinate of the ROC curve, and the ROC curve above the diagonal line represents the feature classification ability of the feature. Calculate the area AUC below the ROC curve to judge the performance of each target feature in turn.

[0213] If the ROC curve of a feature has an AUC less than 0.65 for both pairwise classifications, the feature is deleted. If the AUC for both pairwise classifications is not less than 0.65, the feature is retained according to the inclusion principle.

[0214] Among them, FP represents the number of negative samples mistakenly identified as positive samples;

[0215] TN represents the number of correctly identified negative samples;

[0216] TP represents the number of correctly identified positive samples;

[0217] FN represents the number of positive samples that are incorrectly identified as negative samples.

[0218] In more preferred embodiments, the module M3 calculates the inter-group difference of each feature. and intra-group differences Calculate the F value between each feature and the class

[0219] The P value is calculated based on the F value and degrees of freedom. When P < 0.05, the difference between the features is considered significant, and the feature importance is ranked from large to small based on the F value.

[0220] Where n represents the number of samples;

[0221] n i Represents the number of samples in the i-th category;

[0222] k represents the number of categories;

[0223] represents the sample mean of the i-th category;

[0224] represents the average value of all samples;

[0225] x i,j represents the jth sample of the i-th category;

[0226] SSA represents the mean difference between different groups;

[0227] SSE represents the difference between samples in the same group.

[0228] Those skilled in the art know that, in addition to realizing the system and its various devices, modules, and units provided by the present invention in a purely computer-readable program code, it is entirely possible to realize the same functions in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers by logically programming the method steps. Therefore, the system and its various devices, modules, and units provided by the present invention can be considered as a hardware component, and the devices, modules, and units included therein for realizing various functions can also be regarded as structures within the hardware component; the devices, modules, and units for realizing various functions can also be regarded as both software modules for realizing the method and structures within the hardware component.

[0229] The above describes the specific embodiments of the present invention. It should be understood that the present invention is not limited to the above specific embodiments, and those skilled in the art can make various changes or modifications within the scope of the claims, which does not affect the essence of the present invention. In the absence of conflict, the embodiments of the present application and the features in the embodiments can be combined with each other arbitrarily.

Claims

1. A multidimensional feature selection method for underwater target recognition, characterized in that: include: Step S1: normalize the feature vector to obtain the feature space; Step S2: filter out multiple collinear features and retain features with strong classification capabilities; Step S3: Calculate the feature importance score and obtain the feature importance ranking; Step S4: Add features to the feature space according to the order of feature importance to obtain the best feature subset.

2. The multidimensional feature selection method for underwater target recognition according to claim 1, characterized in that: The targets include three types: UUVs, frogmen, and others; The features are signal echo features and motion behavior features, including tracking scale calculation, azimuth change, azimuth fluctuation, curvature, curvature fluctuation, speed, speed change, speed fluctuation, trajectory fluctuation, trajectory linearity, aspect ratio, area, rectangularity, centroid ratio, peak-to-peak width ratio, first distribution probability and second distribution probability; In the step S1, the signal echo and the motion behavior feature vector of the extracted underwater target are normalized in sequence to obtain a normalized feature space; The filtering of multiple collinear features includes establishing a multivariate linear regression model for each underwater target feature, solving the regression coefficient, and calculating the corresponding variance inflation factor; The retaining of features with strong classification capabilities includes defining multiple thresholds along the feature scale of the underwater target, calibrating the true positive rate and false positive rate under different thresholds, and screening the classification capability of each underwater target feature; In step S3, the F value between each feature and the category is calculated in turn, and the importance ranking of the underwater target features is established according to the correlation between the features and the categories; In step S4, an underwater target classification model is constructed, and the underwater target features are sorted according to their importance, and the features are added to the feature space in sequence. The number of features with the best model performance is retained according to the model training results to obtain the best underwater target feature subset.

3. The multidimensional feature selection method for underwater target recognition according to claim 2 is characterized in that: The normalization calculation is Among them, X represents the selected eigenvalue; X max Indicates the maximum value corresponding to the selected eigenvalue; X min Indicates the minimum value corresponding to the selected eigenvalue; X * represents the normalized result; In the step S4, the multilayer perceptron model uses 5 hidden layers, each hidden layer contains 100 neurons, uses a RELU activation function, the batch_size is set to 64, uses the Adam optimization algorithm, the learning rate is set to 0.001, and iterates 100 times; According to the model training results, when the accuracy reaches 100%, the features with the best model performance are retained to obtain the best feature subset.

4. The multidimensional feature selection method for underwater target recognition according to claim 1, characterized in that: The filtering of multicollinear features comprises: For each underwater target feature X i , and use them as dependent variables in turn, and the remaining features as independent variables to establish a multiple linear regression model X1 = α1 + α2X2 + ... + α N X N +ε; Among them, X i represents the feature of the i-th underwater target; N indicates that there are N underwater target features, i∈N; a i represents the i-th regression coefficient; ε represents the error term; Use the least squares method to solve the regression coefficient and calculate the coefficient of determination in turn Calculate the variance inflation factor for each underwater feature in, represents the predicted value of the jth underwater target sample of the i-th feature; x i,j represents the true value of the jth sample of the i-th underwater target feature; represents the average value of the feature of the i-th underwater target; VIF i represents the variance expansion factor corresponding to the i-th underwater target feature; n represents the sample size; The features that retain strong classification capabilities include: The targets are divided into three categories according to the pairwise classification: frogmen and UUVs, UUVs and other types, and frogmen and other types. The true probability is calculated. False Positive Rate Take FPR as the horizontal coordinate of the ROC curve, and TPR as the vertical coordinate of the ROC curve. The ROC curve above the diagonal represents the feature classification ability. Calculate the area AUC below the ROC curve to judge the performance of each target feature in turn. If the ROC curve of a feature has an AUC less than 0.65 for both pairwise classifications, the feature is deleted. If the AUC for both pairwise classifications is not less than 0.65, the feature is retained according to the inclusion principle. Among them, FP represents the number of negative samples mistakenly identified as positive samples; TN represents the number of correctly identified negative samples; TP represents the number of correctly identified positive samples; FN represents the number of positive samples that are incorrectly identified as negative samples.

5. The multidimensional feature selection method for underwater target recognition according to claim 1, characterized in that: In step S3, the inter-group differences of each feature are calculated. and intra-group differences Calculate the F value between each feature and the class The P value is calculated based on the F value and the degree of freedom. When P < 0.05, the difference between the features is considered significant. The importance of the features is ranked from large to small according to the F value. Where n represents the number of samples; n i Represents the number of samples in the i-th category; k represents the number of categories; represents the sample mean of the i-th category; represents the average value of all samples; x i,j represents the jth sample of the i-th category; SSA represents the mean difference between different groups; SSE represents the difference between samples in the same group.

6. A multi-dimensional feature selection system for underwater target recognition, characterized in that: include: Module M1: normalize the feature vector to obtain the feature space; Module M2: Filter multiple collinear features and retain features with strong classification ability; Module M3: Calculate feature importance scores and get feature importance rankings; Module M4: Add features to the feature space according to the order of feature importance to obtain the best feature subset.

7. The multi-dimensional feature selection system for underwater target recognition according to claim 6, characterized in that: The targets include three types: UUVs, frogmen, and others; The features are signal echo features and motion behavior features, including tracking scale calculation, azimuth change, azimuth fluctuation, curvature, curvature fluctuation, speed, speed change, speed fluctuation, trajectory fluctuation, trajectory linearity, aspect ratio, area, rectangularity, centroid ratio, peak-to-peak width ratio, first distribution probability and second distribution probability; The module M1 sequentially normalizes the extracted signal echo and motion behavior feature vector of the underwater target to obtain a normalized feature space; The filtering of multiple collinear features includes establishing a multivariate linear regression model for each underwater target feature, solving the regression coefficient, and calculating the corresponding variance inflation factor; The retaining of features with strong classification capabilities includes defining multiple thresholds along the feature scale of the underwater target, calibrating the true positive rate and false positive rate under different thresholds, and screening the classification capability of each underwater target feature; In the module M3, the F value between each feature and the category is calculated in turn, and the importance ranking of the underwater target features is established according to the correlation between the features and the categories; In the module M4, an underwater target classification model is constructed, and the underwater target features are sorted according to their importance, and the features are added to the feature space in sequence. The number of features with the best model performance is retained according to the model training results to obtain the best underwater target feature subset.

8. The multi-dimensional feature selection system for underwater target recognition according to claim 7, characterized in that: The normalization calculation is Among them, X represents the selected eigenvalue; X max Indicates the maximum value corresponding to the selected eigenvalue; X min Indicates the minimum value corresponding to the selected eigenvalue; X * represents the normalized result; The multilayer perceptron model in the module M4 uses 5 hidden layers, each hidden layer contains 100 neurons, uses RELU activation function, batch_size is set to 64, uses Adam optimization algorithm, the learning rate is set to 0.001, and iterates 100 times; According to the model training results, when the accuracy reaches 100%, the features with the best model performance are retained to obtain the best feature subset.

9. The multi-dimensional feature selection system for underwater target recognition according to claim 6, characterized in that: The filtering of multicollinear features comprises: For each underwater target feature X i , and use them as dependent variables in turn, and the remaining features as independent variables to establish a multiple linear regression model X1 = α1 + α2X2 + ... + α N X N +ε; Among them, X i represents the feature of the i-th underwater target; N indicates that there are N underwater target features, i∈N; a i represents the i-th regression coefficient; ε represents the error term; Use the least squares method to solve the regression coefficient and calculate the coefficient of determination in turn Calculate the variance inflation factor for each underwater feature in, represents the predicted value of the jth underwater target sample of the i-th feature; x i,j represents the true value of the jth sample of the i-th underwater target feature; represents the average value of the feature of the i-th underwater target; VIF i represents the variance expansion factor corresponding to the i-th underwater target feature; n represents the sample size; The features that retain strong classification capabilities include: The targets are divided into three categories according to the pairwise classification: frogmen and UUVs, UUVs and other types, and frogmen and other types. The true probability is calculated. False Positive Rate Take FPR as the horizontal coordinate of the ROC curve, and TPR as the vertical coordinate of the ROC curve. The ROC curve above the diagonal represents the feature classification ability. Calculate the area AUC below the ROC curve to judge the performance of each target feature in turn. If the ROC curve of a feature has an AUC less than 0.65 for both pairwise classifications, the feature is deleted. If the AUC for both pairwise classifications is not less than 0.65, the feature is retained according to the inclusion principle. Among them, FP represents the number of negative samples mistakenly identified as positive samples; TN represents the number of correctly identified negative samples; TP represents the number of correctly identified positive samples; FN represents the number of positive samples that are incorrectly identified as negative samples.

10. The multi-dimensional feature selection system for underwater target recognition according to claim 6, characterized in that: The module M3 calculates the inter-group differences of each feature and intra-group differences Calculate the F value between each feature and the class The P value is calculated based on the F value and the degree of freedom. When P < 0.05, the difference between the features is considered significant. The importance of the features is ranked from large to small according to the F value. Where n represents the number of samples; n i Represents the number of samples in the i-th category; k represents the number of categories; represents the sample mean of the i-th category; represents the average value of all samples; x i,j represents the jth sample of the i-th category; SSA represents the mean difference between different groups; SSE represents the difference between samples in the same group.

Images