Search

CN-121978665-A - Active sonar target identification method based on multi-pulse accumulation and two-stage fusion

CN121978665ACN 121978665 ACN121978665 ACN 121978665ACN-121978665-A

Abstract

The invention relates to an active sonar target identification method based on multi-pulse accumulation and two-stage fusion; the method comprises the steps of firstly obtaining a multi-dimensional characteristic value of a target, accumulating various characteristics by multiple pulses (ping) to form a characteristic subset, and carrying out characteristic level fusion by a fusion operator to extract stable characteristic representation with strong discrimination. On the basis, the probability that each fusion feature belongs to a 'real target' and a 'non-real target' is calculated based on prior distribution, and decision-level fusion is carried out on probability evidences of all features by adopting a D-S evidence theory, so that the credibility of a final recognition result is obtained. According to the invention, through a time sequence accumulation and two-stage fusion architecture, the multidimensional information and time sequence information of the target are effectively integrated, the recognition accuracy and the decision robustness are obviously improved, and the false alarm rate is greatly reduced.

Inventors

  • Gao Diaolin
  • CAO PENG

Assignees

  • 海鹰企业集团有限责任公司

Dates

Publication Date
20260505
Application Date
20260123

Claims (8)

  1. 1. The active sonar target identification method based on multi-pulse accumulation and two-stage fusion is characterized by comprising the following steps of: step S1, acquiring characteristic values of a plurality of characteristics of an active sonar tracking target, wherein the characteristics at least comprise scale characteristics, radial speed characteristics and absolute speed characteristics of the target; Step S2, feature level fusion, namely accumulating the feature values of the tracking target in a plurality of continuous pulse periods aiming at each single feature to form a short-time accumulated feature subset of the feature; and step S3, decision-level fusion, namely respectively predicting the probability that the target belongs to the 'real target' class and the 'non-real target' based on the fusion characteristics corresponding to the single characteristics obtained in the step S2, taking the class probability obtained by predicting the characteristics as input, carrying out decision-level fusion, and outputting the final credibility that the target is the 'real target'.
  2. 2. The active sonar target recognition method according to claim 1, wherein the fusion is realized by the following process: Wherein, the A short-time accumulated feature subset representing a feature, A subset preprocessing operator representing preprocessing the feature subset, Representing a fusion operator that fuses the feature subset, Representing a fusion feature corresponding to the feature.
  3. 3. The active sonar target identification method of claim 2, wherein the subset preprocessing operator Including normalizing, filtering, or extracting statistics for the feature subset.
  4. 4. The active sonar target identification method of claim 2, wherein the fusion operator The method is obtained through training, wherein the training takes a feature subset as input and fusion features capable of effectively representing target categories as expected output.
  5. 5. The active sonar target recognition method according to claim 1, wherein in step S3, the probability that the predicted target belongs to the "real target" class and the "non-real target" specifically includes: Fusion feature value for each feature Probability density function for the feature based on pre-established "real objectives Probability density function with respect to the feature with respect to "non-real objects Calculating the probability of belonging to the class of' real targets Probability of Calculated by the following formula: Wherein, the And The prior probabilities of the "real target" class and the "non-real target" class, respectively.
  6. 6. The method for identifying the active sonar target according to claim 4, wherein the probability density function And Based on active sonar actual measurement data statistics, expert experience or data training, and is used for defining typical numerical regions of the characteristic values under different categories.
  7. 7. The active sonar target recognition method according to claim 1, wherein in step S3, the decision-level fusion is performed by using a D-S evidence theory method, comprising: Will be the first True target class probability obtained by individual feature prediction And "unrealistic target" class probabilities The basic probability distribution function constructed as the feature ; For all of Of individual features Combining, and combining the basic probability distribution functions Calculated from the following formula: Wherein, the The values of the two types of focus elements are { "real target" class }, { "non-real target" class }, or the whole set of the two types of focus elements; is normalized by a factor, finally, in the category of' real target The value is taken as the final confidence.
  8. 8. The method of active sonar target identification of claim 1 wherein the scale feature is an extension of the target in a radial distance dimension.

Description

Active sonar target identification method based on multi-pulse accumulation and two-stage fusion Technical Field The invention relates to the technical field of sonar signal processing, in particular to an active sonar target identification method based on multi-pulse accumulation and two-stage fusion. Background The accurate detection and identification of the underwater target are core tasks of the active sonar system, and have important significance for ocean safety, resource exploration and underwater operation. However, due to the extreme complexity of the underwater acoustic environment (including strong reverberation, complex noise, multipath effects, etc.), and the diversity of the characteristics of the target itself (such as geometry, materials, motion states), active sonar systems generally face two significant challenges in practical applications, namely that the false alarm rate of target detection is high and the accuracy and reliability of target identification are insufficient. And to improve performance, the prior art is directed to extracting multi-dimensional features of scattering, waveform, motion, and scale of the target from the echoes. However, how to effectively fuse these multidimensional features remains a key bottleneck, and the existing methods mainly have the following limitations: And the time sequence information is ignored, namely the characteristics of single detection are directly used, the same characteristics are not effectively accumulated and fused in a plurality of continuous pulse periods, and the characteristic stability is poor. The fusion strategy is single, namely, simple feature splicing or decision weighted average is adopted, information cannot be processed in a layering mode, and differences of different features in dimension, reliability and evidence conflict cannot be effectively processed. The uncertainty treatment is lacking, namely when different feature evidences have conflict or uncertainty (such as motion feature support is a true target and scale feature support is not), the existing fusion method based on deterministic or simple probability models lacks a strict mathematical framework for quantification and reconciliation. In summary, the existing active sonar target recognition technology has the defects of neglect of time sequence accumulation, single fusion strategy, weak uncertainty processing capability, insufficient flow cooperativity and the like in the processing and fusion of multidimensional features. Therefore, an innovative multidimensional feature fusion method is needed to systematically integrate the observation information of the target in the time dimension, adopt a hierarchical fusion architecture and effectively process the uncertainty of feature evidence by using an advanced fusion theory, so that the accuracy, reliability and robustness of active sonar target identification are fundamentally improved, and the false alarm rate is remarkably reduced. Disclosure of Invention In order to solve the technical problems, the invention provides an active sonar target identification method based on multi-pulse accumulation and two-stage fusion, which comprises the following steps: step S1, acquiring characteristic values of a plurality of characteristics of an active sonar tracking target, wherein the characteristics at least comprise scale characteristics, radial speed characteristics and absolute speed characteristics of the target; Step S2, feature level fusion, namely accumulating the feature values of the tracking target in a plurality of continuous pulse periods aiming at each single feature to form a short-time accumulated feature subset of the feature; and step S3, decision-level fusion, namely respectively predicting the probability that the target belongs to the 'real target' class and the 'non-real target' based on the fusion characteristics corresponding to the single characteristics obtained in the step S2, taking the class probability obtained by predicting the characteristics as input, carrying out decision-level fusion, and outputting the final credibility that the target is the 'real target'. In one embodiment of the invention, the fusion is achieved by the following process: Wherein, the A short-time accumulated feature subset representing a feature,And a subset preprocessing operator for preprocessing the feature subset, a fusion operator for fusing the feature subset and a fusion feature corresponding to the feature are represented. In one embodiment of the present invention, wherein the subset preprocessing operatorIncluding normalizing, filtering, or extracting statistics for the feature subset. In one embodiment of the invention, the fusion operatorThe method is obtained through training, wherein the training takes a feature subset as input and fusion features capable of effectively representing target categories as expected output. In one embodiment of the present invention, in step S3, the probabilities that the