Active sonar target identification method based on multi-pulse accumulation and two-stage fusion
By employing a multi-pulse accumulation and two-level fusion method, combined with a feature-level and decision-level fusion architecture, the shortcomings of multi-dimensional feature fusion in active sonar systems are addressed, thereby improving the accuracy and robustness of underwater target identification and reducing the false alarm rate.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HAIYING ENTERPRISE GROUP
- Filing Date
- 2026-01-23
- Publication Date
- 2026-05-05
AI Technical Summary
Existing active sonar systems suffer from high false alarm rates and low recognition accuracy in underwater target detection. Furthermore, their multi-dimensional feature fusion strategies are limited and their uncertainty handling capabilities are weak, making it difficult to effectively improve recognition accuracy and robustness.
We employ a multi-pulse accumulation and two-level fusion method, using a fusion architecture that combines feature level and decision level, and utilize DS evidence theory to handle uncertainty, integrate multi-dimensional feature information, and improve feature stability and decision credibility.
It significantly improves the accuracy and reliability of underwater target identification, while greatly reducing the false alarm rate and enhancing the identification performance of active sonar systems.
Smart Images
Figure CN121978665A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of sonar signal processing technology, and in particular to an active sonar target recognition method based on multi-pulse accumulation and two-level fusion. Background Technology
[0002] Accurate detection and identification of underwater targets is the core task of active sonar systems, and it is of vital importance to marine security, resource exploration, and underwater operations. However, due to the extreme complexity of the underwater acoustic environment (including strong reverberation, complex noise, multipath effects, etc.) and the diversity of target characteristics (such as geometry, material, and motion state), active sonar systems generally face two major challenges in practical applications: a persistently high false alarm rate for target detection and insufficient accuracy and reliability in target identification.
[0003] Furthermore, to improve performance, existing technologies focus on extracting multi-dimensional features of the target from the echo, such as scattering, waveform, motion, and scale. However, effectively fusing these multi-dimensional features remains a key bottleneck, and existing methods mainly have the following limitations: Ignoring timing information: Features from a single detection are often used directly without effectively accumulating and fusing the same feature over multiple consecutive pulse cycles, resulting in poor feature stability.
[0004] The fusion strategy is too simplistic: it often uses simple feature splicing or decision weighted averaging, fails to process information hierarchically, and cannot effectively handle the differences in scale, reliability, and evidence conflict among different features.
[0005] Lack of uncertainty handling: When there are conflicts or uncertainties in evidence from different features (such as motion features supporting a true target while scale features do not), existing fusion methods based on deterministic or simple probability models lack a rigorous mathematical framework for quantification and reconciliation.
[0006] In summary, existing active sonar target recognition technologies suffer from several shortcomings in processing and fusing multidimensional features, including neglecting temporal accumulation, employing a single fusion strategy, lacking the ability to handle uncertainty, and exhibiting insufficient process coordination. Therefore, there is an urgent need for an innovative multidimensional feature fusion method that can systematically integrate target observation information across the temporal dimension, adopt a hierarchical fusion architecture, and utilize advanced fusion theory to effectively handle the uncertainty of feature evidence. This would fundamentally improve the accuracy, reliability, and robustness of active sonar target recognition, and significantly reduce the false alarm rate. Summary of the Invention
[0007] To address the aforementioned technical problems, this invention provides an active sonar target recognition method based on multi-pulse accumulation and two-level fusion, comprising the following steps: Step S1: Obtain feature values of multiple features of the target tracked by active sonar, wherein the features include at least the target's scale features, radial velocity features, and absolute velocity features; Step S2: Feature-level fusion: For each single feature, the feature value of the tracked target is accumulated over multiple consecutive pulse cycles to form a short-term accumulated feature subset of the feature; based on the feature subsets of each feature, they are fused using the corresponding fusion operator to obtain the fused feature corresponding to each single feature; Step S3: Decision-level fusion: Based on the fusion features corresponding to each individual feature obtained in step S2, predict the probability that the target belongs to the "real target" class and the "non-real target" class respectively; take the class probabilities predicted by each feature as input, perform decision-level fusion, and output the final credibility that the target is the "real target".
[0008] In one embodiment of the present invention, the fusion is achieved through the following process:
[0009] in, Represents a short-term accumulated feature subset of a certain feature. This represents a subset preprocessing operator that preprocesses the subset of features, a fusion operator that fuses the subset of features, and a fusion feature corresponding to the feature.
[0010] In one embodiment of the invention, the subset preprocessing operator... The operations include standardizing, filtering, or extracting statistics from a subset of features.
[0011] In one embodiment of the present invention, the fusion operator It is obtained through training, where the training takes a subset of features as input and the fused features that can effectively represent the target category as the expected output.
[0012] In one embodiment of the present invention, step S3, predicting the probability that the target belongs to the "real target" class or the "non-real target" class, specifically includes: For each feature, the fused feature value Based on the pre-established probability density function of the "real target" with respect to this feature The probability density function of the "non-real target" with respect to this feature Calculate the probability that it belongs to the "real target" class. And probability Calculated using the following formula:
[0013] in, and These are the prior probabilities for the "real target" class and the "non-real target" class, respectively.
[0014] In one embodiment of the present invention, the probability density function and The features are obtained based on statistical analysis of active sonar measured data, expert experience, or data training, and are used to define typical numerical ranges of the features under different categories.
[0015] In one embodiment of the present invention, step S3, "performing decision-level fusion," employs the DS evidence theory method, including: In one embodiment of the present invention, the first The probability of the "true target" class obtained from feature prediction and the probability of "non-real target" Construct the basic probability assignment function for this feature. ; For all Features Combining, the basic probability assignment function after combination Calculated by the following formula:
[0016] in, The value is a focal element, which can be {"real target" class}, {"non-real target" class}, or their entire set; As a normalization factor; ultimately, based on the "real target" class The value serves as the final credibility score.
[0017] In one embodiment of the invention, the scale feature is the extended scale of the target in the radial distance dimension.
[0018] Compared with existing technologies, the above-mentioned technical solution of the present invention has the following advantages: The active sonar target identification method of the present invention improves feature stability through multi-pulse (ping) accumulation; it adopts a two-layer fusion architecture combining feature level and decision level to effectively integrate multi-dimensional heterogeneous information; and it uses DS evidence theory to perform rigorous mathematical fusion of uncertain or conflicting evidence. Ultimately, it significantly improves the accuracy and reliability of underwater "real targets" identification, while greatly reducing the false alarm rate. Attached Figure Description
[0019] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings.
[0020] Figure 1 This is a schematic diagram of the active sonar multi-dimensional feature fusion process of the present invention; Figure 2 This is a distribution map of the scale features of the active sonar "real target" described in this invention; Figure 3 This is a distribution map of the scale features of the active sonar "non-real target" described in this invention; Figure 4 This is a distribution map of a certain feature value of the "real target" in the active sonar described in this invention; Figure 5 This is a distribution map of a certain feature value of the active sonar "non-real target" described in this invention; Figure 6 This is a joint distribution map of a certain feature value of the active sonar "real target" and "non-real target" described in this invention. Detailed Implementation
[0021] This invention provides an active sonar multidimensional feature fusion method based on multi-pulse accumulation and two-level fusion. Its core lies in employing a two-level fusion architecture combining "feature-level fusion" and "decision-level fusion" to comprehensively utilize the multidimensional and temporal features of the target tracked by the active sonar. For example... Figure 1 As shown, the overall process of this method includes three stages: multi-dimensional feature acquisition, single feature multi-pulse (ping) accumulation and feature-level fusion, and multi-dimensional feature decision-level fusion.
[0022] Furthermore, the active sonar multidimensional feature fusion method includes the following steps: Step 1: Obtain the feature values of the target tracked by the active sonar.
[0023] In this embodiment, the system uses an active sonar device to detect and track underwater targets, acquiring echo data of the target in real time over multiple consecutive detection cycles (pings). From this data, multiple feature values characterizing the target's attributes are extracted. The features used include, but are not limited to: Absolute velocity characteristics: reflect the velocity of the target relative to a fixed coordinate system.
[0024] Radial velocity characteristics: Reflects the velocity component of the target along the sonar beam direction.
[0025] Scale characteristics: Reflects the physical extension scale of the target in the radial distance dimension, for example, estimated by the echo envelope width. According to statistical data, the scale characteristics of real targets (such as submarines) are usually distributed in a small range (e.g., 0~150 meters), while the scale of non-real targets (such as large schools of fish, geological clutter) may be larger or more dispersed.
[0026] Step 2: Single feature multi-pulse accumulation and feature-level fusion.
[0027] This step aims to condense and refine the information of each individual feature over time to suppress random errors from single measurements and enhance the stability and characterization of the features.
[0028] The short-term accumulation forms a feature subset: it is assumed that the target's motion state and scattering characteristics are quasi-stationary over a short period of time (e.g., within T minutes). The system receives a sequence of feature values of the tracked target over N consecutive pulse periods, such as a scale feature value sequence. These N values are organized to form a short-term accumulated feature subset of this feature, denoted as a vector. This process corresponds to Figure 1 The "Feature Subset Formation" module in the document.
[0029] Specific feature-level fusion: feature subsets for each feature Perform feature-level fusion operations to obtain a more representative fused feature value. This process can be formally represented as:
[0030] in: This represents a subset preprocessing operator that preprocesses the feature subset; its function is to preprocess the original feature subset, such as standardizing to eliminate dimensions, performing smoothing filtering to suppress noise, or calculating statistics (such as mean and variance). In this embodiment, it may be to calculate the arithmetic mean of the scale feature subset.
[0031] This represents a fusion operator that fuses the feature subset; its function is to fuse preprocessed information into a scalar value.
[0032] This can be achieved through training. For example, a shallow neural network can be designed, taking preprocessed features as input and fused features that better distinguish target categories as output, for training. After training, the parameters... That is, it is fixed for online fusion. Simple linear weighting or other methods can also be used. This process corresponds to... Figure 1 The "Feature-level Fusion" module in the text.
[0033] Repeat the steps of short-time accumulation to form feature subsets and feature-level fusion for features such as absolute velocity, radial velocity, and scale to obtain fused features corresponding to each feature. wait.
[0034] Step 3: Decision-level fusion of multi-dimensional features of active sonar targets like Figure 1As shown, the specific steps of decision-level fusion of multi-dimensional features of active sonar targets are as follows: ① Assuming that active sonar targets are divided into two categories, "real targets" and "non-real targets," the fused features of each single feature subset are used as input to predict the probability of the target echo belonging to a certain category. For example, the probability of the target echo belonging to a "real target" predicted based on the radial velocity fusion feature is... The probability that the target echo predicted based on the target scale fusion features belongs to the "real target" is: ② Decision-level fusion discrimination is performed on the discrimination probabilities of each single category feature.
[0035] 1) Predict the probability of the target echo belonging to a certain category Based on active sonar measured data, the distribution functions of various characteristic values for "real targets" and "non-real targets" are statistically analyzed. Using these characteristic value distribution functions as a basis, the probability that a measured characteristic value belongs to a "real target" or "non-real target" is calculated. For example, based on expert experience and data training, the scale characteristic distribution function for "real targets" is summarized as follows: Figure 2 As shown, the scale eigenvalue range of the "real target" is generally [0, 150m], that is... Region; Distribution function of "non-real target" scale characteristics, such as Figure 3 As shown, the scale eigenvalue of "non-realistic targets" is generally greater than 150m, that is... area.
[0036] a) Distribution of a certain characteristic value of the "real target" Based on expert experience and data training, a certain characteristic value of the "true target" was summarized. Distribution function, assumed region Represents the "real target" and a certain characteristic value of the "real target". The probability density function is as follows Figure 4 As shown. When the eigenvalue At that time, it fell Region; when eigenvalues At that time, it fell Region; when eigenvalues At that time, it fell Region. Eigenvalues Falling When the region is defined, the probability of it belonging to the "real target" is: eigenvalues Falling When the region is defined, the probability of it belonging to the "real target" is: .
[0037] b) Distribution of a certain characteristic value of a "non-real target" Based on expert experience and data training, a certain characteristic value of "non-realistic target" was summarized. Distribution function, assumed region This represents a "non-real target" or a certain characteristic value of the "non-real target". The probability density function is as follows Figure 5 As shown. When the eigenvalue At that time, it fell Region; when eigenvalues At that time, it fell Region; when eigenvalues At that time, it fell Region. Eigenvalues Falling When the region is defined, the probability of it belonging to a "non-real target" is: eigenvalues Falling When the region is defined, the probability of it belonging to a "non-real target" is: .
[0038] c) Joint distribution of a certain characteristic value of "real target" and "non-real target" The joint probability density function of a certain eigenvalue of "real target" and "non-real target" is as follows: Figure 6 As shown.
[0039] like Figure 6 As shown, Indicates the "real target". Let "non-real target" be an example. In the joint distribution of a certain characteristic value of the two target classes, the probabilities of the two target classes are as follows:
[0040]
[0041] d) A certain eigenvalue Probability calculation of falling in different regions According to the multiplication formula of probability, a certain eigenvalue... The probabilities of belonging to the respective categories are as follows: When a certain characteristic value or At that time, the probability of belonging to the "real target" is:
[0042] When a certain characteristic value or When the target is "not a real target", the probability is:
[0043] When a certain characteristic value At that time, the probability of belonging to the "real target" is:
[0044] When a certain characteristic value When the target is "not a real target", the probability is:
[0045] 2) Active sonar multi-dimensional feature decision-level fusion When distinguishing between "real targets" and "non-real targets" in active sonar target echoes, ,in Indicates the "real target". Indicates "non-real target", and has , , , , This indicates a situation that could be either a "real target" or a "non-real target".
[0046] Active sonar multidimensional feature decision-level fusion takes the probability of each feature of the target echo belonging to its category as input, and performs pairwise combination processing of the features. First, it calculates all possible intersections of two features and their product of mass functions, i.e. and Multiply each pairwise; then calculate the normalization factor. Finally, calculate the combined mass function:
[0047] By fusing the features according to the above method and process, the credibility of whether the active sonar echo signal is a "real target" can be obtained.
[0048] The active sonar target identification method described in this embodiment, through the above two-stage fusion process, has the following beneficial effects: Feature-level fusion effectively smooths out random interference and extracts more stable and discriminative feature representations by accumulating and fusing multiple pulses for each feature.
[0049] Among them, decision-level fusion utilizes DS evidence theory to integrate evidence from different features (such as scale and velocity) in a rigorous mathematical form. It can properly handle the uncertainty and partial conflict between various feature evidences and arrive at identification conclusions with higher overall credibility.
[0050] Combined with appendix Figures 2-6 The prior knowledge of the feature distribution shown makes the probability prediction reliable and improves the system's ability to distinguish between "real targets" and "non-real targets" in complex underwater environments, thereby significantly reducing the false alarm rate while maintaining a high recognition rate.
[0051] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.
Claims
1. An active sonar target recognition method based on multi-pulse accumulation and two-level fusion, characterized in that, Includes the following steps: Step S1: Obtain feature values of multiple features of the target tracked by active sonar, wherein the features include at least the target's scale features, radial velocity features, and absolute velocity features; Step S2: Feature-level fusion: For each single feature, the feature value of the tracked target is accumulated over multiple consecutive pulse cycles to form a short-term accumulated feature subset of the feature; based on the feature subsets of each feature, they are fused using the corresponding fusion operator to obtain the fused feature corresponding to each single feature; Step S3: Decision-level fusion: Based on the fusion features corresponding to each individual feature obtained in step S2, predict the probability that the target belongs to the "real target" class and the "non-real target" class respectively; take the class probabilities predicted by each feature as input, perform decision-level fusion, and output the final credibility that the target is the "real target".
2. The active sonar target identification method according to claim 1, characterized in that: Among them, fusion This is achieved through the following process: in, Represents a short-term accumulated feature subset of a certain feature. This represents a subset preprocessing operator that preprocesses the aforementioned feature subset. This represents the fusion operator that fuses the subset of features. This indicates the fusion feature corresponding to this feature.
3. The active sonar target identification method according to claim 2, characterized in that, Subset preprocessing operator The operations include standardizing, filtering, or extracting statistics from a subset of features.
4. The active sonar target identification method according to claim 2, characterized in that, The fusion operator The training process takes a subset of features as input and a fusion feature that can effectively represent the target category as the desired output.
5. The active sonar target identification method according to claim 1, characterized in that: In step S3, predicting the probability that the target belongs to the "real target" category or the "non-real target" category specifically includes: For each feature, the fused feature value Based on the pre-established probability density function of the "real target" with respect to this feature The probability density function of "non-real target" with respect to this feature Calculate the probability that it belongs to the "real target" class. And probability Calculated using the following formula: in, and These are the prior probabilities for the "real target" class and the "non-real target" class, respectively.
6. The active sonar target identification method according to claim 4, characterized in that: The probability density function and The features are obtained based on statistical analysis of active sonar measured data, expert experience, or data training, and are used to define typical numerical ranges of the features under different categories.
7. The active sonar target identification method according to claim 1, characterized in that: In step S3, "performing decision-level fusion" employs the DS evidence theory approach, including: The first The probability of the "true target" class obtained from feature prediction and the probability of "non-real target" Construct the basic probability assignment function for this feature. ; For all Features Combining, the basic probability assignment function after combination Calculated by the following formula: in, The value is a focal element, which can take the values of {"real target" class}, {"non-real target" class}, or their entirety; As a normalization factor; ultimately, in the case of "real target" class The value serves as the final credibility score.
8. The active sonar target identification method according to claim 1, characterized in that: The scale feature is the extended scale of the target in the radial distance dimension.