Underwater target management method and system based on neural network archive

Abstract

The invention relates to an underwater target management method and system based on a neural network archive, and the method comprises the following steps: S1, carrying out the beam forming of an array signal, generating an azimuth history diagram, detecting a target through an energy maximum value tracking method, and starting the autonomous tracking; s2, extracting a power spectrum of each tracking target and surrounding noise thereof as a neural network training sample; and S3, the deep neural network learns each tracking target, and the learning result of each target forms a neural network parameter set. According to the method, strong hypothesis of a target motion model is avoided, effective features do not need to be manually extracted, each tracked target is automatically converted into a neural network parameter set for representation, storage and indexing are facilitated, the parameter sets are uniformly managed through establishment and a neural network archive so as to finally realize target management, and the method has the advantage of result quantification and is suitable for large-scale popularization and application. And the interpretability of the target management method is improved.

Description

Technical Field:

[0001] The present invention belongs to the technical field of sonar signal tracking, and particularly relates to an underwater target management method and system based on a neural network archive. Background Art:

[0002] At present, domestic underwater tracking technologies mainly adopt a real-time processing method based on on-site data streams, discarding a lot of historical data, making very insufficient use of target information, and lacking the ability to manage on-site target and historical target information. In underwater multi-target tracking, the importance of target management cannot be ignored. In view of situations such as the temporary disappearance, occlusion, or sudden speed change of underwater targets, real-time target management can timely adjust the tracking strategy to avoid confusing the trajectories of different targets. On the other hand, by effectively utilizing the target information obtained during past tracking processes and through reasonable storage and index management of historical data, the target re-identification process can be accelerated, avoiding starting from scratch each time and improving the real-time performance of the tracking system. Existing tracking methods are difficult to be used for target management. The association method based on a preset motion model is difficult to deal with highly maneuverable targets; the association algorithm based on feature matching has poor adaptability to different environments and different noises, and it is difficult to determine effective features. Summary of the Invention:

[0003] The technical problem to be solved by the present invention is to provide an underwater target management method and system based on a neural network archive. This method uses the advantages of neural network methods in the field of artificial intelligence in dealing with complex and non-linear tasks to solve the problem of underwater acoustic target management. It not only avoids strong assumptions about the target motion model but also eliminates the need for manual extraction of effective features. Each tracked target is automatically converted into a neural network parameter set for easy storage and indexing. By establishing a neural network archive to uniformly manage each parameter set, target management is ultimately achieved. The present invention uses the target matching degree output by the neural network as the association basis, has the advantage of result quantification, and improves the interpretability of the target management method.

[0004] The technical solution of the present invention is to provide an underwater target management method based on a neural network archive, including the following steps:

[0005] S1. Perform beamforming on the array signal to generate an azimuth history diagram, detect the target using the energy maximum tracking method, and start autonomous tracking;

[0006] S2. Extract the power spectra of each tracked target and the surrounding noise as the neural network training samples;

[0007] S3. The deep neural network learns each tracked target separately, and the learning result of each target forms a neural network parameter set;

[0008] S4. Establish a neural network archive composed of various parameter sets in the background. Each time a tracking task is carried out, a real-time archive is established to record on-site targets, and the real-time archives of all historical tracking tasks are sorted into a general archive by the administrator;

[0009] S5. When a new signal appears, collect new signal samples and match them with on-site targets or historical data to distinguish whether it is a previously appeared target or a new target, and perform target association;

[0010] S6. Continue tracking, sampling, and learning, and file the new parameter set according to the matching result;

[0011] S7. When the current tracking task ends, the administrator sorts out the real-time archive and updates the general archive.

[0012] Preferably, the S1 specifically includes the following processes:

[0013] S1.1. Perform beamforming on the time-domain signals collected by the hydrophone array to generate an azimuth history diagram; the azimuth history diagram records the angle change and energy intensity information of the target relative to the array over time;

[0014] The calculation process of the maximum value tracking method is as follows:

[0015] Define BTR(t,θ) to represent the energy value at a given time t and beam angle θ, track_positions(t) to represent the beam position of the target tracked at time t, and track_energies(t) to represent the energy of the target tracked at time t. initial_beam represents the beam position of the target at the initial moment, and an energy threshold named energy_threshold is set to judge the appearance or disappearance of the target; taking a single target as an example, at the beginning t start = 1, traverse each beam angle θ to find the initial target position p start = θ i and the initial target energy E start = BTR(1,p start );

[0016]

[0017]

[0018] For each subsequent time frame t = 2, 3,..., t end First, obtain the target position at the previous moment:

[0019] p prev = track_positions(t - 1) (3)

[0020] Based on the position p at the previous moment prev , calculate the search window range at the current moment:

[0021] search_start = max(1, p prev - search_window) (4)

[0022] search_end = min(θ num, p prev + search_window) (5)

[0023] Extract the energy values within the search range at the current time point:

[0024] current_slice = BTR(t, search_start:search_end) (6)

[0025] Find the maximum energy value and its corresponding position within the current search range:

[0026] max(E current ) = max(current_slice) (7)

[0027] relative_position = argmax(current_slice) (8)

[0028] If the maximum energy E at the current moment current is greater than or equal to the energy threshold energy_threshold, update the target position and energy:

[0029] track_positions(t) = search_start + relative_position - 1 (9)

[0030] track_energies(t) = E current (10)

[0031] If the maximum energy is less than the threshold, it is considered that the target has disappeared and is set to NaN:

[0032] track_positions(t) = NaN (11)

[0033] track_energies(t) = NaN (12)

[0034] For all time steps t, the finally obtained target positions and energy values can be arranged in chronological order:

[0035] track_positions = {p1, p2,..., p T} (13)

[0036] track_energies = {E1, E2,..., E T} (14)

[0037] As long as the target energy is greater than the set energy threshold, the energy maximum tracking method continuously tracks the target and obtains the corresponding position of the target on the azimuth history diagram.

[0038] Preferably, the S2 specifically includes the following processes:

[0039] S2.1. According to the target position (time, beam) obtained in S1, extract the corresponding beam domain power spectrum, set it as a row vector, and add a label 1 at the end of the row;

[0040] S2.2. For each target, the power spectrum samples are successively superimposed as time is updated, and finally a vector group with the number of rows being the number of time frames and the number of columns being the number of power spectrum frequency points plus one column of label 1 is obtained;

[0041] S2.3. Outside the range of 10 beams to the left and right of the coordinate where the target is located, extract the beam domain power spectrum corresponding to the noise, set it as a row vector, and add a label 0 at the end of the row;

[0042] S2.4. For each target, the surrounding noise power spectrum samples are also successively superimposed as time is updated, and finally a vector group with the number of rows being the number of time frames and the number of columns being the number of power spectrum frequency points plus one column of label 0 is obtained;

[0043] S2.5. Finally, each target has a set of sample sets, and each set of sample sets only labels a unique target as label 1.

[0044] Preferably, the S3 specifically includes the following processes:

[0045] S3.1. Build a deep convolutional neural network as a classifier;

[0046] S3.2. Input each set of sample sets into the neural network for training respectively;

[0047] S3.3. Package and store the neural network parameters obtained by training each set of sample sets.

[0048] Among them, the deep neural network model is a multi-layer convolutional structure, as shown in the appendix Figure 1 shown. This model adopts a one-dimensional convolutional neural network architecture, and the input is one-dimensional power spectrum data. First, the deep convolutional and pooling layers extract high-dimensional features from the input power spectrum, and the output of the final fully connected layer is the probability of the target.

[0049] Preferably, S4 specifically includes the following processes:

[0050] S4.1. Establish a neural network archive in the background to manage all neural network parameter sets, and file the parameter sets obtained in S3 according to different targets;

[0051] S4.2. Name the neural network archive that records on-site targets as the real-time archive, and the administrator organizes the real-time archives of all historical tracking tasks into an archive general library.

[0052] Preferably, S5 specifically includes the following processes:

[0053] S5.1. When the target energy weakens and the trajectory is not obvious on the azimuth history diagram, the maximum value tracking method will lose the target. At this time, continue to retain each parameter set;

[0054] S5.2. The data is updated with the time frame. When the maximum value tracking method detects the appearance of a new signal, collect the beam domain power spectrum of the new signal;

[0055] S5.3. Input the power spectrum of the new signal into the neural network classifier. At this time, the parameters of the neural network are replaced by each parameter set in the neural network archive; input the neural network constructed by the parameter set representing target i, and the similarity with target i will be output;

[0056] S5.4. Input and traverse the neural networks composed of each parameter set in the neural network archive to obtain the similarities with each known target, so as to distinguish whether the new signal is a target that has appeared before or a new target, and perform target association or establish a new parameter set.

[0057] Preferably, the steps of S5.4 are specifically as follows:

[0058] First, traverse the parameter sets in the real-time archive. If no matching target is found, then call the archive general library for matching; by inputting and traversing the network composed of N parameter sets, N outputs are obtained, which respectively correspond to the similarity rates between the new signal and the previous N tracked targets; and set an association threshold of 0.001. If the new signal cannot be matched with each old target in the archive general library, that is, all probability values are very low, then this new signal is considered a new target that has not appeared before, and it is tracked and sampled, and the new learned parameters are put into the neural network archive; if the similarity probability between the target in the archive general library and the new signal meets the set threshold, then associate the new signal with the old target corresponding to the maximum probability.

[0059] Preferably, S6 specifically includes the following processes:

[0060] S6.1. Continuously track the new signal, repeat S2 for sampling, put the training samples into the neural network for learning, and obtain a new parameter set.

[0061] S6.2. According to the matching result of S5.4, if the new signal matches the target recorded in the real-time archive, retain the existing parameter set of the target, and add the latest parameter set on this basis. For weak targets with discontinuous trajectories, if the maximum value tracking method tracks several trajectories, the corresponding number of parameter sets will be stored in the real-time archive under this target category; if the new signal does not match the target in the real-time archive but is a target recorded in the overall archive, add the new parameter set to a new target category in the real-time archive, and the target name should be the same as that recorded in the overall archive; if the new signal fails to match all the targets on file, it is considered a new target, and the new parameter set is also placed in a new target category in the real-time archive.

[0062] Preferably, S7 specifically includes the following process:

[0063] When the current tracking task ends, the administrator sorts out the real-time archive, updates the parameter sets of the targets already recorded in the overall archive, and supplements the targets that have not been recorded yet.

[0064] Furthermore, the present invention also provides a system for implementing the above-mentioned underwater target management method based on a neural network archive. The system includes

[0065] A system construction module that constructs an underwater target tracking system based on sonar signals, and performs beamforming on the time-domain data collected by the hydrophone array to obtain an azimuth history diagram.

[0066] A detection and tracking module that uses the energy maximum value tracking method to detect and autonomously track the targets on the azimuth history diagram.

[0067] A training sample acquisition module that, according to the tracked target positions, acquires the power spectra of the targets and the surrounding noise as training samples and performs automatic label annotation.

[0068] A real-time neural network archive construction module that conducts on-site learning for each tracked target, and the learning results of each target form corresponding neural network parameter sets, and finally establishes a real-time neural network archive in the background.

[0069] A neural network overall archive calling module that collates and integrates the real-time archives of all historical tracking tasks into an overall archive, and establishes a remote connection each time a tracking task is started for convenient calling.

[0070] The target management module, when a new signal appears, first matches it with on-site targets. If there is no corresponding target, it matches with historical targets and automatically associates or adds a new set of target parameters according to the matching result.

[0071] Compared with the prior art, the present invention has the following advantages:

[0072] The method proposed by the present invention utilizes the advantages of neural networks in dealing with complex and non-linear tasks, converts each tracked target into a neural network parameter set for easy storage and indexing. Compared with the existing methods, the present invention does not require manual extraction of effective features and avoids the influence caused by strong assumptions of motion models. At the same time, by establishing a neural network archive, unified management of on-site targets and historical targets can be realized to solve the problem of insufficient application of historical data;

[0073] The present invention forms a corresponding neural network parameter set for each target, and each set of parameters only focuses on the feature learning of one tracked target, simplifying the learning process of the neural network;

[0074] In addition, the present invention uses the target matching degree output by the neural network as the association basis, has the advantage of result quantification, and improves the interpretability of the target management method;

[0075] The neural network archive proposed by the present invention includes two modules: a real-time archive and a general archive. The real-time archive requires on-site learning, but the recorded targets are more targeted, suitable for quickly associating intermittent trajectories during the tracking process, expanding short-term trajectories to long-term trajectories composed of multiple segments, and facilitating the analysis of target motion trends; The advantage of the general archive is that the training step is omitted. Even if there is a long time span between different tasks, as long as the on-file targets appear, they will be accurately identified; at the same time, the general archive effectively utilizes historical data and can continuously improve target feature information over time; after being sorted out, the general archive further subdivides events for each target, and combines information obtained by other means to completely record what type of target appeared at what time and place, and is expected to become an effective intelligence analysis means. Description of the drawings:

[0076] Figure 1 is the multi-layer convolutional neural network structure of the present invention;

[0077] Figure 2 is the comparison chart of the association effect of the method proposed by the present invention for Swellex-96 sea trial data. Detailed implementation manners:

[0078] The following further describes the present invention in detail with reference to the drawings:

[0079] The present invention first performs beamforming on array signals to generate an azimuth history diagram, detects the target using the energy maximum tracking method, and initiates autonomous tracking; to obtain neural network training samples, the power spectra of each tracked target and the surrounding noise are extracted; the deep neural network learns for each tracked target separately, and the learning result of each target forms a neural network parameter set; a neural network archive composed of each parameter set is established in the background; when a new signal appears, the similarity is calculated with known targets through the neural network archive to distinguish whether it is a target that has appeared before or a new target, and target association is performed or a new target parameter set is established.

[0080] Furthermore, the underwater target management method based on the neural network archive provided in this embodiment includes the following specific implementation steps:

[0081] S1. Perform beamforming on the time-domain signals collected by the hydrophone array to generate an azimuth history diagram. The azimuth history diagram records the angular change of the target relative to the array over time. Different colors of each beam represent the strength of the signal energy at the corresponding position. Detect the target using the energy maximum tracking method and initiate autonomous tracking;

[0082] S2. According to the target position (time, beam) obtained in S1, extract the corresponding beam-domain power spectrum, set it as a row vector, and add a label 1 at the end of the row. For each target, the power spectrum samples are updated and superimposed sequentially over time, and finally a vector group with the number of rows equal to the number of time frames and the number of columns equal to the number of power spectrum frequency points plus one column of label 1 is obtained. Outside the range of 10 beams to the left and right of the coordinate where the target is located, extract the beam-domain power spectrum corresponding to the noise, set it as a row vector, and add a label 0 at the end of the row. For each target, the surrounding noise power spectrum samples are also updated and superimposed sequentially over time, and finally a vector group with the number of rows equal to the number of time frames and the number of columns equal to the number of power spectrum frequency points plus one column of label 0 is obtained. Finally, each target has a set of sample sets, and each set of sample sets only labels a unique target as label 1;

[0083] S3. Build a deep convolutional neural network as a classifier. The deep neural network learns for each tracked target separately, and the learning result of each target forms a corresponding neural network parameter set. The obtained neural network parameters are packaged and stored, named as the parameter set of target x; in this embodiment, the deep neural network model is a multi-layer convolutional structure, as shown in the appendix Figure 1 shown. This model adopts a one-dimensional convolutional neural network architecture, and the input is one-dimensional power spectrum data. First, the deep convolutional and pooling layers extract high-dimensional features from the input power spectrum, and the output of the final fully connected layer is the probability of the target;

[0084] S4. For the parameter sets learned on-site, establish a real-time neural network archive composed of each parameter set in the background to manage all neural network parameter sets;

[0085] S5. When the target energy weakens and the trajectory is not obvious on the azimuth history diagram, the maximum value tracking method will lose the target. At this time, each parameter set is continued to be retained. As the time frame is updated, when the maximum value tracking method detects the appearance of a new signal, the beam domain power spectrum of the new signal is collected, and the power spectrum of the new signal is input into the neural network classifier. At this time, the parameters of the neural network are replaced by each parameter set in the neural network archive. Input the neural network constructed by the parameter set representing target i, and the similarity with target i will be output. First, traverse the parameter sets in the real-time archive. If there is no matching target, then call the total archive for matching.

[0086] S6. Continue to track, sample the new signal, and put the training samples into the neural network for learning to obtain a new parameter set. According to the matching result, if the new signal is the target recorded in the real-time archive, retain the existing parameter set of the target, and add this new parameter set on this basis. That is, for a weak target with discontinuous trajectories, if the maximum value tracking method tracks several trajectories, the real-time archive will store as many parameter sets as there are under this target category; if the new signal is not in the real-time archive but is a target recorded in the total archive, then put the new parameter set under the new target category in the real-time archive, and the target name should be the same as that recorded in the total archive; if the new signal fails to match all the targets on file, it is considered a new target, and the new parameter set is also put under the new target category in the real-time archive.

[0087] S7. When the current tracking task ends, the administrator sorts out the real-time archive and updates the parameter sets of the targets already recorded in the total archive, and supplements the targets that have not been recorded yet.

[0088] The following further verifies the target management ability of the present invention in combination with the Swellex-96 sea trial test data:

[0089] Use the data received by the horizontal south line array in the Swellex-96 sea trial event S59 for processing. The array sampling rate is 3276.8 Hz, use the data of 28 array elements, the processing bandwidth is 20 - 1000 Hz, the number of FFT points is 2048 points, and use conventional beamforming to obtain the azimuth history diagram. Use the energy maximum value tracking method and the method proposed in this paper to process it, as Figure 2 shown.

[0090] Event S59 records a scenario where two targets approach each other first, then their paths cross and then move away. Figure 2 Shows the verification results of two data segments. Figure 2 (a). Figure 2 (b). Figure 2(c) respectively shows the original azimuth history diagram of the first data segment, the tracking results before association, and the tracking results after association. Figure 2 (d), Figure 2 (e), Figure 2 (f) shows the verification effect of the second data segment. It can be found that before using this method for target management, each time the signal is redetected, it is regarded as a new target and marked with different colors. After using this method for target management, the mutual relationship between trajectories is restored, and it is easy to identify which trajectories come from the same target, which is beneficial to automated trajectory management.

[0091] In addition, the system for implementing the above-mentioned underwater target management method based on a neural network archive specifically includes

[0092] A system construction module that constructs an underwater target tracking system based on sonar signals and performs beamforming on the time-domain data collected by the hydrophone array to obtain an azimuth history diagram;

[0093] A detection and tracking module that uses the energy maximum tracking method to detect and autonomously track targets on the azimuth history diagram;

[0094] A training sample acquisition module that collects the power spectra of the target and surrounding noise as training samples and performs automated label annotation according to the tracked target positions;

[0095] A real-time neural network archive construction module that performs on-site learning for each tracked target, and the learning results of each target form corresponding neural network parameter sets, and finally establishes a real-time neural network archive in the background;

[0096] A neural network archive general library calling module that organizes and aggregates the real-time archives of all historical tracking tasks into an archive general library, and establishes a remote connection each time a tracking task is started for convenient calling;

[0097] A target management module that, when a new signal appears, preferentially matches it with on-site targets. If there is no corresponding target, it matches it with historical targets, and performs automatic association or adds new target parameter sets according to the matching results.

[0098] The above is only an illustration of the preferred embodiments of the present invention, and it should not be construed as a limitation to the claims. All equivalent process transformations made using the specification of the present invention are included in the patent protection scope of the present invention.

Claims

1. An underwater target management method based on a neural network archive, characterized in that: It includes the following steps: S1. Perform beamforming on the array signals to generate an azimuth history diagram, detect the target using the energy maximum tracking method, and initiate autonomous tracking; S2. Extract the power spectra of each tracking target and the surrounding noise as the neural network training samples; S3. The deep neural network learns each tracking target separately, and the learning result of each target forms a neural network parameter set; S4. Establish a neural network archive library composed of each parameter set in the background. During each tracking task, a real-time archive library is established to record the on-site targets, and the real-time archive libraries of all historical tracking tasks are sorted into an archive general library by the administrator; S5. When a new signal appears, collect new signal samples and match them with on-site targets or historical data to distinguish whether it is a target that has appeared before or a new target, and perform target association; S6. Continue tracking, sampling, and learning, and file the new parameter set according to the matching result; S7. When the current tracking task ends, the administrator sorts out the real-time archive library and updates the archive general library.

2. The underwater target management method based on a neural network archive according to claim 1, characterized in that: The specific process of S1 includes the following: S1.

1. Perform beamforming on the time-domain signals collected by the hydrophone array to generate an azimuth history diagram; the azimuth history diagram records the angle change and energy intensity information of the target relative to the array over time; S1.

2. The calculation process of the maximum value tracking method is as follows: Define \(BTR(t,\theta)\) to represent the energy value at a given time \(t\) and beam angle \(\theta\), \(track\_positions(t)\) to represent the beam position of the target tracked at time \(t\), \(track\_energies(t)\) to represent the energy of the target tracked at time \(t\), and \(initial\_beam\) to represent the beam position of the target at the initial time. Set an energy threshold, named \(energy\_threshold\), to determine the appearance or disappearance of the target; initially \(t\) start = 1, traverse each beam angle \(\theta\) to find the initial target position \(p\) start = \(\theta\) i and the initial target energy \(E\) start = \(BTR(1,p\) start ); For each subsequent time frame t = 2, 3, ..., t end First, obtain the target position at the previous moment: p prev = track_positions(t - 1) (3) Based on the position p at the previous moment prev , calculate the search window range at the current moment: search_start = max(1, p prev - search_window) (4) search_end=min(θ num, p prev +search_window) (5) Extract the energy values within the search range at the current time point: current_slice = BTR(t, search_start:search_end) (6) Find the maximum energy value and its corresponding position within the current search range: max(E current ) = max(current_slice) (7) relative_position = argmax(current_slice) (8) If the maximum energy E at the current moment current is greater than or equal to the energy threshold energy_threshold, then update the target position and energy: track_positions(t) = search_start + relative_position - 1 (9) track_energies(t) = E current (10) If the maximum energy is less than the threshold, it is considered that the target has disappeared and is set to NaN: track_positions(t) = NaN (11) track_energies(t) = NaN (12) For all time steps t, the finally obtained target positions and energy values can be arranged in chronological order: track_positions={p1,p2,...,p T} (13) track_energies = {E1, E2,..., E T} (14) When the target energy is greater than the set energy threshold, the energy maximum tracking method continuously tracks the target and obtains the corresponding position of the target on the azimuth history diagram.

3. The underwater target management method based on a neural network archive according to claim 1, characterized in that: The specific process of S2 includes the following: S2.

1. According to the target position obtained in S1, extract the corresponding beam domain power spectrum, set it as a row vector, and add the label 1 at the end of the row; S2.

2. For each target, the power spectrum samples are updated and superimposed over time. Finally, a vector group with the number of rows equal to the number of time frames and the number of columns equal to the number of power spectrum frequency points plus one column of label 1 is obtained; S2.

3. Extract the beam domain power spectrum corresponding to the noise outside the 10 beam ranges on the left and right of the coordinates where the target is located, set it as a row vector, and add the label 0 at the end of the row; S2.

4. For each target, the surrounding noise power spectrum samples are also successively superimposed as time is updated, and finally a vector group is obtained with the number of rows being the number of time frames and the number of columns being the number of power spectrum frequency points plus a column of label 0. S2.

5. Finally, each target has a set of sample sets, and each set of sample sets only labels a unique target as label 1.

4. The underwater target management method based on a neural network archive according to claim 1, characterized in that: The specific steps of S3 include the following processes: S3.

1. Build a deep convolutional neural network as a classifier. S3.

2. Input each set of sample sets into the neural network for training respectively. S3.

3. Package and store the neural network parameters obtained from training each set of sample sets.

5. The underwater target management method based on a neural network archive according to claim 1, characterized in that: The specific steps of S4 include the following processes: S4.

1. Establish a neural network archive in the background to manage all neural network parameter sets, and file the parameter sets obtained in S3 according to different targets. S4.

2. Name the neural network archive that records on-site targets as a real-time archive, and the administrator organizes the real-time archives of all historical tracking tasks into an archive master library.

6. The underwater target management method based on a neural network archive according to claim 1, characterized in that: The specific steps of S5 include the following processes: S5.

1. When the target energy weakens and the maximum value tracking method loses the target, the parameter sets are still retained at this time. S5.

2. As time frames are updated, when the maximum value tracking method detects the emergence of a new signal, collect the beam domain power spectrum of the new signal. S5.

3. Input the power spectrum of the new signal into the neural network classifier. At this time, the parameters of the neural network are replaced by each parameter set in the neural network archive; input the neural network constructed by the parameter set representing target i, and the similarity with target i will be output. S5.

4. Input the neural network formed by traversing each parameter set in the neural network archive to obtain the similarities with each known target, so as to distinguish whether the new signal is a target that has appeared before or a new target, and perform target association or establish a new parameter set.

7. The underwater target management method based on a neural network archive according to claim 6, characterized in that: The specific steps of S5.4 are as follows: First, traverse the parameter sets in the real-time archive. If no matching target is found, then call the archive master library for matching; by inputting the neural network formed by traversing N parameter sets, N outputs are obtained, which respectively correspond to the similarity rates of the new signal with the previous N tracked targets. And set an association threshold. If the new signal cannot be matched with each old target in the archive master library, then perform tracking and sampling on it, learn new parameters and put them into the neural network archive; if the similarity probability between the target in the archive master library and the new signal meets the set threshold, then associate the new signal with the old target corresponding to the maximum probability.

8. The underwater target management method based on a neural network archive according to claim 7, characterized in that: The specific steps of S6 include the following processes: S6.

1. Continue to track the new signal, repeat S2 for sampling, and put the training samples into the neural network for learning to obtain a new parameter set. S6.

2. Based on the matching results in S5.4, if the new signal matches the target recorded in the real-time archive, the existing parameter set of the target is retained, and the latest parameter set is added on this basis. For weak targets with intermittent trajectories, if the maximum value tracking method tracks several trajectories, the corresponding number of parameter sets will be stored in the real-time archive under this target category; if the new signal does not match the target in the real-time archive but is a target recorded in the total archive, the new parameter set is added to a new target category in the real-time archive, and the target name should be the same as that recorded in the total archive. If the new signal fails to match all the on-file targets, it is considered a new target, and the new parameter set is also placed in a new target category in the real-time archive.

9. The underwater target management method based on a neural network archive according to claim 1, characterized in that: The above-mentioned S7 specifically includes the following processes: When the current tracking task ends, the administrator sorts out the real-time archive, updates the parameter sets of the targets already recorded in the total archive, and supplements the targets that have not been recorded yet.

10. A system for implementing the underwater target management method based on a neural network archive as described in claim 1, characterized in that: including System construction module, constructing an underwater target tracking system based on sonar signals, and performing beamforming on the time-domain data collected by the hydrophone array to obtain an azimuth history diagram; Detection and tracking module, detecting and autonomously tracking the targets on the azimuth history diagram using the energy maximum value tracking method; Training sample acquisition module, collecting the power spectra of the target and the surrounding noise as training samples and performing automatic label annotation according to the tracked target position; Real-time neural network archive construction module, performing on-site learning for each tracked target, and the learning results of each target form corresponding neural network parameter sets, and finally establishing a real-time neural network archive in the background; Neural network total archive call module, sorting out the real-time archives of all historical tracking tasks and integrating them into a total archive, and establishing a remote connection every time a tracking task is started for easy calling; Target management module, when a new signal appears, it first matches with the on-site targets. If there is no corresponding target, it matches with the historical targets, and automatically associates or adds new target parameter sets according to the matching results.

Images