A sonar pulse interception method and product based on MCP architecture

Abstract

The present application belongs to the technical field of signal interception and passive sonar system, and relates to a sonar pulse interception method and product based on MCP architecture, which comprises the following steps: obtaining a sonar pulse signal to be analyzed; based on the MCP architecture, using a trained Mamba model as the core of time series data processing to extract the long-time dependence features of the sonar pulse signal to be analyzed, and then inputting the long-time dependence features into a large language model to obtain higher-level feature understanding; converting the generated feature understanding into features and fusing the features with the long-time dependence features to obtain the waveform category and signal parameters of the sonar pulse signal to be analyzed, so as to realize the interception and classification of the sonar pulse signal. The present application has strong operability and good application prospect, and can be applied to incoming wave interception tasks in other fields.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of signal interception and passive sonar system, and particularly relates to a sonar pulse interception method based on MCP architecture and a product. BACKGROUND

[0002] To realize the interception of sonar pulse signals, the tasks should include waveform classification, waveform parameter estimation and interception result analysis, and the parameters to be estimated include pulse frequency, bandwidth, pulse width and repetition period. When the relative motion between the sonar platform and the receiving platform cannot be known, the parameters that can be estimated are only the frequency, bandwidth, pulse width and repetition period affected by Doppler. The traditional method is to adopt multiple signal processing methods for step-by-step processing. First, the time period in which the outgoing wave signal exists is found through energy detection; then the frequency with the strongest energy is obtained according to the spectral peak value by converting to the frequency domain; the instantaneous frequency is extracted by using Hilbert transform, or parameter estimation is performed by time-frequency analysis, such as short-time Fourier transform (STFT), fractional Fourier transform (FRFT), Wigner-Ville distribution (WVD) and the like. In the processing process, multiple signal processing methods need to be combined, and multiple thresholds are set based on various rules for comprehensive judgment. The problems caused by the cascading classification and estimation parameter tasks are: 1. error accumulation: if the classification is wrong, the parameter estimation will also be affected; the order estimation between multiple parameters will also cause the transmission of errors between errors. 2. calculation redundancy: the classification and regression models run independently, and the calculation amount is large. 3. task independent optimization, insufficient information utilization, and in fact there is coupling between the classification and regression tasks, and multiple parameters also affect each other, and the order processing will certainly cause the loss of information.

[0003] In order to solve these problems, researchers have begun to explore more intelligent multi-task optimization strategies, and task cooperative processing (MCP, Multi-Task Cooperative Processing) is a computing framework developed on the basis of multi-task learning (MTL, Multi-Task Learning). It emphasizes the cooperative learning and optimization of multiple related tasks, improves the overall task performance by sharing information and features. In the fields of deep learning and signal processing, MCP is usually used to perform multiple tasks such as classification, regression, detection, segmentation, etc. at the same time, so that different tasks promote each other to achieve better results. SUMMARY

[0004] The purpose of the present application is to overcome the defects of the prior art and provide a sonar pulse interception method based on MCP architecture and a product.

[0005] Therefore, the present application provides a sonar pulse interception method based on MCP architecture, which comprises:

[0006] obtaining a sonar pulse signal to be analyzed;

[0007] Based on the MCP architecture, the trained Mamba model is used as the core of time series data processing to extract long-time dependence features of the sonar pulse signal to be analyzed, and then input into a large language model to obtain higher-level feature understanding.

[0008] The generated feature understanding is converted into a feature after fusion with the long-time dependence feature, and the waveform category and signal parameter of the sonar pulse signal to be analyzed are obtained, so as to realize the interception and classification of the sonar pulse signal.

[0009] Preferably, the input of the Mamba model is the sonar pulse signal to be analyzed, and the output is an embedding representation.

[0010] Preferably, the input of the large language model is the embedding representation, and the output is a text description through feature projection and normalization.

[0011] Preferably, the waveform category includes CW, LFM, HFM and combined waveform; and the signal parameter includes frequency, bandwidth, pulse width and time delay.

[0012] Preferably, the method further comprises a training step of the Mamba model, comprising:

[0013] Step 1: constructing a data set of sonar pulse signals through simulation and generating labels to obtain a training set;

[0014] Step 2: preprocessing the data in the training set;

[0015] Step 3: randomly shuffling the preprocessed data into several Epochs, setting the size of Batch according to the computing power in each Epoch, traversing the training data, inputting the Mamba model, inputting the embedding representation output by the Mamba model into the large language model, calculating the gradient and updating the parameters layer by layer using the joint loss function, performing supervised learning until a trained Mamba model is obtained.

[0016] Preferably, the step 1 comprises:

[0017] Step 1-1: for a transmitted signal s(t), the data received by a hydrophone is x(t), and a signal model containing multipath, reverberation and noise effects is established:

[0018]

[0019] wherein, α i is a complex propagation coefficient, τ i is the time delay of the i-th echo, K is the total number of echoes, ζ iis Doppler coefficient, r(t) is reverberation, n(t) is white noise, and t represents time domain;

[0020] Step 1-2: Randomly set the waveform type, subject to a discrete uniform distribution;

[0021] Set the original frequency in the possible frequency range, set the bandwidth, pulse width and period subject to a discrete uniform distribution;

[0022] Set the radial velocity in the possible range subject to a uniform distribution, and get the corresponding Doppler coefficient;

[0023] According to the original frequency, bandwidth, pulse width, period and Doppler coefficient, the frequency, bandwidth, pulse width and period affected by Doppler are calculated;

[0024] Randomly generate the number of multipath and delay time and sort them, let the first time delay be the direct wave and have a normalized amplitude, and the attenuation factor of other paths is set in a certain range;

[0025] Randomly set the signal-to-noise ratio to 10, 20 or 30 dB, and the signal-to-mixture ratio to 10, 20 or 30 dB;

[0026] Step 1-3: Randomly generate a number of samples, and construct training set, validation set and test set according to a certain proportion;

[0027] Step 1-4: Set the storage structure, including: file name, waveform category, sampling rate, original starting frequency, bandwidth, pulse width, period, Doppler velocity, frequency, bandwidth, pulse width, period affected by Doppler, number of multipath and time delay; For the case where there is no pulse signal, the waveform category is set to 0, and all other parameters are set to empty except the sampling rate, so as to generate a label.

[0028] Preferably, the step 2 comprises:

[0029] Step 2-1: Use a uniform sampling frequency, and the data is normalized according to the maximum possible length specification. For data that is not long enough, Gaussian noise is added at the end to realize time domain normalization;

[0030] Step 2-2: According to the frequency range of the pulse signal, the effective frequency range is cut off according to the set coefficient to realize frequency domain normalization;

[0031] Step 2-3: Perform short-time Fourier transform to obtain the time-frequency spectrum of the signal, and normalize the time domain waveform and spectrum data, and use logarithmic scale to compress the fast Fourier transform result and the short-time Fourier transform result to enhance the feature expression ability, thereby realizing the normalization of time-frequency domain;

[0032] Step 2-4: Based on the processed time-frequency features in step 2-3, further normalization processing is performed on key parameters including frequency, bandwidth, pulse width and time delay signal, the dimension is unified and the consistency of parameters is improved, so as to realize the standardization of label parameters.

[0033] Preferably, the joint loss function in step 3 is:

[0034]

[0035] wherein, is the total loss function, and are the loss functions of classification, regression and inference respectively, λ1, λ2 and λ3 are corresponding weighting coefficients, and λ3 is a variable coefficient.

[0036] In another aspect, the present application provides a computer program product comprising a computer program which, when executed by a processor, implements the steps of the above method.

[0037] Compared with the prior art, the present application has the following advantages:

[0038] The present application innovatively proposes a sonar pulse signal interception method based on a multi-task cooperative processing framework, which can realize classification and parameter estimation of pulse signals. The method effectively extracts multi-layer features of waveforms through the Mamba model, is more efficient, realizes sharing of features and reduces calculation amount: more deeply mines information, multiple tasks cooperate and promote each other: unifies optimization of the whole task flow, reduces error accumulation and realizes end-to-end optimization. This working mode has strong operability and good application prospect, and can be popularized and applied to incoming wave interception tasks in other fields. BRIEF DESCRIPTION OF DRAWINGS

[0039] Figure 1 is a method flowchart of a sonar pulse interception method based on the MCP framework of the present application;

[0040] Figure 2 is a training process, wherein Figure 2(a) is the change of training loss and validation loss with training rounds (Epoch), Figure 2(b) is the change of training accuracy and validation accuracy with training rounds, Figure 2(c) is the prediction accuracy of different parameters (such as fo_acc, B_acc, T_acc, delay_acc) with training rounds, and Figure 2(d) is the relative error of different parameters (such as fo_rel, B_rel, T_rel, delay_rel) with training rounds;

[0041] Figure 3 is a confusion matrix (Confusion Matrix);

[0042] Figure 4 is an estimation of parameters, showing the relationship between the true value and the predicted value of the model when predicting different parameters, wherein Figure 4(a) is the prediction result of the f0 parameter, Figure 4(b) is the prediction result of the B parameter, Figure 4(c) is the prediction result of the T parameter, and Figure 4(d) is the prediction result of the delay parameter. DETAILED DESCRIPTION

[0043] In order to better solve the problem of pulse signal interception, a multi-task framework is proposed, which shares the same feature extraction network for classification, regression and reasoning tasks. Classification can help regression reduce the search range and improve accuracy.

[0044] The present application is directed to the interception of sonar pulses, and based on a multi-task framework, a classification and parameter estimation method for unknown sonar pulse signals is designed, comprising:

[0045] Obtain the sonar pulse signal to be analyzed;

[0046] Based on the MCP architecture, the trained Mamba model is used as the core of time series data processing to extract the long-time dependence features of the sonar pulse signal to be analyzed, and then input into the large language model to obtain higher-level feature understanding;

[0047] The generated feature understanding is converted into a feature after fusion with the long-time dependence feature, and the waveform category and signal parameter of the sonar pulse signal to be analyzed are obtained, thereby realizing the interception and classification of the sonar pulse signal.

[0048] The technical solutions of the present application will be described in detail below in conjunction with the drawings and embodiments.

[0049] Embodiment 1

[0050] Embodiment 1 of the present application provides a sonar pulse interception method based on MCP architecture, as shown in Figure 1 The steps are as follows:

[0051] MCP extracts the global time series information of the signal by adopting a structure of shared feature layer and task-specific branch, and then realizes three branch tasks: classification branch (output waveform category), parameter regression branch (output signal parameter), and interception analysis branch (output whether it meets the interception condition).

[0052] I. Data set construction

[0053] a) Construct the data set of sonar pulse signals by simulation

[0054] Suppose the received data is x(t). Considering the environment of receiving the sonar pulse, a signal model is established, which contains multipath, reverberation, and noise effects, as follows,

[0055]

[0056] The received data is superimposed by several path signals, reverberation r(t), and white noise n(t), while the pulse signal echoes are all from the transmitted signal s(t) of a sonar platform through different paths.α i is the complex propagation coefficient, whose amplitude is the attenuation of the path echo, and whose phase contains the phase of the high-frequency carrier and the phase change of the signal in the propagation process.τ i is the time delay of the path echo. Due to the radial velocity between the transmitting platform and the receiving platform, the pulse signals received by each path are affected by the Doppler effect, and there is a Doppler coefficientζ i .

[0057] b) Waveform type

[0058] According to the commonly used pulses of sonar, support CW, LFM, HFM, and combined waveforms, and generate different samples according to random parameters. According to the possible bandwidth, pulse width, and period, set the parameters of the waveform.

[0059] Considering the possible relative motion between the sound source and the receiving platform, set the Doppler coefficient according to the relative velocity v where c is the speed of sound. And adjust the frequency to match different motion conditions.

[0060] c) Multipath effect

[0061] Multipath includes direct wave, path through sea surface, path through sea bottom, etc. Multiple propagation paths with different delays and amplitudes simulate the arrival of signals on different paths. The time delay of the direct wave should be the smallest, and the amplitude should be the largest; the time delay of other paths can be set according to the simulated ocean environment, the depth of the sound source and the receiving platform. The amplitude difference between other paths and the direct wave can be set according to the Lambert empirical formula, etc., which mainly depends on the grazing angle, the flatness of the interface, the seabed bottom, etc.

[0062] d) Reverberation

[0063] For the listening mode, after receiving the direct wave, due to the influence of various scatterers in the sea surface, sea bottom, and water body, long-time reverberation will continue to be received. Due to the possible influence of the last cycle, reverberation may continue to exist. Reverberation can be constructed by convolving random coherent noise, spectrum expansion, and transmitted pulses. Reverberation can also be constructed according to a spatial multiple scattering point model.

[0064] e) Noise

[0065] Gaussian white noise is set to simulate the impact of ocean environmental noise and self-noise on the receiving platform.

[0066] f) Simulation parameters

[0067] Randomly set the waveform type subject to a discrete uniform distribution; set the original frequency, bandwidth, pulse width, and period subject to a discrete uniform distribution in the constraint condition; set the radial velocity subject to a uniform distribution in the possible range and obtain the corresponding Doppler coefficient; calculate the frequency, bandwidth, pulse width, and period affected by the Doppler according to the original frequency, bandwidth, pulse width, period, and Doppler coefficient. Randomly generate the number of multipaths and delay time and sort them, with the first delay being the direct wave and having a normalized amplitude. Set the attenuation factor of other paths to be between 0.2 and 0.7 (which can be adjusted according to actual conditions). Randomly set the signal-to-noise ratio to be 10 / 20 / 30 dB and the signal-to-mixing ratio to be 10 / 20 / 30 dB.

[0068] g) Generate data

[0069] Randomly generate 10,000 samples and construct training set, validation set, and test set in the ratio of 8:1:1. Save each sample as a bin file.

[0070] h) Generate labels

[0071] Store the structure containing the file name, waveform category, sampling rate, original starting frequency, bandwidth, pulse width, period, Doppler velocity, frequency affected by Doppler, bandwidth, pulse width, period, number of multipaths, delay, etc. as a.csv file.

[0072] For CW signals, the bandwidth is set to the inverse of the pulse width. Since the pulse width range is set to 20ms-1s, the bandwidth range is 1Hz-50Hz. For non-CW signals, the minimum bandwidth is 200Hz.

[0073] For the case where there is no pulse signal, the waveform category is set to 0 and other parameters are set to empty except for the sampling rate.

[0074] II. Data preprocessing

[0075] a) Time domain normalization

[0076] Normalize the data according to the maximum possible length with a uniform sampling frequency. For data that is not long enough, supplement it at the end with very low Gaussian noise.

[0077] b) Frequency domain normalization

[0078] Assuming the frequency range of the pulse signal is f1-f2, after FFT, the effective frequency range is truncated according to f1*0.5-f2*1.2.

[0079] c) Time-frequency domain normalization

[0080] In order to balance the resolution of frequency and time, a higher overlap coefficient is set, and a short-time Fourier transform is performed to obtain a time-frequency spectrum.

[0081] The normalized time-domain waveform and spectrum data are processed to facilitate the processing of large models and deep learning models.

[0082] The FFT and STFT results are processed on a logarithmic scale to enhance the dynamic range.

[0083] d) Label parameter normalization

[0084] Due to the large difference in the numerical range of the four parameters, the frequency is of the order of thousands, while the time delay may be from microseconds to tens of milliseconds. The numerical value of the bandwidth is also related to the signal type. For CW signals, it is within tens of Hz (depending on the pulse width); for non-CW signals, it is greater than 200 Hz. According to the statistical characteristics of different parameters, the normalization method is designed respectively, and the distribution characteristics of the bandwidth parameter with different signal types need to be considered.

[0085] III. MCP task construction

[0086] For sonar pulse signals, the Mamba model is used as the core architecture for time series data processing to extract long-time dependence features. In addition to the interception result analysis task, which should be relatively late, the classification and parameter estimation tasks should complement each other.

[0087] a) Define the MCP task flow

[0088] Create MCP tasks, including signal classification, parameter estimation, and reasoning.

[0089] b) Joint optimization of multi-branch tasks

[0090] In the multi-task collaborative framework, Mamba is used as the core of time series data processing. Large models such as DeepSeek are introduced to enhance decision-making: the embedding generated by Mamba is input into the large model to obtain a higher level of feature understanding. The waveform pattern is analyzed through a natural language reasoning mechanism, and auxiliary decision-making is generated. A joint loss function is used to simultaneously optimize classification, regression, and reasoning tasks:

[0091]

[0092] where, is the total loss function, and are the loss functions of classification, regression, and reasoning, respectively, and λ1, λ2, and λ3 are their weighting coefficients. Among them, λ3 is a variable coefficient, which is set to 0 in the first several epochs of training.

[0093] IV. Mamba Processing

[0094] Mamba model is used as the core architecture for time series data processing to extract long-term dependency features.

[0095] a) Mamba Model Feature Extraction

[0096] Input: Time-domain waveform, FFT spectrum, STFT time-frequency spectrum.

[0097] Task: Perform feature extraction and generate embedding representation.

[0098] b) Training Mamba Model:

[0099] Supervised Learning: Train waveform classification tasks (such as CW, LFM, HFM, combined waveform) using simulation data sets.

[0100] Parameter Regression: Train Mamba model to predict parameters such as frequency, bandwidth, arrival time, pulse width, etc.

[0101] Output head includes classification head (predict waveform type) and parameter head (predict signal parameters f0, B, T, delay)

[0102] V. LLM (Large Language Model) Interface Processing

[0103] a) Feature Extraction

[0104] Receive features extracted by Mamba model, then perform feature projection and normalization.

[0105] b) LLM Call

[0106] Construct prompt, call external or local LLM API, and process returned text description.

[0107] c) Feature Fusion

[0108] Convert LLM-generated description to features and fuse with Mamba features.

[0109] VI. Training Process

[0110] a) Batch Processing

[0111] Randomly shuffled, trained for 100 Epochs. In each Epoch, set the size of Batch according to computing power, and traverse the training data.

[0112] b) Forward Propagation

[0113] Finally produce prediction output through each component of the model.

[0114] c) Backpropagation

[0115] Backpropagation starts with loss calculation, then calculates gradients layer by layer and updates parameters.

[0116] 7. Testing and Evaluation

[0117] a) Performance evaluation

[0118] Comprehensively evaluate performance through confusion matrix analysis, accuracy of each category, misclassification analysis, and parameter prediction performance

[0119] b) Visual analysis

[0120] Visualize training history, including loss curves, accuracy curves, and parameter prediction errors; visualize prediction results, including waveform type predictions, parameter prediction scatter plots, and error distribution plots.

[0121] The results of classification and parameter estimation are given based on the test data.

[0122] Example 2

[0123] Example 2 of the present invention provides a sonar pulse interception method based on the MCP architecture, specifically including: constructing a simulation experiment: constructing a simulation data set according to the above method, and performing training, verification, and testing. The results are as follows.

[0124] The estimates for the two groups of samples in the test set are as follows:

[0125] Table 1

[0126] Signal type Frequency Bandwidth Pulse width Latency Actual value CW+LFM 4919.85 1297.85 0.812343 0.001946 Estimated value CW+LFM 4887.33 1289.69 0.831973 0.002267 0.66％ 0.63％ 2.42％ 16.48％

[0127] Table 2

[0128] Signal type Frequency Bandwidth Pulse width Latency Actual value CW 5663.12 2.16 0.462445 0.004700 Estimated value CW 5645.56 2.11 0.474499 0.004917 0.315％ 2.61％ 2.61％ 4.61％

[0129] Figure 2 shows the training process, where Figure 2(a) is the change of training loss and validation loss with the training epoch, which is used to evaluate the fitting ability and generalization ability of the model; Figure 2(b) is the change of training accuracy and validation accuracy with the training epoch, which reflects the performance of the model in the classification task; Figure 2(c) is the change of the prediction accuracy of different parameters (such as fo_acc, B_acc, T_acc, delay_acc) with the training epoch, which evaluates the performance of the model in the multi-target prediction task; Figure 2(d) is the change of the relative error of different parameters (such as fo_rel, B_rel, T_rel, delay_rel) with the training epoch, which measures the accuracy of the model prediction.

[0130] Figure 3 shows the confusion matrix, where the horizontal axis represents the predicted value, and the vertical axis represents the true value. The value in each cell represents the classification result of the model on the corresponding category, and the dark blue color represents a higher sample number. Specifically, the values on the diagonal line (such as 2538, 2416, 2419, 2512) represent the number of samples correctly classified by the model, while the values on the other non-diagonal lines (such as 71, 43, 1) represent the classification errors of the model. Figure 3

[0131] Figure 4 shows the estimation of parameters, which shows the relationship between the true value and the predicted value of the model when predicting different parameters, where Figure 4(a) is the prediction result of the f0 parameter, and the scatter points in the figure represent the corresponding relationship between the true value and the predicted value, and the red dashed line is the ideal consistency line (i.e. the predicted value is equal to the true value); Figure 4(b) is the prediction result of the B parameter, which also shows the accuracy of the prediction through scatter points and consistency line; Figure 4(c) is the prediction result of the T parameter, which further evaluates the prediction ability of the model on this parameter; Figure 4(d) is the prediction result of the delay parameter, which reflects the performance of the model in the delay prediction task through scatter point distribution and consistency line.

[0132] Example 3

[0133] The embodiments of the present application can also provide a computer program product, which includes a computer program. When the computer program is executed by a processor, it can implement each step in the above-mentioned method embodiments.

[0134] ​Finally, it should be noted that the above examples are merely used to illustrate the technical solutions of the present application but not to limit. Although the present application is explained in detail with reference to the examples, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced equivalently without departing from the spirit and scope of the present application, and all of them should be covered in the scope of the claims of the present application.

Claims

1. A sonar pulse interception method based on the MCP architecture, comprising: Obtaining sonar pulse signals to be analyzed; Based on the MCP architecture, the trained Mamba model is used as the core of time series data processing to extract the long-term dependency features of the sonar pulse signal to be analyzed, and then input into the large language model to obtain a higher level of feature understanding; The generated feature understanding is converted into features and then fused with long-term dependent features to obtain the waveform category and signal parameters of the sonar pulse signal to be analyzed, thereby achieving the interception and classification of the sonar pulse signal; The method also includes a Mamba model training step, including: Step 1: Construct a dataset of sonar pulse signals through simulation and generate labels to obtain a training set; Step 2: Preprocess the data in the training set; Step 3: Randomly shuffle the preprocessed data and divide it into several epochs. In each epoch, set the batch size based on computing power. Iterate over the training data and input it into the Mamba model. Then, input the embedded representation output by the Mamba model into the large language model. Using a joint loss function, calculate the gradient layer by layer and update the parameters for supervised learning until a trained Mamba model is obtained. The step 1 comprises: Step 1-1: For a transmitted signal s(t), the data received by a hydrophone is x(t). Build a signal model that includes the effects of multipath, reverberation, and noise: Among them, α i is the complex propagation coefficient, τ i is the time delay of the i-th echo, K is the total number of echoes, ζ i is the Doppler coefficient, r(t) is reverberation, n(t) is white noise, and t represents the time domain; Step 1-2: Randomly set the waveform type to obey the discrete uniform distribution; Set the original frequency to the possible frequency range, and set the bandwidth, pulse width and period to obey discrete uniform distribution; Set the radial velocity to obey uniform distribution in the possible range and obtain the corresponding Doppler coefficient; The frequency, bandwidth, pulse width and period affected by Doppler are calculated based on the original frequency, bandwidth, pulse width, period and Doppler coefficient; The number of multipaths and delay times are randomly generated and sorted, with the first delay being the direct wave and having a normalized amplitude. The attenuation factors of the other paths are set within a set range. Randomly set the signal-to-noise ratio to 10, 20 or 30dB, and the signal-to-mixing ratio to 10, 20 or 30dB; Steps 1-3: Randomly generate several samples and construct training sets, validation sets, and test sets according to a certain ratio; Steps 1-4: Set up the storage structure, including: file name, waveform type, sampling rate, original starting frequency, bandwidth, pulse width, period, Doppler velocity, frequency affected by Doppler, bandwidth, pulse width, period, number of multipaths, and delay; if there is no pulse signal, the waveform type is set to 0, and all parameters except the sampling rate are set to empty, thus generating a label; The step 2 includes: Step 2-1: Use a unified sampling frequency and standardize the data according to the maximum possible length. For data with insufficient length, add Gaussian noise at the end to achieve time domain normalization; Step 2-2: According to the frequency range of the pulse signal, the effective frequency range is intercepted according to the set coefficient to achieve frequency domain normalization; Step 2-3: Perform short-time Fourier transform to obtain the time-frequency spectrum of the signal. At the same time, normalize the time domain waveform and spectrum data, and compress the fast Fourier transform results and short-time Fourier transform results using a logarithmic scale to enhance the feature expression capability, thereby achieving normalization of the time-frequency domain. Step 2-4: Based on the time-frequency features processed in step 2-3, further normalize the key parameters of the signal, including frequency, bandwidth, pulse width, and delay, to unify the dimensions and improve the consistency of the parameters, thereby achieving standardization of the label parameters; The joint loss function in step 3 is: in, is the total loss function, and are the loss functions for classification, regression, and inference respectively, λ1, λ2, and λ3 are the corresponding weighting coefficients, and λ3 is a variable coefficient.

2. The sonar pulse interception method based on the MCP architecture according to claim 1 is characterized in that: The input of the Mamba model is the sonar pulse signal to be analyzed, and the output is the embedded representation.

3. The sonar pulse interception method based on the MCP architecture according to claim 1 is characterized in that: The input of the large language model is the embedding representation, and through feature projection and normalization, the output is a text description.

4. The sonar pulse interception method based on the MCP architecture according to claim 1 is characterized in that: The waveform categories include: CW, LFM, HFM and combined waveforms; the signal parameters include: frequency, bandwidth, pulse width and delay.

5. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method according to claim 1 are implemented.

Images