LFM frequency hopping signal long-distance target echo detection method, terminal device and storage medium

Abstract

The invention discloses an LFM frequency hopping signal long-distance target echo detection method, terminal equipment and a storage medium, and the method comprises the steps: constructing the two-dimensional time-frequency features and one-dimensional matched filtering features of a target echo, extracting the two-dimensional time-frequency features through employing a CNN network, and extracting one-dimensional time sequence features through an SE-CNN-GRU network; carrying out weighted fusion on the multi-dimensional features through a self-adaptive weighted fusion module; a multi-head attention mechanism is introduced to enhance the fused features; and outputting a classification result through the full connection layer and the Softmax layer to obtain a detection result of the target echo. According to the method, one-dimensional matched filtering and a two-dimensional time-frequency graph are selected as input features of the artificial intelligence detection method, the advantages of different artificial intelligence networks are fully utilized to perform multi-dimensional feature mining, adaptive weighted fusion is performed through complementarity and correlation of the multi-dimensional features, and the fused features are further enhanced through an attention mechanism, so that the detection accuracy of the artificial intelligence detection method is improved. Therefore, the target echo detection capability is improved, and high detection accuracy and robustness can be realized in a low signal-to-noise ratio environment.

Description

Technical Field

[0001] The present invention relates to the technical field of underwater acoustic target recognition, and in particular to an LFM frequency hopping signal long-distance target echo detection method, terminal equipment and storage medium. Background Art

[0002] In recent years, underwater target echo detection technology has gained increasing attention. Matched filtering is one of the most commonly used methods for target echo detection. However, due to underwater acoustic channel distortion, the correlation between target echoes and transmitted signals is significantly weakened, often limiting the signal matching processing gain and resulting in low target detection rates with traditional methods. Consequently, traditional target echo detection methods are no longer sufficient for long-range underwater target detection.

[0003] The development of artificial intelligence technology has provided new insights for improving underwater target echo detection capabilities. Existing intelligent underwater acoustic target recognition methods can be roughly divided into two categories: intelligent underwater acoustic target recognition based on single feature input and intelligent underwater acoustic target recognition based on multidimensional feature fusion. Intelligent underwater acoustic target recognition based on single feature input is more common in the field of passive underwater target detection. This method primarily relies on a single input feature and is relatively simple in data processing. However, since it can only describe target characteristics from a single dimension, this can lead to incomplete information, affecting the accuracy and reliability of target echo detection. In contrast, intelligent underwater acoustic target recognition based on multidimensional feature fusion fully utilizes the complementary properties of target echo features from different dimensions. The fusion of multiple features helps increase the difference between target echoes and noise interference, thereby improving detection accuracy and robustness. However, most current research on multidimensional feature fusion focuses on the field of passive underwater target detection. In active underwater detection scenarios, target echoes not only face complex time-varying channel distortion but are also affected by various interference factors. Compared with passive detection, the active detection process is more complex, making it difficult to directly apply these research results to the field of active underwater detection. Therefore, there is an urgent need to conduct in-depth exploration and research on long-distance active detection target echo detection technology based on artificial intelligence.

[0004] In recent years, a comparison of invention patent applications in the field of underwater long-distance active detection of target echoes is as follows: "Active sonar detection of target bearing based on phase space reconstruction" (Publication No.: CN118171234A, Publication Date: 2024.06.11). By extracting and fusing the multidimensional features of target echo samples, it can effectively improve the detection accuracy of active sonar in low signal-to-noise ratio environments. However, the fusion weighting coefficients used in its model are set based on engineering experience, resulting in low generalization performance of the model, making it difficult to be widely applied to various practical scenarios. "A method, system, device, and storage medium for intelligent training of underwater acoustic signals" (Publication No.: CN1116150599A, Publication Date: 2023.05.23). It uses raw data as input features of the CNN network, avoiding the phenomenon of some important features being lost when acoustically processing the sonar signal and converting it into various images. However, the data quality deteriorates under low signal-to-noise ratios, making it difficult to extract effective features of the target echo, affecting the detection effect. "A method for underwater acoustic target recognition based on an adaptive multi-feature fusion model" (publication number: CN 115909040A, publication date: 2023.04.04) provides more discriminative features for subsequent target recognition through a multi-dimensional feature extraction network structure and adaptive weighting, effectively improving recognition accuracy. However, the adaptive weighted network designed is built based on a relatively simple SE attention mechanism, which can only consider attention in the channel dimension and is not easy to capture the features of different types of information for multimodal data fusion. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a method, terminal device and storage medium for detecting long-distance target echoes using an LFM frequency hopping signal, in view of the shortcomings of the existing technology, so as to improve the detection capability of the target echoes.

[0006] To solve the above technical problems, the technical solution adopted by the present invention is: a method for detecting a long-distance target echo using an LFM frequency hopping signal, comprising the following steps:

[0007] S1, generate two-dimensional time-frequency map and one-dimensional matched filtering features of target echo;

[0008] S2. Use the two-dimensional time-frequency map as the input of the CNN branch network to extract the two-dimensional time-frequency change feature x STFT ; Use the one-dimensional matched filter feature as the input of the SE-CNN-GRU branch network to extract the one-dimensional time series information x MF ;

[0009] S3, through the adaptive weighted fusion module to the two-dimensional time-frequency change feature x STFT and one-dimensional time series information x MF Perform weighted fusion;

[0010] S4, using the multi-head attention mechanism to the fused feature vector x new Perform feature enhancement;

[0011] S5. The enhanced feature vector is passed through the fully connected layer and the Softmax layer to output the detection result. The classification detection error is calculated based on the cross entropy loss function, and the network weight is updated by relying on the classification detection error back propagation mechanism.

[0012] The network includes a CNN branch network and an SE-CNN-GRU branch network; the CNN branch network and the SE-CNN-GRU branch network are connected to an adaptive weighted fusion module; the adaptive weighted fusion module is connected to a feature enhancement module, and the feature enhancement module is connected to a classification module.

[0013] The two-dimensional time-frequency diagram is obtained by performing short-time Fourier transform on the result after the target echo beam is formed, and the one-dimensional matched filter feature is obtained by performing matched filtering on the result after the target echo beam is formed.

[0014] Extract one-dimensional time series information x MF The specific implementation process includes:

[0015] The one-dimensional matched filter features are converted into a data format suitable for the CNN network through a sequence folding operation, and then input into a CNN network with two convolutional layers;

[0016] The local feature U1 of the one-dimensional underwater acoustic signal extracted by the first convolution layer is input into the second convolution layer and the SE attention mechanism module at the same time. The attention weight S learned by the SE attention mechanism module is multiplied and weighted on the output U2 of the second convolution layer to obtain the enhanced feature.

[0017] Will After sequence expansion and Flatten feature dimensionality reduction operations, it is converted into a data format suitable for the GRU network and learned through the GRU network. The changing law of , we can get the one-dimensional time series information x contained in the one-dimensional time series feature. MF .

[0018] The fused feature vector x new The expression is: in,

[0019] x STFT-MF =w STFT-MF V MF , x MF-STFT =w MF-STFT V STFT , Q MF By x MF The calculated query vector, KSTFT By x STFT The calculated key vector, is the compression factor, Q STFT By x STFT The calculated query vector, K MF By x MF The computed key vector.

[0020] Use the multi-head attention mechanism to focus on the fused feature vector x new The specific implementation process of feature enhancement includes:

[0021] The fused feature vector H = [x new1 ,x new2 ,···,x newm ] Input feature enhancement module, use the formula Q = HW for H Q =[β1,β2,···,β μ ], K = HW K =[k1,k2,···,k μ ] and V = HW V =[v1,v2,···,v μ ] perform linear transformation and map H to query space Q, key space K and value space V respectively, where m is the number of feature vectors, W Q , W K , W V are the linear transformation parameters of query space, key space, and value space respectively, β is the element vector in the query space, k1, k2, ···, k μ is the element vector in the key space, v1,v2,···,v μ is the element vector in the value space, and u is the number of dimensions after linear transformation;

[0022] Calculate the similarity between Q and K, use the Softmax function to normalize, and obtain each attention distribution f softmax [s(β i, k j )], and perform weighted summation on the input value space V according to the attention distribution to obtain the corresponding attention output Y i : Where s(β i ,k j ) represents the query vector β i With key vector k j The scaled dot product similarity of is calculated as follows: k a Represents the key vectors at all positions in the input sequence; T is the transpose transformation, D k A matrix with the number of dimensions for each key;

[0023] Concatenate multiple attention outputs into Y to obtain the feature-enhanced result.

[0024] The expression of the cross entropy loss function Loss is: Among them, N represents the number of input samples, y (i) Indicates the true value of the classification detection result of the i-th sample, which is 1 if the true sample is the target echo, otherwise it is 0; p (i) Represents the probability of Softmax on the classification detection result of the i-th sample.

[0025] As an inventive concept, the present invention also provides a terminal device, including a memory, a processor, and a computer program stored in the memory; the processor executes the computer program to implement the steps of the above method.

[0026] As an inventive concept, the present invention also provides a computer-readable storage medium having a computer program / instruction stored thereon; the computer program / instruction implements the steps of the above method when executed by a processor.

[0027] Compared with the prior art, the present invention has the following beneficial effects: the present invention constructs two-dimensional time-frequency features and one-dimensional matched filter features of the target echo; uses a CNN network to extract two-dimensional time-frequency features, and a SE-CNN-GRU network to extract one-dimensional time series features; performs weighted fusion of multi-dimensional features through an adaptive weighted fusion module; introduces a multi-head attention mechanism to enhance the fused features; and outputs classification results through a fully connected layer and a Softmax layer to obtain the detection results of the target echo. The present invention selects one-dimensional matched filtering and two-dimensional time-frequency maps as input features of the artificial intelligence detection method, fully utilizes the advantages of different artificial intelligence networks to mine multi-dimensional features, performs adaptive weighted fusion through the complementarity and correlation of multi-dimensional features, and further enhances the fused features through an attention mechanism, thereby improving the detection capability of the target echo, and can have high detection accuracy and robustness in a low signal-to-noise ratio environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 Schematic diagram of the structure of a target echo detection network based on multi-dimensional feature adaptive weighted fusion and feature enhancement according to an embodiment of the present invention;

[0029] Figure 2 A flow chart for generating data samples in one embodiment of the present invention;

[0030] Figure 3 This is a schematic diagram of the CNN network two-dimensional time-frequency feature extraction process;

[0031] Figure 4Schematic diagram of the CNN-GRU network module results based on the SE attention mechanism;

[0032] Figure 5 It is a schematic diagram of the structure of the multi-feature adaptive weighted fusion module;

[0033] Figure 6 This is a line graph of recognition accuracy under different signal-to-noise ratios and different methods in one embodiment of the present invention;

[0034] Figure 7 This is a line chart of the recognition accuracy of different methods under time-varying channels. DETAILED DESCRIPTION

[0035] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0036] Example 1

[0037] like Figure 1 As shown, an embodiment of the present invention provides a method for target echo detection based on multi-dimensional feature adaptive weighted fusion and feature enhancement, comprising the following steps:

[0038] Step 1: simulate the time-varying underwater acoustic channel and combine it with real lake and sea test noise data to obtain a data set.

[0039] In order to obtain sample data with higher confidence, BELLHOP is used to simulate the time-varying underwater acoustic channel for a typical shallow sea acoustic environment. The simulated underwater acoustic channel, the propagation parameters of the transmitted signal and the measured lake and sea test noise data are comprehensively considered to generate target echo samples under the time-varying channel. By considering different seabed geology and typical sound velocity profiles to simulate the complex and time-varying real underwater acoustic detection scene; at the same time, combined with the real lake and sea test noise data, target echo data samples are generated to approach the real detection situation, thereby improving the credibility of the data. Finally, the time-frequency spectrum of the target echo sample and the matched filtering results are selected as the two-dimensional features and one-dimensional features of the target echo, respectively. It is difficult for a single feature to fully describe the target echo characteristics. The method of the embodiment of the present invention makes full use of the complementary characteristics of the two-dimensional features to highlight the difference between the target echo and noise interference.

[0040] The ocean environment significantly impacts underwater sound wave propagation, with key factors including seabed sediments, sound velocity profiles, sea surface conditions, and seabed topography. Therefore, typical seabed sediments and sound velocity profiles are selected as input parameters for underwater acoustic channel simulation, thereby generating a complex, time-varying underwater acoustic channel scenario.

[0041] The seafloor serves as the lower boundary of the ocean's acoustic channel. Changes in the acoustic properties of seafloor sediments can significantly affect the propagation of sound waves in seawater. Assuming a flat seafloor, eight representative seafloor sediment types were selected based on Hamilton's classification: coarse sand, fine sand, very fine sand, muddy sand, sandy mud, mud, sand-mud-clay, and clayey mud.

[0042] The data sample generation process is as follows Figure 2 As shown in the figure. First, the target echo signal is generated by convolution operation between the transmitted signal and the time-varying underwater acoustic channel simulated by BELLHOP. Secondly, assuming that the target is located at an azimuth of 45°, the generated target echo is superimposed on the lake and sea test noise data measured by the hydrophone array to simulate the signal received by the array. Then, beamforming is used to preprocess the array received signal to obtain the target echo signal at an azimuth of 45°. The signal-to-noise ratio of the target echo is adjusted by changing the amplitude of the target echo at this azimuth, and the signal-to-noise ratios are -35dB, -30dB, -25dB, -20dB and -15dB respectively. Finally, the target echo signal at this azimuth is short-time Fourier transformed to obtain its two-dimensional time-frequency diagram, and a matched filtering operation is performed on it and the transmitted signal to obtain its one-dimensional matched filtering characteristics. Similarly, the pure noise data received by the hydrophone is beamformed according to the azimuth of 45° to obtain the same amount of pure noise data as the target echo.

[0043] Step 2: Build and train a target echo detection network based on multi-dimensional feature adaptive weighted fusion and feature enhancement.

[0044] First, different feature extraction networks are built for the one-dimensional matched filtering results and two-dimensional time-frequency diagrams of the target echo. For the one-dimensional matched filtering results, the SE-CNN-GRU branch network is used for feature extraction to obtain the one-dimensional time series information x MF ; For the two-dimensional time-frequency graph, the CNN branch network is used to extract the two-dimensional time-frequency change features x STFT . Secondly, by calculating x STFT with x MF The correlation between the two factors is used to assign higher weights to features that need to be focused on, and vice versa. For example, high weights are assigned to areas in the time-frequency diagram that are strongly correlated with the one-dimensional matched filter time series (such as signal mutation points), while low weights are assigned to noise. The correlation coefficient is then normalized using the Softmax function to obtain the weight coefficient. The obtained weight coefficient is used to calculate the x MF and xSTFT After weighted decision making, we get the new feature x STFT-MF and x MF-STFT . STFT_MF and x MF_STFT Perform feature summing and fusion to obtain the new fused feature x new Then, x new The multi-head attention mechanism module is used to enhance key features. The enhanced fusion features are input into the fully connected layer and the softmax layer to output the detection results. The classification detection error is calculated based on the cross-entropy loss function, and the network weights are updated by the classification detection error backpropagation mechanism.

[0045] The network in this embodiment uses a CNN network to extract the time-frequency variation features x of the input two-dimensional time-frequency graph STFT The two-dimensional time-frequency map is used as the input of the CNN network, and the local perception of the CNN network is used to merge the extracted local information into the global two-dimensional time-frequency change feature, avoiding the perception of each pixel point, thereby reducing the complexity of feature extraction. The specific process is as follows Figure 3 shown.

[0046] In order to extract features from one-dimensional matched filtering results, a CNN-GRU network based on SE attention mechanism is designed. Its structure is as follows Figure 4 As shown. The model uses the time series signal of one-dimensional matched filtering as the model input feature parameter. The input one-dimensional matched filter is converted into a data format suitable for the CNN network through a sequence folding operation and input into a CNN network with two convolutional layers. Then, the local feature U1 of the underwater acoustic signal extracted by the first convolutional layer is input into the second convolutional layer and the SE attention mechanism module at the same time. The attention weight S learned by the SE attention mechanism module is multiplied and weighted on the output U2 of the second convolutional layer to obtain the enhanced feature. Finally, After sequence expansion and Flatten feature dimensionality reduction operations, it is converted into a data format suitable for the GRU network and learned through the GRU network. The comprehensive time series information x contained in the one-dimensional time series characteristics is obtained by the changing law of MF .

[0047] In order to better integrate different modal information, the feature vector x output by 1D-SE-CNN-GRU and 2D-CNN is MF and x STFT The importance weights of different modal data are adaptively obtained, and weighted fusion is performed to form new features. The flow chart is shown in Figure 5.

[0048] The multi-feature adaptive weighted fusion module is based on the idea of cross-attention mechanism to obtain the interaction characteristics between different modalities. First, by calculating the correlation between different features, the features that need to be focused on are given higher weights, and vice versa. Then, the correlation coefficients are normalized by the softmax function to obtain the weight coefficients. STFT The weight w of the feature needs to be paid attention to MF-STFT for:

[0049]

[0050] Where Q MF By x MF The calculated query vector, K STFT By x STFT The calculated key vector, is the compression factor, and T is the transposition transformation. Similarly, x MF The weight w of the feature needs to be paid attention to STFT-MF for:

[0051]

[0052] Where Q STFT By x STFT The calculated query vector, K MF By x MF The computed key vector.

[0053] Then, we can get x MF and x STFT New feature x after weighted decision STFT-MF and x MF-STFT :

[0054] x STFT-MF =w STFT-MF V MF

[0055] x MF-STFT =w MF-STFT V STFT

[0056] Where V MF By x MF The calculated value vector, V STFT By x STFT The vector of values to be calculated.

[0057] Finally, x STFT_MF and x MF_STFT Perform feature summing and fusion to obtain the new fused feature x new :

[0058]

[0059] The fused feature H = [x new1 ,x new2 ,···,x newm ] Input the adaptive feature fusion module and use the formula Q = HW for H Q =[β1,β2,···,β μ ], K = HW K =[k1,k2,···,k μ ] and V = HW V =[v1,v2,···,v μ ] perform linear transformation and map H to query space Q, key space K and value space V respectively, where m is the number of feature vectors, W Q , W K , W V are the linear transformation parameters of the query space, key space, and value space respectively, β is the element vector in the query space, k is the element vector in the key space, v is the element vector in the value space, and u is the number of dimensions after linear transformation.

[0060] By scaling the dot product operation, we can find the similarity between Q and K and use the Softmax function to normalize them to obtain each attention distribution f softmax [s(β i, k j )], and perform weighted summation on the input value space V according to the attention distribution to obtain the corresponding attention output Y i , as shown below:

[0061]

[0062] Where s(β i ,k j ) represents the query vector β i With key vector k j The scaled dot product similarity of is calculated as follows: k a Represents the key vectors at all positions in the input sequence; T is the transpose transformation, D k A matrix with the number of dimensions for each key.

[0063] Use the splicing method to splice multiple output results into Y (Y = f Concat (Y1,Y2,···,Y μ ), f Concat (·) is the feature concatenation function), thereby extracting key information from the fused feature data.

[0064] In the above embodiment, the network weights are iteratively updated by relying on the back propagation mechanism of the classification detection error. The calculation of the classification detection error follows the cross entropy criterion, and its specific calculation formula is expressed as follows:

[0065]

[0066] Where N represents the number of input samples, y (i) Indicates the true value of the classification result of the i-th sample, which is 1 if the true sample is the target echo, otherwise it is 0; p (i) Represents the probability of Softmax on the classification detection result of the i-th sample.

[0067] In a specific embodiment, to verify the effectiveness of the proposed model, the generated one-dimensional matched filter time series signal and two-dimensional time-frequency diagram were used as input, and experimental verification was carried out in scenarios with different signal-to-noise ratios and time-varying channels. In the experiment, the original sound wave data must first be segmented. The segmented frame length is 14 seconds, and the frame shift is 7 seconds. The segmented sound wave data must also undergo certain preprocessing, including DC removal and normalization, to ensure the consistency of the feature distribution and facilitate rapid convergence of the model.

[0068] In order to demonstrate the improvement of the method of the embodiment of the present invention compared with the previous method, an experiment of the traditional method was also designed.

[0069] In the BP algorithm, the network activation functions were 'logsig' and 'purelin', the training function was 'trainlm', the learning rate was 0.0001, the 'gauss' kernel function was selected for the SVM algorithm, the kernel parameter was set to 2, the penalty factor C = 0.2, and the number of hidden nodes in the BP algorithm was 100. To ensure more stable results, we ran each algorithm 50 times and averaged the results as the final experimental results.

[0070] Model Architecture Parameters: The proposed network model, the SE-CNN-GRU model, is designed to process one-dimensional matching sequence data and extract temporal features from it. The model architecture consists of two convolutional layers, one GRU layer, one attention layer, and a fully connected output layer. The two convolutional layers contain 32 and 64 convolutional filters, respectively, with a kernel size of 3 for each filter. These convolutional layers primarily capture local features in the input sequence, helping the model better understand the structure and patterns within the sequence. The GRU layer contains 100 hidden units and is configured to return the complete sequence, capturing temporal correlations and long-term dependencies within the sequence. To reduce the risk of overfitting, a dropout layer with a dropout rate of 0.1 is applied after the GRU layer. The CNN model is designed to process time-frequency image data and extract time-frequency features from it. The model architecture consists of two convolutional layers and a fully connected output layer. The two convolutional layers contain 64 and 8 convolutional filters, respectively, with a kernel size of 3 for each filter. To reduce the risk of overfitting, a dropout layer with a dropout rate of 0.1 is applied after the convolutional layers. In addition, the number of heads in the multi-head attention module is 8, and the Concat linear transformation parameter is 256.

[0071] The model output layer contains one fully connected layer (with two neurons), and then uses the Softmax activation function to perform category classification, thereby completing the target echo detection task. In addition, the network model optimizer is set to Sgdm, the loss function is set to Categorical_crossentropy, the batch size is set to 37, the number of training generations is set to 50, and the initial learning rate is set to 0.0001. In addition, the comparison methods mentioned in this chapter, including CNN networks, GRU networks, SE-CNN-GRU networks, and equal-weighted multi-feature fusion networks, all have the same parameter settings as the proposed method, which ensures fairness and comparability of experimental conditions, thereby enabling a more accurate evaluation of the performance of different network models.

[0072] The methods discussed are all coded using the Matlab platform, and experiments are conducted on a server equipped with an Intel i913900k@3GHz and an NVIDIAGeForce RTX4090 GPU.

[0073] To verify the proposed method's superiority in target echo recognition accuracy, 80% of the samples were randomly selected for training and 20% for testing in the generated data sets with different signal-to-noise ratios. The data samples were classified as follows: target: pure noise = 1:1 in the training set, with 471 samples in each; target: noise = 1:1 in the test set, with 117 samples in each. The data samples with different signal-to-noise ratios were used as input one by one, and the processing results of different methods are shown in the figure below. Figure 6As shown in Figure 2, TF represents the two-dimensional time-frequency graph feature and MF represents the one-dimensional matched filter feature.

[0074] Depend on Figure 6 As can be seen, first, when the SNR is -15dB, the detection accuracy of all methods is high, approximately 1. When the SNR is less than -15dB, the accuracy of CFAR detection is the lowest. When the SNR is -20dB, the detection accuracy is almost 0, indicating that when the SNR is lower than -20dB, conventional CFAR detection is no longer applicable. Therefore, it is necessary to introduce artificial intelligence technology to solve the problem of target echo detection under low SNR conditions.

[0075] When using a single feature input, traditional machine learning methods such as SVM and BP networks achieve detection accuracy exceeding 90% when the signal-to-noise ratio (SNR) is greater than -25dB, demonstrating good classification results. However, when the SNR is less than -25dB, the accuracy drops sharply, indicating that the detection performance of SVM and BP networks is limited. When using a single deep network structure, such as a CNN or GRU network, the detection accuracy is lower than that of SVM or BP networks when the SNR is less than -15dB. This may be because as the SNR decreases, the quality of the data sample deteriorates, making it difficult for a single CNN or GRU network to fully exploit the deep features of the data sample. When using a composite network, such as the SE-CNN-GRU network, the accuracy is higher than that of SVM, BP, CNN, and GRU when the SNR is less than -15dB, significantly improving detection accuracy. This demonstrates that the composite network can fully utilize the advantages of the non-network architecture, effectively mining the characteristics of the input signal and improving the accuracy of target echo detection. However, when the signal-to-noise ratio is less than -25dB, the accuracy drops significantly. This may be because the use of a single feature can no longer fully represent the target characteristics.

[0076] When multiple input features are fused with equal weights using a composite network, when the signal-to-noise ratio is greater than -25dB, the detection accuracy of this network is significantly higher than that of CFAR, SVM, BP, CNN, GRU, and SE-CNN-GRU networks. However, when SNR≤-30dB, the detection accuracy of this network is lower than that of its branch SE-CNN-GRU network, indicating that when the signal-to-noise ratio is low, features of different dimensions contribute differently to the accuracy of target echo recognition. When equal weights are used to fuse the input features, the complementarity between features of different dimensions cannot be fully utilized. However, when the proposed adaptive weighted fusion method is used, the detection accuracy is significantly improved. In addition, by Figure 6 It can be clearly seen that when the SNR is less than -20dB, the detection accuracy of the proposed method is the highest, and the average detection accuracy is improved by about 8%. Therefore, the proposed method can effectively retain and enhance complementary information, suppress redundant features, and effectively improve the accuracy of target detection.

[0077] To verify the generalization performance of the proposed method, the generalization performance of the proposed method is further tested in a real-world time-varying channel scenario, that is, when the underwater acoustic channels of the training and test data do not match. The training data uses the sea trial noise data collected in the first half of the actual measurement as the noise background, and the ocean noise in the second half of the actual measurement as the background noise of the test set. The target echoes under different channel scenarios are superimposed respectively. The data set classification is as follows: training set 1 [-35dB, -30dB, -20dB, -15dB, sea trial pure noise], test set 1 [-25dB, sea trial pure noise]; training set 2 [-35dB, -25dB, -20dB, -15dB, sea trial pure noise], test set 2 [-30dB, sea trial pure noise]; training set 3 [-30dB, -25dB, -20dB, -15dB, sea trial pure noise], test set 3 [-35dB, sea trial pure noise]; the training set: test set = 4:1. The detection results are as follows Figure 7 As shown in Figure 2, TF represents the two-dimensional time-frequency graph feature and MF represents the one-dimensional matched filter feature.

[0078] From the polyline Figure 7 It can be clearly seen that when time-varying channels are present, the feature adaptive weighted fusion model has stronger generalization capabilities for data under time-varying channels that were not trained. In fact, the classification accuracy can be improved by more than 5% compared to other methods. Even when the test set is at -35dB, it still achieves an accuracy of approximately 60%. It can be noted that the multi-dimensional feature equal weight fusion model has the lowest recognition accuracy. This is mainly due to the significant difference in recognition accuracy between the CNN branch network and the SE-CNN-GRU branch network, resulting in different weights for features of different dimensions. In addition, the proposed method's SE-CNN-GRU branch has similar recognition accuracy to BP and SVM, but both are lower than the proposed multi-dimensional feature adaptive weighted fusion model. Therefore, the proposed multi-dimensional feature adaptive weighted fusion model has better robustness and generalization in scenarios with time-varying underwater acoustic channels.

[0079] Example 2

[0080] Embodiment 2 of the present invention provides a terminal device corresponding to the above-mentioned embodiment 1. The terminal device can be a processing device for a client, such as a mobile phone, a laptop computer, a tablet computer, a desktop computer, etc., to execute the method of the above-mentioned embodiment.

[0081] The terminal device of this embodiment includes a memory, a processor, and a computer program stored in the memory; the processor executes the computer program in the memory to implement the steps of the method in the above-mentioned embodiment 1.

[0082] In some implementations, the memory may be a high-speed random access memory (RAM), and may also include a non-volatile memory, such as at least one disk storage.

[0083] In other implementations, the processor may be a central processing unit (CPU), a digital signal processor (DSP), or other general-purpose processors, which are not limited herein.

[0084] Example 3

[0085] Embodiment 3 of the present invention provides a computer-readable storage medium corresponding to the above-mentioned embodiment 1, on which a computer program / instruction is stored. When the computer program / instruction is executed by a processor, the steps of the method of the above-mentioned embodiment 1 are implemented.

[0086] Computer readable storage media can be tangible devices that hold and store instructions used by instruction execution devices. Computer readable storage media can be, for example, but not limited to, electronic storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any combination thereof.

[0087] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code. The scheme in the embodiment of the present application can be implemented in various computer languages, for example, object-oriented programming language Java and literal translation scripting language JavaScript, etc.

[0088] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the steps in the flowchart and / or block diagram. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0089] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0090] Although the preferred embodiments of the present application have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present application.

[0091] Obviously, those skilled in the art may make various changes and modifications to this application without departing from the spirit and scope of this application. Thus, if these modifications and variations of this application fall within the scope of the claims of this application and their equivalents, this application is intended to include these modifications and variations.

Claims

1. A method for detecting distant target echoes using an LFM frequency hopping signal, characterized in that: The following steps are involved: S1, generate two-dimensional time-frequency map and one-dimensional matched filtering features of target echo; S2. Use the two-dimensional time-frequency map as the input of the CNN branch network to extract the two-dimensional time-frequency change feature x STFT ; Use the one-dimensional matched filter feature as the input of the SE-CNN-GRU branch network to extract the one-dimensional time series information x MF ; S3, through the adaptive weighted fusion module to the two-dimensional time-frequency change feature x STFT and one-dimensional time series information x MF Perform weighted fusion; S4, using the multi-head attention mechanism to the fused feature vector x new Perform feature enhancement; S5. The enhanced feature vector is passed through the fully connected layer and the Softmax layer to output the detection result. The classification detection error is calculated based on the cross entropy loss function, and the network weight is updated by relying on the classification detection error back propagation mechanism. The network includes a CNN branch network and an SE-CNN-GRU branch network; the CNN branch network and the SE-CNN-GRU branch network are connected to an adaptive weighted fusion module; the adaptive weighted fusion module is connected to a feature enhancement module, and the feature enhancement module is connected to a classification module.

2. The LFM frequency hopping signal long-distance target echo detection method according to claim 1, characterized in that: The two-dimensional time-frequency diagram is obtained by performing short-time Fourier transform on the result after the target echo beam is formed, and the one-dimensional matched filter feature is obtained by performing matched filtering on the result after the target echo beam is formed.

3. The LFM frequency hopping signal long-distance target echo detection method according to claim 1, characterized in that: Extract one-dimensional time series information x MF The specific implementation process includes: The one-dimensional matched filter features are converted into a data format suitable for the CNN network through a sequence folding operation, and then input into a CNN network with two convolutional layers; The local feature U1 of the one-dimensional underwater acoustic signal extracted by the first convolution layer is input into the second convolution layer and the SE attention mechanism module at the same time. The attention weight S learned by the SE attention mechanism module is multiplied and weighted on the output U2 of the second convolution layer to obtain the enhanced feature. Will After sequence expansion and Flatten feature dimensionality reduction operations, it is converted into a data format suitable for the GRU network and learned through the GRU network. The changing law of , we can get the one-dimensional time series information x contained in the one-dimensional time series feature. MF .

4. The LFM frequency hopping signal long-distance target echo detection method according to claim 1, characterized in that: The fused feature vector x new The expression is: in, x STFT-MF =w STFT-MF V MF , x MF-STFT =w MF-STFT V STFT , Q MF By x MF The calculated query vector, K STFT By x STFT The calculated key vector, is the compression factor, Q STFT By x STFT The calculated query vector, K MF By x MF The computed key vector.

5. The LFM frequency hopping signal long-distance target echo detection method according to claim 1, characterized in that: Use the multi-head attention mechanism to focus on the fused feature vector x new The specific implementation process of feature enhancement includes: The fused feature vector H = [x new1 ,x new2 ,···,x newm ] Input feature enhancement module, use the formula Q = HW for H Q =[β1,β2,···,β μ ], K = HW K =[k1,k2,···,k μ ] and V = HW V =[v1,v2,···,v μ ] perform linear transformation and map H to query space Q, key space K and value space V respectively, where m is the number of feature vectors, W Q , W K , W V are the linear transformation parameters of query space, key space, and value space respectively, β is the element vector in the query space, k1, k2, ···, k μ is the element vector in the key space, v1,v2,···,v μ is the element vector in the value space, and u is the number of dimensions after linear transformation; Calculate the similarity between Q and K, use the Softmax function to normalize, and obtain each attention distribution f softmax [s(β i, k j )], and perform weighted summation on the input value space V according to the attention distribution to obtain the corresponding attention output Y i : Where s(β i ,k j ) represents the query vector β i With key vector k j The scaled dot product similarity of is calculated as follows: k a Represents the key vectors at all positions in the input sequence; T is the transpose transformation, D k A matrix with the number of dimensions for each key; Concatenate multiple attention outputs into Y to obtain the feature-enhanced result.

6. The LFM frequency hopping signal long-distance target echo detection method according to claim 1, characterized in that: The expression of the cross entropy loss function Loss is: Among them, N represents the number of input samples, y (i) Indicates the true value of the classification detection result of the i-th sample, which is 1 if the true sample is the target echo, otherwise it is 0; p (i) Represents the probability of Softmax on the classification detection result of the i-th sample.

7. A terminal device comprising a memory, a processor, and a computer program stored in the memory; characterized in that: The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 6.

8. A computer-readable storage medium having a computer program / instruction stored thereon; characterized in that: When the computer program / instructions are executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images