A small-sample underwater target recognition method based on deep transfer learning of multi-platform auditory perception features

Abstract

This invention proposes a method for underwater target recognition based on deep transfer learning of multi-platform auditory perception features. This method establishes a multi-deep learning joint decision model consisting of an MFCC feature extraction network, a GFCC feature extraction network, a CFCC feature extraction network, and a fusion recognition network. Furthermore, a multi-platform data transfer learning system is established. This system leverages target data from other sonar platforms, simulated sonar target data, speaker voice data, and other data with referenceable features to perform transfer learning training on the deep learning model, specifically targeting the small sample data conditions of the method's application platform. Compared to traditional target recognition methods based on physical features, this method offers more in-depth and comprehensive feature mining and utilization. Furthermore, compared to intelligent recognition models driven solely by platform data, this method utilizes a wider range of data, and the model training process is less susceptible to overfitting. This method can improve sonar target recognition capabilities under small sample data conditions.

Description

Technical Field

[0001] The present invention belongs to the field of underwater target recognition and artificial intelligence technology, and specifically relates to a small-sample underwater target recognition method based on multi-platform auditory perception feature deep transfer learning. Background Art

[0002] Underwater target recognition technology can provide sonar operators with target feature information and identify target types. It is an important basis for making comprehensive decisions and is also one of the most important research directions in sonar signal processing.

[0003] The core research topic of underwater target recognition is the extraction and expression of target signal acoustic features. Traditional underwater acoustic target feature acquisition methods primarily build models based on existing target data and characteristic knowledge. However, due to factors such as multi-target interference, spatiotemporal variations in ocean channels, and platform and environmental noise, obtaining clean, clear, and realistic target features is difficult, creating a bottleneck that restricts the development of traditional recognition technology.

[0004] In recent years, deep learning technology has developed rapidly and has been widely used in fields such as speech and images. It has also attracted the attention of many scholars at home and abroad in the field of underwater acoustics. However, due to factors such as experimental methods, it is difficult to obtain high-quality target data from most sonar platforms, resulting in poor generalization ability of intelligent models directly driven by these small-sample sonar platform target data. Summary of the Invention

[0005] The purpose of the present invention is to provide a small-sample underwater target recognition method based on deep transfer learning of multi-platform auditory perception features, so as to solve the problem of weak generalization ability of intelligent models driven by small-sample sonar platform target data in the background technology.

[0006] A method for recognizing underwater targets with a small sample size based on deep transfer learning of multi-platform auditory perception features is disclosed. The method comprises the following steps: constructing a multi-depth learning joint decision model for multi-dimensional underwater target auditory perception features; generating a multi-platform sample set based on multi-platform data; conducting multi-level transfer learning training on the multi-depth learning joint decision model based on the multi-platform sample set; preprocessing radiation noise data of unknown underwater targets, performing target recognition on the preprocessing results using the trained multi-depth learning joint decision model, and outputting the recognition results.

[0007] Preferably, the method for constructing the multi-depth learning joint decision model is: constructing an MFCC feature extraction network, a GFCC feature extraction network and a CFCC feature extraction network, placing the above three feature extraction networks in parallel and adding a fusion recognition network and a support vector machine classifier on top in sequence; the fusion recognition network is a two-layer perceptron network; initializing the parameters of the multi-depth learning joint decision model; constructing the loss function J of the multi-depth learning joint decision model, and setting the training parameters during training.

[0008] Preferably, the construction method of the MFCC feature extraction network, the GFCC feature extraction network or the CFCC feature extraction network is: adding a data input layer; adding a first convolutional layer, the parameters of the first convolutional layer are (48×1, 144, 4); adding an lrn layer; adding a second convolutional layer, the parameters of the second convolutional layer are (16×1, 192, 2); adding a third convolutional layer, the parameters of the third convolutional layer are (7×1, 256, 1); adding an lrn layer; adding a first pooling layer, the parameters of the first pooling layer are (3×1, 2); adding three basic ResNet-Inception modules; adding a fourth convolutional layer, the parameters of the fourth convolutional layer are (3×1, 384, 1); adding a fifth convolutional layer, the parameters of the fifth convolutional layer are (3×1, 512, 1); adding a second pooling layer, the parameters of the second pooling layer are (3×1, 2); adding a fully connected layer, and the number of output nodes is 256.

[0009] Preferably, three parallel branches are added after the data input layer, branch 1 is a direct branch; branch 2 includes two convolutional layers, and the parameters of each convolutional layer are (1×1, 128, 1) and (3×3, 128, 1) respectively; branch 3 includes two convolutional layers, and the parameters of each convolutional layer are (1×3, 128, 1) and (3×1, 128, 1) respectively; a feature dimension expansion operation is added after branches 2 and 3, and the output features of the two branches are integrated to obtain a comprehensive feature set output, with a total number of features of 256; a direct addition and summation operation is added after the output features of branch 1 and the comprehensive feature set to obtain the top-level features of the basic ResNet-Inception module, and then a ReLU activation function is added to output the final convolution feature.

[0010] Preferably, the initialization of the multi-deep learning joint decision model parameters includes generating 48-channel auditory perception filter coefficients according to the set frequency band based on the MFCC, GFCC, and CFCC auditory perception filter construction method, each auditory perception filter is a 48-dimensional vector, and all auditory perception filters are integrated to form a 48×144 matrix, and the parameters of the first convolutional layer of the MFCC feature extraction network, the GFCC feature extraction network, and the CFCC feature extraction network are initialized respectively.

[0011] Preferably, the multi-platform sample set is generated based on multi-platform data by utilizing a speaker speech dataset, a simulated sonar target dataset, other sonar platform target datasets, and a small sample sonar platform target dataset to be applied, based on MFCC, GFCC, and CFCC auditory perception filters, setting the upper and lower limit frequencies of data processing according to requirements, generating corresponding auditory perception feature samples, and constructing a multi-platform sample set.

[0012] Preferably, the multi-level transfer learning training of multiple deep learning joint decision models based on multi-platform sample sets includes the following steps: initializing training of multiple deep learning joint decision models; conducting primary transfer learning training on multiple deep learning joint decision models; and conducting advanced transfer learning training on multiple deep learning joint decision models.

[0013] Preferably, the initialization training of the multi-deep learning joint decision model includes: a loop optimization step, randomly selecting a fixed number of samples from the speaker speech sample set and optimizing the loss function J for several rounds based on the gradient descent method, and then randomly selecting a fixed number of samples from the simulated sonar target sample set and optimizing the loss function J for several rounds based on the gradient descent method; repeating the above loop optimization steps until the number of loop optimizations reaches a preset number of loop rounds, until the loss function converges to the expected threshold.

[0014] Preferably, the primary transfer learning training of the multi-deep learning joint decision model includes: retaining the structural parameters of the MFCC feature extraction network, GFCC feature extraction network, and CFCC feature extraction network during initialization training, randomly initializing the structural parameters of the fusion recognition network, randomly selecting a fixed number of samples from other sonar target sample sets, and optimizing the loss function J for several rounds based on the gradient descent method until the loss function converges to the expected threshold.

[0015] Preferably, the advanced transfer learning training for the multi-depth learning joint discrimination model includes the following steps: extracting the first local network structure of the multi-depth learning joint discrimination model and retaining the network structure parameters to obtain the feature extraction model S E ={s M , s G , s C}, the first local network structure includes the corresponding three basic ResNet-Inception modules in the MFCC feature extraction network, the GFCC feature extraction network, and the CFCC feature extraction network and the previous network structure s M 、s G 、s C ; Extract the second local network structure of the multi-deep learning joint discriminant model and form it into a feature recognition model S R , construct feature recognition model S RThe loss function J m The second local network structure is the corresponding fourth convolutional layer l in the MFCC feature extraction network, GFCC feature extraction network, and CFCC feature extraction network M 、l G 、l C To the network structure of softmax classifier; through S E The feature sample set x is obtained by extracting features from all samples in the target sample set of the small sample sonar platform. E ={x E，1 , x E，2 ,…,x E，n ,(x E，n =(x M，n , x G，n , x C，n ), n∈N * )}, where the feature sample x E，n Contains a set of three features x M，n 、x G，n and x C，n , corresponding to s M 、s G 、s C The feature extraction results of model S R The structural parameters are randomly initialized; from x E A fixed number of feature samples are randomly selected and the loss function J is calculated based on the gradient descent method. m Perform several rounds of optimization until the loss function converges to the expected threshold; reintegrate S E and S R , get the trained multi-deep learning joint decision model S D .

[0016] The present invention proposes a small sample underwater target recognition method based on multi-platform auditory perception feature deep transfer learning. First, according to the characteristics of underwater acoustic target noise data, a deep learning method is combined with Mel-Frequency Cepstral Coefficients (MFCC), Gammatone Frequency Cepstral Coefficients (GFCC), Cochlear Frequency Cepstral Coefficients (CFCC) and other factors to identify underwater targets. Coefficients (CFCC) and other auditory perception theories, a multi-deep learning joint decision model consisting of MFCC feature extraction network, GFCC feature extraction network, CFCC feature extraction network and fusion recognition network is constructed. The model structure parameters are initialized based on MFCC, GFCC and CFCC filter coefficients. Secondly, standardized MFCC samples, GFCC samples and CFCC samples are generated for multi-source and multi-platform data such as speaker speech data, other sonar platform target data, target data of the sonar platform to be applied, and simulated sonar target data. Then, multi-level transfer learning training is carried out on the multi-deep learning joint decision model based on the multi-platform sample set to fully utilize the available data to optimize the model structure parameters. Finally, unknown underwater acoustic target data is recognized based on the optimized multi-deep learning joint decision model.

[0017] Compared with the prior art, the present invention has the following beneficial effects:

[0018] The present invention comprehensively uses auditory perception, deep networks and transfer learning. On the one hand, samples are constructed and deep networks are initialized based on auditory perception filters, so that the network structure parameters are more adapted to the samples. At the same time, by establishing a multi-deep learning joint decision model driven by multiple auditory perception feature samples, the separability information of the target features in different feature dimensions can be fully utilized. On the other hand, a multi-platform data transfer learning system is established. In view of the small sample data conditions of the method application platform, multi-platform data is fully utilized to optimize the deep learning model, which is expected to improve the tolerance of target recognition.

[0019] Compared with traditional target recognition methods based on physical features, feature mining and utilization are more in-depth and comprehensive; compared with intelligent recognition models that are simply driven by data on this platform, the scope of data utilization is wider, the model training process is not prone to overfitting, and the sonar target recognition capability under small sample conditions can be improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 This is the overall solution for small-sample underwater target recognition based on multi-platform auditory perception feature deep transfer learning of the present invention.

[0021] Figure 2 This is a flowchart of small-sample underwater target recognition based on multi-platform auditory perception feature deep transfer learning in this embodiment.

[0022] Figure 3 t-SNE visualization projection results for original features and deep learning features.

[0023] Figure 4 Schematic diagram of the framework structure of the basic ResNet-Inception module.

[0024] Figure 5 Schematic diagram of the framework structure of the MFCC feature extraction network.

[0025] Figure 6 The following are the statistical results of the recognition accuracy of the method proposed in this invention and various comparative methods. DETAILED DESCRIPTION

[0026] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. 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 any creative efforts shall fall within the scope of protection of the present invention.

[0027] Figure 1 For the overall solution of the present invention, refer to Figure 1 , a small-sample underwater target recognition method based on deep transfer learning of multi-platform auditory perception features, the method includes the following four steps.

[0028] Step 1: Construct a multi-deep learning joint decision model for multi-dimensional underwater target auditory perception features;

[0029] Step 2: Generate a multi-platform sample set based on multi-platform data;

[0030] Step 3: Conduct multi-level transfer learning training on multiple deep learning joint decision models based on multi-platform sample sets;

[0031] Step 4: Preprocess the radiation noise data of unknown underwater targets, use the trained multi-deep learning joint decision model to perform target recognition on the preprocessing results, and output the recognition results.

[0032] Reference Figure 2 As part of the stage, the specific process of constructing a multi-depth learning joint decision model for multi-dimensional underwater target auditory perception features in step 1 is as follows.

[0033] Step 1.1, build the basic ResNet-Inception module, refer to Figure 4As shown in the figure, three parallel branches are added after the data input layer. Branch 1 is a direct branch and does not add any operations. Branch 2 includes two convolutional layers. The parameters of convolutional layer 1 are (1×1, 128, 1), that is, the convolution kernel size is 1×1, the number of convolution kernels is 128, and the convolution step is 1. The representation method is the same below. The parameters of convolutional layer 2 are (3×3, 128, 1). Branch 3 includes two convolutional layers. The parameters of convolutional layer 1 and convolutional layer 2 in this branch are (1×3, 128, 1) and (3×1, 128, 1) respectively. A feature dimension expansion operation is added after branches 2 and 3. By integrating the output features of the two branches, a comprehensive feature set output is obtained, with a total number of features of 256. A direct addition and summation operation is added after the output features of branch 1 and the comprehensive feature set to obtain the top-level features of the basic module. The ReLU activation function is further added to output the final convolution features.

[0034] Step 1.2: construct the MFCC feature extraction network, the GFCC feature extraction network, and the CFCC feature extraction network respectively. The construction methods of the above three feature extraction networks are the same. Figure 5 As shown in Figure 2, the construction method of the MFCC feature extraction network is as follows: add a data input layer, the number of nodes in the data input layer is consistent with the length of the MFCC sequence; add the first convolutional layer with parameters (48×1, 144, 4), that is, the convolution kernel size is 48×1, the number of convolution kernels is 144, and the convolution step is 4, the representation method is the same below; add the lrn layer; add the second convolutional layer with parameters (16×1, 192, 2); add the third convolutional layer with parameters (7×1, 256, 1); add the lrn layer; add the first pooling layer with parameters of (3×1, 2), that is, the pooling kernel size is 3×1 and the stride is 2; add three basic ResNet-Inception modules connected in sequence, namely basic ResNet-Inception module 1, basic ResNet-Inception module 2 and basic ResNet-Inception module 3; add the fourth convolutional layer with parameters of (3×1, 384, 1); add the fifth convolutional layer with parameters of (3×1, 512, 1); add the second pooling layer with parameters of (3×1, 2); add a fully connected layer with 256 output nodes.

[0035] Step 1.3: Build a fusion recognition network, that is, add a two-layer perceptron network on top of the MFCC feature extraction network, GFCC feature extraction network, and CFCC feature extraction network. The number of nodes in the two-layer perceptron network is 512 and 128 respectively.

[0036] In step 1.4, add a Softmax classifier on top of the two-layer perceptron network.

[0037] The present invention completes the construction of the overall framework of the multi-depth learning joint decision model through steps 1.1-1.4.

[0038] In step 1.5, based on the MFCC, GFCC, and CFCC auditory perceptual filter construction methods, 48 channels of auditory perceptual filter coefficients are generated for the specified frequency bands. Each auditory perceptual filter is a 48-dimensional vector. All auditory perceptual filters are integrated to form a 48×144 matrix. The parameters of the first convolutional layer of the MFCC, GFCC, and CFCC feature extraction networks are initialized. The remaining network layers and training methods of the multi-deep learning joint decision model are set according to general convolutional neural network parameter initialization and training methods. Here, "initializing the parameters of the first convolutional layer of the MFCC, GFCC, and CFCC feature extraction networks" refers to initializing the convolution kernel weights.

[0039] In step 1.5 of the present invention, based on the MFCC filter coefficients - Mel-Frequency Cepstral Coefficients (MFCC), the GFCC filter coefficients - Gammatone Frequency Cepstral Coefficients (GFCC) and the CFCC filter coefficients - Cochlear Frequency Cepstral Coefficients (CFCC), samples are constructed and the network structure parameters of the above three feature extraction networks are initialized, so that the network structure parameters are more adapted to the samples.

[0040] In step 1.6, the cross entropy method is used to construct the loss function J between the output results of the multi-deep learning joint decision model and the corresponding labels of the input samples, and the training parameters such as the optimizer, learning rate, and number of training times during iterative training are set.

[0041] Reference Figure 2 In the second phase of the process, the specific process for generating a multi-platform sample set based on multi-platform data in step 2 is as follows. Using a speaker speech dataset, a simulated sonar target dataset, other sonar platform target datasets, and a small sample sonar platform target dataset to be applied, the multi-platform sample set is constructed by generating corresponding auditory perceptual feature samples based on MFCC, GFCC, and CFCC auditory perceptual filters, setting the upper and lower frequency limits of data processing as required.

[0042] Reference Figure 2The three-stage process involves multi-level transfer learning training of the multi-deep learning joint decision model based on multi-platform sample sets in step 3, which mainly includes three interacting sub-steps: initialization training of the multi-deep learning joint decision model, primary transfer learning training of the multi-deep learning joint decision model, and advanced transfer learning training of the multi-deep learning joint decision model. Specifically, the training of the multi-deep learning joint decision model includes the following process.

[0043] Step 3.1, initialize the training of the multi-deep learning joint decision model. Let the speaker speech sample set and the simulated sonar target sample set used for training be x1={x 1，1 , x 1，2 ,…,x 1，n ,(n∈N * )} and x2={x 2，1 , x 2，2 ,…,x 2，n ,(n∈N * )}. The number of input samples each time when training the multi-deep learning joint decision model is N Batch , randomly select a fixed number (N Batch ) samples and optimize the loss function J for several rounds based on the gradient descent method. Then, a fixed number of samples are randomly selected from x2 and the loss function J is optimized for several rounds based on the gradient descent method. Based on the above method, x1 and x2 are used in turn to carry out model optimization training until a number of cycles set as required are completed. If the loss value of the loss function has not converged to the expected threshold after the specified number of cycles, the optimization is continued in the above method until the loss value of the loss function converges to the expected threshold.

[0044] Step 3.2: Based on the initial training of the multi-deep learning joint decision model, the model is then subjected to primary transfer learning training. The network structure parameters of the MFCC feature extraction network, GFCC feature extraction network, and CFCC feature extraction network after initial training are retained, and the network structure parameters of the fusion recognition network (two-layer perceptron network) are randomly initialized. The other sonar target sample set used for training is denoted as x3 = {x 3，1 , x 3，2 ,…,x 3，n ,(n∈N * )}. Randomly select a fixed number (N Batch ) samples and optimize the loss function J for several rounds based on the gradient descent method until the loss value of the loss function converges to the expected threshold.

[0045] Step 3.3: Based on the primary transfer learning training of the multi-deep learning joint decision model, the model is trained with advanced transfer learning. The target sample set of the small sample sonar platform to be used for training is x4 = {x 4，1 , x 4，2 ,…,x 4，n ,(n∈N * )}; The fourth convolutional layer of the MFCC feature extraction network is also recorded as l M , The fourth convolutional layer of the GFCC feature extraction network is l G , the fourth convolutional layer of the CFCC feature extraction network is l C In the multi-depth learning joint discriminant model after primary transfer learning, the first local network structure is extracted and the network structure parameters are retained to obtain the feature extraction model S E ={s M , s G , s C}, the first local network structure includes the corresponding basic ResNet-Inception module 3 in the MFCC feature extraction network, GFCC feature extraction network, CFCC feature extraction network and the previous network structure s M 、s G 、s C Specifically, the feature extraction model consists of three independent network structures s M 、s G 、s C The network structure s M The network structure of the MFCC feature extraction network from the data input layer to the basic ResNet-Inception module 3 is selected from the multi-depth learning joint discrimination model obtained in step 3.2. M 、s G 、s C Corresponding to MFCC, GFCC and CFCC feature extraction functions respectively. Extract the second local network structure of the multi-deep learning joint discrimination model and form a feature recognition model S R , construct feature recognition model S R The loss function J m The second local network structure is the corresponding fourth convolutional layer l in the MFCC feature extraction network, GFCC feature extraction network, and CFCC feature extraction network M 、l G 、l C To the network structure of the softmax classifier, that is, to extract the three features from the network M 、l G 、l C(Including) the top-level network part, the fusion recognition network (two-layer perceptron network) and the softmax classifier are integrated and reorganized into the feature recognition model S R . Apply feature extraction model S E Perform feature extraction on all samples in x4 to obtain the feature sample set x E ={x E，1 , x E，2 ,…,x E，n ,(x E，n =(x M，n , x G，n , x C，n ), n∈N * )}, where the feature sample x E，n Contains a set of three features x M，n 、x G，n and x C，n , corresponding to s M 、s G 、s C The feature extraction results of the feature recognition model S R The structural parameters of are randomly initialized. E Randomly select a fixed number (N Batch ) and the loss function J is calculated based on the gradient descent method. m Perform several rounds of optimization until the loss function converges to the expected threshold. Reintegrate S E and S R , get the trained multi-deep learning joint decision model S D , that is, forming a multi-deep learning joint decision model for small sample platforms.

[0046] The fixed number in step 3.1 of the present invention is consistent with the fixed number in steps 3.2 and 3.3, which are all N Batch indivual.

[0047] Reference Figure 2 The four stages of the process are as follows:

[0048] In step 4.1, the underwater target radiation noise data is preprocessed to generate several frames of MFCC, GFCC, and CFCC samples with temporal consistency.

[0049] Step 4.2, based on the trained multi-deep learning joint decision model S D Process the MFCC, GFCC and CFCC samples and output the final recognition results.

[0050] The present invention first carries out preprocessing on the radiation noise of underwater targets, constructs MFCC, GFCC and CFCC auditory perception filters, generates MFCC, GFCC and CFCC auditory perception features, and based on the deep learning method, respectively constructs deep network models for MFCC, GFCC, CFCC feature extraction, and at the same time uses the auditory perception filter coefficients to initialize the initial convolution layer of the above deep network model, so that the model structure parameters are more compatible with the auditory perception features to be processed. On this basis, combined with the multi-layer perceptron method, a fusion recognition network is constructed to achieve comprehensive target recognition. By using the above multiple networks, an overall multi-depth learning joint decision model is formed, which can comprehensively utilize MFCC, GFCC and CFCC features, improve the comprehensiveness of feature utilization, and help enhance target recognition tolerance. Furthermore, in view of the constraints of the small sample data conditions of the sonar platform to be applied, full use is made of the data of other platforms to carry out model transfer learning training to make up for the current situation of insufficient data on this platform. First, the multi-deep learning joint decision model is initialized and trained using speaker voice data and simulated sonar target data, and the model structure parameter migration and training strategy are set. Then, the model is trained on primary transfer learning based on other sonar platform target data with relatively sufficient sample data, and the model structure parameter migration and training strategy is set again. Finally, the model is trained on advanced transfer learning based on the small sample sonar platform target data to be applied, so as to optimize the model structure parameters as much as possible.

[0051] Reference Figure 3 For the three types of target data obtained from a surface ship platform, combined with large-scale speaker voice and simulated sonar target data, as well as shore-based sonar platform target data (the duration is more than 5 times that of the surface ship platform target data), the method proposed in this invention is used for processing. It can be seen that the separability of deep learning features is much higher than that of the original features.

[0052] Figure 6 The figure shows the statistical results of the recognition accuracy of the method proposed in the present invention and multiple comparison methods. It can be seen that the method proposed in the present invention has the best recognition effect, which verifies the effectiveness of intelligent target recognition under small sample conditions.

Claims

1. A small sample underwater target recognition method based on multi-platform auditory perception feature deep transfer learning, characterized by: The method comprises the following steps: A multi-depth learning joint decision model for multi-dimensional underwater target auditory perception features is constructed using the following method: Constructing MFCC feature extraction network, GFCC feature extraction network and CFCC feature extraction network, placing the above three feature extraction networks in parallel and adding fusion recognition network and support vector machine classifier on top in sequence; the fusion recognition network is a two-layer perceptron network; The construction method of the MFCC feature extraction network, the GFCC feature extraction network or the CFCC feature extraction network is: Add data input layer; Add the first convolutional layer with parameters of (48×1, 144, 4); add the Irn layer; Add a second convolutional layer with parameters of (16×1, 192, 2); Add the third convolutional layer with parameters of (7×1, 256, 1); add the lrn layer; Add the first pooling layer with parameters of (3×1, 2); Add three basic ResNet-Inception modules; Add the fourth convolutional layer with parameters of (3×1, 384, 1); Add the fifth convolutional layer with parameters of (3×1, 512, 1); Add a second pooling layer with parameters of (3×1, 2); Add a fully connected layer with 256 output nodes; The construction method of the basic ResNet-Inception module is: Three parallel branches are added after the data input layer, with branch 1 being a direct branch. Branch 2 includes two convolutional layers, with the parameters of each convolutional layer being (1×1, 128, 1) and (3×3, 128, 1) respectively. Branch 3 includes two convolutional layers, with the parameters of each convolutional layer being (1×3, 128, 1) and (3×1, 128, 1) respectively. A feature dimension expansion operation is added after branches 2 and 3, and the output features of the two branches are integrated to obtain a comprehensive feature set output with a total number of 256 features. A direct addition and summation operation is added after the output features of branch 1 and the comprehensive feature set to obtain the top-level features of the basic ResNet-Inception module, and then a ReLU activation function is added to output the final convolutional features. Initialize the parameters of the multi-deep learning joint decision model, including: Based on the MFCC, GFCC, and CFCC auditory perception filter construction methods, 48 channels of auditory perception filter coefficients are generated according to the set frequency band. Each auditory perception filter is a 48-dimensional vector. All auditory perception filters are integrated to form a 48×144 matrix, and the parameters of the first convolutional layer of the MFCC feature extraction network, GFCC feature extraction network, and CFCC feature extraction network are initialized respectively. Construct the loss function J of the multi-deep learning joint judgment model and set the training parameters during training; Based on multi-platform data, a multi-platform sample set is generated as follows: using the speaker speech dataset, simulated sonar target dataset, other sonar platform target dataset and the small sample sonar platform target dataset to be applied, based on MFCC, GFCC and CFCC auditory perception filters, the upper and lower limit frequencies of data processing are set according to requirements, and the corresponding auditory perception feature samples are generated to construct a multi-platform sample set; Multi-level transfer learning training of multiple deep learning joint decision models based on multi-platform sample sets includes the following steps: Initialize and train the multi-deep learning joint decision model, including: A loop optimization step randomly selects a fixed number of samples from the speaker speech sample set and performs several rounds of optimization on the loss function J based on the gradient descent method. Then, a fixed number of samples are randomly selected from the simulated sonar target sample set and the loss function J is optimized for several rounds based on the gradient descent method. Repeat the above optimization steps until the number of optimization cycles reaches the preset number of rounds, or until the loss function converges to the expected threshold; Conducting primary transfer learning training for multiple deep learning joint decision models, including: The structural parameters of the MFCC feature extraction network, GFCC feature extraction network, and CFCC feature extraction network during initial training are retained, the structural parameters of the fusion recognition network are randomly initialized, a fixed number of samples are randomly selected from other sonar target sample sets, and the loss function J is optimized for several rounds based on the gradient descent method until the loss function converges to the expected threshold; Conducting advanced transfer learning training on multiple deep learning joint decision models includes the following steps: Extract the first local network structure of the multi-depth learning joint discriminant model and retain the network structure parameters to obtain the feature extraction model S E ={s M , s G , s C }, the first local network structure includes the corresponding three basic ResNet-Inception modules in the MFCC feature extraction network, GFCC feature extraction network, and CFCC feature extraction network and the previous network structure s M 、s G 、s C ; Extract the second local network structure of the multi-deep learning joint discriminant model and form it into a feature recognition model S R , construct feature recognition model S R The loss function Jm; the second local network structure is the corresponding fourth convolution layer l in the MFCC feature extraction network, GFCC feature extraction network, and CFCC feature extraction network M 、l G 、l C To the network structure of softmax classifier; By S E The feature sample set x is obtained by extracting features from all samples in the target sample set of the small sample sonar platform. E ={x E，1 , x E，2 ,…,x E，n ,(x E，n =(x M，n , x G，n , x C，n ), n∈N * )}, where the feature sample x E，n Contains a set of three features x M，n 、x G，n and x C，n , corresponding to s M 、s G 、s C Feature extraction results; Model S R The structural parameters are randomly initialized; From x E A fixed number of feature samples are randomly selected and the loss function J is calculated based on the gradient descent method. m Perform several rounds of optimization until the loss function converges to the expected threshold; Reintegrate S E and S R , get the trained multi-deep learning joint decision model S D ; The radiation noise data of unknown underwater targets is preprocessed, and the preprocessing results are used to identify the targets using the trained multi-deep learning joint decision model, and the recognition results are output.

Images