An intelligent underwater noise recognition and classification method based on improved domain adaptation

Abstract

This invention addresses the difficulty of feature transfer in underwater noise signals by proposing an intelligent underwater noise recognition and classification method based on improved domain adaptation. The method comprises the following steps: First, the collected underwater noise signals are divided into a source domain and a target domain, and a domain-shared 1D-CNN is established to extract data features from both the source and target domains. Second, in the domain adaptation module, this method combines a multi-layer multi-kernel MMD (ML-MK-MMD) with a domain discriminator. By minimizing the total loss of the domain adaptation module and the network classifier, it effectively extracts the maximum domain-invariant features from both the source and target domains. Finally, the trained model is used to test the target domain samples. This network learns source domain knowledge to complete the transfer task of target domain samples. This invention can effectively extract domain-invariant features, reduce the distribution differences between the source and target domains, improve cross-domain testing performance, and achieve the goal of cross-domain transfer recognition.

Description

Technical Field

[0001] This invention belongs to the field of intelligent underwater signal recognition transfer learning, and specifically relates to an intelligent underwater noise recognition and classification method based on improved domain adaptation. Background Technology

[0002] The development of intelligent methods relies on the long-term accumulation of data from various fields, resulting in relatively complete datasets. Therefore, intelligent identification methods are data-driven algorithms. However, the signals collected by underwater vehicles are a special type of signal, and the acquisition process is complex, noisy, and involves confidentiality issues. These data-driven intelligent diagnostic methods achieve high fault diagnosis accuracy by using deep learning networks instead of manual extraction of the time-frequency features of fault signals. However, because they rely heavily on a large number of fault sample labels, these intelligent diagnostic methods may experience accuracy degradation when label data is difficult to obtain. Furthermore, these problems can lead to insufficient datasets. To address this issue, a transfer learning method is proposed to prevent overfitting effects from small datasets.

[0003] Transfer learning methods can use knowledge learned from one environment to help with learning tasks in a new environment. It provides a new way to solve the problem of insufficient data and allows the transfer of existing knowledge to solve learning problems in a target domain with only a small amount of labeled sample data. In other words, it uses existing knowledge to solve problems in different but related domains.

[0004] Among them, domain-adaptive intelligent transfer learning methods employ single-layer, single-core MMD to reduce the distributional differences and inter-class distances of the transferable features learned by the network. These methods mostly minimize the distributional differences between the source and target domains by constructing feature extraction networks, thereby learning transferable features more effectively. While this method achieves a certain degree of transfer learning capability, it suffers from the drawback of relatively singular features in the learned sample data. ML-MK MMD, with its better domain adaptability, is an option for addressing this problem; however, its high computational burden consumes a significant amount of training time.

[0005] Unlike the methods mentioned above, transfer learning methods can also employ domain adversarial approaches to achieve cross-domain fault diagnosis. This method extracts domain-invariant features by using adversarial training with labeled samples from the source domain and unlabeled samples from the target domain. This adversarial training method can effectively reduce training time when processing samples from both the source and target domains. However, these methods still have some drawbacks, including relatively limited feature extraction and insufficient feature extraction from the input domain. Summary of the Invention

[0006] The technical problem addressed by this invention is: to solve the problems of large amounts of interference in existing underwater noise signals and data acquisition, the difficulty in extracting transfer features in existing transfer learning methods, the insufficient feature extraction caused by using single-layer single-core MMD, and the long network training time caused by using domain discriminators. This invention proposes an intelligent underwater noise recognition and classification method based on improved domain adaptation. By adopting an improved domain adaptation method, multi-layer multi-core MMD is combined with a domain classifier. By minimizing the established total loss function, domain-invariant features can be effectively extracted, distribution differences between domains can be reduced, and the model can improve the recognition and classification of unknown type samples in the target domain.

[0007] An intelligent underwater noise identification and classification method based on improved domain adaptation includes the following steps:

[0008] Step 1: Classify the collected underwater data samples into labeled source domain samples and unlabeled target domain samples;

[0009] Step 2: Perform Fast Fourier Transform (FFT) on the source domain and target domain sample data to extract the frequency domain features of the source and target domain samples and create a training data sample set;

[0010] Step 3: Establish a 1D-CNN feature extraction network with domain-shared parameters;

[0011] Step 4: Initialize the network parameters of the domain-shared parameter feature extraction network 1D-CNN established in Step 3. The initialization parameters include the weight coefficients of the convolutional layers, pooling layers, and fully connected layers in the network, as well as the initialization of the network optimizer and the network training learning rate.

[0012] Step 5: For the domain-shared parameter 1D-CNN established in Step 4, establish a new total loss function, which includes three parts: the classification loss in the source domain, the domain discriminator loss function, and the multi-layer multi-kernel MMD in the feature extraction network.

[0013] Step 6: Calculate the total loss function;

[0014] Step 7: Optimize the total loss function of the established network structure and further update the network parameters;

[0015] Step 8: Judge the training results in Step 7 to determine whether the maximum number of training epochs or the total loss function has converged to the minimum. If it has, it means that the transfer model training is complete. Input the target domain test set samples into the trained samples for recognition and classification to obtain the target domain prediction results. If the training is not complete, the training process will continue from Step 5 until the maximum number of training epochs or the total loss function converges to the minimum.

[0016] Furthermore, in step 2, the formula for performing a Fast Fourier Transform (FFT) on the domain samples is as follows:

[0017]

[0018] x(i) (i = 0, 1, ..., N-1) is a time-domain signal of finite length, where i represents an integer from 0 to N-1, and k = 0, 1, ..., N-1. X(k) represents the frequency domain signal after FFT transformation, where X(k) is the weighting parameter.

[0019] Furthermore, in step 3, the feature extraction network 1D-CNN is constructed sequentially including convolutional layer 1, max pooling layer 1, convolutional layer 2, max pooling layer 2, convolutional layer 3, max pooling layer 3, fully connected layer 1, and fully connected layer 2; at the same time, a domain discriminator is established in the network, including fully connected layer 3 (FD layer) and an output layer.

[0020] Furthermore, in step 5:

[0021] (1) The classification loss function in the source domain is:

[0022]

[0023] Where n is the batch size of the network training samples, It is the label probability distribution of the output samples of the source domain samples in the second fully connected layer of the established network. These are the corresponding labels of the source domain samples, where s indicates that the input is a source domain sample, and T represents the transpose symbol in mathematical operations;

[0024] (2) The loss function of the domain classifier is:

[0025]

[0026] in, and L represents the high-level features learned from the source domain data and target domain data of the domain discriminator output layer, respectively. a The domain classification loss is calculated using the following formula:

[0027]

[0028] Among them, g i It is a real domain label, d(x) i ) is the domain output of the i-th sample, x i These are samples from the source and target domains, and m is the batch size of the network training samples.

[0029] (3) The summation loss function of the multi-layer multi-kernel MMD in the feature extraction network is:

[0030]

[0031] in It is the set of parameters for convolutional layers 1, 2, and 3, and fully connected layers 1 and 2 in the constructed network. These represent the multi-layer transfer features of the source and target domains, respectively. Indicates the index of each level;

[0032]

[0033] Furthermore, in step 6, the formula for calculating the total loss function is:

[0034]

[0035] Where λ and β are weighting parameters, and θ f θ c θ d These are the parameter sets for each module of the feature extraction network, the category classifier, and the domain classifier.

[0036] Furthermore, in the total loss function, the coefficients λ and β are trade-off parameters, representing the degree of influence of the domain discriminator and the multi-layer multi-kernel MMD during training, respectively. Their changes affect the transfer performance of the model. The trade-off parameters λ and β are varied from 0 to 1 using the formula 2 / (1+exp(-10×p))-1, where p is a parameter that changes linearly from 0 to 1 with the number of training epochs = 50.

[0037] Furthermore, in step 7, the model uses the SGD optimizer to optimize the total loss function L. total Perform backpropagation training. This training updates the parameter sets of each part of the network module. The expression for the parameter update process is as follows:

[0038]

[0039]

[0040]

[0041] Where μ is the learning rate in the network.

[0042] Invention Effects

[0043] The technical advantages of this invention are as follows:

[0044] 1. First, the collected underwater noise data is divided into two domains: the labeled source domain and the unlabeled target domain. Second, to overcome the difficulty of feature extraction from underwater data, the sample data in both domains are first processed using FFT to fully extract the frequency domain features of the source and target domains and the sample data. The converted frequency domain signal is then used to create a dataset as input to the network. FFT can overcome the difficulty of feature extraction in time-domain analysis of underwater noise signals. By converting the time-domain signal to the frequency domain using FFT, the frequency domain features of the noise signal can be extracted effectively, providing preprocessing for the subsequent creation of the sample dataset.

[0045] 2. Establish a new domain-shared feature network with a new total loss function. This loss function includes the classification loss of the classifier, the distribution difference between the source and target domains obtained through ML-MK MMD, and the domain discriminator loss. These three factors are combined to establish the new loss function, and the network is trained using this as the optimization objective. The method of combining ML-MK MMD and the domain discriminator can reduce the distribution difference between cross-domain samples through network training, fully extracting and maximizing domain-invariant features.

[0046] This method overcomes the shortcomings of the original method, such as insufficient feature extraction due to the use of single-layer single-core MMD and excessive network training time due to the use of domain discriminator. It uses multi-layer multi-core MMD to calculate the distribution difference between the source domain and the target domain, and combines it with domain discriminator to fully reduce the cross-domain distribution difference of samples, extract the cross-domain invariant features to the maximum extent, and at the same time reduce the network training time.

[0047] 3. This invention can extract domain-invariant features hidden in the source and target domains as much as possible. This improved method has higher classification and recognition accuracy than other migration fault diagnosis methods. Attached Figure Description

[0048] Figure 1 The following are the overall steps of an improved domain-adaptive intelligent underwater noise identification and classification method according to the present invention;

[0049] Figure 2 This is a structural diagram of an improved domain-adaptive intelligent underwater noise recognition and classification network model according to the present invention;

[0050] Figure 3 This is a graph showing the changes in weight parameters in the total loss function of the feature network in an improved domain-adaptive intelligent underwater noise recognition and classification method of the present invention. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of the embodiments of this invention will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0052] See Figures 1-3 An intelligent underwater noise identification and classification method based on improved domain adaptation is proposed, the steps of which are as follows:

[0053] Step 1: First, the collected underwater data samples are divided into a source domain and a target domain. Source domain samples are labeled, while target domain samples are unlabeled. In the training set, each sample has 4096 sampling points, with 4000 labeled samples in the source domain and 4000 unlabeled samples in the target domain. The test set contains 400 unlabeled test samples.

[0054] Step 2: Perform FFT on the source and target domain samples to extract the frequency domain features of the samples. The time domain features of the acquired time domain signal are extracted using the FFT formula shown in formula (1):

[0055]

[0056] xn (n = 0, 1, ..., N-1) is a time-domain signal of finite length. Where k = 0, 1, ..., N-1, Let X(k) be the weighting parameter, and X(k) represent the frequency domain signal after FFT transformation. After FFT preprocessing, in the training set, each sample has 2048 sampling points, with 4000 labeled samples in the source domain and 4000 unlabeled samples in the target domain. The test set contains 400 unlabeled test samples.

[0057] Step 3: Establish a domain-shared feature extraction network (1D-CNN). The network structure parameters are shown in the table below:

[0058] The first part consists of the feature extraction network parameters, which are as follows:

[0059]

[0060] The second part consists of the following parameters for the domain discriminator:

[0061]

[0062] Step 4: Initialize network parameters.

[0063] Step 5: In the domain adaptation module, a new total loss function is established. This total loss function consists of three parts: the classification loss in the source domain, the domain discriminator loss, and the multi-layer multi-kernel MMD value calculated in the feature extraction network layer. Backpropagation is used to optimize this total loss function and update the network parameters.

[0064] Step 5 specifically includes the following: Constructing a new total loss function. This loss function consists of the following three parts.

[0065] 5.1 Optimization Objective 1

[0066] Transferable features learned from cross-domain data are affected by domain distribution differences. Existing research generally suggests that high-level distribution differences are concentrated in fully connected layers; however, this leads to a limited range of domain-invariant features learned by the network. This invention proposes a multi-layer domain adaptation method, training the parameters of a domain-shared CNN by minimizing multi-layer multi-kernel MMD. In addition to the last fully connected layer, distribution difference calculations are also incorporated into other convolutional layers and another fully connected layer to adjust the distribution of higher-level transferable features learned in the network.

[0067] During the training of a domain-sharing CNN, the distribution of learned features changes as the parameters of the domain-sharing network are updated. To avoid this phenomenon, the distribution of transferable features in convolutional layers C1, C2, and C3 is adjusted. The multi-layer MMD calculation of the learned transferable features is as follows:

[0068]

[0069] in These represent the multi-layer transfer features of the source and target domains, respectively. This represents the index of each layer. Domain-shared CNNs can help learn transferable features with specific attributes, and the values ​​of multiple MMD layers will change during training.

[0070] In equation (10), the kernel bandwidth and time complexity significantly affect the value of MMD when calculating multi-layer MMD. In this paper, we choose a Gaussian kernel, the expression of which is shown below:

[0071]

[0072] Where y is the bandwidth of the Gaussian kernel.

[0073] The above calculations show that the distributional differences of transferable features learned during training are ultimately achieved by minimizing the sum of multi-layer MMD, which can be expressed by the following formula.

[0074]

[0075] in It is the parameter set of layers C1, C2, C3, F1, and F2.

[0076] 5.2 Optimization Objective 2

[0077] Optimization objective 2 is to minimize the classification error of the training samples. This paper uses the cross-entropy function as the loss function for the sample set in the source domain, which is defined as follows:

[0078]

[0079] Where n is the batch size of the network training samples, It is the label probability distribution of the output samples of the source domain samples in layer F2. These are the corresponding labels for the source domain samples.

[0080] 5.3 Optimization Objective 3

[0081] The third optimization objective of this network is to maximize the domain classification error of the source-target domain data. The domain classification loss is defined as...

[0082]

[0083] Among them, g i It is a real domain label, d(x) i ) is the domain output of the i-th sample, x i These are samples from the source and target domains.

[0084] During the training phase, the training dataset consists of n source domain samples and n target domain samples. Therefore, (14) can also be written in the following form:

[0085]

[0086] in and These represent the high-level features learned from the source and target domain data of the FD layer, respectively.

[0087] 5.4 Total Loss Function

[0088] With the above optimization objectives, the total loss of the network established in this paper is expressed by the following formula:

[0089]

[0090] Where λ and β are weight parameters, determining the weights of the domain classifier and the multi-layer multi-kernel MMD component in domain adaptation, respectively. Where θ f θ c θ dThese are the parameter sets for the feature extraction network, the fault type classifier, and the domain classifier, respectively.

[0091] In this invention, the coefficients of the multilayer multi-kernel MMD and the domain discriminator loss in the total loss function are trade-off parameters affecting the transfer performance of the model. In the total loss function of this invention, these trade-off parameters λ and β are transformed from 0 to 1 using the equation 2 / (1+exp(-10×p))-1, where p is a linear transformation parameter that changes from 0 to 1 with each training epoch. They represent the degree of influence of the domain discriminator and the multilayer multi-kernel MMD during training, respectively, and their changes affect the transfer performance of the model. The changes of the trade-off parameters in the total loss function during training are as follows: Figure 3 As shown.

[0092] The model in this invention uses the SGD optimizer to optimize the total loss function L. total Training is performed using this optimization algorithm to accelerate the training of the network parameters and minimize the network's loss function. This method updates the parameter sets in the network using the SGD optimization algorithm; the expressions for each part of the parameter update are as follows.

[0093]

[0094]

[0095]

[0096] Where μ is the learning rate in the network.

[0097] In this example, in the domain-sharing network, convolutional layer 1 has a kernel size of 64, a stride of 1, and padding of 64. Convolutional layers 2 and 3 have kernel sizes of 5, strides of 1, and padding of 2. The ReLU function is used as the activation function after the convolutional layers. The pooling layer has a kernel size of 2 and a stride of 2. The max-pooling layer has a kernel size of 2 and a stride of 2.

[0098] In this example, the bandwidth parameters for the ML-MK MMD are chosen as 1, 2, 4, 8, and 16, with their weights remaining constant. This example uses the above bandwidth parameter values ​​and a multi-layer MMD in a multi-core computing domain-sharing network. During the training of the transfer learning network, the batch size is set to 128. In this example, the network training optimizer is trained using an SGD optimizer with a learning rate of 0.0001.

[0099] Step 6: The transfer model training is complete when the total loss function converges to the minimum or reaches the maximum number of training iterations.

[0100] In step 6, in the proposed method, the coefficients λ and β in the total loss function are trade-off parameters that affect the model's transfer performance. In the total loss function, the trade-off parameters λ and β are determined by the formula...

[0101] 2 / (1+exp(-10×p))-1 changes from 0 to 1, where p is a linearly changing parameter, changing with the number of training epochs from 0 to 1. In the experiment, epoch = 50 was chosen. The change process of the tradeoff parameter in the total loss function during training is as follows: Figure 3 As shown.

[0102] The training method for this network is to minimize L total To optimize the objective and minimize the distributional differences of transferable features learned layer by layer, the network learns and maximizes domain-invariant features. When the network converges, it indicates that the network has learned the maximized domain-invariant features between the two domains, ultimately enabling the diagnostic task of identifying fault categories in the target domain by learning diagnostic knowledge from the source domain.

[0103] Step 7: Input the target domain test set samples into the trained samples for recognition and classification.

[0104] In summary, the embodiments of the present invention can minimize the distribution difference between the source domain and target domain sample sets and fully extract the maximum domain-invariant features through the above steps, which can effectively improve performance.

[0105] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of a preferred embodiment, and the sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0106] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A smart underwater noise identification and classification method based on improved domain adaptation, characterized in that, Includes the following steps: Step 1: Classify the collected underwater data samples into labeled source domain samples and unlabeled target domain samples; Step 2: Perform Fast Fourier Transform (FFT) on the source domain and target domain sample data to extract the frequency domain features of the source and target domain samples and create a training data sample set; Step 3: Establish a 1D-CNN feature extraction network with domain-shared parameters, which includes convolutional layer 1, max pooling layer 1, convolutional layer 2, max pooling layer 2, convolutional layer 3, max pooling layer 3, fully connected layer 1, and fully connected layer 2 in sequence; at the same time, establish a domain discriminator in this network, including fully connected layer 3 (FD layer) and output layer; Step 4: Initialize the network parameters of the domain-shared parameter feature extraction network 1D-CNN established in Step 3. The initialization parameters include the weight coefficients of the convolutional layers, pooling layers, and fully connected layers in the network, as well as the initialization of the network optimizer and the network training learning rate. Step 5: For the domain-shared parameter 1D-CNN established in Step 4, establish a new total loss function, which includes three parts: the classification loss in the source domain, the domain discriminator loss function, and the multi-layer multi-kernel MMD in the feature extraction network. (1) The classification loss function in the source domain is: Where n is the batch size of the network training samples, It is the label probability distribution of the output samples of the source domain samples in the second fully connected layer of the established network. These are the corresponding labels of the source domain samples, where s indicates that the input is a source domain sample, and T represents the transpose symbol in mathematical operations; (2) The loss function of the domain classifier is: in, and These represent the high-level features learned from the source and target domain data, respectively, from the output layer of the domain discriminator. The domain classification loss is calculated using the following formula: in, These are real domain tags. It is the domain output of the i-th sample. These are samples from the source and target domains, and m is the batch size of the network training samples. (3) The summative loss function of the multi-layer multi-kernel MMD in the feature extraction network is: ； in It is the set of parameters for convolutional layers 1, 2, and 3, and fully connected layers 1 and 2 in the established network. , These represent the multi-layer transfer features of the source and target domains, respectively. Indicates the index of each level; Step 6: Calculate the total loss function; Step 7: Optimize the total loss function of the established network structure and further update the network parameters; Step 8: Judge the training results in Step 7 to determine whether the maximum number of training epochs or the total loss function has converged to the minimum. If it has, it means that the transfer model training is complete. Input the target domain test set samples into the trained samples for recognition and classification to obtain the target domain prediction results. If the training is not complete, the training process will continue from Step 5 until the maximum number of training epochs or the total loss function converges to the minimum.

2. The intelligent underwater noise identification and classification method based on improved domain adaptation as described in claim 1, characterized in that, In step 2, the formula for performing a Fast Fourier Transform (FFT) on the domain samples is as follows: Let be a time-domain signal of finite length, where i represents an integer from 0 to N-1. , For weight parameters, This represents the frequency domain signal after FFT transformation.

3. The intelligent underwater noise identification and classification method based on improved domain adaptation as described in claim 1, characterized in that, In step 6, the formula for calculating the total loss function is: in and These are weight parameters. , , These are the parameter sets for each module of the feature extraction network, the category classifier, and the domain classifier.

4. The intelligent underwater noise identification and classification method based on improved domain adaptation as described in claim 3, characterized in that, In the total loss function, the coefficients and These are tradeoff parameters, representing the relative importance of the domain discriminator and the multi-layer multi-kernel MMD during training; their variations affect the model's transfer performance. and Through formula To achieve a change from 0 to 1, p is a parameter that changes linearly from 0 to 1 along with the number of training epochs (Epoch=50).

5. The intelligent underwater noise identification and classification method based on improved domain adaptation as described in claim 1, characterized in that, In step 7, the model optimizes the total loss function using the SGD optimizer. Perform backpropagation training; update the parameter sets of each part of the network module through this training. The expression for the parameter update process of each part is as follows: in, It is the learning rate in the network.

Images