A method for underwater vehicle self-noise feature extraction, recognition and separation based on multi-task learning

Abstract

The present invention relates to a method for extracting, identifying and separating underwater vehicle self-noise features based on multi-task learning. A multi-task learning model is adopted to separately estimate the category energy ratio of a signal and the time domain waveform of the signal. Compared with a single separation task, the present invention only slightly increases network overhead, shares the encoder and high-dimensional feature extractor in steps 2 and 3, and uses the category information of the label to guide the training of the network, thereby achieving noise separation performance that is superior to that of a single separation task. At the same time, the energy estimator structure introduced in the present invention can more effectively estimate the energy proportion of a single-source noise component in mixed noise. Compared with extracting energy information from the separated noise, this structure improves the accuracy of separation and identification, and can better meet the requirements for separation and identification of underwater vehicle self-noise.

Description

Technical Field

[0001] The present invention belongs to the field of underwater vehicle self-noise feature extraction, identification and separation, and in particular relates to a method for underwater vehicle self-noise feature extraction, identification and separation based on multi-task learning. Background Art

[0002] The ocean covers two-thirds of the Earth's surface, harboring vast resources and energy, and possessing crucial strategic significance. Underwater vehicles, as a crucial technological means of developing and utilizing marine resources and ensuring maritime safety, play an irreplaceable role in building a strong maritime nation and enhancing maritime competitiveness.

[0003] Underwater vehicle self-noise refers to the noise generated by the underwater vehicle platform during its motion. This noise affects the platform's transducer array. Underwater vehicle self-noise primarily includes vibration noise, flow noise, propeller noise, and electrical noise. Vibration noise is primarily generated by the periodic collisions between the platform's hull, frame, and various functional devices during underwater motion, and is transmitted to the transducer array through the platform's structural vibrations. Flow noise is primarily caused by the excitation of the underwater vehicle's structural surface by turbulent fluctuating pressure and the interaction of turbulent fluctuating pressure with cavities, appendages, and protrusions. Propeller noise is the noise generated by the platform's propeller rotation. Electrical noise refers to the electromagnetic noise generated by the vehicle's motor and drive system. To improve underwater vehicle long-range target detection and multi-device collaborative communication capabilities, enhance information transmission and processing capabilities, and reduce the harmful effects of self-noise on the platform's transducer array and data processing systems, research on self-noise feature extraction, identification, and separation is essential.

[0004] Generally speaking, when processing underwater acoustic signals, researchers at home and abroad typically extract the following five types of features: time-domain waveform features, power spectrum, line spectrum, and modulation spectrum features, joint time-frequency features, nonlinear dynamic features, and auditory features. Time-domain waveform features include signal mean, variance, peak value, kurtosis and skewness, zero crossings, wavelength difference, and peak-to-peak amplitude. Some researchers have combined these time-domain waveform features with power spectrum and STFT analysis to classify sonar signals. Joint time-frequency features are typically two-dimensional features derived through methods such as short-time Fourier transform and wavelet transform. While using classic waveform, power spectrum, and time-frequency analysis to extract underwater acoustic signal features, some researchers at home and abroad have also used newer methods such as fractal, chaos, and limit cycle analysis to investigate underwater acoustic signal characteristics. Auditory features have also been studied in recent years. Wang Yang et al. have achieved a series of research results on the auditory spectrum and speech characteristics of underwater acoustic targets. Li Zhaohui et al. have demonstrated the practicality of their proposed auditory model for underwater acoustics.

[0005] Noise component separation methods are often used in tasks such as audio separation, medical signal separation, and mechanical vibration noise separation. These problems can be summarized as blind source separation. Before the development of deep learning, researchers generally used blind source separation methods based on ICA and EMD algorithms, as well as their variants, to solve the problem of noise component extraction.

[0006] In recent years, with the development of deep learning technology, the use of deep learning technology for noise recognition and separation and extraction of noise components has gradually become a research direction. Huang et al. applied forward DNN and RNN to separation tasks. Some scholars combined deep learning technology with time-frequency features and used DNN training to estimate the logarithmic power spectrum of the target signal in the mixed signal; also combining time-frequency features with deep learning technology, some scholars used DNN training to directly map the target and interference noise spectra from the mixed signal. Zhang and Wang proposed a deep integration network for separation applications. Luo et al. adopted an end-to-end separation network and proposed a new method that goes beyond the idealized time-frequency mask. Chen et al. integrated RNN into Transformer to form a new separation network. Subakan et al. introduced a self-attention mechanism in 1-D full convolution to extract features for separation applications.

[0007] However, these deep learning-based separation tasks target signals like speech and music that don't have significant frequency overlap or interference. However, underwater vehicle self-noise is complex and subject to significant interference between different components. Existing deep learning networks are unable to effectively separate these components. Furthermore, these methods primarily focus on isolating the target signal's time-domain waveform and are unable to effectively estimate the energy contribution of each component within the mixed signal. Summary of the Invention

[0008] The technical problem addressed by this invention is to provide a multi-task learning method for self-noise separation and identification, effectively addressing the severe interference and frequency band overlap between various types of complex self-noise encountered by underwater vehicles, as well as the resulting difficulty in separating mixed noise and identifying energy contributions. This method first establishes a dataset that can be used for both separation and identification tasks. A shallow mixed representation of the signal is extracted using a 1D convolutional block initialized with specific parameters. This representation information is then fused through deep representation extraction. Finally, an energy estimator and decoder are used to estimate the energy and reconstruct the waveform of the desired separated signal.

[0009] A method for extracting, identifying, and separating underwater vehicle self-noise features based on multi-task learning includes the following steps:

[0010] Step 1: Use an online synthesis method to mix the noises of different types and stages collected to obtain a mixed noise dataset composed of multiple noises. The noise category and the proportion of mixed energy constitute the input label of the noise;

[0011] Step 2: Use the encoder to preliminarily process the mixed noise to obtain a shallow mixed representation;

[0012] Step 3: For the shallow mixed representation, use a high-dimensional feature extractor to extract the deep representation;

[0013] Step 4: Use the energy estimator to process the deep representation and obtain the category and energy proportion output of each noise component in the mixed noise. This output is optimized using the input label; at the same time, use the mask estimator to obtain the separation mask of each noise component in the mixed noise;

[0014] Step 5: Multiply the separation mask of each noise component with the shallow mixed representation to obtain the separation representation of each noise component. Input the separation representation into the decoder corresponding to the encoder structure to obtain the time domain waveform of each noise component.

[0015] Step 6: Multiply the noise category and energy proportion with the recovered time domain waveform to generate the estimated signal of each type of noise.

[0016] Furthermore, in step 1, the collected noise is divided into different categories according to the type of the collecting sensor and the navigation stage of the underwater vehicle, and each category is assigned a label;

[0017] According to the initial distribution, the signal-to-noise ratio is sampled within a certain range. For one of the signal-to-noise ratio sampling results, the data of any two labels is mixed using this signal-to-noise ratio to obtain the corresponding mixed noise. For each mixed noise sample, a different signal-to-noise ratio sampling result is used to mix the data to obtain the corresponding dataset.

[0018] Furthermore, in step 2, the encoding process is as shown in formula (1):

[0019] Y=H(x) (1)

[0020] Where x is the one-dimensional mixed noise input with a length of T, Y is the encoder output, and H is the encoder;

[0021] The encoding process of encoder H is shown in formula (2):

[0022]

[0023] in, and Respectively represent different frequencies θ iThe convolution block initialized by the cosine function and wavelet function under and denotes the convolution output at the i-th frequency, cat(·) denotes the concatenation in the frequency dimension, and Conv2d(·) is a two-dimensional convolution block for weighted summation.

[0024] Furthermore, in step 3, the extraction process is to use one-dimensional convolution blocks multiple times to perform deep feature extraction according to the two dimensions of time and frequency. The formula is:

[0025]

[0026] Among them, Conv1d T (·) and Conv1d F (·) are one-dimensional convolution blocks for time dimension reduction and frequency dimension reduction, respectively. and They represent the dimensionality reduction results in the b-th feature extraction step, and B is the total number of feature extraction steps;

[0027] Repeat the process of formula (3) above to compress the shallow mixed representation from both time and frequency dimensions, and generate a deep representation M in the compressed time-frequency dimension. d , that is, M (B) The above steps are called high-dimensional feature extractor.

[0028] Furthermore, in step 4, the energy estimator is a plurality of multi-layer perceptron (MLP) layers, which output the result of the estimated category. Here, the label information obtained in step 1 and the signal-to-noise ratio corresponding to the label are used to optimize the network parameters of the MLP layer using the cross entropy loss function. The formula is as follows:

[0029] P={p1,...,p C}=MLP(M d ) (4)

[0030]

[0031] Where P = {p1,...,p C} is the C output of the recognition task, p c is the energy proportion of the Cth noise component estimated by the network, M d For deep representation, label c is the true value of the recognition task, and loss1 is the loss value of the recognition task.

[0032] At the same time, a mask estimator is used to estimate the separation mask of each noise component, as follows:

[0033]

[0034] Mask is the sequence of separation masks that generates the desired number of components, C is the number of components that are desired to be separated, Generate upper-level representation of separation mask; M d For deep representation.

[0035] Furthermore, in step 5, the separation of the cth noise is characterized as The data decoded by the decoder is the waveform recovery signal Will Compared with pure separation noise without energy information, the network parameters are optimized using the scale-invariant signal-to-noise ratio (SI-SNR) as the loss function:

[0036]

[0037]

[0038] Among them, Mask(c) is the c-th noise separation mask, is the signal recovered from the cth waveform, x (c) is the time domain waveform of the cth noise, and loss2 is the loss value of the separation task.

[0039] We further obtain the loss function of the global network, and finally use the Adam optimizer to optimize the global parameters:

[0040] loss=loss1 2 +loss2 2 (9)

[0041] Furthermore, in step 6, the noise separated in step 5 is weighted according to the mixed energy ratio obtained in step 4 to obtain a final estimated signal.

[0042]

[0043] Among them, p c is the energy proportion of the cth noise component estimated by the network, is the signal recovered from the cth waveform, is the c-th estimated signal.

[0044] Effects of the Invention

[0045] The technical effect of the present invention is that: the present invention adopts a multi-task learning model to separately estimate the category energy ratio of the signal and the time domain waveform of the signal. Compared with a single separation task, the present invention only adds a little network overhead, shares the encoder and high-dimensional feature extractor in steps 2 and 3, and uses the category information of the label to guide the training of the network, thereby obtaining noise separation performance that is better than that of a single separation task; at the same time, the energy estimator structure introduced by the present invention can more effectively estimate the energy proportion of a single source noise component in the mixed noise. Compared with extracting energy information from the separated noise, this structure improves the accuracy of separation and recognition, and can better meet the requirements for separation and recognition of underwater vehicle self-noise. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 This is a general diagram of the network structure of a method for underwater vehicle self-noise feature extraction, identification and separation based on multi-task learning in the present invention;

[0047] Figure 2 The network structure shared by two tasks in the method for underwater vehicle self-noise feature extraction, recognition and separation based on multi-task learning of the present invention: the encoder and the high-dimensional feature extractor part structure;

[0048] Figure 3 It is an output module for separating and identifying respective tasks in a method for extracting, identifying and separating underwater vehicle self-noise features based on multi-task learning of the present invention;

[0049] Figure 4 This is a sample in the data set of a method for underwater vehicle self-noise feature extraction, identification and separation based on multi-task learning in the present invention;

[0050] Figure 5 This is a flow chart of a method for extracting, identifying and separating underwater vehicle self-noise features based on multi-task learning in the present invention. DETAILED DESCRIPTION

[0051] In the description of the present invention, it should be understood that the terms "center", "longitudinal", "lateral", "length", "width", "thickness", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", "clockwise", "counterclockwise" and the like to indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as limiting the present invention.

[0052] See also Figure 1-Figure 5The present invention proposes a method for extracting, identifying and separating underwater vehicle self-noise features based on multi-task learning, the method comprising the following steps:

[0053] Step 1: Use the online synthesis method to mix the noises of different types and stages collected to obtain a mixed noise dataset composed of multiple noises. The noise category and the proportion of mixed energy constitute the input label of the noise.

[0054] Step 2: Use an encoder composed of multiple 1D convolutional blocks to perform preliminary processing on the mixed noise and obtain a shallow mixed representation;

[0055] Step 3: For the shallow mixed representation, use a high-dimensional feature extractor to extract the deep representation;

[0056] Step 4: Use the energy estimator to process the aforementioned deep representation to obtain the category and energy proportion output of each noise component in the mixed noise. This output is optimized using the input label; at the same time, use the mask estimator to obtain the separation mask of each noise component;

[0057] Step 5: Multiply the separation mask with the shallow mixed representation to obtain the separation representation of each noise component, which is input into a decoder with a similar structure to the encoder to obtain the signal waveform of each noise component.

[0058] Step 6: Combine the energy with the estimated signal to obtain the final estimated signal.

[0059] The following is a detailed description of each of the above steps:

[0060] In step 1, the collected noise is classified into different categories based on the type of sensor used and the underwater vehicle's navigation phase, with each category assigned a label. For example, based on the type of sensor, noise can be divided into flow noise, vibration noise, and electrical noise collected by hydrophones, accelerometers, and oscilloscopes. Based on the different navigation phases, noise can also be divided into categories such as powered navigation, unpowered surfacing, and surface floating. The signal-to-noise ratio is sampled according to a certain distribution, and the data from any two labeled categories is mixed using this signal-to-noise ratio to obtain a mixed noise that approximates the actual noise. This mixed noise is used as the input data for the network, and a dataset is created. The distribution here is set according to the actual situation and can generally be set to a uniform distribution or a Gaussian distribution.

[0061] In step 2: multiple 1D convolution blocks are used to form an encoder to encode the input data and obtain a shallow mixed representation of the mixed data. The encoding process is shown in formula (1).

[0062] Y=H(x) (11)

[0063] Where x is the one-dimensional mixed noise input with a length of T, Y is the encoder output, and H is the encoder.

[0064] The above encoder H is composed of multiple one-dimensional convolution blocks, and the cosine function at different frequencies and different wavelet functions are used to initialize the parameters of the convolution blocks. The outputs of the one-dimensional convolution blocks with different initialization parameters are spliced together, and the output results of the cosine function and wavelet function with similar frequencies are spliced together. The outputs of the one-dimensional convolution blocks initialized with the cosine function and the wavelet function are weighted and summed using a two-dimensional convolution block to obtain a shallow mixed representation similar to the time-frequency feature. C H is the number of channels of the encoder with a single parameter initialization method. The encoding process of encoder H is shown in formula (2).

[0065]

[0066] in, and Respectively represent different frequencies θ i The convolution block initialized by the cosine function and wavelet function under and denotes the convolution output at the i-th frequency, cat(·) denotes the concatenation in the frequency dimension, and Conv2d(·) is a two-dimensional convolution block for weighted summation.

[0067] In step 3, the shallow mixed representation obtained in step 2 is further processed by using multiple one-dimensional convolution blocks to extract deep features in the time and frequency dimensions. The above process is shown in formula (3).

[0068]

[0069] Among them, Conv1d T (·) and Conv1d F (·) are one-dimensional convolution blocks for time dimension reduction and frequency dimension reduction, respectively. and They represent the dimensionality reduction results in the b-th feature extraction step, and B is the total number of feature extraction steps.

[0070] Repeat the process of formula (3) above to compress the shallow mixed representation from both time and frequency dimensions, and generate a deep representation M in the compressed time-frequency dimension. d , that is, M (B) The above steps are called a high-dimensional feature extractor.

[0071] In step 4: Use the energy estimator to represent the deep layer M in step 3 dProcessing is performed to obtain the category and mixed energy ratio output of the mixed data. The energy estimator is a multi-layer perceptron (MLP) layer that outputs the estimated category. Here, the label information obtained in step 1 and the signal-to-noise ratio corresponding to the label are used to optimize the MLP layer, using the cross-entropy loss function. The estimation and optimization of the first task - identifying the component energy task is shown in formulas (5) and (6).

[0072] P={p1,...,p C}=MLP(M d ) (14)

[0073]

[0074] While the energy estimator is running, the mask estimator is used to estimate the separation mask of each noise component. The length of the feature sequence is made consistent with the length of the shallow mixed representation mentioned above by transposed convolution, and finally a 1×1 convolution block is used to compress the feature dimension back to C×C H ×T (C is the number of components to be separated) to generate the desired number of separation mask sequences Mask. This process is named Tr. The process of generating the mask sequence is shown in formula (4).

[0075]

[0076] In step 5: multiply the separation mask with the shallow mixed representation to obtain the separation representation of each noise component The separation of the cth noise is characterized by The decoder Ht with a structure similar to that of the encoder in step 2 is used to decode the above separation representation. The number of decoders is determined by the desired number of separated noises. The decoded data is the waveform recovery signal. Will The network is compared with pure separated noise without energy information, and the scale-invariant signal-to-noise ratio (SI-SNR) is used as the loss function to optimize the network.

[0077]

[0078]

[0079] When updating the global network, loss1 in step 4 and loss2 in step 5 are first descaled, and then the square sum is calculated to obtain the loss function of the global network. Finally, the Adam optimizer is used to optimize the global parameters.

[0080] loss=loss1 2 +loss2 2 (19)

[0081] In step 6: the separated noise in step 5 is weighted according to the mixed energy ratio obtained in step 4 to obtain the final estimated signal.

[0082]

[0083] In the above steps, the two tasks share the encoder and deep representation extraction structure. The parameters of the shared parts are shown in Tables 1 and 2. The parameters of the energy estimator and mask estimator are shown in Tables 3 and 4.

[0084] Table 1 Encoder parameter table

[0085]

[0086] Table 2 Single deep layer representation extraction parameters

[0087]

[0088]

[0089] Table 3 Energy estimator MLP layer parameters

[0090] Layer Category parameter Activation Function Output feature size Input / / (1×2800) Dense1 (2800,1200) PReLU (1×1200) Dense2 (1200,256) PReLU (1×2800) Dense3 (256,C) Softmax (1×C)

[0091] Table 4 Mask estimator and decoder parameters

[0092] Layer Category parameter Activation Function Output feature size Input / / (4×750) ConvTrans1d (8,3) PReLU (8,1500) … PReLU ConvTrans1d (256,3) PReLU (256,48000) Conv1d (C,20) Softmax (C,48000)

[0093] Additional aspects and advantages of the present invention will be set forth in part in the description which follows and, in part, will be obvious from the description which follows, or may be learned by practice of the present invention.

[0094] The following describes in detail embodiments of the present invention. The embodiments are exemplary and intended to explain the present invention, but are not to be construed as limiting the present invention.

[0095] like Figure 1 As shown, the method for underwater vehicle self-noise feature extraction, identification and separation based on multi-task learning in this example includes the following steps.

[0096] Step 1 specifically includes: classifying the collected noise into different categories according to the type of sensor and the navigation stage of the underwater vehicle, for example, assigning a label to each category, sampling the signal-to-noise ratio according to a certain distribution, and using the signal-to-noise ratio to mix the data of any two categories of labels to obtain a mixed noise that is close to the real noise, which is used as the input data of the network and creates a dataset. A single sample in the dataset is as follows Figure 4, which is a mixture of vibration noise, flow noise, and electrical noise. The contribution ratio of the three noises is [0.2:0.3:0.5]. At this time, the length of a single mixed noise as the network input is 1s and the dimension is (1,48000).

[0097] Step 2 specifically includes: using Figure 2 The encoder shown processes the input noise, and the encoder parameters are shown in Table 1, obtaining a shallow mixed representation with a dimension of (256,48000).

[0098] Step 3 specifically includes: further processing the shallow mixed representation obtained in step 2, using one-dimensional convolution blocks multiple times to extract deep representations according to the two dimensions of time and frequency. A single process is as follows Figure 2 The feature extraction is shown in Table 2. Step 3 obtains the deep representation through deep representation extraction, and its dimension is (4,750).

[0099] Step 4 specifically includes: using Figure 3 The energy estimator shown processes the deep representation M in step 3 d , we get the two types of labels and mixed energy ratio of the mixed data. The specific parameters of the energy estimator are shown in Table 3. The first recognition task is achieved by Figure 4 The mixing ratio and category are optimized using the cross entropy loss function. At the same time, the deep representation M d Input such as Figure 3 In the mask estimator shown, separation masks of various component noises are obtained. In this example, the separation mask dimension is (3, 256, 48000).

[0100] Step 5 specifically involves multiplying the separation mask with the shallow mixed representation to obtain separate representations of flow noise, vibration noise, and electrical noise. A decoder Ht, similar in structure to the encoder in step 2, decodes these separate representations. The decoded data is the waveform-recovered signal. The loss function is calculated using the scale-invariant signal-to-noise ratio.

[0101] When updating the global network, the cross entropy loss function in step 4 and the scale-invariant signal-to-noise ratio loss function in step 5 are first descaled and then summed up to obtain the loss function of the global network. Finally, the Adam optimizer is used to optimize the global parameters.

[0102] loss=loss1 2 +loss2 2 (twenty one)

[0103] Step 6 specifically includes: weighting the separated noise in step 5 according to the mixed energy ratio obtained in step 4 to obtain the final three estimated signals.

[0104]

Claims

1. A method for underwater vehicle self-noise feature extraction, identification and separation based on multi-task learning, characterized in that: The following steps are involved: Step 1: Use an online synthesis method to mix the noises of different types and stages collected to obtain a mixed noise dataset composed of multiple noises. The noise category and the proportion of mixed energy constitute the input label of the noise; Step 2: Use the encoder to preliminarily process the mixed noise to obtain a shallow mixed representation; Step 3: For the shallow mixed representation, use a high-dimensional feature extractor to extract the deep representation; Step 4: Use the energy estimator to process the deep representation to obtain the category and energy proportion output of each noise component in the mixed noise. This output is optimized using the input label; at the same time, use the mask estimator to obtain the separation mask of each noise component in the mixed noise; Step 5: Multiply the separation mask of each noise component with the shallow mixed representation to obtain the separation representation of each noise component. Input the separation representation into the decoder corresponding to the encoder structure to obtain the time domain waveform of each noise component. Step 6: Multiply the noise category and energy proportion with the recovered time domain waveform to generate the estimated signal of each type of noise.

2. The method for extracting, identifying and separating underwater vehicle self-noise features based on multi-task learning according to claim 1, characterized in that: In step 1, the collected noise is divided into different categories according to the type of the collection sensor and the navigation stage of the underwater vehicle, and each category is assigned a label; According to the initial set distribution, the signal-to-noise ratio is sampled within a certain range. For one of the signal-to-noise ratio sampling results, the signal-to-noise ratio is used to mix the data of any two types of labels to obtain the corresponding mixed noise; each mixed noise sample is mixed with a different signal-to-noise ratio sampling result to obtain the corresponding data set.

3. The method for extracting, identifying and separating underwater vehicle self-noise features based on multi-task learning according to claim 1, characterized in that: In step 2, the encoding process is as shown in formula (1): Y=H(x) (1) Where x is the one-dimensional mixed noise input with a length of T, Y is the encoder output, and H is the encoder; The encoding process of encoder H is shown in formula (2): in, and Respectively represent different frequencies θ i The convolution block initialized by the cosine function and wavelet function under and denotes the convolution output at the i-th frequency, cat(·) denotes the concatenation in the frequency dimension, and Conv2d(·) is a two-dimensional convolution block for weighted summation.

4. The method for extracting, identifying and separating underwater vehicle self-noise features based on multi-task learning according to claim 1, characterized in that: In step 3, the extraction process is to use one-dimensional convolution blocks multiple times to perform deep feature extraction according to the two dimensions of time and frequency. The formula is: Among them, Conv1d T (·) and Conv1d F (·) are one-dimensional convolution blocks for time dimension reduction and frequency dimension reduction, respectively. and They represent the dimensionality reduction results in the b-th feature extraction step, and B is the total number of feature extraction steps; Repeat the process of formula (3) above to compress the shallow mixed representation from both time and frequency dimensions, and generate a deep representation M in the compressed time-frequency dimension. d , that is, M (B) ,The above steps are called high-dimensional feature extractor.

5. The method for extracting, identifying and separating underwater vehicle self-noise features based on multi-task learning according to claim 1, characterized in that: In step 4, the energy estimator is a plurality of multi-layer perceptron (MLP) layers, which output the result of the estimated category. Here, the label information obtained in step 1 and the signal-to-noise ratio corresponding to the label are used to optimize the network parameters of the MLP layer using the cross entropy loss function. The formula is as follows: P={p1,...,p C }=MLP(M d ) (4) Where P = {p1,...,p C } is the C output of the recognition task, p c is the energy proportion of the Cth noise component estimated by the network, M d For deep representation, label c is the true value of the recognition task, loss1 is the loss value of the recognition task; At the same time, a mask estimator is used to estimate the separation mask of each noise component, as follows: Mask is the sequence of separation masks that generates the desired number of components, C is the number of components that are desired to be separated, Generate upper-level representation of separation mask; M d For deep representation.

6. The method for extracting, identifying and separating underwater vehicle self-noise features based on multi-task learning according to claim 1, characterized in that: In step 5, the separation of the cth noise is characterized by The data decoded by the decoder is the waveform recovery signal Will Compared with pure separation noise without energy information, the network parameters are optimized using the scale-invariant signal-to-noise ratio (SI-SNR) as the loss function: Among them, Mask(c) is the c-th noise separation mask, is the signal recovered from the cth waveform, x (c) is the time domain waveform of the c-th noise, loss2 is the loss value of the separation task; We further obtain the loss function of the global network, and finally use the Adam optimizer to optimize the global parameters: loss=loss1 2 +loss2 2 (9)。 7. The method for extracting, identifying and separating underwater vehicle self-noise features based on multi-task learning according to claim 1, characterized in that: In step 6, the noise separation in step 5 is weighted according to the mixed energy ratio obtained in step 4 to obtain the final estimated signal. Among them, p c is the energy proportion of the cth noise component estimated by the network, is the signal recovered from the cth waveform, is the c-th estimated signal.

Images