Self-noise separation method for unmanned underwater vehicle based on time-frequency domain characteristics and domain adaptation

Abstract

The invention provides an unmanned underwater vehicle self-noise separation method based on time-frequency domain characteristics and domain self-adaption, and relates to the field of unmanned underwater vehicle self-noise separation, and the method comprises the steps: obtaining a to-be-separated self-noise signal, and carrying out the short-time Fourier transform and coding processing, and obtaining a time-frequency characteristic pattern; time domain features and frequency domain features are extracted from the time-frequency feature map, time domain and frequency domain weights are determined based on the time-frequency features of prior noise and signal weights in the to-be-separated self-noise signals, and then time-frequency domain fusion features are obtained; and performing domain adaptive normalization processing on the time-frequency domain fusion features, and finally performing decoding and inverse short-time Fourier transform to obtain a separation result. Through the method, the weights of the time domain and frequency domain features can be adaptively adjusted, so that the precision of multi-source noise separation is improved; and meanwhile, the normalization intensity can be adaptively adjusted according to different working conditions, and original features are reserved, so that the problem of inaccurate self-noise separation of the UUV under different working conditions is solved.

Description

Technical Field

[0001] This invention relates to the field of self-noise separation for unmanned underwater vehicles, and more particularly to a self-noise separation method for unmanned underwater vehicles based on time-frequency domain characteristics and domain adaptation. Background Technology

[0002] As a device capable of acquiring and processing underwater information data, Unmanned Underwater Vehicles (UUVs) shoulder the mission of developing and utilizing marine resources and ensuring national maritime security. However, in practical applications, due to platform size limitations, the underwater acoustic signal receiving system is affected by the noise of the UUV itself, impacting its receiving performance. UUV self-noise is noise generated by the platform itself during movement, affecting the platform's transducer array. The magnitude of UUV self-noise determines the performance indicators of the UUV acoustic payload and has become crucial for improving related technical performance. UUVs have complex structures and diverse noise sources, and the interference effects of self-noise from different sources on the acoustic payload system vary. Extracting and separating different types of self-noise from UUV self-noise and implementing targeted noise control are effective ways to improve UUV receiving capabilities.

[0003] Generally, self-noise originates from three main sources: vibration noise, flow noise, and electrical noise. Vibration noise primarily arises from the periodic collisions of functional equipment during underwater movement, transmitting the noise through the platform's own structural vibrations to the transducer array. Flow noise mainly stems from the noise generated on the surface of the underwater vehicle due to turbulent pulsating pressure, and the interaction between this pressure and the vehicle's cavities, appendages, and protrusions; its generation is closely related to the underwater vehicle's structural characteristics and motion state. Electrical noise primarily originates from the underwater vehicle's internal electronic equipment and electrical systems during operation. These noise sources include, but are not limited to, electromagnetic interference from the power motor, armature and magnetic systems, and brush noise. Furthermore, in received data, the noise level of self-noise is generally much higher than that of ambient noise, and it partially overlaps with the target signal in multiple characteristic dimensions, affecting the system's signal detection and judgment. Depending on the UUV's structure, motion state, and mission, the contribution and energy proportion of the three types of self-noise to the overall self-noise vary, and the difficulty of researching suppression and cancellation methods also differs. Conducting a general study of self-noise may not yield satisfactory results.

[0004] Self-noise separation technology is generally used as a preprocessing method to decompose the total self-noise into single-source noise components, enabling self-noise characteristic analysis and noise generation mechanism research, thereby providing prior guidance for suppressing or canceling self-noise of interest. UUV self-noise separation extracts the single-source components from the total self-noise. During extraction, each single-source component has equal weight, without excessive focus on any one component. For example, when operating in a quiet background environment, the data collected by a UUV through a hydrophone transducer can be considered as its own generated noise, which includes vibration noise, flow noise, and electrical noise—three different sources mentioned earlier. Self-noise separation extracts these three self-noise components separately from the total self-noise, obtaining a time-domain sequence with characteristics approximating the true components.

[0005] In the field of noise signal separation, deep learning methods have gradually become the mainstream solution for noise signal separation tasks in recent years, thanks to the rapid development of computing power and machine learning technology. Among them, attention mechanisms, as a key deep learning technique, have attracted much attention because they can dynamically focus on important features in the input data. Attention mechanisms can effectively capture temporal or high-dimensional dependencies in complex data, and are particularly suitable for tasks such as audio signal processing.

[0006] However, directly applying attention mechanisms to noise signal separation tasks, especially in underwater acoustic environments involving unmanned underwater vehicles (UUVs), faces two major challenges. First, attention-based models rely primarily on learned general representations, making it difficult to fully exploit the unique properties of different signals. This limitation in feature extraction leads to poor separation performance. Second, signals generated by UUVs at different speeds exhibit significant differences in statistical properties, such as spectral distribution, amplitude, or temporal dynamics. These differences in the signal domain challenge the model's generalization ability. For example, a model trained at a specific speed may not be effective in adapting to signals at other speeds, resulting in decreased separation accuracy and affecting the model's reliability in real-world underwater environments. Summary of the Invention

[0007] To address the issues of low separation accuracy and poor adaptability to different speeds in existing UUV self-noise separation methods, this invention proposes a self-noise separation method for unmanned underwater vehicles based on time-frequency domain characteristics and domain adaptation. The method includes:

[0008] Step 1: Obtain the noise signal to be separated;

[0009] Step 2: Perform short-time Fourier transform and convolutional neural network-based encoding processing on the self-noise signal to be separated to obtain a time-frequency feature map;

[0010] Step 3: Perform feature extraction processing on the time-frequency feature map based on the attention mechanism to obtain the time-domain and frequency-domain features of the noise signal to be separated;

[0011] Step 4: Based on the time-frequency characteristics of the three preset prior noise signals and the weights of the three noise signals in the noise signal to be separated, determine the time-domain weights corresponding to the time-domain features and the frequency-domain weights corresponding to the frequency-domain features.

[0012] Step 5: Perform weighted calculation on the time-domain features and the frequency-domain features based on the time-domain weights and the frequency-domain weights to obtain the time-frequency domain fusion features of the noise signal to be separated;

[0013] Step 6: Normalize the time-frequency domain fusion features to obtain normalized fusion features;

[0014] Step 7: Perform weighted calculation on the time-frequency domain fusion feature and the normalized fusion feature based on the preset normalized weight coefficient to obtain the output feature;

[0015] Step 8: Decode the output features based on a convolutional neural network and perform an inverse short-time Fourier transform to obtain the separation result of the noise signal to be separated.

[0016] Optionally, the three prior noise signals include prior electrical noise signal, prior flow noise signal, and prior vibration noise signal;

[0017] Step 4 includes:

[0018] Step 41: Perform softmax-based classification processing on the noise signal to be separated to obtain the weights of the three noise signals in the noise signal to be separated.

[0019] Step 42: Calculate the phase offset corresponding to the three preset prior noise signals according to the following formula:

[0020] ,

[0021] in, This represents the phase offset corresponding to the i-th type of noise signal. This represents the total energy of the i-th prior noise signal in the time domain. Let represent the total energy of the i-th prior noise signal in the frequency domain; Represents the hyperbolic tangent function;

[0022] Step 43: Based on the weights of the three noise signals in the noise signal to be separated and the phase offsets corresponding to the three noise signals respectively, calculate the overall phase offset according to the following formula:

[0023] ,

[0024] in, Indicates the overall phase shift; , , These represent the weights of electrical noise, flow noise, and vibration noise in the noise signal to be separated, respectively. , , These represent the phase offsets corresponding to the electrical noise signal, the flow noise signal, and the vibration noise signal, respectively.

[0025] Step 44: Calculate the time-domain weights corresponding to the time-domain features and the frequency-domain weights corresponding to the frequency-domain features according to the following formulas:

[0026]

[0027] in, This represents the temporal weights corresponding to the temporal features. This represents the frequency domain weight corresponding to the frequency domain feature.

[0028] Optionally, step 5 includes:

[0029] Step 51: Calculate the time-domain features and the frequency-domain features based on the time-domain weights and the frequency-domain weights according to the following formula to obtain the initial fused features:

[0030] ,

[0031] in, X represents the initial fusion feature; X represents the time-frequency feature map; This represents the time-domain feature. This represents the frequency domain characteristics;

[0032] Step 52: Based on the time-frequency feature map and the initial fusion features, calculate the time-frequency domain fusion features according to the following formula:

[0033] ,

[0034] Where Y represents the time-frequency domain fusion feature, " represents the Hadamard product operator.

[0035] Optionally, step 6 includes:

[0036] Step 61: Normalize the time-frequency domain fusion features according to the following formula to obtain normalized fusion features:

[0037] ,

[0038] in, This represents the value at position (h, w) of the nth sample, cth channel, in the normalized fusion feature; This represents the value at position (h, w) of the nth sample, cth channel, in the time-frequency domain fusion feature. This represents the mean of the nth sample and the cth channel in the time-frequency domain fusion feature. This represents the variance of the nth sample and the cth channel in the time-frequency domain fusion feature; It is a constant.

[0039] Optionally, step 7 includes:

[0040] Step 71: According to the following formula, the time-frequency domain fusion feature and the normalized fusion feature are weighted and calculated based on preset normalized weight coefficients to obtain the output feature:

[0041] ,

[0042] in, This represents the value at position (h, w) of the nth sample, cth channel, in the output feature. The preset normalized weighting coefficients are determined according to the following formula:

[0043] ,

[0044] in, This represents a preset scalar parameter.

[0045] Optionally, step 8 includes:

[0046] Step 81: Perform decoding processing on the output features based on a convolutional neural network to obtain the separation masks corresponding to the three noise signals respectively;

[0047] Step 82: Multiply the separation mask element-wise with the amplitude spectrum of the noise signal to be separated to obtain the separation amplitude spectra corresponding to the three noise signals respectively; wherein, the amplitude spectrum is obtained by performing a short-time Fourier transform on the noise signal to be separated;

[0048] Step 83: Perform an inverse short-time Fourier transform based on the separation amplitude spectrum and the phase information of the noise signal to be separated to obtain the separation result of the noise signal to be separated; wherein, the phase information is obtained by performing a short-time Fourier transform on the noise signal to be separated.

[0049] Optionally, step 2 includes:

[0050] The noise signal to be separated is subjected to a short-time Fourier transform based on a Hanning window of length 256 and an overlap rate of 50% to obtain a two-dimensional noise signal.

[0051] The two-dimensional noise signal is encoded using a convolutional neural network to obtain a time-frequency feature map.

[0052] Optionally, step 1 includes:

[0053] Acquire the original mixed noise signal;

[0054] The original mixed noise signal is preprocessed by standardization to obtain the noise signal to be separated.

[0055] The embodiments described in this invention have the following advantages:

[0056] This invention provides a method for separating self-noise from unmanned underwater vehicles based on time-frequency domain features and domain adaptation. After acquiring the self-noise signal to be separated, a short-time Fourier transform and a convolutional neural network-based encoding process are performed on the self-noise signal to be separated to obtain a time-frequency feature map. Then, the time-domain and frequency-domain features of the self-noise signal to be separated are extracted from it. Based on the time-frequency characteristics of three preset prior noise signals and the weights of the three noise signals in the self-noise signal to be separated, the time-domain weights and frequency-domain weights are determined, and then weighted calculations are performed to obtain the time-frequency domain fusion features of the self-noise signal to be separated. The time-frequency domain fusion features are normalized and weighted to obtain the output features. Finally, the output features are decoded based on a convolutional neural network and subjected to an inverse short-time Fourier transform to obtain the separation result of the self-noise signal to be separated.

[0057] By employing the above methods, attention mechanisms are used to extract noise features from both the time and frequency domains. A phase-locked mechanism is then used to adaptively adjust the attention weights in the time and frequency domains based on the time-frequency characteristics of electrical, current, and vibration noise, as well as the signal proportion in the self-noise signal to be separated. This significantly improves the accuracy of UUV multi-source noise separation. Simultaneously, a domain-adaptive normalization process is used to dynamically balance the noise signal features and normalized features, enabling the model to adaptively adjust the normalization intensity based on data under specific operating conditions while preserving the original content features. This addresses the problem of inaccurate UUV self-noise separation under different operating conditions, further improving separation accuracy.

[0058] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0059] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which:

[0060] Figure 1 A flowchart illustrating the steps of an embodiment of the self-noise separation method for unmanned underwater vehicles based on time-frequency domain features and domain adaptation proposed in this invention is provided.

[0061] Figure 2 A flowchart illustrating the steps of another embodiment of the self-noise separation method for unmanned underwater vehicles based on time-frequency domain features and domain adaptation provided by the present invention is given.

[0062] Figure 3 A flowchart illustrating the steps of an embodiment of the phase-locking mechanism provided by the present invention is given. Detailed Implementation

[0063] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0064] Reference Figure 1 The present invention provides a flowchart of the steps of an embodiment of a self-noise separation method for unmanned underwater vehicles based on time-frequency domain features and domain adaptation, the method comprising:

[0065] Step 1: Obtain the noise signal to be separated.

[0066] Among them, the noise signal to be separated refers to the mixed noise signal, including electrical noise, flow noise and vibration noise, collected during the operation of UUV. Figure 2 A flowchart of another embodiment of the self-noise separation method for unmanned underwater vehicles based on time-frequency domain features and domain adaptation provided by the present invention is given, wherein the one-dimensional time-domain noise signal is the self-noise signal to be separated in step 1.

[0067] Optionally, step 1 includes:

[0068] Acquire the original mixed noise signal;

[0069] The original mixed noise signal is preprocessed by standardization to obtain the noise signal to be separated.

[0070] To ensure data consistency, the acquired raw mixed noise signal can be standardized to achieve a mean of zero and a variance of one, thereby reducing errors caused by data variability. For example, the standardized preprocessed signal can be divided into fixed-length segments, each containing 2000 or 3000 sampling points, the number of which can be determined based on the actual situation.

[0071] Step 2: Perform short-time Fourier transform and convolutional neural network-based encoding processing on the self-noise signal to be separated to obtain a time-frequency feature map.

[0072] like Figure 2 The method shown is to perform a Short Time Fourier Transform (STFT) on the noise signal to be separated, which can convert the time-domain signal into a time-frequency domain representation, thereby obtaining a two-dimensional time-frequency signal. The STFT calculation can simultaneously preserve the amplitude spectrum and phase spectrum, providing rich feature representations for subsequent processing and improving the accuracy of signal separation.

[0073] The two-dimensional time-frequency signal obtained from the STFT is then encoded using a convolutional neural network. This process consists of K two-dimensional convolutional layers, each performing a downsampling operation, followed by a residual block to enhance feature extraction and mitigate the vanishing gradient problem. The residual block comprises two 3×3 convolutional layers and batch normalization, with a dropout rate of 0.2 to prevent overfitting. The number of convolutional layers K can be adjusted according to specific requirements, such as 2 or 4.

[0074] Optionally, step 2 includes:

[0075] The noise signal to be separated is subjected to a short-time Fourier transform based on a Hanning window of length 256 and an overlap rate of 50% to obtain a two-dimensional noise signal.

[0076] The two-dimensional noise signal is encoded using a convolutional neural network (CNN) to obtain a time-frequency feature map.

[0077] Using a Hanning window of length 256 and an overlap ratio of 50% when performing STFT calculations helps to achieve a balance between time resolution and frequency resolution.

[0078] In addition, to improve the generalization ability and robustness of the model, after performing STFT on the self-noise signal to be separated and before encoding, data augmentation operations can be performed on the time-frequency domain signal obtained by STFT calculation, such as Hilbert transform and introducing linear phase shift.

[0079] Step 3: Perform feature extraction processing on the time-frequency feature map based on the attention mechanism to obtain the time-domain and frequency-domain features of the noise signal to be separated.

[0080] In order to capture the unique characteristics of noise types, such as Figure 2As shown, this invention proposes two attention mechanisms: frequency domain attention is achieved through a self-attention layer along the frequency axis, focusing on harmonic structure and formant features; time domain attention is achieved through a self-attention layer along the time axis, focusing on rhythm and transient features.

[0081] Specifically, in the temporal attention part, the time dimension of the time-frequency feature map obtained after CNN encoding is used as the sequence length, and it is processed through an attention mechanism to obtain temporal features. Specifically, a linear transformation is performed on the input... That is, the time-frequency feature map in the time dimension generates query (Q), key (K), and value (V) matrices, which are represented as follows:

[0082] (1)

[0083] in, This is the weight matrix obtained through training. Subsequently, the temporal attention output, i.e., the temporal features, is obtained by calculating the scaled dot product and applying the softmax function to normalize it. The calculation formula is as follows:

[0084] (2)

[0085] in, This represents the dimension of the key vector, used to scale the dot product to mitigate the effects of large values. This indicates the temporal features obtained after processing by the temporal attention mechanism.

[0086] Similarly, in the frequency domain attention part, the frequency dimension of the time-frequency feature map is used as the sequence length, and a similar self-attention mechanism is employed to capture the frequency domain features of the self-noise signal to be separated. The calculation process can be expressed as follows:

[0087] (3)

[0088] in, This represents the frequency domain features obtained after processing by the frequency domain attention mechanism. Represents a time-frequency diagram in the frequency domain.

[0089] By integrating time-domain and frequency-domain attention, the features of noise in the time and frequency dimensions can be effectively extracted, thereby enhancing the model's ability to separate different types of noise and improving separation accuracy.

[0090] Step 4: Based on the time-frequency characteristics of the three preset prior noise signals and the weights of the three noise signals in the noise signal to be separated, determine the time-domain weights corresponding to the time-domain features and the frequency-domain weights corresponding to the frequency-domain features.

[0091] The three noise signals include electrical noise, flow noise, and vibration noise. Correspondingly, the three a priori noise signals include a priori electrical noise, a priori flow noise, and a priori vibration noise.

[0092] like Figure 2 As shown, this invention proposes a phase-locked mechanism that dynamically adjusts the weights of two attention mechanisms based on the characteristics of electrical noise, flow noise, and vibration noise signals and the proportions of the three types of noise in the noise signal to be separated, so as to separate the signal according to the prominent features of the signal in the time domain or frequency domain.

[0093] Specifically, a priori noise signal from a single source can be introduced first. By calculating the ratio of its time-domain characteristics to its frequency-domain characteristics, such as the ratio of envelope energy to high-frequency energy, the prominence of this type of noise in the time or frequency domain can be obtained. Then, combined with the proportion of various types of noise in the noise signal to be separated, the weights of the time-domain attention and frequency-domain attention of the noise signal to be separated can be further derived. This allows for flexible adjustment of the processing strategy. That is, for different types and proportions of noise signals to be separated, the processing method that emphasizes the time domain, frequency domain, or a balance of both can be adaptively selected according to the prominence of the signal in the time or frequency domain. This achieves more efficient mixed noise analysis and suppression, and improves signal separation accuracy.

[0094] Optionally, step 4 includes:

[0095] Step 41: Perform softmax-based classification processing on the noise signal to be separated to obtain the weights of the three noise signals in the noise signal to be separated.

[0096] Step 42: Calculate the phase offset corresponding to the three preset prior noise signals according to the following formula:

[0097] (4)

[0098] in, This represents the phase offset corresponding to the i-th type of noise signal. This represents the total energy of the i-th prior noise signal in the time domain. Let represent the total energy of the i-th prior noise signal in the frequency domain; Represents the hyperbolic tangent function;

[0099] Step 43: Based on the weights of the three noise signals in the noise signal to be separated and the phase offsets corresponding to the three noise signals respectively, calculate the overall phase offset according to the following formula:

[0100] (5)

[0101] in, Indicates the overall phase shift; , , These represent the weights of electrical noise, flow noise, and vibration noise in the noise signal to be separated, respectively. , , These represent the phase offsets corresponding to the electrical noise signal, the flow noise signal, and the vibration noise signal, respectively.

[0102] Step 44: Calculate the time-domain weights corresponding to the time-domain features and the frequency-domain weights corresponding to the frequency-domain features according to the following formulas:

[0103]

[0104] in, This represents the temporal weights corresponding to the temporal features. This represents the frequency domain weight corresponding to the frequency domain feature.

[0105] Reference Figure 3 This paper presents a flowchart illustrating the steps of an embodiment of a phase-locked mechanism provided by the present invention, namely, the steps of calculating time-domain weights and frequency-domain weights based on the time-frequency characteristics of prior noise and the weight ratio of the noise signal to be separated. The core steps include the weight calculation of the three types of noise in the noise signal to be separated, the determination of phase offset, the synthesis of overall phase offset, and the dynamic allocation of time-frequency weights.

[0106] Specifically, a classification method based on the Softmax function can be used to calculate normalized weights for the three noise types in the input signal to be separated from noise. For example... Figure 2 As shown, the specific process may include: adaptive average pooling, two-dimensional convolution, activation process based on the Leaky ReLU activation function, two-dimensional convolution, and softmax classification. The number of channels for the three processes—two-dimensional convolution, activation process, and two-dimensional convolution—are 64, 64, and 3, respectively, and the weight calculation formula is shown below:

[0107] (8)

[0108] Where X refers to the input signal. Let represent the normalized weights of electrical noise, flow noise, and vibration noise, respectively, and satisfy . ;V ( () represents a feature compression and mapping function, where the input signal is... This means there are C feature channels, each with T time points of data, and the output feature... This represents the weights of the three types of noise (electrical, current, and vibration) after mapping; "" indicates that the function V receives a C×T input feature and finally compresses and maps it into a three-dimensional vector.

[0109] To accurately characterize the time-frequency characteristics of each noise type, this invention uses prior noise signals to calculate phase offset values ​​for each noise type. Specifically, by analyzing the statistical characteristics of the noise signal in the time and frequency domains, its energy ratio is extracted, and nonlinear mapping is performed, i.e., step 42, to obtain the phase offset parameter that reflects the essential characteristics of the noise.

[0110] In terms of temporal feature extraction, this invention utilizes envelope energy. As a time-domain characteristic of noise, envelope energy can effectively reflect the amplitude modulation characteristics and low-frequency variation trend of noise, and is particularly suitable for characterizing noise with periodic vibration characteristics. The calculation steps are as follows:

[0111] First, apply the Hilbert transform to the original prior noise signal x[n] to obtain the analytic signal z[n];

[0112] Secondly, the envelope is obtained by taking the modulus of the analytic signal z[n] according to the following formula:

[0113] a[n]=|z n | (9)

[0114] Then, the energy of the envelope a[n] is calculated according to the following formula, which is used as a time-domain feature:

[0115] E t,i = 1 N ∑ n=0 N - 1 a [n ] 2 (10)

[0116] In terms of frequency domain feature extraction, this invention utilizes high-frequency energy... Characterizing the frequency domain properties of noise, especially the intensity and variation trend of high-frequency components, is suitable for describing noise with high-frequency harmonic characteristics, such as electrical noise. The calculation process includes:

[0117] First, a short-time Fourier transform is performed on the prior noise signal to obtain its complex spectral representation X[k,n], where k is the frequency index and n is the time frame index;

[0118] Secondly, calculate the amplitude spectrum for each frame. P[k,n] :

[0119] P[k,n]=|X[k,n] | 2 (11)

[0120] Then, the energy of the frequency components within the high-frequency band is summed to obtain the high-frequency energy. The frequency index range for the high-frequency band is from k1 = 1kHz to k2 = , The preset sampling rate is:

[0121] E f,i = 1 N ∑ k1 k 2 P [k ] 2 (12)

[0122] In obtaining time domain features Frequency domain characteristics Then, by calculating the energy ratio of the two and applying a logarithmic operation to amplify the dynamic range, as shown in step 42, the tanh function is used to map this ratio to the interval [−1, 1]. This mapping result is the phase bias parameter that reflects the characteristics of the noise type.

[0123] After obtaining the phase offsets of various types of noise, the overall phase offset Δϕ is calculated by weighted summation, i.e., step 43.

[0124] Steps 41-43 calculate the overall phase shift Δϕ by integrating the weights of noise types with their time-frequency distribution characteristics, thereby comprehensively characterizing the multi-source components in the mixed noise signal and fully extracting the unique features of the signal. Finally, based on the obtained Δϕ, the time-domain weights are constructed according to step 44 using the periodicity and smoothness of trigonometric functions. and frequency domain weights Where Δϕ∈[-1,1]. This constraint utilizes the monotonicity of the sin and cos functions in [0, π / 2] to guarantee... and The values ​​of are monotonically varying within the range [0, 1], and simultaneously satisfy the normalization condition. .

[0125] Therefore, by using the method described in steps 41-44, the prior information of the three noise types—flow noise, electrical noise, and vibration noise—is utilized to obtain the time-frequency characteristics, and combined with the weight ratio of the three noise types in the noise signal to be separated, weight coefficients for dynamically adjusting the attention in the time and frequency domains are obtained, thereby helping to improve the accuracy and targeting of noise separation.

[0126] Step 5: Perform weighted calculations on the time-domain features and the frequency-domain features based on the time-domain weights and the frequency-domain weights to obtain the time-frequency domain fusion features of the noise signal to be separated.

[0127] Through steps 21-44 and step 5, the integration of time-domain and frequency-domain attention is achieved, which can effectively extract the features of noise in the time and frequency dimensions, thereby enhancing the model's ability to separate different types of noise.

[0128] Optionally, step 5 includes:

[0129] Step 51: Calculate the time-domain features and the frequency-domain features based on the time-domain weights and the frequency-domain weights according to the following formula to obtain the initial fused features:

[0130] (13)

[0131] in, X represents the initial fusion feature; X represents the time-frequency feature map; This represents the time-domain feature. This represents the frequency domain characteristics;

[0132] Step 52: Based on the time-frequency feature map and the initial fusion features, calculate the time-frequency domain fusion features according to the following formula:

[0133] (14)

[0134] Where Y represents the time-frequency domain fusion feature, " represents the Hadamard product operator.

[0135] After fusing the time-domain and frequency-domain features of the attention output according to formula (13), the time-frequency feature map X can also be combined with the weighted features. Element-wise multiplication is performed to obtain the output feature Y, which is then used to fuse attention information from different sources and highlight the time-frequency characteristics of the noise signal.

[0136] Step 6: Normalize the time-frequency domain fusion features to obtain normalized fusion features.

[0137] After obtaining the time-frequency domain fusion characteristics, in order to adapt to the differences in data distribution under different UUV operating conditions, such as flight speed, etc. Figure 2 As shown, a domain adaptive normalization process, including steps 6 and 7, is performed. Step 6 performs a normalization process on the time-frequency domain fused features. Specifically, step 6 includes:

[0138] Step 61: Normalize the time-frequency domain fusion features according to the following formula to obtain normalized fusion features:

[0139] (15)

[0140] in, This represents the value at position (h, w) of the nth sample, cth channel, in the normalized fusion feature; This represents the value at position (h, w) of the nth sample, cth channel, in the time-frequency domain fusion feature. This represents the mean of the nth sample and the cth channel in the time-frequency domain fusion feature. This represents the variance of the nth sample and the cth channel in the time-frequency domain fusion feature; It is a constant, specifically a small constant used to prevent division by zero.

[0141] and Decibels can be expressed as:

[0142] (16)

[0143] (17)

[0144] Where H and W represent the dimensions of the feature map of the nth sample and the cth channel in the time-frequency domain fusion feature.

[0145] The normalization process described in step 61 reduces the statistical differences in input data among different samples by standardizing features to zero mean and unit variance, which is particularly important for tasks that require cross-domain generalization.

[0146] Step 7: Perform weighted calculations on the time-frequency domain fusion features and the normalized fusion features based on preset normalized weight coefficients to obtain output features.

[0147] Specifically, step 7 includes:

[0148] Step 71: According to the following formula, the time-frequency domain fusion feature and the normalized fusion feature are weighted and calculated based on preset normalized weight coefficients to obtain the output feature:

[0149] (18)

[0150] in, This represents the value at position (h, w) of the nth sample, cth channel, in the output feature. The preset normalized weighting coefficients are determined according to the following formula:

[0151] (19)

[0152] in, This represents a preset scalar parameter.

[0153] Preset normalized weighting coefficients It is a learnable weight parameter used to dynamically control the degree of normalization, enabling the model to adaptively adjust the normalization strength according to the specific task and data. The scalar parameter b can be calculated by applying the sigmoid function as shown in formula (19). The initial value of b is 0.5. During forward propagation, b is converted into a weight value between 0 and 1 by the sigmoid activation function, thus allowing flexible control over the fusion of time-frequency domain fusion features and normalized fusion features. Of course, other functions that can map parameters to [0,1] can also be chosen depending on the specific situation.

[0154] Therefore, through the domain adaptive normalization process described in steps 6-7, the normalization weight coefficients can be learned to dynamically determine the normalization strength based on the characteristics of the training data. When the noise signal data comes from different flight speed conditions, the model can reduce inter-domain statistical differences and improve the model's generalization ability by adjusting α. Integrating domain adaptive normalization into the model and optimizing parameters during the training process can significantly improve the performance and robustness of the signal separation task.

[0155] Step 8: Decode the output features based on a convolutional neural network and perform an inverse short-time Fourier transform to obtain the separation result of the noise signal to be separated.

[0156] Among them, such as Figure 2 As shown, the convolutional neural network consists of K convolutional layers. Its output layer generates three separation masks, corresponding to electrical noise, flow noise, and vibration noise, respectively. The masks are calculated using the softmax activation function, ensuring that the sum of the mask values ​​at each time-frequency point is 1, representing the relative contribution of each noise component. Finally, the signal is recovered from the two-dimensional time-frequency domain to a one-dimensional signal through inverse short-time Fourier transform, thus obtaining the separation results: the separated electrical noise signal, flow noise signal, and vibration noise signal.

[0157] Optionally, step 8 includes:

[0158] Step 81: Perform decoding processing on the output features based on a convolutional neural network to obtain the separation masks corresponding to the three noise signals respectively;

[0159] Step 82: Multiply the separation mask element-wise with the amplitude spectrum of the noise signal to be separated to obtain the separation amplitude spectra corresponding to the three noise signals respectively; wherein, the amplitude spectrum is obtained by performing a short-time Fourier transform on the noise signal to be separated;

[0160] Step 83: Perform an inverse short-time Fourier transform based on the separation amplitude spectrum and the phase information of the noise signal to be separated to obtain the separation result of the noise signal to be separated; wherein, the phase information is obtained by performing a short-time Fourier transform on the noise signal to be separated.

[0161] To improve the separation effect and the reconstruction accuracy of the separated signal, the generated separation mask is multiplied element-wise with the amplitude spectrum of the noise signal to be separated to obtain the separated amplitude spectrum of each noise component. Then, combined with the phase information stored in the STFT, the separated time-frequency domain signal is reconstructed into three time-domain signals through inverse short-time Fourier transform. The reconstruction process uses the same Hanning window and overlap parameters as the forward STFT to ensure signal continuity and high-quality output.

[0162] In summary, this invention provides a method for separating self-noise from unmanned underwater vehicles based on time-frequency domain features and domain adaptation. After acquiring the self-noise signal to be separated, a short-time Fourier transform and a convolutional neural network-based encoding process are performed on the self-noise signal to be separated to obtain a time-frequency feature map. Then, the time-domain and frequency-domain features of the self-noise signal to be separated are extracted from it. Based on the time-frequency characteristics of three preset prior noise signals and the weights of the three noise signals in the self-noise signal to be separated, the time-domain weights and frequency-domain weights are determined, and then weighted calculations are performed to obtain the time-frequency domain fusion features of the self-noise signal to be separated. The time-frequency domain fusion features are normalized and weighted to obtain the output features. Finally, the output features are decoded based on a convolutional neural network and subjected to an inverse short-time Fourier transform to obtain the separation result of the self-noise signal to be separated.

[0163] The above method utilizes an attention mechanism to extract noise features from both the time and frequency domains. A phase-locked mechanism is then employed to adaptively adjust the attention weights in the time and frequency domains based on the time-frequency characteristics of electrical, current, and vibration noise, as well as the signal proportion within the noise signal to be separated. This significantly improves the accuracy of multi-source noise separation. Simultaneously, a domain-adaptive normalization process dynamically balances the noise signal's features and the normalized features, enabling the model to adaptively adjust the normalization intensity based on data under specific operating conditions while preserving the original content features. This addresses the problem of inaccurate UUV self-noise separation under different operating conditions, further enhancing separation accuracy.

[0164] Optionally, the method may further include:

[0165] The root mean square error (RMSE) of the energy proportions of each signal before and after separation is used to evaluate the separation effect and then optimize it. Specifically, the RMSE can be calculated based on the true energy proportion of the self-noise signal and the energy proportion of the separated signal predicted by the method described in this invention. The smaller the RMSE value, the closer the prediction result is to the true value, and the better the performance of the model of the method described in this invention.

[0166] In addition, it is understood that the method described in this invention requires model training before separating the self-noise signal to be separated.

[0167] Specifically, mixed noise signals containing electrical noise, flow noise, and vibration noise can be collected during UUV operation. To ensure data consistency, these signals are standardized and then data augmentation is performed to generate training samples with different noise mixing ratios. For each segment of the collected signal... x ̃ [ n ] Constructing analytic signals using the Hilbert transform h [n]=H x ̃ [n] Then introduce linear phase shift The enhanced signal is obtained:

[0168] x k , aug [ n ]= R { h k [ n ] e j Φ [ n ] } ⋅ s + η k [ n ], k ∈ { e , f , v } (20)

[0169] in This is the scaling factor. η k [ n ] ~N (0, σ 2 ) For random noise, "These correspond to electrical noise, flow noise, and vibration noise, respectively." This is the angular frequency offset; This represents the real part. To simulate the influence of different noise sources in the mixed signal, weighting coefficients are introduced for each noise component. These correspond to electrical noise, flow noise, and vibration noise, respectively, and satisfy the following conditions: In enhancing the signal x aug [n] Based on this, the final training samples are obtained after assigning a mixing ratio. x final [n] :

[0170] x final [n]= r e x e [n]+ r f x f [n]+ r v x v [n] ,(twenty one)

[0171] in, x e [n] , x f [n] and x v [n] These represent the original electrical noise, flow noise, and vibration noise, respectively. In addition, 66 samples with different ratios can be retained for testing to ensure the independence of training and testing. The augmented data is then input into the model for training. The training set and test set can be divided in a 7:3 ratio. Of course, to ensure data consistency and improve separation, the data augmentation methods described in formulas (20) and (21) can also be applied to the model application process; this invention does not limit this application.

[0172] To verify the effectiveness of the method described in this invention, comparative experiments were conducted. Table 1 shows the RSME results of UUV self-noise separation based on different method models. The first row represents different noise separation models, and the second row represents the RSME results corresponding to each model. It can be seen that the RSME values ​​of this invention are lower than those of the CNN model, Transformer model, LSTM model, Conv-TasNet model, and Att-TasNet model, indicating better separation performance.

[0173] Table 1

[0174]

[0175] Table 2 shows the ablation experiment results of the method described in this invention. As can be seen from the RSME corresponding to the removal of different components, each part of the method described in this invention is important and can only achieve such a good UUV self-noise separation effect as a whole.

[0176] Table 2

[0177]

[0178] Table 3 shows the RSME results under different phase offset conditions corresponding to the three types of noise. By comparing the phase offset parameters of each group, the effectiveness of the method for calculating time-domain weights and frequency-domain weights based on phase offset described in steps 41-44 of this invention can be demonstrated.

[0179] Table 3

[0180]

[0181] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," and "counterclockwise," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0182] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0183] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention without departing from the principles and spirit of the present invention.

Claims

1. A method for separating self-noise from unmanned underwater vehicles based on time-frequency domain features and domain adaptation, characterized in that, The method includes: Step 1: Obtain the noise signal to be separated; Step 2: Perform short-time Fourier transform and convolutional neural network-based encoding processing on the self-noise signal to be separated to obtain a time-frequency feature map; Step 3: Perform feature extraction processing on the time-frequency feature map based on the attention mechanism to obtain the time-domain and frequency-domain features of the noise signal to be separated; Step 4: Based on the time-frequency characteristics of the three preset prior noise signals and the weights of the three noise signals in the noise signal to be separated, determine the time-domain weights corresponding to the time-domain features and the frequency-domain weights corresponding to the frequency-domain features. Step 5: Perform weighted calculation on the time-domain features and the frequency-domain features based on the time-domain weights and the frequency-domain weights to obtain the time-frequency domain fusion features of the noise signal to be separated; Step 6: Normalize the time-frequency domain fusion features to obtain normalized fusion features; Step 7: Perform weighted calculation on the time-frequency domain fusion feature and the normalized fusion feature based on the preset normalized weight coefficient to obtain the output feature; Step 8: Decode the output features based on a convolutional neural network and perform an inverse short-time Fourier transform to obtain the separation result of the self-noise signal to be separated.

2. The method according to claim 1, characterized in that, The three types of prior noise signals include prior electrical noise signal, prior flow noise signal and prior vibration noise signal; Step 4 includes: Step 41: Perform softmax-based classification processing on the noise signal to be separated to obtain the weights of the three noise signals in the noise signal to be separated. Step 42: Calculate the phase offset corresponding to the three preset prior noise signals according to the following formula: ， in, This represents the phase offset corresponding to the i-th type of noise signal. This represents the total energy of the i-th prior noise signal in the time domain. Let represent the total energy of the i-th prior noise signal in the frequency domain; Represents the hyperbolic tangent function; Step 43: Based on the weights of the three noise signals in the noise signal to be separated and the phase offsets corresponding to the three noise signals respectively, calculate the overall phase offset according to the following formula: ， in, Indicates the overall phase shift; , , These represent the weights of electrical noise, flow noise, and vibration noise in the noise signal to be separated, respectively. , , These represent the phase offsets corresponding to electrical noise signals, flow noise signals, and vibration noise signals, respectively. Step 44: Calculate the time-domain weights corresponding to the time-domain features and the frequency-domain weights corresponding to the frequency-domain features according to the following formulas: in, This represents the temporal weights corresponding to the temporal features. This represents the frequency domain weight corresponding to the frequency domain feature.

3. The method according to claim 2, characterized in that, Step 5 includes: Step 51: Calculate the time-domain features and the frequency-domain features based on the time-domain weights and the frequency-domain weights according to the following formula to obtain the initial fused features: ， in, X represents the initial fusion feature; X represents the time-frequency feature map; This represents the time-domain feature. This represents the frequency domain characteristics; Step 52: Based on the time-frequency feature map and the initial fusion features, calculate the time-frequency domain fusion features according to the following formula: ， Where Y represents the time-frequency domain fusion feature, " represents the Hadamard product operator.

4. The method according to claim 1, characterized in that, Step 6 includes: Step 61: Normalize the time-frequency domain fusion features according to the following formula to obtain normalized fusion features: ， in, This represents the value at position (h, w) of the nth sample, cth channel, in the normalized fusion feature; This represents the value at position (h, w) of the nth sample, cth channel, in the time-frequency domain fusion feature. This represents the mean of the nth sample and the cth channel in the time-frequency domain fusion feature. This represents the variance of the nth sample and the cth channel in the time-frequency domain fusion feature; It is a constant.

5. The method according to claim 4, characterized in that, Step 7 includes: Step 71: According to the following formula, the time-frequency domain fusion feature and the normalized fusion feature are weighted and calculated based on preset normalized weight coefficients to obtain the output feature: ， in, This represents the value at position (h, w) of the nth sample, cth channel, in the output feature. The preset normalized weighting coefficients are determined according to the following formula: ， in, This represents a preset scalar parameter.

6. The method according to claim 1, characterized in that, Step 8 includes: Step 81: Perform decoding processing on the output features based on a convolutional neural network to obtain the separation masks corresponding to the three noise signals respectively; Step 82: Multiply the separation mask element-wise with the amplitude spectrum of the noise signal to be separated to obtain the separation amplitude spectra corresponding to the three noise signals respectively; wherein, the amplitude spectrum is obtained by performing a short-time Fourier transform on the noise signal to be separated; Step 83: Perform an inverse short-time Fourier transform based on the separation amplitude spectrum and the phase information of the noise signal to be separated to obtain the separation result of the noise signal to be separated; wherein, the phase information is obtained by performing a short-time Fourier transform on the noise signal to be separated.

7. The method according to claim 1, characterized in that, Step 2 includes: The noise signal to be separated is subjected to a short-time Fourier transform based on a Hanning window of length 256 and an overlap rate of 50% to obtain a two-dimensional noise signal. The two-dimensional noise signal is encoded using a convolutional neural network to obtain a time-frequency feature map.

8. The method according to claim 1, characterized in that, Step 1 includes: Acquire the original mixed noise signal; The original mixed noise signal is standardized and preprocessed to obtain the noise signal to be separated.

Images