A self-supervised signal denoising method and system based on adaptive signal features and learnable loss weights
By employing a self-supervised method that adapts to signal features and uses learnable loss weights, the problem of unstable performance of traditional signal denoising methods in complex environments is solved. This method achieves adaptive denoising without external labels, improving generalization ability and robustness in mixed noise environments.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SHANGHAI UNIV
- Filing Date
- 2026-04-10
- Publication Date
- 2026-06-23
Smart Images
Figure CN122027043B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of modern wireless communication technology, and specifically to a self-supervised signal denoising method and system based on adaptive signal features and learnable loss weights. Background Technology
[0002] In modern wireless communication systems, signals are inevitably subject to various noise interferences during transmission, including thermal noise, interference noise, and multipath fading. These noises severely degrade signal quality, affecting the bit error rate, spectral efficiency, and coverage of the communication system. Therefore, signal denoising technology has always been an important research direction in the field of communications.
[0003] Traditional signal denoising methods mainly include:
[0004] Filter-based methods, such as Wiener filtering and Kalman filtering, require accurate noise statistics models, are sensitive to model mismatch, and have limited performance in complex channel environments.
[0005] Wavelet transform-based methods separate signal and noise through wavelet decomposition. However, the selection of wavelet basis functions depends on experience and has limited effectiveness for non-stationary signals.
[0006] Supervised methods based on deep learning involve training with a large number of clean-noisy signal pairs. However, in practical applications, obtaining perfectly matched clean signals is extremely difficult and costly.
[0007] In addition, the main drawbacks of existing technologies include:
[0008] Filter-based and wavelet transform-based methods are highly dependent on the accuracy of prior models. When faced with complex, unknown, and non-stationary noise in the real world, the performance of these methods often drops sharply, and their parameters require careful manual design and tuning, lacking adaptive capabilities.
[0009] Deep learning-based supervised methods: Supervised denoising requires large-scale, high-quality paired training data. In many practical applications, obtaining absolutely pure "Ground Truth" signals is extremely difficult, costly, or even impossible. This "data pair dependency" problem has become the biggest obstacle restricting the deployment and application of supervised denoising methods in the real world.
[0010] To address the aforementioned issues, there is an urgent need for a self-supervised signal denoising method and system based on adaptive signal features and learnable loss weights to solve the problems existing in traditional methods. Summary of the Invention
[0011] The purpose of this invention is to provide a self-supervised signal denoising method and system based on signal feature adaptation and learnable loss weights. This self-supervised denoising method does not require clean signal labels, can achieve sample-level adaptation, integrates time-frequency domain information, and enhances generalization ability through learnable loss weights.
[0012] To achieve the above objectives, the technical solution adopted by the present invention is as follows:
[0013] A self-supervised signal denoising method based on adaptive signal features and learnable loss weights includes:
[0014] Step 1: Extract multi-dimensional statistical features from the noisy signal, and weight the multi-dimensional statistical features using learnable weight parameters to generate a weighted feature vector. The learnable weight parameters are the CLIP alignment loss weight and the noise target ratio in the historical noisy signal.
[0015] Step 2: Input the weighted feature vector into the pre-trained prediction network to predict the CLIP alignment loss weights and noise target ratio of the noisy signal;
[0016] Step 3: Decompose the noisy signal into odd-numbered sampling sequences and even-numbered sampling sequences, and perform time-domain and frequency-domain feature encoding respectively to obtain time-domain features and frequency-domain features;
[0017] Step 4: Construct the CLIP alignment loss based on time-domain features, frequency-domain features, and CLIP alignment loss weights;
[0018] Step 5: Construct a residual signal prediction model based on the U-Net encoder-decoder structure, construct a total loss function based on CLIP alignment loss and noise target ratio, and train the residual signal prediction model based on the total loss function;
[0019] Step 6: Input the noisy signal to be tested into the trained residual signal prediction model to obtain the residual signal;
[0020] Step 7: Subtract the residual signal from the noisy signal to be measured to obtain the denoised signal output.
[0021] Furthermore, in step 1, the multidimensional statistical features include at least one of Shannon entropy, power spectral entropy, amplitude variance, phase jump, peak-to-average power ratio, and frequency domain energy distribution standard deviation.
[0022] Further, in step 2, the weighted feature vector is input into the pre-trained prediction network to predict the CLIP alignment loss weights and noise target ratio of the noisy signal, specifically as follows:
[0023] The weighted feature vectors are input into a deep neural network for encoding. The CLIP alignment loss weights and the proportion of noise target are predicted by a multilayer perceptron structure. The prediction process employs logarithmic space parameterization and temperature regulation mechanisms to ensure the numerical stability of the prediction weights.
[0024] Further, in step 3, the noisy signal is decomposed into odd-numbered sampling sequences and even-numbered sampling sequences, and time-domain and frequency-domain feature encodings are performed respectively to obtain time-domain features and frequency-domain features, specifically as follows:
[0025] The noisy signal is decomposed into odd-numbered sampling sequences and even-numbered sampling sequences. Time-domain feature encoding and frequency-domain spectral transformation are performed on each odd-numbered sampling sequence and even-numbered sampling sequence respectively to obtain time-domain features and frequency-domain features. The frequency-domain spectral transformation adopts short-time Fourier transform and introduces a learnable frequency weight vector to adaptively weight different frequency components.
[0026] Furthermore, in step 4, the CLIP alignment loss is constructed based on the time-domain features, frequency-domain features, and CLIP alignment loss weights, specifically as follows:
[0027] The time-domain and frequency-domain features are mapped to the same embedding space. The similarity between the time and frequency domains is represented by a learnable temperature parameter. CLIP alignment loss is constructed for odd-numbered and even-numbered sampling sequences respectively.
[0028] Furthermore, in step 5, the total loss function includes residual reconstruction loss, consistency loss, CLIP alignment loss, noise estimation loss, and weight regularization term.
[0029] Furthermore, the noise estimation loss is constructed using the noise target ratio.
[0030] This invention also provides a self-supervised signal denoising system based on adaptive signal features and learnable loss weights, applied to the aforementioned self-supervised signal denoising method based on adaptive signal features and learnable loss weights, comprising:
[0031] The signal feature extraction and weighting module is used to extract multi-dimensional statistical features from noisy signals and perform learnable weighting to generate weighted feature vectors.
[0032] An adaptive weight prediction module is used to receive the weighted feature vector and predict the CLIP alignment loss weight and noise target ratio of the noisy signal.
[0033] The time-frequency domain feature coding module is used to decompose the noisy signal into odd-numbered sampling sequences and even-numbered sampling sequences, and perform time-domain and frequency-domain feature coding respectively, outputting time-domain features and frequency-domain features;
[0034] The multimodal alignment and loss calculation module is used to construct the CLIP alignment loss based on the time-domain features, frequency-domain features, and CLIP alignment loss weights.
[0035] The U-Net residual prediction module is used to build a residual signal prediction model based on the U-Net encoder-decoder structure.
[0036] A multi-objective loss function construction module is used to form the total loss function;
[0037] The model training and optimization module is used to train the residual signal prediction model using the total loss function;
[0038] The output module is used to calculate and output the denoised signal based on the residual signal predicted by the trained residual prediction model and the noisy signal to be tested.
[0039] In summary, the present invention has at least one of the following beneficial technical effects:
[0040] 1. No external labels required, true self-supervised denoising: By extracting statistical features from the signal itself to replace the SNR (Signal-to-noise ratio) label, fully self-supervised adaptive denoising is achieved, reducing system complexity and improving practicality.
[0041] 2. Sample-level adaptive capability: Each signal sample obtains independent denoising strategy parameters based on its own characteristics, which can finely adapt to changes in signal quality and significantly improve the generalization capability in mixed noise environments and under unknown signal-to-noise ratio conditions.
[0042] 3. Conservative learnable loss weight strategy: Keep the core loss weights fixed to maintain stability, and only make the CLIP scaling factor learnable. This increases adaptive flexibility while avoiding training instability and overfitting risks, making it more robust than the fully learnable approach.
[0043] 4. Fully integrated multimodal information: The consistency of time-frequency domain representation is enhanced through CLIP alignment mechanism, which makes full use of the complementary information of the two domains, enhances the understanding of the essential characteristics of the signal, and improves the denoising robustness.
[0044] 5. End-to-end joint optimization: All module parameters are trained together to automatically discover the optimal configuration, avoiding the tedious manual parameter tuning process and improving development efficiency.
[0045] 6. Strong generalization ability: The adaptive mechanism based on signal features enables the model to handle noise types and signal-to-noise ratio conditions that were not seen during training, and has good cross-scene generalization ability. Attached Figure Description
[0046] Figure 1This is a flowchart of the present invention;
[0047] Figure 2 This is a schematic diagram of the signal feature extraction process;
[0048] Figure 3 This is a schematic diagram of the adaptive weight prediction process;
[0049] Figure 4 This is a schematic diagram of the time-frequency domain feature encoding process;
[0050] Figure 5 This is a schematic diagram of the multimodal alignment process;
[0051] Figure 6 This is a schematic diagram of the residual signal prediction model structure;
[0052] Figure 7 This is a schematic diagram of the loss function calculation process;
[0053] Figure 8 This is a schematic diagram of the configuration of some learnable loss weights. Detailed Implementation
[0054] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.
[0055] like Figure 1 As shown, this invention provides a self-supervised signal denoising method based on adaptive signal features and learnable loss weights, comprising:
[0056] Step 1: Extract multi-dimensional statistical features from the noisy signal, and weight the multi-dimensional statistical features using learnable weight parameters to generate a weighted feature vector. The learnable weight parameters are the CLIP alignment loss weight and the noise target ratio in the historical noisy signal.
[0057] Step 2: Input the weighted feature vector into the pre-trained prediction network to predict the CLIP alignment loss weights and noise target ratio of the noisy signal;
[0058] Step 3: Decompose the noisy signal into odd-numbered sampling sequences and even-numbered sampling sequences, and perform time-domain and frequency-domain feature encoding respectively to obtain time-domain features and frequency-domain features;
[0059] Step 4: Construct the CLIP alignment loss based on time-domain features, frequency-domain features, and CLIP alignment loss weights;
[0060] Step 5: Construct a residual signal prediction model based on the U-Net encoder-decoder structure, construct a total loss function based on CLIP alignment loss and noise target ratio, and train the residual signal prediction model based on the total loss function;
[0061] Step 6: Input the noisy signal to be tested into the trained residual signal prediction model to obtain the residual signal;
[0062] Step 7: Subtract the residual signal from the noisy signal to be measured to obtain the denoised signal output.
[0063] Next, the above method will be explained in detail with reference to specific embodiments:
[0064] like Figure 2 As shown, in step 1, multi-dimensional statistical features are extracted from the noisy in-phase-orthogonal signal. These multi-dimensional statistical features are then weighted using learnable weight parameters to generate a weighted feature vector. Specifically:
[0065] The multi-dimensional statistical features include at least one of Shannon entropy, power spectral entropy, amplitude variance, phase jump, peak-to-average power ratio, and standard deviation of frequency domain energy distribution. The calculation methods for each are described below:
[0066] 1. Shannon entropy
[0067] Shannon entropy is used to measure the uncertainty of signal amplitude distribution. The greater the noise, the more dispersed the amplitude distribution, and the higher the entropy value.
[0068] 2. Power spectral entropy
[0069] Power spectral entropy is used to measure the uniformity of energy distribution in the frequency domain. Clean signals have a concentrated spectrum, while noisy signals have a flat spectrum.
[0070] 3. Amplitude Variance
[0071] This reflects the degree of fluctuation in signal amplitude. Noise causes amplitude instability and increases variance. The statistical variance of the complex amplitude sequence is calculated directly.
[0072] 4. Phase jump
[0073] Detecting phase discontinuities, noise causes irregular phase jumps.
[0074] 5. Peak to Average Power Ratio (PAPR)
[0075] The dynamic range of signal power is measured, and noise affects the peak power distribution of the signal.
[0076] 6. Standard deviation of frequency domain energy distribution
[0077] The dispersion of spectral energy is quantified, as noise makes the energy distribution more irregular. The standard deviation of the energy at each frequency point is calculated after FFT (Fast Fourier Transform).
[0078] The six features described above reflect the noise characteristics of the signal in the time, frequency, amplitude, and phase domains, respectively, providing a comprehensive noise characterization. Each feature dimension is configured with a learnable weight parameter, and the importance of each feature is automatically learned through training.
[0079] like Figure 3 As shown, in step 2, the weighted feature vector is input into the pre-trained prediction network to predict the CLIP alignment loss weights and noise target ratio of the noisy signal, specifically as follows:
[0080] The weighted feature vectors are input into a deep neural network for encoding. The CLIP alignment loss weights and noise target ratio of the noisy signal are predicted by a multilayer perceptron structure. The prediction process uses logarithmic space parameterization and temperature adjustment mechanism to ensure the numerical stability of the prediction weights.
[0081] Step 2 can be summarized into three stages, each corresponding to one of three modules:
[0082] 1. Feature Encoding Module: Employs a two-layer fully connected network structure, with each layer followed by a normalization layer, ReLU (Rectified Linear Unit) activation layer, and a Dropout layer, mapping 6-dimensional features to a 64-dimensional latent space. Layer normalization stabilizes the training process, and Dropout prevents overfitting.
[0083] 2. CLIP Weight Predictor: It adopts a two-layer fully connected network and finally outputs sample-level CLIP weights through logarithmic space transformation and temperature adjustment mechanism. Logarithmic space parameterization ensures that the weights are always positive, and the temperature parameter controls the smoothness of the prediction.
[0084] 3. Noise target predictor: The structure is similar to the CLIP weight predictor, but it uses the Sigmoid activation function to output the noise ratio.
[0085] Step 2 performs weight prediction entirely based on signal features, without the need for external SNR labels, thus achieving true self-supervised adaptation.
[0086] In step 3, the noisy signal is decomposed into odd-numbered sampling sequences and even-numbered sampling sequences, and time-domain and frequency-domain feature encoding is performed respectively to obtain time-domain features and frequency-domain features, specifically as follows:
[0087] like Figure 4As shown, this step decomposes the noisy signal into odd-numbered sampling sequences and even-numbered sampling sequences. Time-domain feature encoding and frequency-domain spectrum transformation are performed on each odd-numbered sampling sequence and even-numbered sampling sequence to obtain time-domain features and frequency-domain features. The frequency-domain spectrum transformation adopts short-time Fourier transform and introduces a learnable frequency weight vector to adaptively weight different frequency components.
[0088] In terms of frequency domain coding, a learnable spectrum transformation module is introduced: First, short-time Fourier transforms are performed on the I (In-phase, I) and Q (Quadrature, Q) signals respectively to obtain the amplitude spectrum, and learnable weight vectors are used to weight different frequency components to automatically learn the frequency importance; then the weighted I / Q spectra are stitched together to form a dual-channel time-frequency map, which is then processed by a two-dimensional convolutional layer and 8×8 adaptive average pooling, flattened, and output as a 2048-dimensional frequency domain feature vector through a fully connected layer.
[0089] In step 4, CLIP alignment loss is constructed based on time-domain features, frequency-domain features, and CLIP alignment loss weights.
[0090] like Figure 5 As shown, time-domain features and frequency-domain features are mapped to the same embedding space. The similarity between the time and frequency domains is represented based on the learnable temperature parameter. CLIP alignment loss is constructed for odd-numbered sampling sequences and even-numbered sampling sequences respectively.
[0091] In detail, this step achieves deep alignment of time-frequency features through embedded spatial mapping and contrastive learning mechanisms. To adjust the hardness or softness of the similarity distribution, a learnable temperature parameter is introduced to scale the matrix. Subsequently, a symmetric cross-entropy loss function is used to maximize the similarity of the same real sample under different domain representations, resulting in the final CLIP total loss.
[0092] In step 5, a residual signal prediction model is constructed based on the U-Net encoder-decoder structure. A total loss function is constructed based on the CLIP alignment loss and the noise-to-target ratio. The residual signal prediction model is then trained using this total loss function. Specifically:
[0093] First, this step uses the classic U-Net architecture, such as... Figure 6As shown, the system mainly consists of three parts: an encoder, a bottleneck layer, and a decoder. The encoder path includes four downsampling stages. Each stage extracts features through two consecutive convolutional layers, followed by a max-pooling layer with a stride of 2 to reduce the spatial dimensionality, gradually expanding the number of feature channels from the initial 2 channels to 512 channels. The bottleneck layer further increases the number of channels to 1024 to extract deep abstract features. The decoder path performs progressive upsampling through transposed convolutions and introduces a skip connection mechanism to concatenate the upsampled features with the output of the corresponding encoder layer, achieving effective fusion of high-level semantic information and low-level detail information. After processing through four decoding blocks, the number of channels gradually recovers from 1024 to 64. Finally, the output layer maps the features to 2 channels through a 1×1 convolution, generating a residual signal with the same dimension as the input, thus preserving details while ensuring the precision of denoising.
[0094] Secondly, such as Figure 7 As shown, the present invention constructs a total loss function, which includes residual reconstruction loss, consistency loss, CLIP alignment loss, noise estimation loss, and weight regularization term.
[0095] In step 6, the noisy signal to be tested is input into the trained residual signal prediction model to obtain the residual signal.
[0096] In step 7, the residual signal is subtracted from the noisy signal to be measured to obtain the denoised signal output.
[0097] The entire inference process requires no external labels or human intervention; the model automatically adjusts the denoising strategy based on the signal characteristics.
[0098] Furthermore, this invention also provides some learnable loss weight configurations, as illustrated in the diagram below. Figure 8 As shown, the parameters are first propagated forward, then grouped into network parameters, adaptive weight parameters, and loss weight parameters, and their learning rate is controlled. Gradient pruning and learning rate scheduling are then performed, and finally backpropagation and parameter updates are performed.
[0099] This invention also provides a self-supervised signal denoising system based on adaptive signal features and learnable loss weights, applied to the aforementioned self-supervised signal denoising method based on adaptive signal features and learnable loss weights, comprising:
[0100] The signal feature extraction and weighting module is used to extract multi-dimensional statistical features from noisy signals and perform learnable weighting to generate weighted feature vectors.
[0101] An adaptive weight prediction module is used to receive the weighted feature vector and predict the CLIP alignment loss weight and noise target ratio of the noisy signal.
[0102] The time-frequency domain feature coding module is used to decompose the noisy signal into odd-numbered sampling sequences and even-numbered sampling sequences, and perform time-domain and frequency-domain feature coding respectively, outputting time-domain features and frequency-domain features;
[0103] The multimodal alignment and loss calculation module is used to construct the CLIP alignment loss based on the time-domain features, frequency-domain features, and CLIP alignment loss weights.
[0104] The U-Net residual prediction module is used to build a residual signal prediction model based on the U-Net encoder-decoder structure.
[0105] A multi-objective loss function construction module is used to form the total loss function;
[0106] The model training and optimization module is used to train the residual signal prediction model using the total loss function;
[0107] The output module is used to calculate and output the denoised signal based on the residual signal predicted by the trained residual prediction model and the noisy signal to be tested.
[0108] Embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0109] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0110] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1The function specified in one or more boxes.
[0111] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0112] Contents not described in detail in this specification are prior art known to those skilled in the art. It is hereby indicated that the above description is intended to help those skilled in the art understand this invention, but does not limit the scope of protection of this invention. Any equivalent substitutions, modifications, improvements, or simplifications of the above descriptions that do not depart from the essential content of this invention fall within the scope of protection of this invention.
Claims
1. A self-supervised signal denoising method based on adaptive signal features and learnable loss weights, characterized in that, include: Step 1: Extract multi-dimensional statistical features from the noisy signal, and weight the multi-dimensional statistical features using learnable weight parameters to generate a weighted feature vector. The learnable weight parameters are CLIP alignment loss weight and noise target ratio in the historical noisy signal. The multi-dimensional statistical features include at least one of Shannon entropy, power spectral entropy, amplitude variance, phase jump, peak-to-average power ratio, and frequency domain energy distribution standard deviation. Step 2: Input the weighted feature vector into the pre-trained prediction network to predict the CLIP alignment loss weights and noise target ratio of the noisy signal; specifically: The weighted feature vectors are input into a deep neural network for encoding. The CLIP alignment loss weights and noise target ratio of the noisy signal are predicted by a multilayer perceptron structure. The prediction process uses logarithmic space parameterization and temperature adjustment mechanism to ensure the numerical stability of the prediction weights. Step 3: Decompose the noisy signal into odd-numbered sampling sequences and even-numbered sampling sequences, and perform time-domain and frequency-domain feature encoding respectively to obtain time-domain features and frequency-domain features; specifically: The noisy signal is decomposed into odd-numbered sampling sequences and even-numbered sampling sequences. Time-domain feature encoding and frequency-domain spectrum transformation are performed on each odd-numbered sampling sequence and even-numbered sampling sequence respectively to obtain time-domain features and frequency-domain features. The frequency-domain spectrum transformation adopts short-time Fourier transform and introduces a learnable frequency weight vector to adaptively weight different frequency components. Step 4: Construct the CLIP alignment loss based on time-domain features, frequency-domain features, and CLIP alignment loss weights; specifically: The time-domain features and frequency-domain features are mapped to the same embedding space. The similarity between the time and frequency domains is represented by a learnable temperature parameter. CLIP alignment loss is constructed for odd-numbered sampling sequences and even-numbered sampling sequences respectively. Step 5: Construct a residual signal prediction model based on the U-Net encoder-decoder structure, construct a total loss function based on CLIP alignment loss and noise target ratio, and train the residual signal prediction model based on the total loss function; Step 6: Input the noisy signal to be tested into the trained residual signal prediction model to obtain the residual signal; Step 7: Subtract the residual signal from the noisy signal to be measured to obtain the denoised signal output.
2. The self-supervised signal denoising method based on adaptive signal features and learnable loss weights according to claim 1, characterized in that, In step 5, the total loss function includes residual reconstruction loss, consistency loss, CLIP alignment loss, noise estimation loss, and weight regularization term.
3. The self-supervised signal denoising method based on adaptive signal features and learnable loss weights according to claim 2, characterized in that, The noise estimation loss is constructed using the noise target ratio.
4. A self-supervised signal denoising system based on adaptive signal features and learnable loss weights, applied to the self-supervised signal denoising method based on adaptive signal features and learnable loss weights as described in any one of claims 1-3, characterized in that, include: The signal feature extraction and weighting module is used to extract multi-dimensional statistical features from noisy signals and perform learnable weighting to generate weighted feature vectors. An adaptive weight prediction module is used to receive the weighted feature vector and predict the CLIP alignment loss weight and noise target ratio of the noisy signal. The time-frequency domain feature coding module is used to decompose the noisy signal into odd-numbered sampling sequences and even-numbered sampling sequences, and perform time-domain and frequency-domain feature coding respectively, outputting time-domain features and frequency-domain features; The multimodal alignment and loss calculation module is used to construct the CLIP alignment loss based on the time-domain features, frequency-domain features, and CLIP alignment loss weights. The U-Net residual prediction module is used to build a residual signal prediction model based on the U-Net encoder-decoder structure. A multi-objective loss function construction module is used to form the total loss function; The model training and optimization module is used to train the residual signal prediction model using the total loss function; The output module is used to calculate and output the denoised signal based on the residual signal predicted by the trained residual prediction model and the noisy signal to be tested.