Signal noise suppression based on deep learning to ensure signal integrity

Abstract

This invention discloses a deep learning-based method for signal noise suppression while ensuring signal integrity, belonging to the field of noise suppression processing technology. The suppression processing steps are as follows: acquiring noise; analyzing the noise based on spectral analysis; comparing and analyzing noise types based on big data analysis to determine the current noise type; constructing a neural network model; training and optimizing the model; setting an adaptive learning mechanism to analyze different noise types; and inputting the extracted noise signal into the constructed model accordingly to suppress the noise source while ensuring signal integrity. This invention analyzes and determines the source of noise, uses different processing methods for different types of noise, optimizes the neural network model, and sets an adaptive learning mechanism to suppress the classified noise signal, improving the smoothness of noise processing and better ensuring signal integrity and original characteristics after noise suppression.

Description

Technical Field

[0001] This invention relates to the field of noise suppression technology, specifically to a deep learning-based method for signal noise suppression to ensure signal integrity. Background Technology

[0002] Noise is a widely existing concept in many fields such as signal processing, communication, acoustics, and electronic engineering. It generally refers to all unwanted, interfering signals or fluctuations outside the useful signal. Noise sources are diverse, including natural environmental sources such as wind, rain, and electromagnetic radiation, as well as man-made sources such as machine noise and electromagnetic interference from electronic equipment. In signal processing, noise is usually considered a random signal, with characteristics including random variations in amplitude, frequency, and phase. The presence of noise can interfere with or mask useful signals, making signal reception, identification, and analysis difficult. Therefore, noise suppression is an important task in signal processing. Existing technologies suppress sound waves by removing noise components from the signal through filtering. However, filtering often results in unsatisfactory noise smoothing and fails to guarantee signal integrity after noise suppression, affecting the original signal characteristics and signal quality. To address this, we propose a deep learning-based signal noise suppression method that ensures signal integrity. Summary of the Invention

[0003] To address the aforementioned technical problems, this technical solution provides a deep learning-based method for signal noise suppression to ensure signal integrity, thus resolving the problems described above.

[0004] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0005] A deep learning-based method for signal noise suppression to ensure signal integrity involves the following suppression steps:

[0006] Noise is acquired, analyzed using spectral analysis, and compared with other noise types based on big data analysis to determine the type of noise.

[0007] Build a neural network model, train and optimize the model, set up an adaptive learning mechanism, and analyze different types of noise;

[0008] The extracted noise signals are input into the constructed model accordingly, suppressing noise sources while ensuring signal integrity.

[0009] Preferably, noise acquisition is based on the collection of audio acquisition equipment, and the acquired noise is uploaded to a computer for processing; the spectrum analysis method pre-samples the continuous-time signal, performs Fourier transform processing on the sampled discrete-time signal, and then calculates the spectrum to calculate the amplitude and phase of the signal in different frequency domains, and plots the calculated spectrum in graphical form.

[0010] Preferably, the formula for calculating signal sampling is:

[0011] x(n)=x(nT s )

[0012] Where T s Let x(n) be the sampling period, x(n) be the discrete-time signal, and n be an integer representing the index of the sampling point.

[0013] The Fourier transform formula is:

[0014]

[0015] Where N is the total number of sample points, k is the frequency index with a value range of 0≤k≤N-1, j is the imaginary unit, X(k) is the frequency domain, and e is the base of the natural logarithm. The amplitude and phase of the signal in different frequency domains are calculated from the results obtained from the Fourier transform.

[0016] Preferably, multiple sets of noise spectrum diagrams are obtained based on big data, and the multiple sets of noise spectrum diagrams are compared with the current noise spectrum diagram to determine whether they are consistent. The calculation formula is as follows:

[0017]

[0018] Where o is the sample size, X i With Y i These are the sample point observations of the two variables at point i on the spectrum, respectively. and Both are the sample mean values ​​of X and Y, and r is the similarity between the two sets of spectrograms. A value of r of 1 indicates complete correlation, meaning the two sets of spectrograms are identical, while other values ​​indicate inconsistency. The closer the value of r is to 1, the more identical the two sets of spectrograms are, and vice versa. The method is to calculate whether the currently collected noise signal is consistent with the noise spectrogram obtained from the big data, so as to obtain the type of the currently collected noise signal.

[0019] Preferably, the noise signal types include natural noise and man-made noise. Natural noise includes atmospheric noise and wind noise. Natural noise originates from noise generated in the natural environment. Man-made noise is generated based on human activities. Man-made noise includes industrial noise, mechanical noise, and electronic product noise. The noise signal categories of both the natural environment and man-made noise are obtained through big data. By comparing the noise signal obtained from the big data, the type of current noise can be determined.

[0020] Preferably, the model training method updates the weights and bias terms based on gradient descent, and optimizes the model based on regularization to prevent overfitting. The Adam algorithm is selected as the adaptive learning mechanism, the algorithm parameters are configured in the model, and iterative optimization is continuously performed to improve the performance of the neural network model.

[0021] Preferably, when inputting the noise signal of the current classification, the neural network model is equipped with a dual-input frame, which inputs different corresponding noise signals into the corresponding input ports of the neural network model. The dual-input frame of the neural network model includes a natural noise input port and an artificial noise input port. After judging the noise of the current input, the corresponding input is sent to the different input ports.

[0022] Preferably, when training the model, the parameters and structure of the model are dynamically adjusted according to the model's performance on the validation set. When training the model, two sets of training sets are prepared: one set is an artificial noise dataset and the other set is a natural noise dataset. Each set of datasets is preprocessed. The model contains two parallel sub-networks: one for processing artificial noise data and the other for processing natural noise data.

[0023] Preferably, the iterative training formula is:

[0024]

[0025] Where θ1, θ2 and θ n Let L(θ) be the model parameters and L(θ) be the loss function. The parameters are updated using the following formula:

[0026]

[0027] Where η is the learning rate, the step parameters are updated according to the formula in each iteration of the model to optimize the performance of the neural network model.

[0028] Preferably, the extracted noise signal is input into the corresponding input port of the trained model. The model suppresses the noise based on the learned knowledge. While suppressing the noise, the model ensures that the integrity of the signal is not affected and retains useful information in the signal, such as frequency components and phase information.

[0029] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0030] The method proposed in this invention analyzes and judges the source of noise, adopts different processing methods for different types of noise, optimizes the neural network model, and sets up an adaptive learning mechanism to suppress the classified noise signal, improve the smoothness of noise processing, and better protect the integrity and original characteristics of the signal after noise suppression. Attached Figure Description

[0031] Figure 1 This is a flowchart of the noise suppression processing steps of the present invention;

[0032] Figure 2 This is a flowchart of the noise type identification and analysis process of the present invention. Detailed Implementation

[0033] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0034] Reference Figure 1 As shown, the signal noise suppression method based on deep learning to ensure signal integrity involves the following suppression steps:

[0035] Noise is acquired, analyzed using spectral analysis, and compared with other noise types based on big data analysis to determine the type of noise.

[0036] Build a neural network model, train and optimize the model, set up an adaptive learning mechanism, and analyze different types of noise;

[0037] The extracted noise signal is input into the constructed model to suppress the noise source while ensuring the integrity of the signal.

[0038] Spectral analysis of noise signals can clearly reveal their frequency distribution characteristics, revealing their spectral features and helping to distinguish different types of noise. Based on big data analytics, massive amounts of noise data can be deeply mined to identify noise sources, distribution, and trends. By comparing and analyzing noise data from different sources, the type of noise can be more accurately determined, providing a strong basis for subsequent noise control. Constructing and training neural network models enables them to learn and recognize different types of noise signals. These models possess powerful pattern recognition capabilities and can adaptively adjust parameters to cope with complex noise environments. An adaptive learning mechanism allows the neural network model to continuously optimize its performance in practical applications. When encountering new noise types or changes in the noise environment, the model can automatically adjust its parameters and strategies to better suppress noise sources. The application of spectral analysis, big data analysis, and neural network models can promote continuous innovation and development in noise control technology.

[0039] Reference Figure 2 As shown, noise acquisition is based on the collection of audio acquisition equipment, and the acquired noise is uploaded to the computer for processing; the spectrum analysis method pre-samples the continuous-time signal, performs Fourier transform processing on the sampled discrete-time signal, and then calculates the spectrum to calculate the amplitude and phase of the signal in different frequency domains, and plots the calculated spectrum in graphical form.

[0040] The formula for calculating signal sampling is as follows:

[0041] x(n)=x(nT s )

[0042] Where T s Let x(n) be the sampling period, x(n) be the discrete-time signal, and n be an integer representing the index of the sampling point.

[0043] The Fourier transform formula is:

[0044]

[0045] Where N is the total number of sample points, k is the frequency index with a value range of 0≤k≤N-1, j is the imaginary unit, X(k) is the frequency domain, and e is the base of the natural logarithm. The amplitude and phase of the signal in different frequency domains are calculated from the results obtained from the Fourier transform.

[0046] Based on big data, multiple sets of noise spectrum diagrams are obtained. These multiple noise spectrum diagrams are then compared with the current noise spectrum diagram to determine their consistency. The calculation formula is as follows:

[0047]

[0048] Where o is the sample size, X i With Y i These are the sample point observations of the two variables at point i on the spectrum, respectively. and Both are the sample mean values ​​of X and Y, and r is the similarity between the two sets of spectrograms. A value of r of 1 indicates complete correlation, meaning the two sets of spectrograms are identical, while other values ​​indicate inconsistency. The closer the value of r is to 1, the more identical the two sets of spectrograms are, and vice versa. The method is to calculate whether the currently collected noise signal is consistent with the noise spectrogram obtained from the big data, so as to obtain the type of the currently collected noise signal.

[0049] Spectrum calculation is an existing technology and will not be elaborated upon further. The spectrum plotting method is based on using MATLAB plotting tools to draw spectrum diagrams. The plotting steps are as follows:

[0050] Set the drawing parameters;

[0051] The horizontal axis represents frequency, in Hz or other appropriate frequency units;

[0052] The vertical axis represents amplitude and power. On a logarithmic scale, decibels (dB) are usually used as the unit to more clearly show the relative magnitude of different frequency components.

[0053] Add grid lines as needed to read the data more accurately;

[0054] Add a title, horizontal and vertical axis labels, and necessary legends to the spectrogram;

[0055] Based on the calculated spectral data and the set plotting parameters, a spectrum diagram is drawn. The spectrum diagram can show the distribution of the signal at different frequencies, thereby helping to analyze the frequency domain characteristics of the signal.

[0056] A spectrum diagram is a graphical representation of a signal at various frequencies, using ripples on the horizontal and vertical axes. It is a commonly used analytical tool in signal processing and communication.

[0057] Noise signal types include natural noise and man-made noise. Natural noise includes atmospheric noise and wind noise, which originate from noise generated in the natural environment. Man-made noise is generated by human activities and includes industrial noise, mechanical noise, and electronic product noise. The categories of noise signals in the natural environment and man-made noise are obtained through big data. By comparing the noise signals obtained from big data, the type of noise can be determined.

[0058] The model training method is based on gradient descent to update the weights and biases, and regularization is used to optimize the model to prevent overfitting. The Adam algorithm is selected as the adaptive learning mechanism. The algorithm parameters are configured in the model and iterative optimization is performed continuously to improve the performance of the neural network model.

[0059] The gradient descent and regularization methods used in this application are existing technologies and will not be elaborated upon here. The core idea of ​​the Adam algorithm is to estimate the learning rate using the mean of the first moment and the variance of the second moment of the gradient, and adaptively update the model parameters during training. It maintains two dynamically adjusted learning rates: one is a momentum term based on the first moment estimation of the gradient, used to smooth the direction of gradient updates; the other is an adaptive learning rate based on the second moment estimation of the gradient, used to dynamically adjust the size of the learning rate. The Adam algorithm can adaptively adjust the learning rate according to the gradient characteristics of different parameters. For parameters with large gradients, the learning rate will decrease accordingly to avoid oscillations caused by excessively rapid parameter updates; for parameters with small gradients, the learning rate will increase accordingly to accelerate convergence. By combining the momentum method, the Adam algorithm obtains a more stable and consistent update direction in the parameter space, thus achieving a faster convergence speed. The Adam algorithm can better adapt to sparse gradients through its adaptability, reducing oscillations during training. Due to the introduction of momentum and second moment estimation, the Adam algorithm has a certain degree of anti-interference ability against noise during training, making the model more stable.

[0060] When inputting the noise signal of the current classification, the neural network model is set with a dual-input frame, which inputs different corresponding noise signals into the corresponding input ports of the neural network model. The dual-input frame of the neural network model includes a natural noise input port and an artificial noise input port. After judging the noise of the current input, the corresponding input is sent to the different input ports.

[0061] In this application, by setting separate input points for natural noise and man-made noise, the model can handle these two different types of noise more specifically. Each type of noise has its unique characteristics and effects. Processing them separately allows the model to better learn and identify the characteristics of each type of noise, thereby improving the processing effect. Inputting different types of noise separately can reduce mutual interference between them. The dual-input framework allows the model to learn natural noise and man-made noise independently. This learning method helps the model to understand the nature and variation patterns of each type of noise more deeply, thereby enhancing the model's generalization ability. Even when encountering new types of noise, the model can reason and make judgments based on its existing learning experience.

[0062] During model training, the model's parameters and structure are dynamically adjusted based on its performance on the validation set. Two training sets are prepared for model training: one set is an artificial noise dataset, and the other set is a natural noise dataset. Each dataset undergoes data preprocessing. The model contains two parallel sub-networks: one for processing artificial noise data and the other for processing natural noise data.

[0063] The iterative training formula is:

[0064]

[0065] Where θ1, θ2 and θ n Let L(θ) be the model parameters and L(θ) be the loss function. The parameters are updated using the following formula:

[0066]

[0067] Where η is the learning rate, the step parameters are updated according to the formula in each iteration of the model to optimize the performance of the neural network model.

[0068] In this application, iterative training allows the model to learn and optimize gradually over multiple iterations. Each iteration is adjusted based on the results of the previous iteration, enabling the model to more accurately capture patterns and rules in the data. Through iterative training, the complexity and parameters of the model can be gradually adjusted to reduce the risk of overfitting. Overfitting is a phenomenon where a model performs well on training data but poorly on new data. Techniques such as cross-validation in iterative training can help identify and prevent overfitting.

[0069] The extracted noise signal is input into the corresponding input port of the trained model. The model suppresses the noise based on the learned knowledge. While suppressing the noise, the model ensures that the integrity of the signal is not affected and retains useful information in the signal, such as frequency components and phase information.

[0070] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A signal noise suppression method based on deep learning to ensure signal integrity, characterized in that, The suppression process is as follows: Noise is acquired, analyzed using spectral analysis, and compared with other noise types based on big data analysis to determine the type of noise. Build a neural network model, train and optimize the model, set up an adaptive learning mechanism, and analyze different types of noise; The extracted noise signals are input into the constructed model to suppress noise sources while ensuring signal integrity. Noise acquisition is based on the collection of audio acquisition equipment, and the acquired noise is uploaded to the computer for processing; the spectrum analysis method pre-samples the continuous-time signal, performs Fourier transform processing on the sampled discrete-time signal, calculates the spectrum, calculates the amplitude and phase of the signal in different frequency domains, and plots the calculated spectrum in graphical form. The formula for calculating signal sampling is as follows: x(n)=x(nT s ) Where T s Let x(n) be the sampling period, x(n) be the discrete-time signal, and n be an integer representing the index of the sampling point. The Fourier transform formula is: Where N is the total number of sample points, k is the frequency index with a value range of 0≤k≤N-1, j is the imaginary unit, X(k) is the frequency domain, and e is the base of the natural logarithm. The amplitude and phase of the signal in different frequency domains are calculated from the results obtained from the Fourier transform. Noise signal types include natural noise and man-made noise. Natural noise includes atmospheric noise and wind noise, which originate from noise generated in the natural environment. Man-made noise is generated based on human activities and includes industrial noise, mechanical noise, and electronic product noise. The categories of noise signals in the natural environment and man-made noise signals are obtained through big data. By comparing the noise signals obtained with the big data, the type of current noise can be determined. The model training method is based on gradient descent to update the weights and biases, and regularization is used to optimize the model to prevent overfitting. The Adam algorithm is selected as the adaptive learning mechanism. The algorithm parameters are configured in the model and iterative optimization is continuously performed to improve the performance of the neural network model. When inputting the noise signal of the current classification, the neural network model is set with a dual-input frame, which inputs different corresponding noise signals into the corresponding input ports of the neural network model. The dual-input frame of the neural network model includes a natural noise input port and an artificial noise input port. After judging the noise of the current input, the corresponding input is sent to the different input ports.

2. The signal noise suppression and signal integrity protection method based on deep learning according to claim 1, characterized in that, Based on big data, multiple sets of noise spectrum diagrams are obtained. These multiple noise spectrum diagrams are then compared with the current noise spectrum diagram to determine their consistency. The calculation formula is as follows: Where o is the sample size, X i With Y i These are the sample point observations of the two variables at point i on the spectrum, respectively. and Both are the sample mean values ​​of X and Y, and r is the similarity between the two sets of spectrograms. A value of r of 1 indicates complete correlation, meaning the two sets of spectrograms are identical, while other values ​​indicate inconsistency. The closer the value of r is to 1, the more identical the two sets of spectrograms are, and vice versa. The method is to calculate whether the currently collected noise signal is consistent with the noise spectrogram obtained from the big data, so as to obtain the type of the currently collected noise signal.

3. The signal noise suppression and signal integrity protection method based on deep learning according to claim 1, characterized in that, During model training, the model's parameters and structure are dynamically adjusted based on its performance on the validation set. Two training sets are prepared for model training: one set is an artificial noise dataset, and the other set is a natural noise dataset. Each dataset undergoes data preprocessing. The model contains two parallel sub-networks: one for processing artificial noise data and the other for processing natural noise data.

4. The signal noise suppression and signal integrity protection method based on deep learning according to claim 1, characterized in that, The iterative training formula is: Where θ1, θ2, ..., θ n Let L(θ) be the model parameters and L(θ) be the loss function. The parameters are updated using the following formula: θ←θ-η▽L(θ) Where η is the learning rate, the step parameters are updated according to the formula in each iteration of the model to optimize the performance of the neural network model.

5. The signal noise suppression and signal integrity protection method based on deep learning according to claim 1, characterized in that, The extracted noise signal is input into the corresponding input port of the trained model. The model suppresses the noise based on the learned knowledge. While suppressing the noise, the model ensures that the integrity of the signal is not affected and retains the useful information in the signal, including frequency components and phase information.

Images