Acousto-optic signal enhancement method based on neural network

Abstract

The invention discloses an acousto-optic signal enhancement method based on a neural network, and belongs to the technical field of signal processing. According to the invention, the hidden mapping relation between the enhanced signal and the original signal is constructed through deep learning based on the neural network, and the acousto-optic signal enhancement model is continuously trained and updated, so that the signal-to-noise ratio and the voice quality of the original signal are remarkably improved, and convenient output is realized. The phase loss Lphase and the amplitude loss LMagnitude are respectively output into the acousto-optic signal enhancement model to update corresponding parameters, so that data test correction and parameter updating of the acousto-optic signal enhancement model are realized; meanwhile, model parameters of the acousto-optic signal enhancement model are continuously updated and iterated by adopting an AdamW optimization algorithm, a loss function is minimized by calculating gradients and updating the parameters, the training efficiency and performance of the model are improved, and the voice quality and the signal-to-noise ratio of output signals of the acousto-optic signal enhancement model can be continuously improved.

Description

Technical Field

[0001] The invention relates to the technical field of signal processing, in particular to an acoustic and optical signal enhancement method based on a neural network. Background Art

[0002] Optical interferometry has become an important tool in precision measurement and micro-nanotechnology research due to its high precision and non-contact measurement characteristics. The basic principle of using optical interferometry to measure vibration is to reflect the vibration of the surface of an object through the phase change of the interference signal. Commonly used interferometers include Michelson interferometers, Fabry-Perot interferometers, and fiber interferometers. These interferometers use beam splitting and beam combining to cause the reflected light from the surface of the object to interfere with the reference light, generating interference fringes. The vibration of the surface of the object will cause a change in the optical path difference, which will lead to a phase change in the interference fringes. Vibration can be accurately measured by detecting the phase change.

[0003] At present, there are many signal distortion problems when using the principle of optical interference to demodulate signals, including nonlinear distortion, phase noise, frequency noise, environmental interference, optical path mismatch and noise superposition. Nonlinear distortion is caused by the fact that the intensity of the optical interference signal is the square of the amplitude of the light wave, resulting in signal nonlinearity. Due to the high sensitivity of phase noise to phase information, any slight phase change will affect the accurate demodulation of the signal. Frequency noise is related to the frequency stability of the light source, and unstable frequency will cause signal distortion. Environmental interference such as temperature, vibration and air flow will affect the stability of the optical path, resulting in interference signal distortion. Optical path mismatch and misalignment will cause the light beams to not completely overlap during interference, increasing signal distortion. During the demodulation process, system noise and background light noise will also be superimposed on the signal, affecting the clarity and accuracy of the signal. In order to reduce and eliminate these distortion problems, it is usually necessary to optimize the optical system design, improve the stability of the light source, improve the demodulation algorithm, and take environmental control measures. Summary of the invention

[0004] In view of the above problems, a method for enhancing acoustic and optical signals based on neural networks is provided. The present invention constructs an implicit mapping relationship between the enhanced signal and the original signal through deep learning based on neural networks, continuously trains and updates the acoustic and optical signal enhancement model, and thus significantly improves the signal-to-noise ratio and speech quality of the original signal. L phase and amplitude loss L MagnitudeThe data are respectively output to the acoustic-optical signal enhancement model to update the corresponding parameters, so as to realize data test correction and parameter update of the acoustic-optical signal enhancement model; and the AdamW optimization algorithm is used to continuously update and iterate the model parameters of the acoustic-optical signal enhancement model. The optimization algorithm of momentum and adaptive learning rate is combined to minimize the loss function by calculating the gradient and updating the parameters, thereby improving the training efficiency and performance of the model, and continuously improving the speech quality and signal-to-noise ratio of the output signal of the acoustic-optical signal enhancement model.

[0005] To achieve the above object, the technical solution adopted by the present invention is.

[0006] A method for enhancing acoustic and optical signals based on a neural network comprises the following steps: Step 1: Obtain standard audio as a reference signal through an audio acquisition device, and obtain the micro-vibration signal generated by the target object as an original signal through an optical sensing device; Step 2: Align the reference signal obtained in step 1 with the original signal, and use a filter to remove abnormal signals; Step 3: Randomly split the filtered reference signal and the original signal into 1-2s signals, and divide them into training data set, test data set and validation data set according to a certain ratio; Step 4: Convert the original signal and the reference signal in the training data set and the verification data set into a phase spectrum and an amplitude spectrum respectively by short-time Fourier transform, and output them to the acousto-optic signal enhancement model; Step 5: The encoder in the acoustic-optical signal enhancement model extracts features from the phase spectrum and amplitude spectrum in step 4 respectively and outputs them to the decoder; Step 6: The decoder reconstructs the phase spectrum and amplitude spectrum after the feature extraction in step 5, and outputs the reconstructed amplitude spectrum and the reconstructed phase spectrum respectively; Step 7: The reconstructed amplitude spectrum and the reconstructed phase spectrum in step 6 are integrated into a final amplitude spectrum by reconstruction, and the final amplitude spectrum is subjected to inverse short-time Fourier transform to output an enhanced signal; Step 8: Calculate the time domain loss using the enhanced signal in step 7 and the reference signal in step 4 L time , Amplitude loss L Magnitude and phase loss L phase , and the phase loss L phase and amplitude loss L Magnitude The signals are respectively output to the acoustic and optical signal enhancement model to update the corresponding parameters; Step 9: Based on the time domain loss in step 8 L time , Amplitude lossL Magnitude and phase loss L phase Calculate the overall loss of the final generated signal L g To evaluate the effect of the acoustic-optical signal enhancement model, and input the test data set in step 3 into the acoustic-optical signal enhancement model for data testing and correction; Step 10: Repeat steps 1 to 9, continuously obtain new reference signals and original signals, input the new original signals and reference signals into the acoustic-optical signal enhancement model again, and use the AdamW optimization algorithm to continuously update and iterate the model parameters of the acoustic-optical signal enhancement model. In this process, the voice quality and signal-to-noise ratio of the output signal of the acoustic-optical signal enhancement model are continuously improved, and finally the construction and training of the acoustic-optical signal enhancement model are completed.

[0007] Preferably, the audio acquisition device is one of a microphone, a recorder or an acoustic camera.

[0008] Preferably, the split ratio of the training data set, the test data set and the validation data set is 8:1:1.

[0009] Preferably, the abnormal signal includes one or more combinations of white noise, Gaussian white noise, colored noise and clipping distortion.

[0010] Preferably, the time domain loss in step 8 is L time , Amplitude loss L Magnitude and phase loss L phase The calculation formula is as follows: in, L ip , L gd , L iaf They are in-phase loss, gradient delay loss, and integrated absolute frequency loss; y is the reference signal; f ( x ) is the prediction signal; N Represents the total number of batches of training samples.

[0011] Preferably, the overall loss of the signal finally generated in step 9 is L g The formula is as follows: in, w 1. w 2 and w 3 are the time domain losses L time , Amplitude loss L Magnitude and phase loss L phase The weight of .

[0012] Preferably, the time domain loss L time , Amplitude loss L Magnitude and phase loss L phase The weights are 0.1, 0.8, and 0.1 respectively.

[0013] A method for using an acoustic-optical signal enhancement model comprises the following steps: Step S1: acquiring a micro-vibration signal generated by a target object as an original signal through an optical sensing device; Step S2: using a filter to filter out abnormal signals from the original signal in step S1; Step S3: randomly divide the filtered original signal into 1-2s signals, convert the divided original signal into a phase spectrum through short-time Fourier transform, and output it to the trained acoustic-optical signal enhancement model; Step S4: After the original signal is processed by the trained acoustic-optical signal enhancement model, a predicted enhanced signal is output.

[0014] Due to the adoption of the above technical solution, the present invention has the following beneficial effects.

[0015] (1) The present invention uses a neural network model of a codec architecture to extract signal features through short-time Fourier transform, reconstructs the enhanced signal through inverse short-time Fourier transform, and calculates the time domain loss of the enhanced signal and the reference signal respectively. L time , Amplitude loss L Magnitude , Phase loss L phase and the overall loss of the resulting signal L g To evaluate and judge the signal enhancement effect of the acoustic-optical signal enhancement model, and at the same time L phase and amplitude loss L MagnitudeThe corresponding parameters are updated by outputting them to the acoustic-optic signal enhancement model respectively, realizing data test correction and parameter update of the acoustic-optic signal enhancement model, and continuously reducing the loss function. With the increase of the input original signal and reference signal, the enhancement capability of the original signal can be continuously improved, and the signal enhancement and prediction accuracy are continuously improved.

[0016] (2) The present invention uses the AdamW optimization algorithm to continuously update and iterate the model parameters of the acoustic-optical signal enhancement model, and combines the optimization algorithm of momentum and adaptive learning rate to minimize the loss function by calculating the gradient and updating the parameters, thereby improving the training efficiency and performance of the model, and continuously improving the voice quality and signal-to-noise ratio of the output signal of the acoustic-optical signal enhancement model. The gradient and current parameter values ​​calculated during the model training process are used to update the model parameters, so that the loss function is gradually reduced. As the number of iterations increases, the optimization algorithm will gradually find the local minimum or global minimum of the loss function, so that the acoustic-optical signal enhancement capability of the model and the quality and signal-to-noise ratio of the output signal are continuously improved.

[0017] (3) The present invention constructs an implicit mapping relationship between the enhanced signal and the original signal through deep learning based on neural networks, and continuously trains and updates the acoustic-optical signal enhancement model to achieve a significant improvement in the signal-to-noise ratio and voice quality of the original signal. At the same time, indicators such as PESQ and SSNR are used to objectively evaluate the enhanced signal, and the performance of the acoustic-optical signal enhancement model is verified by objective indicators. It can be concluded that the enhanced signal output by the acoustic-optical signal enhancement model of the present invention is significantly improved over the original signal in terms of evaluation indicators such as speech quality perception evaluation (PESQ) and segmented signal-to-noise ratio (SSNR), which can greatly improve the voice quality and signal-to-noise ratio of the original signal.

[0018] (4) The present invention builds and trains an acoustic-optical signal enhancement model based on a neural network architecture, and relies on the acoustic-optical signal enhancement model to directly predict the original signal obtained by the optical sensor device and output the enhanced signal. The neural network has strong scalability, can process real-time signals, has high robustness in the face of different types and degrees of noise, has strong nonlinear modeling capabilities, adaptability and feature extraction capabilities, and can perform multi-task learning and end-to-end training, which significantly improves the performance, scope of application and convenience of the acoustic-optical signal enhancement model. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] The following is a detailed discussion of the making and application of the preferred embodiments of the present invention. However, it should be understood that the present invention provides many applicable inventive concepts that can be embodied in various specific environments. The specific embodiments discussed are only for the purpose of illustrating the specific ways of making and using the present invention, and do not limit the scope of the present invention. For those of ordinary skill in the art, other drawings can also be obtained based on these drawings without creative work.

[0020] Figure 1 It is a flow chart of the present invention.

[0021] Figure 2 Schematic diagram of signal acquisition of the acoustic-optical signal enhancement model of the present invention.

[0022] Figure 3 It is a comparison diagram of the evaluation indexes of the enhanced signal and the original signal of the present invention. DETAILED DESCRIPTION

[0023] The following is a detailed discussion of the making and application of the preferred embodiments of the present invention. However, it should be understood that the present invention provides many applicable inventive concepts that can be embodied in various specific environments. The specific embodiments discussed are only intended to illustrate specific ways to make and use the present invention and do not limit the scope of the present invention.

[0024] The present invention uses a neural network model of a codec architecture to extract signal features through short-time Fourier transform, reconstructs the enhanced signal through inverse short-time Fourier transform, and calculates the time domain loss of the enhanced signal and the reference signal respectively. L time , Amplitude loss L Magnitude , Phase loss L phase and the overall loss of the resulting signal L g To evaluate and judge the signal enhancement effect of the acoustic-optical signal enhancement model, and at the same time L phase and amplitude loss L Magnitude The corresponding parameters are updated by outputting them to the acoustic-optic signal enhancement model respectively, realizing data test correction and parameter update of the acoustic-optic signal enhancement model, and continuously reducing the loss function. With the increase of the input original signal and reference signal, the enhancement capability of the original signal can be continuously improved, and the signal enhancement and prediction accuracy are continuously improved.

[0025] In addition, the present invention uses the AdamW optimization algorithm to continuously update and iterate the model parameters of the acoustic-optical signal enhancement model, combines the optimization algorithm of momentum and adaptive learning rate, and minimizes the loss function by calculating the gradient and updating the parameters to improve the training efficiency and performance of the model, which can continuously improve the voice quality and signal-to-noise ratio of the output signal of the acoustic-optical signal enhancement model, and updates the model parameters by the gradient and current parameter values ​​calculated during the training model, so that the loss function is gradually reduced. With the increase in the number of iterations, the optimization algorithm will gradually find the local minimum or global minimum of the loss function, so that the acoustic-optical signal enhancement capability of the model and the quality and signal-to-noise ratio of the output signal are continuously improved.

[0026] The following is combined with Figure 1-3 Further elaboration.

[0027] like Figure 1 The method for enhancing acoustic and optical signals based on a neural network includes the following steps: Step 1: obtaining standard audio as a reference signal through an audio acquisition device, wherein the audio acquisition device is a microphone, a recorder or an acoustic camera. The micro-vibration signal generated by the target object is obtained through an optical sensing device as an original signal.

[0028] Step 2: Align the reference signal and the original signal obtained in step 1, and use a filter to remove abnormal signals. The abnormal signals include one or more combinations of white noise, Gaussian white noise, colored noise and clipping distortion.

[0029] Step 3: Randomly split the filtered reference signal and the original signal into 1-2s signals, and divide them into training data set, test data set and validation data set in a split ratio of 8:1:1.

[0030] Step 4: The original signal and the reference signal in the training data set and the validation data set are respectively converted into a phase spectrum and an amplitude spectrum by short-time Fourier transform, and output to the acousto-optic signal enhancement model.

[0031] Step 5: The encoder in the acoustic-optical signal enhancement model extracts features from the phase spectrum and amplitude spectrum in step 4 and outputs them to the decoder.

[0032] Step 6: After the decoder reconstructs the phase spectrum and amplitude spectrum after the feature extraction in step 5, it outputs the reconstructed amplitude spectrum and the reconstructed phase spectrum respectively.

[0033] Step 7: The reconstructed amplitude spectrum and the reconstructed phase spectrum in step 6 are integrated into a final amplitude spectrum through reconstruction. The final amplitude spectrum is subjected to inverse short-time Fourier transform to output an enhanced signal.

[0034] Step 8: Calculate the time domain loss using the enhanced signal in step 7 and the reference signal in step 4 L time , Amplitude loss L Magnitude and phase loss L phase , and the phase loss L phase and amplitude loss L Magnitude The signals are respectively output to the acoustic and optical signal enhancement model to update the corresponding parameters.

[0035] The time domain loss L time , Amplitude lossL Magnitude and phase loss L phase The calculation formula is as follows: in, L ip , L gd , L iaf They are in-phase loss, gradient delay loss, and integrated absolute frequency loss; y is the reference signal; f ( x ) is the prediction signal; N Represents the total number of batches of training samples.

[0036] Step 9: Based on the time domain loss in step 8 L time , Amplitude loss L Magnitude and phase loss L phase Calculate the overall loss of the final generated signal L g To evaluate the effect of the acoustic-optical signal enhancement model, and input the test data set in step 3 into the acoustic-optical signal enhancement model for data testing and correction.

[0037] The overall loss of the final generated signal L g The formula is as follows: in, w 1. w 2 and w 3 are the time domain losses L time , Amplitude loss L Magnitude and phase loss L phase The weights are 0.1, 0.8 and 0.1 respectively.

[0038] Step 10: Repeat steps 1 to 9, continuously obtain new reference signals and original signals, input the new original signals and reference signals into the acoustic-optical signal enhancement model again, and use the AdamW optimization algorithm to continuously update and iterate the model parameters of the acoustic-optical signal enhancement model. In this process, the voice quality and signal-to-noise ratio of the output signal of the acoustic-optical signal enhancement model are continuously improved, and finally the construction and training of the acoustic-optical signal enhancement model are completed.

[0039] The training model uses the AdamW optimization algorithm. The gradients and current parameter values ​​calculated during the training process are used to update the model parameters, so that the loss function is gradually reduced. As the number of iterations increases, the optimization algorithm will gradually find the local minimum or global minimum of the loss function, so that the model's enhancement capabilities are continuously improved.

[0040] The AdamW optimization algorithm includes the following steps: Step 1: Initialize parameters θ , learning rate η , momentum attenuation coefficient β 1. Decay rate of second-order second-moment estimate β 2. Weight decay coefficient w t .

[0041] Step 2: Initialize the first-order moment estimate m 0=0, second-order moment estimate v 0=0.

[0042] Step 3: For each time step, calculate the gradient g t = θLg ( θ t ),in θ t Indicates time t Parameters, L g is the overall loss of the final generated signal.

[0043] Step 2: Calculate the first moment estimate : in g t It's in time t The gradient of β 1 is the decay rate of the first-order moment estimate.

[0044] Step 3: Calculate the second moment estimate : in β2 is the decay rate of the second-order moment estimate, g t It's in time t gradient.

[0045] Step 4: Update parameters θ t+1 : in η is the learning rate, w t is the weight decay rate, ϵ is a small constant added for numerical stability.

[0046] like Figure 2 The method for using the acoustic-optical signal enhancement model shown includes the following steps: Step S1: obtaining the micro-vibration signal generated by the target object as the original signal through an optical sensing device.

[0047] Step S2: using a filter to filter out abnormal signals from the original signal in step S1.

[0048] Step S3: randomly divide the filtered original signal into 1-2s signals, convert the divided original signal into a phase spectrum through short-time Fourier transform, and output it to the trained acousto-optic signal enhancement model.

[0049] Step S4: After the original signal is processed by the trained acoustic-optical signal enhancement model, a predicted enhanced signal is output.

[0050] In addition, the results of the acoustic and optical signal enhancement model use indicators such as Perceptual Evaluation of Speech Quality (PESQ) and Segmental Signal-to-Noise Ratio (SSNR) as objective evaluation indicators of the model effect. Among them, the specific mathematical details and implementation process of PESQ are relatively complex, and are usually calculated through special software packages (such as the implementation of the ITU-T P.862 standard). And because it is difficult to directly write the PESQ calculation formula, it is usually recommended to use the existing third-party library PESQ (V0.0.4) for calculation. The calculation formula for the segment signal-to-noise ratio (SSNR) is as follows: Among them, SSNR (Segmental Signal-to-Noise Ratio) is a variant of signal-to-noise ratio. SNR i For the i The original signal of the present invention is processed by the trained acoustic and optical signal enhancement model, and the output of the speech quality perception evaluation (PESQ) and the segment signal-to-noise ratio (SSNR) are compared with the original signal input at the beginning. Figure 3 shown.

[0051] from Figure 3 It can be seen that after the enhanced signal is processed by the trained acousto-optic signal enhancement model, the output prediction result is significantly improved compared with the original signal in the speech quality perception evaluation (PESQ) and the segmented signal-to-noise ratio (SSNR). That is, the present invention constructs an implicit mapping relationship between the enhanced signal and the original signal through deep learning based on neural networks, and continuously trains and updates the acousto-optic signal enhancement model to achieve a significant improvement in the signal-to-noise ratio and speech quality of the original signal. At the same time, indicators such as PESQ and SSNR are used to objectively evaluate the enhanced signal, and the performance of the acousto-optic signal enhancement model is verified by objective indicators. It can be concluded that the enhanced signal output by the acousto-optic signal enhancement model of the present invention is significantly improved compared with the original signal in evaluation indicators such as speech quality perception evaluation (PESQ) and the segmented signal-to-noise ratio (SSNR), which can greatly improve the speech quality and signal-to-noise ratio of the original signal.

[0052] The present invention builds and trains an acoustic-optical signal enhancement model based on a neural network architecture, and relies on the acoustic-optical signal enhancement model to directly predict the original signal acquired by the optical sensor device and output the enhanced signal. The neural network has strong scalability, can process real-time signals, has high robustness in the face of different types and degrees of noise, has strong nonlinear modeling capabilities, adaptability and feature extraction capabilities, and can perform multi-task learning and end-to-end training, which significantly improves the performance, scope of application and convenience of the acoustic-optical signal enhancement model.

[0053] Although the specification has been described in detail, it should be understood that various changes, substitutions and modifications may be made without departing from the spirit and scope of the invention as defined by the appended claims. In addition, the specific embodiments described are not intended to limit the scope of the invention, and those of ordinary skill in the art can easily understand based on the present invention that currently existing or later to be developed processes, machines, manufactures, material compositions, means, methods, or steps can perform substantially the same functions as the embodiments of the present invention or obtain substantially the same results. Therefore, the appended claims are intended to include such processes, machines, manufactures, material compositions, means, methods or steps within their scope.

Claims

1. A method for enhancing acoustic and optical signals based on a neural network, characterized in that: The following steps are involved: Step 1: Obtain standard audio as a reference signal through an audio acquisition device, and obtain the micro-vibration signal generated by the target object as an original signal through an optical sensing device; Step 2: Align the reference signal obtained in step 1 with the original signal, and use a filter to remove abnormal signals; Step 3: Randomly split the filtered reference signal and the original signal into 1-2s signals, and divide them into training data set, test data set and validation data set according to a certain ratio; Step 4: Convert the original signal and the reference signal in the training data set and the verification data set into a phase spectrum and an amplitude spectrum respectively by short-time Fourier transform, and output them to the acousto-optic signal enhancement model; Step 5: The encoder in the acoustic-optical signal enhancement model extracts features from the phase spectrum and amplitude spectrum in step 4 respectively and outputs them to the decoder; Step 6: The decoder reconstructs the phase spectrum and amplitude spectrum after the feature extraction in step 5, and outputs the reconstructed amplitude spectrum and the reconstructed phase spectrum respectively; Step 7: The reconstructed amplitude spectrum and the reconstructed phase spectrum in step 6 are integrated into a final amplitude spectrum by reconstruction, and the final amplitude spectrum is subjected to inverse short-time Fourier transform to output an enhanced signal; Step 8: Calculate the time domain loss using the enhanced signal in step 7 and the reference signal in step 4 L time , Amplitude loss L Magnitude and phase loss L phase , and the phase loss L phase and amplitude loss L Magnitude The signals are respectively output to the acoustic and optical signal enhancement model to update the corresponding parameters; Step 9: Based on the time domain loss in step 8 L time , Amplitude loss L Magnitude and phase loss L phase Calculate the overall loss of the final generated signal L g To evaluate the effect of the acoustic-optical signal enhancement model, and input the test data set in step 3 into the acoustic-optical signal enhancement model for data testing and correction; Step 10: Repeat steps 1 to 9, continuously obtain new reference signals and original signals, input the new original signals and reference signals into the acoustic-optical signal enhancement model again, and use the AdamW optimization algorithm to continuously update and iterate the model parameters of the acoustic-optical signal enhancement model. In this process, the voice quality and signal-to-noise ratio of the output signal of the acoustic-optical signal enhancement model are continuously improved, and finally the construction and training of the acoustic-optical signal enhancement model are completed.

2. The method for enhancing acoustic and optical signals based on a neural network as claimed in claim 1, characterized in that: The audio acquisition device is one of a microphone, a recorder or an acoustic camera.

3. The method for enhancing acoustic and optical signals based on a neural network as claimed in claim 1, characterized in that: The split ratio of the training data set, the test data set, and the validation data set is 8:1:

1.

4. The method for enhancing acoustic and optical signals based on a neural network as claimed in claim 1, characterized in that: The abnormal signal includes one or more combinations of white noise, Gaussian white noise, colored noise and clipping distortion.

5. The method for enhancing acoustic and optical signals based on a neural network as claimed in claim 1, characterized in that: The time domain loss in step 8 L time , Amplitude loss L Magnitude and phase loss L phase The calculation formula is as follows: in, L ip , L gd , L iaf They are in-phase loss, gradient delay loss, and integrated absolute frequency loss; y is the reference signal; f ( x ) is the prediction signal; N Represents the total number of batches of training samples.

6. The method for enhancing acoustic and optical signals based on a neural network as claimed in claim 1, characterized in that: The overall loss of the final signal generated in step 9 L g The formula is as follows: in, w 1. w 2 and w 3 are the time domain losses L time , Amplitude loss L Magnitude and phase loss L phase The weight of .

7. The method for enhancing acoustic and optical signals based on a neural network as claimed in claim 6, characterized in that: The time domain loss L time , Amplitude loss L Magnitude and phase loss L phase The weights are 0.1, 0.8, and 0.1 respectively.

8. A method for using an acoustic-optical signal enhancement model, using the acoustic-optical signal enhancement method based on a neural network as described in any one of claims 1 to 7, characterized in that: The following steps are involved: Step S1: obtaining a micro-vibration signal generated by a target object as an original signal through an optical sensing device; Step S2: using a filter to filter out abnormal signals from the original signal in step S1; Step S3: randomly divide the filtered original signal into 1-2s signals, convert the divided original signal into a phase spectrum through short-time Fourier transform, and output it to the trained acoustic-optical signal enhancement model; Step S4: After the original signal is processed by the trained acoustic-optical signal enhancement model, a predicted enhanced signal is output.

Images