Data-driven based ultra-low noise acoustic measurement optimization method and system

Abstract

This invention discloses a data-driven optimization method and system for ultra-low noise acoustic measurements, comprising: acquiring multi-channel acoustic measurement signals from the same acoustic environment using acoustic sensors; performing in-situ estimation of the electrical noise component in the multi-channel acoustic measurement signals; performing low-rank sparse decomposition based on the multi-channel acoustic measurement signals and electrical noise characteristic parameters to obtain data features of the measurement state; inputting the data features into a data-driven decision model to output weight parameters of the noise state and obtain a decision result; performing noise de-embedding on the multi-channel acoustic measurement signals based on the decision result and weight parameters to reconstruct the acoustic signals, obtaining measurement results with a reduced equivalent noise lower limit; and outputting optimized ultra-low noise acoustic measurement results. This method optimizes ultra-low noise acoustic measurement results and measurement reliability through in-situ noise estimation and meta-learning data-driven decision-making, while also exhibiting good interpretability.

Description

Technical Field

[0001] This invention relates to the field of acoustic measurement, and more particularly to a data-driven method and system for optimizing ultra-low noise acoustic measurements. Background Technology

[0002] Acoustic measurement is the fundamental guarantee for radar equipment to assist in acoustic interference suppression and other applications. As high-end equipment develops towards high precision, lightweight and integration, stringent requirements are placed on the control of the equivalent noise lower limit, the ability to extract weak signals and the adaptability of in-situ measurement in acoustic measurement. Ultra-low noise acoustic measurement solutions mainly rely on hardware stacking optimization, such as building anechoic chambers, selecting ultra-precision low-noise acoustic sensors and electromagnetic shielding links, and configuring high-precision low-noise amplifiers to reduce noise. This results in high cost and large size of measurement equipment. At the same time, fixed-rule filtering algorithms are difficult to deal with the interference of electrical noise and acoustic noise coupling and non-stationary random noise superposition in complex scenarios, which easily leads to problems such as effective signal distortion and incomplete noise suppression.

[0003] With the rapid development of data-driven technologies such as machine learning and deep learning, powerful adaptive modeling and feature learning capabilities have been demonstrated in signal processing and noise suppression, providing new technical paths for acoustic measurement optimization. Data-driven methods have been applied to scenarios such as acoustic signal denoising and fault diagnosis. By learning the feature distribution of acoustic signals through models such as convolutional neural networks and recurrent neural networks, specific noise can be suppressed. Alternatively, blind source separation algorithms can be used to decompose multi-channel signals to separate target signals from interference noise and reduce the impact of hardware errors. In the sub-field of ultra-low noise acoustic measurement, most data-driven solutions optimize a single aspect, making it difficult to resolve the contradiction between noise suppression and signal fidelity in ultra-low noise measurement. The generalization ability of the models is also insufficient. Therefore, how to optimize ultra-low noise acoustic measurement in a low-cost, highly adaptable, and highly accurate manner has become an urgent problem to be solved. Summary of the Invention

[0004] The purpose of this invention is to provide a data-driven method and system for optimizing ultra-low noise acoustic measurements.

[0005] To achieve the above objectives, the present invention is implemented according to the following technical solution: The first aspect of this invention provides a data-driven method for optimizing ultra-low noise acoustic measurements, comprising: Based on the acquisition of multi-channel acoustic measurement signals of the same acoustic environment by acoustic sensors, the electrical noise components of the acoustic sensors and front-end circuits in the multi-channel acoustic measurement signals are estimated in situ through second-order blind identification to obtain electrical noise characteristic parameters. Low-rank sparse decomposition is performed on the multi-channel acoustic measurement signal and electrical noise characteristic parameters to obtain the data characteristics of the measurement state. The data characteristics include one or more of time-domain characteristics, frequency-domain characteristics, and time-frequency-domain characteristics. The data features are input into the data-driven decision model, which outputs the weight parameters of the noise state to obtain the decision result, which characterizes the reliability of the acoustic measurement. Based on the judgment result and weight parameters, noise de-embedding is performed on the multi-channel acoustic measurement signal to reconstruct the acoustic signal, obtain the measurement result with reduced equivalent noise lower limit, and output the optimized ultra-low noise acoustic measurement result.

[0006] Furthermore, the method for obtaining the multi-channel acoustic measurement signal includes: Based on the synchronous sampling of acoustic signals excited by the same sound source using acoustic sensors, multiple channels of original acoustic sampling data with time alignment and independent channels are obtained. The multiple channels of original acoustic sampling data are de-DC offset and abnormal sampling points are removed to obtain multi-channel acoustic measurement signals. The number of acoustic sensors is greater than or equal to two.

[0007] Furthermore, the method for obtaining the electrical noise characteristic parameters includes: The multi-channel acoustic measurement signal is framed and de-meaned to obtain locally stationary data frames with zero mean. The time delay point is determined according to the sampling frequency of the current measurement state of the acoustic sensor. The time delay covariance matrix corresponding to each time delay point is calculated for the locally stationary data frame. A set of time delay covariance matrices containing different time delay points is constructed. The time delay point is an integer multiple of the sampling interval. The time delay point includes short time delay, medium time delay and long time delay. The short time delay is 1 to 5 times the sampling interval. The medium time delay has the same signal period as the center frequency of the target measurement frequency band. The long time delay is greater than the medium time delay. The time delay covariance matrix set is approximately jointly diagonalized to obtain the blind separation matrix of the multi-channel acoustic measurement signal. The inverse of the blind separation matrix is ​​used as the mixing matrix. The acoustic signal array response matrix is ​​calculated based on the arrangement position of the acoustic sensor, the sound velocity and the target measurement frequency band to obtain the theoretical array manifold. The mixing matrix represents the mixing ratio of each independent statistical component in the multi-channel. The theoretical array manifold is a unity gain steering vector. Based on the pure electrical noise data during the system's silent period, a spatial matching detection threshold is obtained using the Neyman-Pearson criterion under a preset false alarm rate constraint. The column vector matching degree is then calculated based on the hybrid matrix and the theoretical array manifold. The formula for calculating the column vector matching degree is as follows: ; in For the first The column vector matching degree of each independent statistical component For the separation matrix, the first Column vectors For the first The total energy of each independent statistical component For the first The electrical noise power estimate of each independent statistical component is greater than zero. For the theoretical array manifold, according to and Perform a correlation scan within the target measurement frequency band to obtain... The largest direction to obtain , The direction of arrival (DOA) It is a very small positive number. It is the conjugate transpose. Equal to the number of array elements; If the column vector matching degree is less than the spatial matching detection threshold, the corresponding independent statistical component is determined to be an electrical noise component. Electrical noise characteristic parameters are obtained based on the time-domain statistics and frequency-domain power spectrum of the electrical noise component. The time-domain statistics include mean, variance, and root mean square.

[0008] Furthermore, the method for obtaining the data features includes: A pre-whitening matrix is ​​constructed based on electrical noise characteristic parameters. The pre-whitening matrix is ​​used to pre-whiten the multi-channel acoustic measurement signals to obtain a pre-processed signal that suppresses electrical noise related components. The observed power spectrum is calculated based on the pre-processed signal. The signal-to-noise ratio (SNR) at each frequency point is calculated based on the observed power spectrum and the noise power spectrum. The high reliability state and low reliability state are divided according to the SNR of the current measurement state. If the preprocessed signal is in the high reliability state, then the preprocessed signal is subjected to low-rank sparse decomposition, and high-dimensional data features are extracted based on the low-rank signal subspace. The high-dimensional data features include the short-time Fourier transform amplitude spectrum, phase spectrum and Mel frequency cepstral coefficients in the time-frequency domain. The low-rank signal subspace represents the principal components of the target acoustic signal. If the preprocessed signal is in the low reliability state, then the preprocessed signal is decomposed into a low-rank approximation to extract low-dimensional data features, which include Hilbert envelope features, zero-crossing rate features and power spectrum centroid features in the frequency domain. The high-dimensional or low-dimensional data features are cascaded with the signal-to-noise ratio to generate data features for the current measurement state.

[0009] Furthermore, the method for obtaining the weight parameters includes: The thermal noise power spectrum of the front-end circuit at different operating temperatures is obtained. A training task set is constructed based on the difference in thermal noise power spectrum. The neural network is optimized and trained in two layers through model-independent element learning based on the training task set to obtain a data-driven decision model. The operating temperature includes the temperature change range from cold start to steady-state operation. The neural network includes a feature encoder, a task adaptation layer and an output layer. The feature encoder inputs the data features of the measurement state. The task adaptation layer includes a parameter subset. The output layer outputs the weight parameters of the noise state. During the inner loop of the bi-layer optimization training, gradient descent is performed on the support set data of each training task using a regression loss function to obtain task adaptation parameters. During the outer loop meta-update of the bi-layer optimization training, forward inference is performed on the query set data based on the task adaptation parameters to calculate the query set loss until the maximum number of iterations is reached. Then, the meta-model parameters are updated using gradient descent. The regression loss function is: ; in Let be the regression loss function, and let be the objective function of the inner loop to optimize the deviation between the predicted weights and the true labels. To support the number of samples in the set, For the first Signal-to-noise ratio of each sample For the first The labeled weights of each sample, For the weighted output layer, the input features The predicted weight parameters, The square of the L2 norm; Based on the fixed meta-model parameters during the online application phase, an online support set is generated according to the data features of the current scene, and scene adaptive parameters are obtained through two-step gradient descent. The iterative update formula for the scene adaptive parameters is as follows: ; in Let these be the scene adaptation parameters after the t-th iteration. For the first Scene adaptation parameters in the next iteration For learning rate, For loss function pairs gradient, For the non-regular coefficient, For meta-model parameters, This is an indicator function that outputs 1 if the condition within the parentheses is true, and 0 otherwise. The error between the model's predicted output and the actual labeled values. The preset residual threshold; Based on scene adaptive parameters, forward reasoning is performed on the input data features. According to the data-driven decision model output layer, weight parameters of noise suppression intensity under the current measurement state are generated. The weight parameters are used as observation samples and sequential probability ratio test is performed to obtain the decision result.

[0010] Furthermore, the method for obtaining the judgment result includes: Based on the historical distribution statistics of the weight parameters during the meta-training phase, probability density functions for high-reliability and low-reliability states are fitted. The likelihood ratio of the current cumulative observations is calculated based on the observed samples. The formula for calculating the likelihood ratio is: ; in For the first The likelihood ratio of frames, For the first Observed values ​​of the weight parameters of the received frame. High reliability status The probability density function value under the following conditions In low reliability state The probability density function value, For the first Frame likelihood ratio, initial value =1; If the likelihood ratio is greater than the upper threshold, the decision result is output as low reliability; if it is less than the lower threshold, the decision result is output as high reliability; if it is between the upper and lower thresholds, the decision result of the previous moment is maintained and the next frame is observed. The formula for calculating the upper threshold is: ; The formula for calculating the lower bound threshold is: ; in This is the upper threshold. This is the lower bound threshold. To preset the false alarm rate, This sets the preset false negative probability.

[0011] Furthermore, the method for obtaining the extremely low noise acoustic measurement results includes: If the decision result is high reliability, Wiener filtering and weighted fusion of the acoustic measurement signals of each channel are performed based on the weight parameters, and the optimized acoustic signal is obtained by reconstructing it through inverse short-time Fourier transform. If the decision result is low reliability, the signal is projected onto the orthogonal complement space of electrical noise characteristic parameters based on the weight parameters, and a conservative estimate of the acoustic signal is obtained through low-rank approximation reconstruction. The reconstructed acoustic signal is subjected to inverse transformation to recover the time-domain waveform, thereby obtaining optimized ultra-low noise acoustic measurement results, wherein the reduction in the equivalent noise lower limit is greater than or equal to 3 dB.

[0012] A second aspect of the present invention provides a data-driven, ultra-low noise acoustic measurement optimization system, comprising: Data acquisition module: used to acquire multi-channel acoustic measurement signals of the same acoustic environment based on acoustic sensors, and to perform in-situ estimation of the electrical noise components of the acoustic sensors and front-end circuits in the multi-channel acoustic measurement signals through second-order blind identification to obtain electrical noise characteristic parameters; Data feature extraction module: used to perform low-rank sparse decomposition based on the multi-channel acoustic measurement signal and electrical noise characteristic parameters to obtain data features of the measurement state, wherein the data features include one or more of time-domain features, frequency-domain features and time-frequency-domain features; Data-driven decision module: used to input the data features into the data-driven decision model, output the weight parameters of the noise state, and obtain the decision result, which characterizes the reliability of the acoustic measurement; Measurement result optimization module: used to perform noise de-embedding on the multi-channel acoustic measurement signal according to the decision result and weight parameters, reconstruct the acoustic signal, obtain measurement results with reduced equivalent noise lower limit, and output optimized ultra-low noise acoustic measurement results.

[0013] Compared with the prior art, the embodiments of the present invention have at least the following advantages or beneficial effects: This invention performs in-situ estimation of electrical noise components of acoustic sensors and front-end circuits in multi-channel acoustic measurement signals through second-order blind identification. It obtains noise statistical characteristics without changing the sensor arrangement, classifies the measurement state into high-reliability and low-reliability states based on the signal-to-noise ratio (SNR), and extracts data features differentially to ensure feature effectiveness and reduce computational complexity in low SNR scenarios. It uses model-independent element learning combined with sequential probability ratio testing for data-driven decision-making, and performs noise de-embedding and signal reconstruction based on weight parameters to balance electrical noise suppression and signal fidelity, thereby lowering the equivalent noise lower limit and obtaining extremely low-noise acoustic measurement results. Attached Figure Description

[0014] Figure 1 This is a flowchart illustrating the steps of a data-driven, ultra-low noise acoustic measurement optimization method in an embodiment of the present invention. Detailed Implementation

[0015] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0016] Reference Figure 1 As shown, this invention provides a data-driven method for optimizing ultra-low noise acoustic measurements, including: Based on the acquisition of multi-channel acoustic measurement signals of the same acoustic environment by acoustic sensors, the electrical noise components of the acoustic sensors and front-end circuits in the multi-channel acoustic measurement signals are estimated in situ through second-order blind identification to obtain electrical noise characteristic parameters. In the actual evaluation, the acoustic signals of the same wind turbine operating sound source were synchronously sampled to obtain four channels of raw acoustic sampling data with time alignment and independent channels, with a total sample size of 48kHz × 10s. A first-order high-pass filter with a cutoff frequency of 10Hz was used to remove the DC component of each channel. The multi-channel signal was then framed with a frame length of 1024, corresponding to a duration of 21.3ms, a frame shift of 512, and 50% overlap, resulting in 937 zero-mean locally stationary data frames. Based on the sampling interval Ts = 1 / fs ≈ 20.83μs, time delay points were set: short time delays of 1Ts, 2Ts, 3Ts, 4Ts, and 5Ts; medium time delays of 1000Hz (center frequency of the target measurement frequency band 20Hz~2kHz, corresponding to a period of 48Ts); and long time delays of 96Ts (corresponding to 2ms). The time delay covariance matrix was calculated for each time delay point. ;in For the first One delay point The time delay covariance matrix, For the first The first frame of locally stationary data A multi-channel measurement signal column vector at each sampling time. For the first Each sampling time shifted forward A multi-channel signal column vector with sampling intervals Using the matrix transpose operator, column vectors are converted into row vectors, resulting in seven 4×4 time delay covariance matrices. These matrices are then set together. The SOBI algorithm is used to approximate the joint diagonalization of this set, yielding a 4×4 blind separation matrix. The inverse of this matrix is ​​taken as the mixing matrix. The mixing ratio of each column's independent statistical components in the four channels is calculated. The acoustic signal array response matrix is ​​then calculated, where the acoustic sensors are arranged in a uniform linear array with four elements and a spacing of 0.05 m. The velocity of sound at room temperature is 340 m / s, and the target frequency band is 20 Hz to 2 kHz. The theoretical array manifold in unity-gain steering vector form is obtained. The system collected 10 seconds of pure electrical noise data during a quiet period when the fan was off and there were no external sound sources. Based on the Neyman-Pearson criterion and a preset false alarm rate of 0.05, a spatial matching detection threshold of 3.2 was calculated. The matching degree of each independent statistical component was calculated, with the minimum positive number being 1×10e-10. The column vector matching degrees of the four independent components were obtained as 2.8, 15.2, 18.7, and 3.1. Components with matching degrees less than the threshold of 3.2 were identified as electrical noise components. Therefore, the first and fourth independent components were electrical noise components, and the second and third were effective acoustic signal components. Feature parameters were extracted from the electrical noise components, including the time-domain variance. ;, The power spectral density from 20 Hz to 2 kHz was estimated using the periodogram method. The noise power spectral density at 1 kHz was... The time-domain variance, time-domain root mean square, time-domain mean, and frequency-domain power spectrum are used as electrical noise characteristic parameters. Low-rank sparse decomposition is performed on the multi-channel acoustic measurement signal and electrical noise characteristic parameters to obtain the data characteristics of the measurement state. The data characteristics include one or more of time-domain characteristics, frequency-domain characteristics, and time-frequency-domain characteristics. In practical evaluation, a 4×4 diagonal pre-whitening matrix is ​​constructed based on the time-domain variance and frequency-domain power spectrum of electrical noise, and pre-whitening is performed frame by frame on the multi-channel acoustic measurement signals. ;in For the pre-whitening matrix, For preprocessing signals, For multi-channel acoustic measurement signals, FFT is performed frame-by-frame on the preprocessed signal to obtain the observed power spectrum, where the frequency resolution Δf = fs / 1024 ≈ 46.9 Hz. Using the electrical noise power spectrum as the noise floor, the signal-to-noise ratio is calculated point-by-point. ;in To observe the power spectrum, Using a noise floor, the average SNR across the entire frequency band of a single frame is taken as the criterion for determining the current frame state. The average SNR for frames 120–350 is 6.2–8.5 dB, indicating a high reliability state. The average SNR for the background interference segment in frames 1–80 and 800–937 is 1.3–3.7 dB, indicating a low reliability state. The preprocessed signal for the high reliability frames undergoes low-rank sparse decomposition using singular values. The first two principal singular values ​​are retained to form a low-rank signal subspace. Sparse noise components are removed, and high-dimensional time-frequency features are extracted based on the STFT amplitude spectrum to obtain a 513-dimensional phase spectrum from 0 to fs / 2. MFCC, which is 13-dimensional and 13-dimensional first-order difference, has a single-frame high-dimensional feature dimension of 513+513+26=1052 dimensions. The preprocessed signal of the low-reliability frame is decomposed into a low-rank decomposition with dimensionality reduction, retaining only the first principal component. The Hilbert envelope mean is extracted, the envelope variance is 2-dimensional, the zero-crossing rate in the time domain is 1-dimensional, the power spectrum centroid and spectral flatness are 2-dimensional, and the single-frame low-dimensional feature dimension is 5-dimensional. The 1052-dimensional high-dimensional features of the high-reliability frame are concatenated with the 1-dimensional average SNR to obtain 1053-dimensional data features. The 5-dimensional low-dimensional features of the low-reliability frame are concatenated with the 1-dimensional average SNR to obtain 6-dimensional data features. The data features are input into the data-driven decision model, which outputs the weight parameters of the noise state to obtain the decision result, which characterizes the reliability of the acoustic measurement. In the actual evaluation, 10 meta-learning tasks were divided according to the operating temperature of the front-end circuit, including {-10, 0, 10, 20, 30, 40, 50, 60, cold start transient 1, cold start transient 2} from T1 to T10. Each task was split into a support set and a query set in an 8:2 ratio, and the sample noise weights and SNR were labeled. A training task set was constructed based on the difference in the power spectrum of thermal noise in the frequency domain. Based on the training task set, the neural network was optimized and trained in two layers through model-independent meta-learning to obtain a data-driven decision model. The feature encoder of the neural network is a 2-layer MLP. The 053 / 6-dimensional feature mapping is converted into a 64-dimensional embedding. The task adaptation layer adapts to different temperature tasks. The output layer is a 1-dimensional neuron, outputting weight parameters for the noise state suppression strength. When the value approaches 1, the noise is extremely strong and the measurement is unreliable; when it approaches 0, the noise is extremely low and the measurement is reliable. The MAML meta-training inner and outer loops iterate 5 times and 1000 times respectively. The preset residual threshold is 0.05, the meta-regularization coefficient is 0.01, and the learning rate is 0.001. After training, the data-driven decision model's weight prediction error is less than 0.03 for each temperature task, which is satisfactory in the online phase. The model parameters are fixed, and an online support set is constructed using current wind turbine measured data. A two-step gradient iteration is performed, where the current weight prediction error is 0.03, less than the preset residual threshold of 0.05, the indicator function is 0, and meta-regularization is not enabled. Scene adaptive parameters are obtained, and forward inference is performed on the input data features based on these parameters. The observed weight parameter values ​​are 0.12 in frame 150, 0.87 in frame 20, and 0.51 in frame 851. These weight parameters are used as observation samples, and a sequential probability ratio test is performed. A high-reliability model is fitted based on the historical weight distribution. The probability density function of the high reliability and low reliability states is calculated, with an upper threshold of 19 and a lower threshold of 0.0526. The SPRT preset false alarm rate is 0.05 and the preset false negative rate is 0.05. The maximum number of observation frames is set to 20 frames. The likelihood ratio of the current cumulative observation is calculated based on the observation samples. The likelihood ratio of the 150th frame is 0.02, which is less than the lower threshold, so the decision is high reliability. The likelihood ratio of the 20th frame is 27.3, which is greater than the upper threshold, so the decision is low reliability. The likelihood ratio of the 850th frame is between the upper and lower thresholds. The previous decision is maintained, and the next frame is observed. Based on the judgment result and weight parameters, noise de-embedding is performed on the multi-channel acoustic measurement signal to reconstruct the acoustic signal, obtain the measurement result with reduced equivalent noise lower limit, and output the optimized ultra-low noise acoustic measurement result.

[0017] In the actual evaluation, a weighted fusion Wiener filter was applied to the 4-channel signal of frame 150 in the high-reliability test based on the weight parameters. The filter transfer function is: The process involves weakly suppressing noise while retaining all effective signal spectral components. The fused and filtered signal is reconstructed using inverse short-time Fourier transform to recover the time-domain waveform and obtain an optimized acoustic signal. Only a small amount of electrical noise is removed. For the 20th frame in the low-reliability test, the signal is projected onto the orthogonal complement space of the electrical noise characteristic parameters using weighted parameters. The projected signal undergoes low-rank SVD decomposition, retaining the first two principal singular values ​​and removing residual sparse interference to obtain a conservatively estimated clean acoustic signal. The reconstructed frame signals are then superimposed with 50% OLA to eliminate frame aliasing, and spliced ​​to obtain a continuous time-domain acoustic signal, completing the entire noise de-embedding and signal reconstruction process. The equivalent noise power spectral density (PSD) is calculated, where the lower limit of the equivalent noise of the original measured signal is [value missing]. The lower limit of the equivalent noise of the optimized measurement signal is: The noise reduction margin was 4.2 dB, which met the equivalent noise reduction lower limit, and extremely low noise acoustic measurement results were obtained.

[0018] In this embodiment, the method for obtaining the multi-channel acoustic measurement signal includes: Based on the synchronous sampling of acoustic signals excited by the same sound source using acoustic sensors, multiple channels of original acoustic sampling data with time alignment and independent channels are obtained. The multiple channels of original acoustic sampling data are de-DC offset and abnormal sampling points are removed to obtain multi-channel acoustic measurement signals. The number of acoustic sensors is greater than or equal to two.

[0019] In this embodiment, the method for obtaining the electrical noise characteristic parameters includes: The multi-channel acoustic measurement signal is framed and de-meaned to obtain locally stationary data frames with zero mean. The time delay point is determined according to the sampling frequency of the current measurement state of the acoustic sensor. The time delay covariance matrix corresponding to each time delay point is calculated for the locally stationary data frame. A set of time delay covariance matrices containing different time delay points is constructed. The time delay point is an integer multiple of the sampling interval. The time delay point includes short time delay, medium time delay and long time delay. The short time delay is 1 to 5 times the sampling interval. The medium time delay has the same signal period as the center frequency of the target measurement frequency band. The long time delay is greater than the medium time delay. The time delay covariance matrix set is approximately jointly diagonalized to obtain the blind separation matrix of the multi-channel acoustic measurement signal. The inverse of the blind separation matrix is ​​used as the mixing matrix. The acoustic signal array response matrix is ​​calculated based on the arrangement position of the acoustic sensor, the sound velocity and the target measurement frequency band to obtain the theoretical array manifold. The mixing matrix represents the mixing ratio of each independent statistical component in the multi-channel. The theoretical array manifold is a unity gain steering vector. Based on the pure electrical noise data during the system's silent period, a spatial matching detection threshold is obtained using the Neyman-Pearson criterion under a preset false alarm rate constraint. The column vector matching degree is then calculated based on the hybrid matrix and the theoretical array manifold. The formula for calculating the column vector matching degree is as follows: ; in For the first The column vector matching degree of each independent statistical component For the separation matrix, the first Column vectors For the first The total energy of each independent statistical component For the first The electrical noise power estimate of each independent statistical component is greater than zero. For the theoretical array manifold, according to and Perform a correlation scan within the target measurement frequency band to obtain... The largest direction to obtain , The direction of arrival (DOA) It is a very small positive number. It is the conjugate transpose. Equal to the number of array elements; If the column vector matching degree is less than the spatial matching detection threshold, the corresponding independent statistical component is determined to be an electrical noise component. Electrical noise characteristic parameters are obtained based on the time-domain statistics and frequency-domain power spectrum of the electrical noise component. The time-domain statistics include mean, variance, and root mean square.

[0020] In this embodiment, the method for obtaining the data features includes: A pre-whitening matrix is ​​constructed based on electrical noise characteristic parameters. The pre-whitening matrix is ​​used to pre-whiten the multi-channel acoustic measurement signals to obtain a pre-processed signal that suppresses electrical noise related components. The observed power spectrum is calculated based on the pre-processed signal. The signal-to-noise ratio (SNR) at each frequency point is calculated based on the observed power spectrum and the noise power spectrum. The high reliability state and low reliability state are divided according to the SNR of the current measurement state. If the preprocessed signal is in the high reliability state, then the preprocessed signal is subjected to low-rank sparse decomposition, and high-dimensional data features are extracted based on the low-rank signal subspace. The high-dimensional data features include the short-time Fourier transform amplitude spectrum, phase spectrum and Mel frequency cepstral coefficients in the time-frequency domain. The low-rank signal subspace represents the principal components of the target acoustic signal. If the preprocessed signal is in the low reliability state, then the preprocessed signal is decomposed into a low-rank approximation to extract low-dimensional data features, which include Hilbert envelope features, zero-crossing rate features and power spectrum centroid features in the frequency domain. The high-dimensional or low-dimensional data features are cascaded with the signal-to-noise ratio to generate data features for the current measurement state.

[0021] In this embodiment, the method for obtaining the weight parameters includes: The thermal noise power spectrum of the front-end circuit at different operating temperatures is obtained. A training task set is constructed based on the difference in thermal noise power spectrum. The neural network is optimized and trained in two layers through model-independent element learning based on the training task set to obtain a data-driven decision model. The operating temperature includes the temperature change range from cold start to steady-state operation. The neural network includes a feature encoder, a task adaptation layer and an output layer. The feature encoder inputs the data features of the measurement state. The task adaptation layer includes a parameter subset. The output layer outputs the weight parameters of the noise state. During the inner loop of the bi-layer optimization training, gradient descent is performed on the support set data of each training task using a regression loss function to obtain task adaptation parameters. During the outer loop meta-update of the bi-layer optimization training, forward inference is performed on the query set data based on the task adaptation parameters to calculate the query set loss until the maximum number of iterations is reached. Then, the meta-model parameters are updated using gradient descent. The regression loss function is: ; in Let be the regression loss function, and let be the objective function of the inner loop to optimize the deviation between the predicted weights and the true labels. To support the number of samples in the set, For the first Signal-to-noise ratio of each sample For the first The labeled weights of each sample, For the weighted output layer, the input features The predicted weight parameters, The square of the L2 norm; Based on the fixed meta-model parameters during the online application phase, an online support set is generated according to the data features of the current scene, and scene adaptive parameters are obtained through two-step gradient descent. The iterative update formula for the scene adaptive parameters is as follows: ; in Let these be the scene adaptation parameters after the t-th iteration. For the first Scene adaptation parameters in the next iteration For learning rate, For loss function pairs gradient, For the non-regular coefficient, For meta-model parameters, This is an indicator function that outputs 1 if the condition within the parentheses is true, and 0 otherwise. The error between the model's predicted output and the actual labeled values. The preset residual threshold; Based on scene adaptive parameters, forward reasoning is performed on the input data features. According to the data-driven decision model output layer, weight parameters of noise suppression intensity under the current measurement state are generated. The weight parameters are used as observation samples and sequential probability ratio test is performed to obtain the decision result.

[0022] In this embodiment, the method for obtaining the judgment result includes: Based on the historical distribution statistics of the weight parameters during the meta-training phase, probability density functions for high-reliability and low-reliability states are fitted. The likelihood ratio of the current cumulative observations is calculated based on the observed samples. The formula for calculating the likelihood ratio is: ; in For the first The likelihood ratio of frames, For the first Observed values ​​of the weight parameters of the received frame. High reliability status The probability density function value under the following conditions In low reliability state The probability density function value, For the first Frame likelihood ratio, initial value =1; If the likelihood ratio is greater than the upper threshold, the decision result is output as low reliability; if it is less than the lower threshold, the decision result is output as high reliability; if it is between the upper and lower thresholds, the decision result of the previous moment is maintained and the next frame is observed. The formula for calculating the upper threshold is: ; The formula for calculating the lower bound threshold is: ; in This is the upper threshold. This is the lower bound threshold. To preset the false alarm rate, This sets the preset false negative probability.

[0023] In this embodiment, the method for obtaining the extremely low noise acoustic measurement results includes: If the decision result is high reliability, Wiener filtering and weighted fusion of the acoustic measurement signals of each channel are performed based on the weight parameters, and the optimized acoustic signal is obtained by reconstructing it through inverse short-time Fourier transform. If the decision result is low reliability, the signal is projected onto the orthogonal complement space of electrical noise characteristic parameters based on the weight parameters, and a conservative estimate of the acoustic signal is obtained through low-rank approximation reconstruction. The reconstructed acoustic signal is subjected to inverse transformation to recover the time-domain waveform, thereby obtaining optimized ultra-low noise acoustic measurement results, wherein the reduction in the equivalent noise lower limit is greater than or equal to 3 dB.

[0024] A second aspect of the present invention also provides a data-driven, ultra-low-noise acoustic measurement optimization system, comprising: Data acquisition module: used to acquire multi-channel acoustic measurement signals of the same acoustic environment based on acoustic sensors, and to perform in-situ estimation of the electrical noise components of the acoustic sensors and front-end circuits in the multi-channel acoustic measurement signals through second-order blind identification to obtain electrical noise characteristic parameters; Data feature extraction module: used to perform low-rank sparse decomposition based on the multi-channel acoustic measurement signal and electrical noise characteristic parameters to obtain data features of the measurement state, wherein the data features include one or more of time-domain features, frequency-domain features and time-frequency-domain features; Data-driven decision module: used to input the data features into the data-driven decision model, output the weight parameters of the noise state, and obtain the decision result, which characterizes the reliability of the acoustic measurement; Measurement result optimization module: used to perform noise de-embedding on the multi-channel acoustic measurement signal according to the decision result and weight parameters, reconstruct the acoustic signal, obtain measurement results with reduced equivalent noise lower limit, and output optimized ultra-low noise acoustic measurement results.

[0025] The above description is merely an example and illustration of the structure of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the structure of the invention or exceed the scope defined in the claims, all of which should fall within the protection scope of the present invention.

Claims

1. A data-driven based method for optimizing very low noise acoustic measurements, characterized in that, Includes the following steps: Based on the acquisition of multi-channel acoustic measurement signals of the same acoustic environment by acoustic sensors, the electrical noise components of the acoustic sensors and front-end circuits in the multi-channel acoustic measurement signals are estimated in situ through second-order blind identification to obtain electrical noise characteristic parameters. Low-rank sparse decomposition is performed on the multi-channel acoustic measurement signal and electrical noise characteristic parameters to obtain the data characteristics of the measurement state. The data characteristics include one or more of time-domain characteristics, frequency-domain characteristics, and time-frequency-domain characteristics. The data features are input into the data-driven decision model, which outputs the weight parameters of the noise state to obtain the decision result, which characterizes the reliability of the acoustic measurement. Based on the judgment result and weight parameters, noise de-embedding is performed on the multi-channel acoustic measurement signal to reconstruct the acoustic signal, obtain the measurement result with reduced equivalent noise lower limit, and output the optimized ultra-low noise acoustic measurement result.

2. The data-driven based ultra-low noise acoustic measurement optimization method of claim 1, wherein, The method for obtaining the multi-channel acoustic measurement signal includes: Based on the synchronous sampling of acoustic signals excited by the same sound source using acoustic sensors, multiple original acoustic sampling data with time alignment and independent channels are obtained. The multiple original acoustic sampling data are de-DC offset to obtain multi-channel acoustic measurement signals. The number of acoustic sensors is greater than or equal to two.

3. The data-driven based ultra-low noise acoustic measurement optimization method of claim 1, wherein, The method for obtaining the electrical noise characteristic parameters includes: The multi-channel acoustic measurement signal is framed and de-meaned to obtain locally stationary data frames with zero mean. The time delay point is determined according to the sampling frequency of the current measurement state of the acoustic sensor. The time delay covariance matrix corresponding to each time delay point is calculated for the locally stationary data frame. A set of time delay covariance matrices containing different time delay points is constructed. The time delay point is an integer multiple of the sampling interval. The time delay point includes short time delay, medium time delay and long time delay. The short time delay is 1 to 5 times the sampling interval. The medium time delay has the same signal period as the center frequency of the target measurement frequency band. The long time delay is greater than the medium time delay. The time delay covariance matrix set is approximately jointly diagonalized to obtain the blind separation matrix of the multi-channel acoustic measurement signal. The inverse of the blind separation matrix is ​​used as the mixing matrix. The acoustic signal array response matrix is ​​calculated based on the arrangement position of the acoustic sensor, the sound velocity and the target measurement frequency band to obtain the theoretical array manifold. The mixing matrix represents the mixing ratio of each independent statistical component in the multi-channel. The theoretical array manifold is a unity gain steering vector. Based on the pure electrical noise data during the system's silent period, a spatial matching detection threshold is obtained using the Neyman-Pearson criterion under a preset false alarm rate constraint. The column vector matching degree is then calculated based on the hybrid matrix and the theoretical array manifold. The formula for calculating the column vector matching degree is as follows: ； in For the first The column vector matching degree of each independent statistical component For the separation matrix, the first Column vectors For the first The total energy of each independent statistical component For the first The electrical noise power estimate of each independent statistical component is greater than zero. For the theoretical array manifold, according to and Perform a correlation scan within the target measurement frequency band to obtain... The largest direction to obtain , The direction of arrival (DOA) It is a very small positive number. This is the conjugate transpose. Equal to the number of array elements; If the column vector matching degree is less than the spatial matching detection threshold, the corresponding independent statistical component is determined to be an electrical noise component. Electrical noise characteristic parameters are obtained based on the time-domain statistics and frequency-domain power spectrum of the electrical noise component. The time-domain statistics include mean, variance, and root mean square.

4. The data-driven ultra-low noise acoustic measurement optimization method according to claim 1, characterized in that, The method for obtaining the data features includes: A pre-whitening matrix is ​​constructed based on electrical noise characteristic parameters. The pre-whitening matrix is ​​used to pre-whiten the multi-channel acoustic measurement signals to obtain a pre-processed signal that suppresses electrical noise related components. The observed power spectrum is calculated based on the pre-processed signal. The signal-to-noise ratio (SNR) at each frequency point is calculated based on the observed power spectrum and the noise power spectrum. The high reliability state and low reliability state are divided according to the SNR of the current measurement state. If the preprocessed signal is in the high reliability state, then the preprocessed signal is subjected to low-rank sparse decomposition, and high-dimensional data features are extracted based on the low-rank signal subspace. The high-dimensional data features include the short-time Fourier transform amplitude spectrum, phase spectrum and Mel frequency cepstral coefficients in the time-frequency domain. The low-rank signal subspace represents the principal components of the target acoustic signal. If the preprocessed signal is in the low reliability state, then the preprocessed signal is decomposed into a low-rank approximation to extract low-dimensional data features, which include Hilbert envelope features, zero-crossing rate features and power spectrum centroid features in the frequency domain. The high-dimensional or low-dimensional data features are cascaded with the signal-to-noise ratio to generate data features for the current measurement state.

5. The data-driven ultra-low noise acoustic measurement optimization method according to claim 1, characterized in that, The method for obtaining the weight parameters includes: The thermal noise power spectrum of the front-end circuit at different operating temperatures is obtained. A training task set is constructed based on the difference in thermal noise power spectrum. The neural network is optimized and trained in two layers through model-independent element learning based on the training task set to obtain a data-driven decision model. The operating temperature includes the temperature change range from cold start to steady-state operation. The neural network includes a feature encoder, a task adaptation layer and an output layer. The feature encoder inputs the data features of the measurement state. The task adaptation layer includes a parameter subset. The output layer outputs the weight parameters of the noise state. During the inner loop of the bi-layer optimization training, gradient descent is performed on the support set data of each training task using a regression loss function to obtain task adaptation parameters. During the outer loop meta-update of the bi-layer optimization training, forward inference is performed on the query set data based on the task adaptation parameters to calculate the query set loss until the maximum number of iterations is reached. Then, the meta-model parameters are updated using gradient descent. The regression loss function is: ； in Let be the regression loss function, and let be the objective function of the inner loop to optimize the deviation between the predicted weights and the true labels. To support the number of samples in the set, For the first Signal-to-noise ratio of each sample For the first The labeled weights of each sample, For the weighted output layer, the input features The predicted weight parameters, The square of the L2 norm; Based on the fixed meta-model parameters during the online application phase, an online support set is generated according to the data features of the current scene, and scene adaptive parameters are obtained through two-step gradient descent. The iterative update formula for the scene adaptive parameters is as follows: ； in These are the scene adaptation parameters after the t-th iteration. For the first Scene adaptation parameters in the next iteration For learning rate, For loss function pairs gradient, For the non-regular coefficient, For meta-model parameters, This is an indicator function that outputs 1 if the condition within the parentheses is true, and 0 otherwise. The error between the model's predicted output and the actual labeled values. The preset residual threshold; Based on scene adaptive parameters, forward reasoning is performed on the input data features. According to the data-driven decision model output layer, weight parameters of noise suppression intensity under the current measurement state are generated. The weight parameters are used as observation samples and sequential probability ratio test is performed to obtain the decision result.

6. The data-driven ultra-low noise acoustic measurement optimization method according to claim 5, characterized in that, The method for obtaining the judgment result includes: Based on the historical distribution statistics of the weight parameters during the meta-training phase, probability density functions for high-reliability and low-reliability states are fitted. The likelihood ratio of the current cumulative observations is calculated based on the observed samples. The formula for calculating the likelihood ratio is: ； in For the first The likelihood ratio of frames, For the first Observed values ​​of the weight parameters of the received frame. High reliability status The probability density function value under the following conditions In low reliability state The probability density function value, For the first Likelihood ratio of frames, initial value =1; If the likelihood ratio is greater than the upper threshold, the decision result is output as low reliability; if it is less than the lower threshold, the decision result is output as high reliability; if it is between the upper and lower thresholds, the decision result of the previous moment is maintained and the next frame is observed. The formula for calculating the upper threshold is: ； The formula for calculating the lower bound threshold is: ； in This is the upper threshold. This is the lower bound threshold. To preset the false alarm rate, The preset false negative probability.

7. The data-driven ultra-low noise acoustic measurement optimization method according to claim 1, characterized in that, A method for obtaining the aforementioned extremely low noise acoustic measurement results includes: If the decision result is high reliability, Wiener filtering and weighted fusion of the acoustic measurement signals of each channel are performed based on the weight parameters, and the optimized acoustic signal is obtained by reconstructing it through inverse short-time Fourier transform. If the decision result is low reliability, the signal is projected onto the orthogonal complement space of electrical noise characteristic parameters based on the weight parameters, and a conservatively estimated acoustic signal is obtained through low-rank approximation reconstruction. The reconstructed acoustic signal is subjected to inverse transformation to recover the time-domain waveform, thereby obtaining optimized ultra-low noise acoustic measurement results, wherein the reduction in the equivalent noise lower limit is greater than or equal to 3 dB.

8. A data-driven ultra-low noise acoustic measurement optimization system for executing the data-driven ultra-low noise acoustic measurement optimization method according to any one of claims 1 to 7, characterized in that, The system includes: Data acquisition module: used to acquire multi-channel acoustic measurement signals of the same acoustic environment based on acoustic sensors, and to perform in-situ estimation of the electrical noise components of the acoustic sensors and front-end circuits in the multi-channel acoustic measurement signals through second-order blind identification to obtain electrical noise characteristic parameters; Data feature extraction module: used to perform low-rank sparse decomposition based on the multi-channel acoustic measurement signal and electrical noise characteristic parameters to obtain data features of the measurement state, wherein the data features include one or more of time-domain features, frequency-domain features and time-frequency-domain features; Data-driven decision module: used to input the data features into the data-driven decision model, output the weight parameters of the noise state, and obtain the decision result, the decision result representing the reliability of the acoustic measurement; Measurement result optimization module: used to perform noise de-embedding on the multi-channel acoustic measurement signal according to the decision result and weight parameters, reconstruct the acoustic signal, obtain measurement results with reduced equivalent noise lower limit, and output optimized ultra-low noise acoustic measurement results.

Images