A multi-channel speech signal noise reduction method and device

Abstract

The present application relates to a kind of multi-channel speech signal noise reduction method and device, the method includes the multi-channel vibration signal and pure noise signal that distributed optical fiber acoustic sensing system is collected;Multi-channel vibration signal is carried out frame processing, and the frequency domain complex frequency spectrum of each channel per frame is obtained;Spectrum entropy and low-frequency signal-to-noise ratio are determined according to pure noise signal, and target over-reduction coefficient is obtained according to the adaptive adjustment of over-reduction coefficient according to spectrum entropy and low-frequency signal-to-noise ratio;According to target over-reduction coefficient, the spectrum subtraction operation of first amplitude spectrum is carried out, and the phase correction of first phase spectrum is carried out based on the time delay of each channel, to obtain the single-channel complex frequency spectrum of each frame, so that by adaptive adjustment over-reduction coefficient and combined with multi-channel phase correction, non-stationary noise interference is effectively inhibited, while signal phase consistency is maintained;Single-channel complex frequency spectrum is carried out phase optimization and superposition restoration, to obtain the time-domain signal after noise reduction reconstruction, so that by phase optimization and superposition restoration, vibration signal feature fidelity is improved.

Description

Technical Field

[0001] This invention relates to the field of distributed fiber optic acoustic sensing technology, and in particular to a method and apparatus for denoising multi-channel speech signals. Background Technology

[0002] In field applications of distributed fiber acoustic sensing systems (DAS), multi-channel vibration signals collected along optical cables are often subject to interference from various background noises, such as light source phase noise, shot noise, and vibrations from the surrounding environment. This results in a decrease in the overall signal-to-noise ratio of the link, masking the characteristics of weak vibration events, and consequently affecting the event detection sensitivity and false alarm rate.

[0003] Existing noise reduction techniques mainly include fixed-parameter spectral subtraction, bandpass / notch filtering, wavelet thresholding, wavelet packet decomposition, singular spectrum analysis, and the increasingly popular deep learning signal enhancement methods. These methods each have their advantages when dealing with single or stationary noise, but they generally have significant limitations: first, they rely on statistically stationary noise or require a large number of samples for training; second, in scenarios with complex noise components that change rapidly with time and distance, the noise reduction effect is prone to fluctuations, and the algorithm parameters need frequent manual adjustment to adapt to new noise conditions. For example, the classic spectral subtraction algorithm is prone to producing dense "spot" artifacts in the time-distance map under conditions of strong light source phase jitter or shot noise dominance; while deep learning methods are highly dependent on the distribution of training data, and their performance degrades sharply when faced with unseen mixed noise types. Therefore, in DAS scenarios with low signal-to-noise ratios and non-stationary noise, existing technologies struggle to simultaneously achieve both noise suppression effectiveness and fidelity in vibration signal characteristics.

[0004] Therefore, there is an urgent need to propose a multi-channel speech signal noise reduction method and device to solve the technical problem that existing noise reduction technologies are unable to simultaneously achieve noise suppression effect and vibration signal feature fidelity in DAS scenarios with low signal-to-noise ratio and non-stationary noise. Summary of the Invention

[0005] In view of this, it is necessary to provide a multi-channel speech signal noise reduction method and apparatus to solve the technical problem that existing noise reduction techniques are unable to simultaneously achieve noise suppression effect and vibration signal feature fidelity in DAS scenarios with low signal-to-noise ratio and non-stationary noise.

[0006] To address the aforementioned problems, in a first aspect, the present invention provides a multi-channel speech signal noise reduction method, comprising: Acquire multi-channel vibration signals and pure noise signals collected by a distributed fiber optic acoustic sensing system; The multi-channel vibration signal is processed by frame segmentation to obtain the frequency domain complex spectrum of each channel per frame; the frequency domain complex spectrum includes a first amplitude spectrum and a first phase spectrum; The spectral entropy and low-frequency signal-to-noise ratio are determined based on the pure noise signal, and the over-attenuation coefficient is adaptively adjusted based on the spectral entropy and the low-frequency signal-to-noise ratio to obtain the target over-attenuation coefficient. The first amplitude spectrum is subjected to spectral subtraction based on the target over-subtraction coefficient, and the first phase spectrum is subjected to phase correction based on the time delay of each channel to obtain the single-channel complex spectrum of each frame. Phase optimization and superposition restoration are performed on the single-channel complex spectrum to obtain the noise-reconstructed time-domain signal.

[0007] In one possible implementation, the step of performing frame-segmentation processing on the multi-channel vibration signal to obtain the frequency domain complex spectrum of each channel per frame includes: Set the fixed length and frame shift length of the time series for each channel; The vibration signal acquired by each channel is framed and windowed according to the fixed length and the frame shift length to obtain the windowed frame signal of each frame; The frequency domain of the windowed frame signal is calculated based on the Fourier transform and the number of sampling points to obtain the complex frequency spectrum of each frame.

[0008] In one possible implementation, the adaptive adjustment of the over-attenuation coefficient based on the spectral entropy and the low-frequency signal-to-noise ratio to obtain the target over-attenuation coefficient includes: When the low-frequency signal-to-noise ratio is less than the preset signal-to-noise ratio, the target over-attenuation coefficient is determined to be the first over-attenuation coefficient; When the low-frequency signal-to-noise ratio is greater than or equal to the preset signal-to-noise ratio, the target over-attenuation coefficient is determined to be the second over-attenuation coefficient; the first over-attenuation coefficient is greater than the second over-attenuation coefficient.

[0009] In one possible implementation, the step of performing spectral subtraction on the first amplitude spectrum based on the target over-subtraction coefficient and performing phase correction on the first phase spectrum based on the time delay of each channel to obtain the single-channel complex spectrum of each frame includes: Set the spectral base coefficient and estimate the time delay for each channel; Based on the time delay, a reference channel is determined, and the ideal channel phase difference relative to the reference channel is determined. Perform spectral subtraction on the first amplitude spectrum based on the target over-subtraction coefficient and the spectral base coefficient to obtain the noise-reduced net signal amplitude spectrum; The net signal amplitude spectrum of each frame is combined with the first phase spectrum to obtain the frame complex spectrum; The first phase spectrum is iteratively adjusted based on the residual of the frame complex spectrum and the difference between the reference channel and the phase difference of the reference channel to obtain the single-channel complex spectrum of each frame.

[0010] In one possible implementation, after combining the net signal amplitude spectrum with the first phase spectrum to obtain the frame complex spectrum, the method further includes: The phase difference of the frame complex spectrum of adjacent frames is calculated to obtain the adjacent phase difference; The phase difference of the frame complex spectrum of different channels at the same frequency is calculated to obtain the same-frequency phase difference; When the adjacent phase difference and the same-frequency phase difference are greater than the preset phase difference threshold, the preset phase spectrum is combined with the net signal amplitude spectrum to obtain the updated frame complex spectrum; the preset phase spectrum is the average of the phase spectra of adjacent stable frequency bands or the phase spectrum of the previous frame.

[0011] In one possible implementation, the step of iteratively adjusting the first phase spectrum based on the residual of the frame complex spectrum and the difference between the reference channel and the phase difference of the reference channel to obtain the single-channel complex spectrum of each frame includes: The measured phase difference is determined based on the difference between the first phase spectrum of the frame complex spectrum and the phase spectrum of the reference channel; The frequency-related residual is determined based on the difference between the measured phase difference and the reference channel phase difference; Determine whether the frequency-related residual is equal to a preset residual threshold; If not, the first phase spectrum of the frame complex spectrum is iteratively adjusted according to the frequency-related residual until the frequency-related residual is the preset residual threshold. If so, the mean of the frame complex spectrum of all channels is calculated to obtain the single-channel complex spectrum.

[0012] In one possible implementation, the step of performing phase optimization and superposition restoration on the single-channel complex spectrum to obtain the noise-reconstructed time-domain signal includes: Phase optimization of the single-channel complex spectrum is performed based on an iterative algorithm to obtain the time-domain frame of each frame; The temporal frames of each frame are superimposed and restored according to the fixed length and the frame shift length to obtain the noise-reconstructed temporal signal.

[0013] In one possible implementation, the step of performing phase optimization on the single-channel complex spectrum based on an iterative algorithm to obtain the time-domain frame of each frame includes: Perform an inverse Fourier transform on the single-channel complex spectrum to obtain the first spectrum; Perform a Fourier transform on the first spectrum to obtain the second spectrum; The amplitude spectrum of the second spectrum is replaced with the amplitude spectrum of the first spectrum to obtain the third spectrum of each frame; Determine whether the second amplitude spectrum and the second phase spectrum of the third spectrum satisfy the convergence condition; If not, then perform inverse Fourier transform and Fourier transform on the second amplitude spectrum again, and smooth the second phase spectrum to obtain the iterated third spectrum; If so, the third spectrum of each frame is adjusted by the iterative algorithm and an inverse Fourier transform is performed to obtain the time-domain frame of each frame.

[0014] In one possible implementation, determining the spectral entropy and low-frequency signal-to-noise ratio based on the pure noise signal includes: The pure noise signal is processed by frame segmentation to obtain the pure noise spectrum of each frame; Based on the pure noise spectrum of all frames, the spectral entropy and low-frequency signal-to-noise ratio are obtained; The step of performing spectral subtraction on the first amplitude spectrum based on the target over-subtraction coefficient and the spectral base coefficient to obtain the denoised net signal amplitude spectrum includes: When the first amplitude spectrum is greater than the pure noise amplitude spectrum of the pure noise spectrum, the first amplitude spectrum is calculated based on the target over-attenuation coefficient and the pure noise amplitude spectrum to obtain the net signal amplitude spectrum; When the first amplitude spectrum is less than or equal to the pure noise amplitude spectrum, the net signal amplitude spectrum is obtained by multiplying the spectrum base coefficient and the pure noise amplitude spectrum.

[0015] Secondly, the present invention also provides a multi-channel speech signal noise reduction device, comprising: The signal acquisition module is used to acquire multi-channel vibration signals and pure noise signals collected by the distributed fiber optic acoustic sensing system; The frame-segmentation processing module is used to perform frame-segmentation processing on the multi-channel vibration signal to obtain the frequency domain complex spectrum of each channel per frame; the frequency domain complex spectrum includes a first amplitude spectrum and a first phase spectrum; An adaptive adjustment module is used to determine the spectral entropy and low-frequency signal-to-noise ratio based on the pure noise signal, and to adaptively adjust the over-attenuation coefficient based on the spectral entropy and the low-frequency signal-to-noise ratio to obtain the target over-attenuation coefficient. The spectrum adjustment module is used to perform spectrum subtraction on the first amplitude spectrum according to the target over-subtraction coefficient, and to perform phase correction on the first phase spectrum based on the time delay of each channel, so as to obtain the single-channel complex spectrum of each frame. The phase optimization module is used to perform phase optimization and superposition restoration on the single-channel complex spectrum to obtain the noise-reconstructed time-domain signal.

[0016] The beneficial effects of this invention are as follows: It acquires multi-channel vibration signals and pure noise signals collected by a distributed fiber optic acoustic sensing system; it performs frame-by-frame processing on the multi-channel vibration signals to obtain the frequency domain complex spectrum of each frame for each channel; the frequency domain complex spectrum includes a first amplitude spectrum and a first phase spectrum; it determines the spectral entropy and low-frequency signal-to-noise ratio based on the pure noise signal, and adaptively adjusts the over-attenuation coefficient based on the spectral entropy and low-frequency signal-to-noise ratio to obtain the target over-attenuation coefficient; it performs spectral subtraction on the first amplitude spectrum based on the target over-attenuation coefficient, and performs phase correction on the first phase spectrum based on the time delay of each channel to obtain the single-channel complex spectrum of each frame. Thus, by adaptively adjusting the over-attenuation coefficient and combining it with multi-channel phase correction, non-stationary noise interference can be effectively suppressed while maintaining signal phase consistency. Furthermore, phase optimization and superposition restoration are performed on the single-channel complex spectrum to obtain the time-domain signal after noise reduction and reconstruction. Thus, phase optimization and superposition restoration can improve the characteristic fidelity of the vibration signal. Attached Figure Description

[0017] Figure 1 A schematic flowchart of an embodiment of the multi-channel speech signal noise reduction method provided by the present invention; Figure 2 For the present invention Figure 1 A schematic flowchart of an embodiment of step S104; Figure 3 For the present invention Figure 2 A schematic flowchart of an embodiment of step S204; Figure 4 A schematic diagram of an embodiment of phase optimization provided by the present invention; Figure 5 This is a schematic diagram of an embodiment of the multi-channel speech signal noise reduction device provided by the present invention. Detailed Implementation

[0018] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.

[0019] like Figure 1 As shown in the figure, a specific embodiment of the present invention discloses a multi-channel speech signal noise reduction method, comprising: S101. Acquire multi-channel vibration signals and pure noise signals collected by the distributed fiber optic acoustic sensing system.

[0020] This technology can be applied to pipeline safety monitoring, such as natural gas pipeline monitoring. It can process signals collected by distributed fiber optic acoustic sensors installed in the pipeline. Through the multi-channel acquisition capability of the distributed fiber optic acoustic sensing system, it can acquire multi-channel vibration signals and pure noise signals. The multi-channel vibration signals are those collected when the target signal appears, while the pure noise signals are a segment of pure noise collected before the target signal appears, excluding the target signal. Assuming there is... Multiple synchronously acquired fiber optic sensing channels (e.g., dividing the sensing fiber into multiple sensing units, each unit independently sampling and transmitting signals), with a sampling frequency of [missing information]. The real-time acoustic signals acquired by each channel are represented as follows: ,in = 1,2,3,..., Channel number, This is the time sampling index.

[0021] S102. Perform frame-segmentation processing on the multi-channel vibration signal to obtain the frequency domain complex spectrum of each channel per frame; the frequency domain complex spectrum includes the first amplitude spectrum and the first phase spectrum.

[0022] Framing refers to dividing a continuous signal into short data segments of fixed length. Specifically, Hanning or Hamming windows can be used for windowing to reduce spectral leakage. The frequency domain complex spectrum contains information from the first amplitude spectrum and the first phase spectrum. Fourier transform converts the time-domain signal into a frequency-domain representation.

[0023] S103. Determine the spectral entropy and low-frequency signal-to-noise ratio based on the pure noise signal, and adaptively adjust the over-attenuation coefficient based on the spectral entropy and low-frequency signal-to-noise ratio to obtain the target over-attenuation coefficient.

[0024] Among them, spectral entropy reflects the complexity of noise energy distribution, and low-frequency signal-to-noise ratio characterizes the energy proportion of the signal in the low-frequency band. The combination of the two can dynamically assess the noise intensity.

[0025] S104. Perform spectral subtraction on the first amplitude spectrum based on the target over-subtraction coefficient, and perform phase correction on the first phase spectrum based on the time delay of each channel to obtain the single-channel complex spectrum of each frame.

[0026] The target over-attenuation coefficient is adjusted in real time according to noise conditions, increasing the spectral attenuation intensity at low signal-to-noise ratios and decreasing it at high signal-to-noise ratios to avoid signal distortion. Phase correction utilizes the phase difference between multiple channels to eliminate phase distortion caused by noise; the phase difference of the reference channel can be obtained through ideal models or historical data calibration.

[0027] S105. Perform phase optimization and superposition restoration on the single-channel complex spectrum to obtain the time-domain signal after noise reduction and reconstruction.

[0028] Specifically, the system first acquires multi-channel vibration signals and pure noise samples. After framing and windowing the vibration signals, the amplitude and phase spectra of each frame are obtained through Fourier transform. Based on the pure noise signal, the spectral entropy and low-frequency signal-to-noise ratio are calculated, and different over-attenuation coefficients are selected according to preset thresholds. During spectral attenuation, when the signal amplitude is higher than the noise reference, the noise component is weakened by the over-attenuation coefficient; when it is lower, the residual energy corresponding to the spectral base coefficient is retained. In the phase correction stage, the phase spectrum is iteratively adjusted to eliminate inconsistencies by comparing the phase difference between each channel and the reference phase. Finally, the optimized amplitude and phase spectra are reconstructed into a time-domain signal, and the complete waveform is recovered using the overlap-addition method.

[0029] Compared with existing technologies, this embodiment provides a method for acquiring multi-channel vibration signals and pure noise signals collected by a distributed fiber optic acoustic sensing system; performing frame-by-frame processing on the multi-channel vibration signals to obtain the frequency domain complex spectrum of each channel per frame; the frequency domain complex spectrum includes a first amplitude spectrum and a first phase spectrum; determining the spectral entropy and low-frequency signal-to-noise ratio based on the pure noise signal, and adaptively adjusting the over-attenuation coefficient based on the spectral entropy and low-frequency signal-to-noise ratio to obtain a target over-attenuation coefficient; performing spectral subtraction on the first amplitude spectrum based on the target over-attenuation coefficient, and performing phase correction on the first phase spectrum based on the time delay of each channel to obtain the single-channel complex spectrum of each frame. Thus, by adaptively adjusting the over-attenuation coefficient and combining it with multi-channel phase correction, non-stationary noise interference can be effectively suppressed while maintaining signal phase consistency. Furthermore, phase optimization and superposition restoration are performed on the single-channel complex spectrum to obtain the noise-reconstructed time-domain signal, thereby improving the feature fidelity of the vibration signal through phase optimization and superposition restoration.

[0030] In some embodiments of the present invention, step S102 includes: Set the fixed length of the time series and the frame shift length for each channel.

[0031] The fixed length refers to the number of sampling points in each frame of signal, which can be implemented using 512 or 1024 sampling points. Its function is to provide a unified signal segmentation benchmark for frame processing. The frame shift length refers to the number of overlapping sampling points between two adjacent frames, which can be implemented using 50% or 75% of the fixed length. For example, when the fixed length is 512, the frame shift length can be set to 256, which is used to suppress spectral leakage through inter-frame overlap.

[0032] The vibration signals acquired from each channel are divided into frames and windowed according to the fixed length and frame shift length to obtain the windowed frame signal of each frame.

[0033] Windowing refers to applying a window function to each frame of signal, which can be implemented using a Hanning window or a Hamming window. Its function is to reduce spectral energy leakage caused by framing.

[0034] The frequency domain of the windowed frame signal is calculated based on the Fourier transform and the number of sampling points to obtain the complex frequency spectrum of each frame.

[0035] Among them, the frequency domain complex spectrum refers to the conversion of the time domain signal into a complex form in the frequency domain through Fourier transform. Specifically, it can be implemented using the Fast Fourier Transform algorithm. Its function is to provide a frequency domain representation containing amplitude and phase information for subsequent noise reduction processing.

[0036] Specifically, first, set the fixed length and frame shift length for the time series of each channel. For example, the fixed length can be set to... Each sampling point, the frame shift length can be set to [number] samples. (50% overlap). Subsequently, based on these two parameters, the vibration signals acquired from each channel were processed into frames, and the first... The first channel A windowing function (such as a Hamming window) is applied to the frame data to obtain the windowed frame signal. Next, a Fast Fourier Transform is performed on the windowed frame signal, and the frequency domain resolution is determined based on the number of sampling points, as shown in formula (1): (1) In the formula, This is the discrete frequency index in the frequency domain. From this, the first amplitude spectrum of the signal frame can be obtained. and the first phase spectrum For example, when the number of sampling points is 512, a complex spectrum containing 256 frequency points can be obtained. The amplitude spectrum represents the energy distribution of each frequency component, and the phase spectrum records the phase information of each frequency component, providing complete frequency domain features for subsequent adaptive noise reduction.

[0037] In some embodiments of the present invention, step S103 includes: The pure noise signal is processed by frame segmentation to obtain the pure noise spectrum of each frame; Based on the pure noise spectrum of all frames, the spectral entropy and low-frequency signal-to-noise ratio are obtained.

[0038] Spectral entropy refers to the degree of dispersion of the energy distribution in the noise spectrum, which characterizes the noise complexity. Specifically, it can be achieved by normalizing the energy at each frequency point of the pure noise spectrum and then calculating the information entropy. A higher spectral entropy value indicates a more complex noise component. Low-frequency signal-to-noise ratio (SNR) refers to the ratio of the energy of the useful component to the noise component in the low-frequency band. Specifically, it can be achieved by integrating the energy of the pure noise spectrum and the vibration signal spectrum in the low-frequency region and then calculating the ratio. This is used to assess the degree of low-frequency noise interference.

[0039] Specifically, the process of framing the pure noise signal is the same as the process of framing the multi-channel vibration signal described above. Specific embodiments of the present invention will not be repeated here. When determining the spectral entropy and low-frequency signal-to-noise ratio, the pure noise signal is first framing, for example, using a framing method with a 20-millisecond frame length and a 10-millisecond frame shift. Windowing is used to reduce spectral leakage, thus obtaining the pure noise spectrum for each frame. The pure noise spectrum can include the corresponding amplitude spectrum and phase spectrum. Subsequently, statistical analysis is performed on the pure noise spectrum of each frame to calculate the dispersion of its energy distribution as the spectral entropy. Simultaneously, the energy ratio of the vibration signal to the noise in the low-frequency band is calculated to obtain the low-frequency signal-to-noise ratio, in order to evaluate the noise level and spectral characteristics of the current frame. The spectral entropy is defined as shown in formula (2): (2) In the formula, , indicating the first j The normalized spectral energy distribution of a frame. A spectral entropy close to its maximum indicates a relatively flat spectrum and noise dominance; while a lower spectral entropy indicates sharp peaks in the spectrum and the dominance of the target signal component. The low-frequency signal-to-noise ratio (SNR) can be defined as the ratio of signal energy to noise energy within a certain low-frequency range. For example, the SNR of the current frame can be obtained by calculating the ratio of signal power to estimated noise power within the 0-300Hz frequency band. In summary, through spectral entropy... and low-frequency SNR The combination of these factors can determine the noise environment of the current frame: when higher and A lower value indicates that the frame is severely overwhelmed by noise; conversely, a higher value indicates that the frame is significantly overwhelmed by noise. lower and A higher value indicates that the target signal component is prominent.

[0040] In some embodiments of the present invention, step S103 includes: When the low-frequency signal-to-noise ratio is less than the preset signal-to-noise ratio, the target over-attenuation coefficient is determined as the first over-attenuation coefficient; When the low-frequency signal-to-noise ratio is greater than or equal to the preset signal-to-noise ratio, the target over-attenuation coefficient is determined as the second over-attenuation coefficient; the first over-attenuation coefficient is greater than the second over-attenuation coefficient.

[0041] Low-frequency signal-to-noise ratio (SNR) refers to the energy ratio of the effective component to the noise component of a signal in the low-frequency band. It is specifically calculated by comparing the signal power to the noise power in the low-frequency band, and is used to characterize the interference level of low-frequency noise in the current signal. Preset SNR refers to a pre-defined SNR threshold, which can be determined through experimental data statistics or empirical values, serving as a boundary standard for judging whether to enhance noise reduction intensity. The first over-attenuation coefficient is a larger spectral attenuation coefficient used when the low-frequency SNR is low. It can be implemented using a fixed value or dynamic calculation, and is used to improve noise suppression capability in strong noise environments. The second over-attenuation coefficient is a smaller spectral attenuation coefficient used when the low-frequency SNR is high. It can be generated through linear interpolation or nonlinear functions, and is used to avoid excessive suppression of the effective signal in weak noise environments.

[0042] Specifically, based on the aforementioned frame feature indices, the over-subtraction coefficient of the adaptive spectral subtraction is calculated. Low residual coefficient of spectrum Over-reduction coefficient The noise reduction level should be controlled, and the value should generally not be less than 1. A larger value indicates that more estimated noise is subtracted from the signal spectrum. Spectrum base coefficient This is used to set the lower limit for residual noise, typically a very small positive value (e.g., 0.01) to ensure that the estimated noise is not completely eliminated, thus avoiding artifact noise. In practical applications of distributed fiber optic acoustic sensing systems, the intensity of low-frequency noise components dynamically changes with environmental vibrations, equipment status, and other factors. By calculating the low-frequency signal-to-noise ratio in real time and comparing it with a preset signal-to-noise ratio threshold, different spectral reduction coefficients can be dynamically selected. When a low-frequency signal-to-noise ratio is detected... When the signal-to-noise ratio is low, such as below the preset signal-to-noise ratio, it indicates that the current noise interference is severe, and a larger signal-to-noise ratio is required. To enhance noise reduction, a larger first over-attenuation coefficient is used to perform a stronger spectral subtraction operation on the amplitude spectrum, thereby effectively suppressing low-frequency noise components. Conversely, when low-frequency signal-to-noise ratios are lowered... A higher value, such as above the preset signal-to-noise ratio, indicates weaker noise interference; a smaller value should be selected. To protect the signal and avoid excessive attenuation, a smaller second over-attenuation coefficient is used for spectral subtraction to prevent excessive attenuation of the effective vibration signal. This dynamic adjustment mechanism allows the noise reduction intensity to adaptively match the current noise level. Specifically, it can be dynamically calculated using a function. As shown in formula (3): (3) In the formula, f Increases with increasing noise level. ,and It can be a constant (such as 0.01), or it can be dynamically adjusted slightly according to the noise energy but always remain very small. For example, when <0dB and When approaching pure noise levels, it can be set to... ;when >10dB and At a lower level, it can be set 1.5, in the intermediate case, The values ​​will be linearly interpolated. This adaptive strategy ensures that spectral subtraction is performed with appropriate force in each frame, that is, to sufficiently suppress noise while avoiding excessive reduction of the target signal.

[0043] In some embodiments of the present invention, such as Figure 2 As shown, step S104 includes: S201. Set the spectral base coefficient and estimate the time delay for each channel; based on the time delay, determine the reference channel and the phase difference between the ideal channel and the reference channel.

[0044] The spectral base coefficient is a compensation coefficient used to prevent excessive signal attenuation during spectral subtraction. It can be implemented using a fixed value or a dynamically adjusted value, such as a preset value within the range of 0.1 to 0.3. Its function is to avoid distortion of the speech signal due to excessive suppression during noise reduction. The reference channel phase difference refers to the theoretical value of the phase difference between different channels under ideal conditions. It can be obtained through pre-calibration or calculation based on a sound field propagation model. Its function is to provide a benchmark for phase correction and ensure phase consistency of multi-channel signals.

[0045] Specifically, the spatial coherence of the signal is optimized by utilizing the phase relationship between multiple channels. Assuming the position of the target sound source relative to the distributed fiber optic sensors is fixed, the time delay of its arrival at each sensing channel is... It is certain, there is a time delay. This can be obtained through calibration or algorithmic estimation; the specific process can be set according to actual conditions, and this embodiment of the invention does not impose any limitations. Through time delay... A reference channel can be identified, which can be used as a reference for each channel. The phase spectrum and amplitude spectrum of the reference channel can be used to reference other data, thereby allowing the phase difference between each channel and the reference channel under ideal conditions to be calculated.

[0046] S202. Perform spectral subtraction on the first amplitude spectrum based on the target over-subtraction coefficient and the spectral base coefficient to obtain the net signal amplitude spectrum after noise reduction.

[0047] Among them, based on the determined and Perform spectral subtraction on each frequency of each frame to calculate the amplitude spectrum of the net signal after noise reduction. .

[0048] S203. Combine the net signal amplitude spectrum of each frame with the first phase spectrum to obtain the frame complex spectrum.

[0049] Among them, the net signal amplitude spectrum of each frame With the first phase spectrum By combining these methods, the denoised frame complex spectrum can be initially reconstructed, as shown in formula (4): (4) S204. The first phase spectrum is iteratively adjusted based on the frame complex spectrum and the residual of the difference between the phase differences of the reference channels to obtain the single-channel complex spectrum of each frame.

[0050] The residual refers to the deviation between the measured phase difference and the reference phase difference. It can be obtained through difference calculation or least squares method. Its role is to quantify the magnitude of phase adjustment to guide the iterative process.

[0051] Specifically, in the noise reduction process, noise suppression is first achieved through spectral subtraction on the amplitude spectrum. The target over-subtraction coefficient controls the noise suppression intensity, while the spectral base coefficient compensates for potentially over-suppressed low-energy speech components. Subsequently, the noise-reduced amplitude spectrum is combined with the original phase spectrum to generate a frame complex spectrum. Since phase information is crucial for signal reconstruction quality, a reference channel phase difference is introduced as a correction benchmark. The phase spectrum is iteratively adjusted multiple times by calculating the residual between the measured phase difference and the reference value until the residual converges to a preset threshold. This process eliminates phase deviations between multiple channels caused by noise interference or propagation path differences. Finally, the fused single-channel complex spectrum is obtained by averaging the results.

[0052] In some embodiments of the present invention, step S202 includes: When the first amplitude spectrum is greater than the pure noise amplitude spectrum of the pure noise spectrum, the first amplitude spectrum is calculated based on the target over-attenuation coefficient and the pure noise amplitude spectrum to obtain the net signal amplitude spectrum. When the first amplitude spectrum is less than or equal to the pure noise amplitude spectrum, the net signal amplitude spectrum is obtained by multiplying the spectrum base coefficient and the pure noise amplitude spectrum.

[0053] Specifically, when performing spectral subtraction, different calculation methods are selected based on the comparison between the current frame amplitude spectrum and the noise amplitude spectrum: when the signal amplitude is higher than the noise amplitude, i.e. If the target signal component is considered to be dominant at that frequency, the noise is subtracted by a given coefficient, as shown in formula (5): (5) When the signal amplitude is less than or equal to the noise, i.e. If the value is negative, it indicates that the frequency is submerged in noise. To avoid negative values ​​or severe distortion, the product of the spectral base coefficient and the noise amplitude spectrum is used as the lower limit protection, and the output amplitude is set to the noise amplitude. The multiple, as shown in formula (6): (6) During the calculation process, it should be ensured that For non-negative values, the above piecewise function is actually equivalent to: (7) This spectral reduction process can weaken most of the noise energy in the frequency domain while retaining a certain noise floor to avoid a hollow feeling in the enhanced signal.

[0054] In some embodiments of the present invention, after step S203, the method further includes: The phase difference between adjacent frames is calculated by calculating the phase difference of the frame complex spectrum. The phase difference of the frame complex spectrum of different channels at the same frequency is calculated to obtain the phase difference at the same frequency; When the adjacent phase difference and the same-frequency phase difference are greater than the preset phase difference threshold, the preset phase spectrum is combined with the net signal amplitude spectrum to obtain the updated frame complex spectrum; the preset phase spectrum is the average of the phase spectra of adjacent stable frequency bands or the phase spectrum of the previous frame.

[0055] The adjacent phase difference refers to the phase difference between two adjacent frames of signals at the same frequency point. This can be achieved by calculating the phase angle difference between corresponding frequency points in the complex spectra of two frames, and is used to capture phase abrupt changes in the time dimension. The same-frequency phase difference refers to the phase difference between different channels at the same frequency. This can be calculated by measuring the phase angle difference at the same frequency point in the complex spectra of multiple channels, and is used to identify phase inconsistencies between channels. The preset phase difference threshold is the critical value for judging abnormal phase fluctuations. This can be determined by experimentally analyzing the typical range of phase jumps under noisy conditions, for example, set to π / 4 or π / 3. The preset phase spectrum is the stable reference phase used to replace the abnormal phase. This can be achieved by using the average phase of adjacent stable frequency bands or the phase spectrum of the previous frame. The adjacent stable frequency bands can be obtained through spectral flatness detection.

[0056] Specifically, in the process of reconstructing the denoised spectrum of each frame, this embodiment of the invention introduces a phase stability judgment mechanism to better distinguish between target signal components and noise components. In the time dimension, the phase change rate of corresponding frequency components in adjacent frames can be compared; in the spatial dimension, the phase difference of the same frequency in different channels can be compared to see if it remains constant. If the phase of a certain frequency sub-band changes slowly between adjacent frames (e.g., the phase difference between consecutive frames is close to 0 or fluctuates only within a small range), and is relatively consistent across multiple channels, then the phase of that frequency band is determined to be stable, indicating that there may be steady-state or periodic target signal components (such as the fundamental frequency component of periodic vibration or its harmonics). For such frequency bands, this embodiment of the invention chooses to retain their original phase information, i.e., it uses... Give This approach protects weak but stable signals from being corrupted by subsequent processing. Conversely, if the phase of a frequency exhibits random and rapid changes in time or space, it is considered noise-dominated, and the phase of that frequency band is unreliable. For these unreliable phase components, smoothing or an initial estimate is needed during reconstruction for subsequent iterative optimization. For example, the phase average of adjacent stable frequency bands can be temporarily used to replace the phase of the noise-dominated frequency band, or the phase value of the previous frame can be directly used as a reasonable starting point for subsequent phase iteration optimization. The above strategies ensure that weak but continuous target signal components are not lost due to noise phase interference during noise reduction and lay the foundation for subsequent fine-grained phase correction.

[0057] In some embodiments of the present invention, such as Figure 3 As shown, step S204 includes: S301. Determine the measured phase difference based on the difference between the first phase spectrum of the frame complex spectrum and the phase spectrum of the reference channel.

[0058] The reference channel phase difference refers to the theoretical value of the phase difference between different channels under ideal conditions. It can be pre-calibrated through theoretical models or experimental data to establish a benchmark for phase correction.

[0059] Specifically, this time delay corresponds to the linear phase difference between channels in the frequency domain: for frequency Ideally, the first i The phase difference between the channel and the reference channel should be: The negative sign indicates hysteresis relative to the reference channel signal. However, the first phase spectrum of the actually measured frame complex spectrum... Phase spectrum of the reference channel phase difference It may deviate from the ideal value due to noise interference and initial phase error.

[0060] S302. Determine the frequency-related residual based on the difference between the measured phase difference and the reference channel phase difference.

[0061] Among them, the frequency-dependent residual refers to the deviation between the measured phase difference and the reference phase difference. Specifically, it can be calculated by phase subtraction in the complex domain and is used to quantify the distribution characteristics of the phase error.

[0062] Specifically, this deviation is assessed by calculating the spatial phase residual (i.e., the difference between the actual phase difference across channels and the ideal linear phase difference). If the fixed delay is... Therefore, the phase difference should change linearly with frequency. By comparing the actual phase difference curve with the ideal linear relationship, the frequency-dependent residual can be obtained, as shown in formula (8): (8) S303. Determine whether the frequency-related residual is within the preset residual threshold.

[0063] The preset residual threshold refers to the maximum allowable phase error range, which can be set according to the signal fidelity requirements and is used to control the termination condition of iterative adjustment.

[0064] S304. If not, then the first phase spectrum of the frame complex spectrum is iteratively adjusted according to the frequency-related residual until the frequency-related residual is the preset residual threshold. Specifically, if the frequency-related residual is not equal to the preset residual threshold, i.e. This indicates that the phase at that frequency is not yet aligned. Based on this, the embodiments of the present invention design an iterative algorithm: for each frequency sub-band, adjust the phase value of each channel at that frequency along the gradient direction of the phase residual. For example, the update formula (9) is shown: (9) In the formula, The appropriate step size is chosen to adjust and gradually approximate the phase consistency condition. After multiple iterations, the phase consistency condition in the main frequency band can be achieved. The convergence to 0 or very small values ​​achieves phase consistency at the sub-band level. This is equivalent to fine-tuning the arrival time difference of each frequency component in each channel to align it with the propagation delay model of the common signal source. After this step, the target signal at each frequency in each channel is phase-aligned, allowing for coherent synthesis of the processed spectra from each channel.

[0065] S305. If so, calculate the mean of the frame complex spectrum of all channels to obtain the single-channel complex spectrum.

[0066] Among them, all channels can be M The average of the complex spectrum is shown in formula (10): (10) This is the synthesized single-channel complex spectrum. Since the target acoustic signal is phase-aligned across channels, its amplitude will be significantly enhanced after superposition, while environmental noise, due to the uncorrelated phases between channels, will be attenuated by average after superposition. This results in... It has a higher signal-to-noise ratio in the frequency domain.

[0067] Specifically, during phase correction, the measured phase difference is first compared with the reference phase difference to calculate the frequency-dependent residual. If the residual exceeds a preset threshold, the phase spectrum of the current frame is iteratively corrected, for example, by adjusting the phase angle using gradient descent, until the residual converges to within the threshold range. Once the phase errors of all channels meet the conditions, the frame complex spectrum of each channel is weighted and averaged to eliminate random phase noise between channels, ultimately outputting a single-channel complex spectrum. This process avoids overcorrection or undercorrection caused by a fixed threshold by dynamically adjusting the phase error boundary.

[0068] In some embodiments of the present invention, step S105 includes: Phase optimization of the single-channel complex spectrum is performed based on an iterative algorithm to obtain the time-domain frame of each frame.

[0069] The iterative algorithm refers to a computational method that alternates between multiple Fourier transforms and amplitude spectrum constraints. Specifically, it can be implemented using the Griffin-Lim algorithm or its improved forms. Through iterative processes, it gradually approximates the optimal phase solution, resolving the problem of missing or distorted phase information. Phase optimization refers to the process of adjusting the phase components of the complex spectrum to conform to the physical constraints of the signal. Specifically, it can be achieved by reconstructing the time-domain signal using inverse Fourier transform and then performing another Fourier analysis, used to eliminate phase distortion introduced during spectral subtraction.

[0070] The temporal frames of each frame are superimposed and restored based on the fixed length and frame shift length to obtain the denoised and reconstructed temporal signal.

[0071] Among them, superposition restoration refers to the process of reassembling the time-domain signal after frame processing along the time axis. Specifically, it can be achieved by superposition addition method. By matching the fixed length and frame shift length, the continuous time-domain waveform is restored, avoiding harmonic distortion introduced by inter-frame jumps.

[0072] Specifically, the phase optimization process first performs an inverse Fourier transform on the single-channel complex spectrum to obtain an initial time-domain frame, and then performs another Fourier transform on this time-domain frame to generate a new spectrum. During this process, the amplitude spectrum of the new spectrum is replaced with the denoised amplitude spectrum, while the phase spectrum retains the calculated result. This process is iterated multiple times, and phase optimization is considered complete when the amplitude and phase changes of the spectrum are below preset thresholds. In the superposition and restoration stage, based on the fixed length and frame shift length parameters set during frame processing, and according to the overlapping method, the time-domain frames are arranged in chronological order, and the overlapping areas are weighted and averaged to finally generate a continuous time-domain signal without phase jumps. .

[0073] In some embodiments of the present invention, such as Figure 4 As shown, phase optimization of the single-channel complex spectrum is performed based on an iterative algorithm to obtain the time-domain frame of each frame, including: S401. Perform an inverse Fourier transform on the single-channel complex spectrum to obtain the first spectrum.

[0074] The inverse Fourier transform refers to the process of converting a complex frequency spectrum into a time-domain signal. Specifically, it can be implemented using the inverse fast Fourier transform algorithm to reconstruct time-domain waveforms.

[0075] S402. Perform a Fourier transform on the first spectrum to obtain the second spectrum; S403. Replace the amplitude spectrum of the second spectrum with the first amplitude spectrum to obtain the third spectrum of each frame; S404. Determine whether the second amplitude spectrum and the second phase spectrum of the third spectrum meet the convergence condition.

[0076] The convergence condition refers to the threshold standard for judging whether the phase optimization has reached a stable state. Specifically, it can be achieved by the root mean square value of the phase difference between adjacent iterations being less than a preset threshold, which is used to control the termination time of phase adjustment.

[0077] S405. If not, perform inverse Fourier transform and Fourier transform on the second amplitude spectrum again, and smooth the second phase spectrum to obtain the iterated third spectrum. S406. If so, the third spectrum of each frame is adjusted by an iterative algorithm and an inverse Fourier transform is performed to obtain the time-domain frame of each frame.

[0078] The iterative algorithm refers to an optimization method that gradually approximates the ideal phase characteristics through multiple transformations and corrections. Specifically, it can be implemented using the Griffin-Lim algorithm or its improved versions to eliminate signal distortion caused by abrupt phase changes. Smoothing refers to the operation of suppressing high-frequency fluctuations in the phase spectrum. Specifically, it can be implemented using sliding window mean filtering or low-pass filtering algorithms to eliminate phase noise generated during the iteration process.

[0079] In this embodiment of the invention, the current frame is synthesized into a spectrum. Inverse FFT yields a new spectrum Then on Perform FFT to obtain a new spectrum In the frequency domain, use the original target amplitude. replace The amplitude is calculated, its phase is preserved, and the updated value is obtained. This process is iterated repeatedly, gradually approximating the desired amplitude spectrum and ensuring a smooth, continuous phase in the reconstructed signal. Convergence is typically achieved after a few dozen iterations, ensuring the reconstructed time-domain waveform matches the target amplitude spectrum as closely as possible. Finally, after adjusting the spectrum of each frame using the Griffin-Lim algorithm, an inverse FFT is performed to obtain the time-domain frame. .

[0080] Specifically, in the phase optimization stage, the single-channel complex spectrum is first converted into a first time-domain spectrum using an inverse Fourier transform. Then, this time-domain signal is subjected to another Fourier transform to obtain a second spectrum. At this point, the amplitude spectrum of the second spectrum is replaced with the original, denoised first amplitude spectrum, forming a new third spectrum. By comparing the phase and amplitude characteristics of the third spectrum to see if they meet the convergence condition, it is determined whether iterative optimization is needed. If the condition is not met, a combination of inverse and forward transform operations is performed on the current spectrum again, while the phase spectrum is smoothed to eliminate abnormal fluctuations. This process is repeated until the phase characteristics stabilize, and finally, the time-domain frame signal is reconstructed from the converged spectrum.

[0081] The adaptive spectral subtraction of frames in this embodiment of the invention: using the spectral entropy and low-frequency signal-to-noise ratio of the vibration signal in each frame as criteria, the over-subtraction factor in the spectral subtraction algorithm is adaptively adjusted. and residual noise figure This enables frame-level noise reduction intensity control. Spectral entropy measures the flatness of the current spectrum, while low-frequency signal-to-noise ratio reflects the noise proportion in the 0-1 kHz segment (or other service-related low-frequency segments). The two together determine the spectral reduction amplitude, which increases the noise reduction intensity in noisy frames and reduces signal distortion in cleaner frames.

[0082] Phase stability assessment and phase spectrum reconstruction: A phase stability assessment mechanism is introduced to analyze the phase characteristics of the short-time speech spectrum. By comparing the phase changes of corresponding frequency bands between adjacent frames or multiple channels, the stability of the frequency band phase is determined. If the phase of a sub-band changes smoothly in time and space, it is considered to contain weak speech components, and this embodiment of the invention retains its original phase to protect weak signals; conversely, for frequency bands with rapid and random phase fluctuations (usually from noise), they are marked as unreliable phases and adjusted or replaced in subsequent processing to guide the reconstruction of the phase spectrum and prevent weak speech signals from being overwhelmed by noise phases.

[0083] Subband phase consistency optimization: A subband phase consistency optimization algorithm is designed, utilizing the spatial phase gradient information of the multi-channel vibration signal and the equivalent delay propagating along the optical cable for phase correction. Specifically, based on the fiber grating length... d Equivalent propagation speed of vibration signal v The theoretical phase difference is calculated, and the actual-theoretical phase residual is obtained within each sub-band. The phase of the corresponding frequency components of each channel is then adjusted iteratively through gradient descent, gradually reducing the residual and bringing the relative phase of each channel within the same sub-band closer to consistency. This process essentially causes the frequency components from the same vibration event to coherently superimpose in space, thereby enhancing the target vibration signal and reducing incoherent noise.

[0084] Griffin-Lim Phase Re-optimization: After achieving sub-band phase consistency, the Griffin-Lim iterative algorithm is executed on the previously marked unreliable phase regions. This step, while keeping the amplitude spectrum unchanged after noise reduction, updates the phase through multiple "time-domain to frequency-domain round trips," further enhancing inter-frame phase continuity and suppressing artifacts, making the reconstructed vibration signal more natural and smooth.

[0085] The beneficial effects of this invention are as follows: By introducing spectral entropy and low-frequency signal-to-noise ratio to adaptively adjust the denoising parameters, and combining phase stability and sub-band phase consistency optimization, this invention can correct the denoising strategy in real time under environments where noise statistical characteristics change drastically over time, significantly suppressing various non-stationary background noises. Experiments show that in scenarios with sudden noise changes, its denoising effect is significantly better than that of fixed-parameter spectral subtraction. The average signal-to-noise ratio of the processed signal is significantly improved, weak energy segments are more discernible, and the spatial gain brought by multi-channel information fusion further weakens background noise, achieving higher clarity than single-channel solutions. Because the algorithm simultaneously focuses on amplitude and phase, this invention maintains the correct phase structure while reducing noise, avoiding the "musical noise" artifacts common in traditional spectral subtraction, resulting in better naturalness and spatial sense in the output signal.

[0086] To better implement the multi-channel speech signal denoising method in the embodiments of the present invention, correspondingly, the embodiments of the present invention also provide a multi-channel speech signal denoising device, such as... Figure 5 As shown, the multi-channel speech signal noise reduction device 500 includes: The signal acquisition module 501 is used to acquire multi-channel vibration signals and pure noise signals collected by the distributed fiber optic acoustic sensing system. The frame processing module 502 is used to perform frame processing on the multi-channel vibration signal to obtain the frequency domain complex spectrum of each channel per frame; the frequency domain complex spectrum includes a first amplitude spectrum and a first phase spectrum; The adaptive adjustment module 503 is used to determine the spectral entropy and low-frequency signal-to-noise ratio based on the pure noise signal, and to adaptively adjust the over-attenuation coefficient based on the spectral entropy and low-frequency signal-to-noise ratio to obtain the target over-attenuation coefficient. The spectrum adjustment module 504 is used to perform spectrum subtraction on the first amplitude spectrum according to the target over-subtraction coefficient, and to perform phase correction on the first phase spectrum based on the time delay of each channel, so as to obtain the single-channel complex spectrum of each frame. The phase optimization module 505 is used to perform phase optimization and superposition restoration on the single-channel complex spectrum to obtain the time-domain signal after noise reduction and reconstruction.

[0087] The multi-channel speech signal noise reduction device 500 provided in the above embodiments can realize the technical solutions described in the above multi-channel speech signal noise reduction method embodiments. The specific implementation principles of each module or unit can be found in the corresponding content in the above multi-channel speech signal noise reduction method embodiments, which will not be repeated here.

[0088] The multi-channel speech signal noise reduction method and apparatus provided by the present invention have been described in detail above. Specific examples have been used to illustrate the principle and implementation of the present invention. The description of the above embodiments is only for the purpose of helping to understand the method and core idea of ​​the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation and application scope based on the idea of ​​the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for denoising multi-channel speech signals, characterized in that, include: Acquire multi-channel vibration signals and pure noise signals collected by a distributed fiber optic acoustic sensing system; The multi-channel vibration signal is processed by frame segmentation to obtain the frequency domain complex spectrum of each channel per frame; the frequency domain complex spectrum includes a first amplitude spectrum and a first phase spectrum; The spectral entropy and low-frequency signal-to-noise ratio are determined based on the pure noise signal, and the over-attenuation coefficient is adaptively adjusted based on the spectral entropy and the low-frequency signal-to-noise ratio to obtain the target over-attenuation coefficient. The first amplitude spectrum is subjected to spectral subtraction based on the target over-subtraction coefficient, and the first phase spectrum is subjected to phase correction based on the time delay of each channel to obtain the single-channel complex spectrum of each frame. Phase optimization and superposition restoration are performed on the single-channel complex spectrum to obtain the noise-reconstructed time-domain signal; The step of performing spectral subtraction on the first amplitude spectrum based on the target over-subtraction coefficient, and performing phase correction on the first phase spectrum based on the time delay of each channel to obtain the single-channel complex spectrum of each frame includes: Set the spectral base coefficient and estimate the time delay for each channel; Based on the time delay, a reference channel is determined, and the ideal channel phase difference relative to the reference channel is determined. Perform spectral subtraction on the first amplitude spectrum based on the target over-subtraction coefficient and the spectral base coefficient to obtain the noise-reduced net signal amplitude spectrum; The net signal amplitude spectrum of each frame is combined with the first phase spectrum to obtain the frame complex spectrum; The first phase spectrum is iteratively adjusted based on the residual of the frame complex spectrum and the difference between the reference channel and the phase difference of the reference channel to obtain the single-channel complex spectrum of each frame.

2. The multi-channel speech signal noise reduction method according to claim 1, characterized in that, The step of performing frame-segmentation processing on the multi-channel vibration signal to obtain the frequency domain complex spectrum of each channel per frame includes: Set the fixed length and frame shift length of the time series for each channel; The vibration signal acquired by each channel is framed and windowed according to the fixed length and the frame shift length to obtain the windowed frame signal of each frame; The frequency domain of the windowed frame signal is calculated based on the Fourier transform and the number of sampling points to obtain the complex frequency spectrum of each frame.

3. The multi-channel speech signal noise reduction method according to claim 1, characterized in that, The step of adaptively adjusting the over-attenuation coefficient based on the spectral entropy and the low-frequency signal-to-noise ratio to obtain the target over-attenuation coefficient includes: When the low-frequency signal-to-noise ratio is less than the preset signal-to-noise ratio, the target over-attenuation coefficient is determined to be the first over-attenuation coefficient; When the low-frequency signal-to-noise ratio is greater than or equal to the preset signal-to-noise ratio, the target over-attenuation coefficient is determined to be the second over-attenuation coefficient; the first over-attenuation coefficient is greater than the second over-attenuation coefficient.

4. The multi-channel speech signal noise reduction method according to claim 1, characterized in that, After combining the net signal amplitude spectrum with the first phase spectrum to obtain the frame complex spectrum, the method further includes: The phase difference between adjacent frames is calculated by calculating the phase difference between the frame complex spectra. The phase difference of the frame complex spectrum of different channels at the same frequency is calculated to obtain the same-frequency phase difference; When the adjacent phase difference and the same-frequency phase difference are greater than the preset phase difference threshold, the preset phase spectrum is combined with the net signal amplitude spectrum to obtain the updated frame complex spectrum; the preset phase spectrum is the average of the phase spectra of adjacent stable frequency bands or the phase spectrum of the previous frame.

5. The multi-channel speech signal noise reduction method according to claim 1, characterized in that, The step of iteratively adjusting the first phase spectrum based on the residual of the frame complex spectrum and the difference between the reference channel and the phase difference of the reference channel to obtain the single-channel complex spectrum of each frame includes: The measured phase difference is determined based on the difference between the first phase spectrum of the frame complex spectrum and the phase spectrum of the reference channel; The frequency-related residual is determined based on the difference between the measured phase difference and the reference channel phase difference; Determine whether the frequency-related residual is equal to a preset residual threshold; If not, the first phase spectrum of the frame complex spectrum is iteratively adjusted according to the frequency-related residual until the frequency-related residual is the preset residual threshold. If so, the mean of the frame complex spectrum of all channels is calculated to obtain the single-channel complex spectrum.

6. The multi-channel speech signal noise reduction method according to claim 2, characterized in that, The step of performing phase optimization and superposition restoration on the single-channel complex spectrum to obtain the noise-reconstructed time-domain signal includes: Phase optimization of the single-channel complex spectrum is performed based on an iterative algorithm to obtain the time-domain frame of each frame; The temporal frames of each frame are superimposed and restored according to the fixed length and the frame shift length to obtain the noise-reconstructed temporal signal.

7. The multi-channel speech signal noise reduction method according to claim 6, characterized in that, The step of performing phase optimization on the single-channel complex spectrum based on an iterative algorithm to obtain the time-domain frame of each frame includes: Perform an inverse Fourier transform on the single-channel complex spectrum to obtain the first spectrum; Perform a Fourier transform on the first spectrum to obtain the second spectrum; The amplitude spectrum of the second spectrum is replaced with the amplitude spectrum of the first spectrum to obtain the third spectrum of each frame; Determine whether the second amplitude spectrum and the second phase spectrum of the third spectrum satisfy the convergence condition; If not, then perform inverse Fourier transform and Fourier transform on the second amplitude spectrum again, and smooth the second phase spectrum to obtain the iterated third spectrum; If so, the third spectrum of each frame is adjusted by the iterative algorithm and an inverse Fourier transform is performed to obtain the time-domain frame of each frame.

8. The multi-channel speech signal noise reduction method according to claim 1, characterized in that, Determining the spectral entropy and low-frequency signal-to-noise ratio based on the pure noise signal includes: The pure noise signal is processed by frame segmentation to obtain the pure noise spectrum of each frame; Based on the pure noise spectrum of all frames, the spectral entropy and low-frequency signal-to-noise ratio are obtained; The step of performing spectral subtraction on the first amplitude spectrum based on the target over-subtraction coefficient and the spectral base coefficient to obtain the denoised net signal amplitude spectrum includes: When the first amplitude spectrum is greater than the pure noise amplitude spectrum of the pure noise spectrum, the first amplitude spectrum is calculated based on the target over-attenuation coefficient and the pure noise amplitude spectrum to obtain the net signal amplitude spectrum; When the first amplitude spectrum is less than or equal to the pure noise amplitude spectrum, the net signal amplitude spectrum is obtained by multiplying the spectrum base coefficient and the pure noise amplitude spectrum.

9. A multi-channel speech signal noise reduction device, characterized in that, The method for performing the multi-channel speech signal noise reduction method according to any one of claims 1 to 8 includes: The signal acquisition module is used to acquire multi-channel vibration signals and pure noise signals collected by the distributed fiber optic acoustic sensing system; The frame-segmentation processing module is used to perform frame-segmentation processing on the multi-channel vibration signal to obtain the frequency domain complex spectrum of each channel per frame; the frequency domain complex spectrum includes a first amplitude spectrum and a first phase spectrum; An adaptive adjustment module is used to determine the spectral entropy and low-frequency signal-to-noise ratio based on the pure noise signal, and to adaptively adjust the over-attenuation coefficient based on the spectral entropy and the low-frequency signal-to-noise ratio to obtain the target over-attenuation coefficient. The spectrum adjustment module is used to perform spectrum subtraction on the first amplitude spectrum according to the target over-subtraction coefficient, and to perform phase correction on the first phase spectrum based on the time delay of each channel, so as to obtain the single-channel complex spectrum of each frame. The phase optimization module is used to perform phase optimization and superposition restoration on the single-channel complex spectrum to obtain the time-domain signal after noise reduction and reconstruction.

Images