A method for suppressing electric clipper noise based on multi-source data fusion processing

Abstract

This invention relates to the field of signal processing and noise suppression technology, and discloses a noise suppression method for electric hair clippers based on multi-source data fusion processing. A multi-channel microphone is arranged on the outer shell of the electric hair clipper for synchronous sampling to acquire the time-domain signal of the coverage area. The signal is then divided into frames according to frame length and frame shift, and windowed to obtain the intra-frame sequence for each channel. Normalized channel weights are calculated based on the time-domain energy of each frame. A discrete Fourier transform is performed on the sequence to obtain the frequency domain complex spectrum and spectral power. The noise power spectrum is estimated frame by frame-by-frame initialization and recursive estimation. The net spectral power of each channel is calculated and fused according to weights to generate a suppressed spectral distribution. The main channel is selected according to weights to extract the phase, and the noise-suppressed frequency domain complex spectrum is reconstructed. Finally, a frame-level inverse transform and overlapping summation are performed to obtain the output signal. This method overcomes the problems of single-channel noise suppression relying on single-point information, difficulty in tracking mechanical noise, and susceptibility to distortion, thereby improving noise suppression performance while preserving speech details.

Description

Technical Field

[0001] This invention relates to the field of signal processing and noise suppression technology, specifically to a method for suppressing noise in electric clipper based on multi-source data fusion processing. Background Technology

[0002] In modern electroacoustic equipment, noise suppression technology is a key factor in improving user experience and equipment performance, especially in noisy environments. Taking electric hair clippers as an example, the noise generated by the motor's rotation and the blades' cutting action during operation often affects the user experience, particularly in situations requiring a quiet environment, such as during hair drying, where users tend to minimize background noise interference. Therefore, effectively suppressing this noise and improving equipment comfort has become an important issue in the design of electroacoustic equipment.

[0003] Most existing noise suppression techniques are based on single-channel signal processing, such as using traditional low-pass or high-pass filters and noise gating algorithms. The core idea of ​​these methods is to enhance signal quality by filtering out high-frequency or low-frequency noise. However, the noise suppression effect of a single channel is often limited by the sound source environment of the device and cannot effectively distinguish between useful and noise signals. For example, when the motor and blades of an electric clipper operate at high frequencies, single-channel noise filtering may not accurately eliminate the noise generated by the blades, and may misinterpret and suppress user speech as noise. To address this issue, recent research has attempted to introduce multi-channel signal processing methods, using multiple microphones to collect noise from different angles, thereby obtaining multi-source audio data for fusion processing. This approach allows for better capture and precise suppression of noise signals from different sources. For example, the noise from the motor and blades is mainly concentrated within a certain frequency range, while the user's speech signal is located in different frequency bands. Through reasonable frequency domain analysis and weighted fusion, noise and speech signals can be separated more accurately, achieving a more efficient noise suppression effect. However, existing multi-channel noise suppression methods still have certain limitations, especially in dynamically changing environments. Existing multi-channel noise suppression methods mostly rely on static spectrum estimation or constant noise models. This makes it difficult to fully improve noise suppression performance when the characteristics of the noise source are unstable or variable. For example, when the operating conditions of the electric clipper change (such as changes in blade speed or motor load), existing technologies may not be able to adjust the noise suppression strategy in time, leading to a decrease in noise suppression performance and even the loss of some useful signals.

[0004] Therefore, this paper aims to propose a noise suppression method for electric hair clippers based on multi-source data fusion processing. First, multiple microphone acquisition channels are arranged on the outer shell of the electric hair clipper to simultaneously acquire multi-channel discrete time-domain signals. Then, a windowed intra-frame sampled value sequence is obtained through frame division and window function weighting. Next, channel weighting factors are constructed based on the time-domain energy distribution to dynamically evaluate the contribution of each channel signal. Subsequently, a discrete Fourier transform is performed on the windowed sequence to extract the frequency domain complex spectrum and calculate the spectral power. Based on this, the noise power spectrum is estimated through initialization and recursive algorithms. The estimated values ​​and weighting factors are used to calculate and fuse the net spectral power of each channel to generate a dual-channel fused suppressed spectral power distribution. Then, the phase of the main channel is selected and the frequency domain complex spectrum is reconstructed. Finally, a frame-level inverse transform is performed, and the signals are overlapped and spliced ​​to output the suppressed time-domain signal. Summary of the Invention

[0005] This invention provides a method for suppressing electric clipper noise based on multi-source data fusion processing, which helps to solve the problems mentioned in the background art.

[0006] This invention provides the following technical solution: a method for suppressing electric clipper noise based on multi-source data fusion processing, comprising:

[0007] Multiple microphone sensors with acquisition channels are arranged on the surface of the electric clipper housing to synchronously sample the acoustic signals generated during the operation of the equipment, thereby acquiring multi-channel original discrete time-domain signals covering the space near the motor cavity and the cutter head.

[0008] The original discrete time-domain signals of each acquisition channel are divided into frames according to the preset frame length and frame shift. The sampling points in each frame are weighted by a window function to obtain the windowed intra-frame sampling value sequence of each acquisition channel.

[0009] Based on the energy distribution of the windowed intra-frame sampled value sequence of each acquisition channel in the time domain, the time domain energy of each acquisition channel in each frame is calculated, and a channel weight factor that satisfies the normalization constraint in each frame is constructed.

[0010] Perform a discrete Fourier transform on the windowed intra-frame sampled value sequence of each acquisition channel to obtain the frequency domain complex spectrum of each frame at discrete frequency points;

[0011] Based on the preset number of noise estimation frames, the spectral power of each discrete frequency point in each frame of each acquisition channel is initialized and recursively updated to obtain a set of noise power spectral estimates for each frame.

[0012] The net spectral power of each acquisition channel at each discrete frequency point in each frame is calculated based on the noise power spectrum estimate; the net spectral power of each acquisition channel is weighted and fused using a channel weighting factor to generate a dual-channel fused suppressed spectral power distribution.

[0013] The main channel of each frame is selected based on the channel weighting factor. Combined with the dual-channel fusion suppression spectrum power distribution, the main channel phase extraction and noise suppression frequency domain complex spectrum reconstruction are completed.

[0014] Perform a frame-level inverse discrete Fourier transform on the noise-suppressed frequency domain complex spectrum, and then overlap and splice the reconstructed time-domain signals of each frame according to the frame shift relationship to obtain the complete noise-suppressed output time-domain signal.

[0015] Optionally, the microphone sensor with multiple acquisition channels arranged on the surface of the electric hair clipper housing, synchronously samples the acoustic signals generated during the operation of the equipment to obtain multi-channel raw discrete-time signals covering the space near the motor cavity and the cutter head, specifically including:

[0016] A first microphone and a second microphone are installed at equal intervals along a preset direction on the outer surface of the electric clipper housing. The first microphone is located on the outer surface of the motor cavity housing, and the second microphone is located on the housing directly in front of the cutter head. The first microphone and the second microphone constitute the first acquisition channel and the second acquisition channel, respectively.

[0017] Under the set system sampling frequency, synchronous sampling is performed on the microphone outputs of the first and second acquisition channels. The continuous sampling duration is set, and the total number of discrete sampling points for each acquisition channel is calculated based on the product of the system sampling frequency and the continuous sampling duration. The time interval between two adjacent discrete sampling points is determined according to the reciprocal of the system sampling frequency.

[0018] The time-domain signal values ​​of each discrete sampling point are acquired sequentially in the first acquisition channel according to the sampling order to form a first original discrete time-domain signal sequence. The time-domain signal values ​​of each discrete sampling point are acquired in the second acquisition channel at the same sampling time as the first acquisition channel to form a second original discrete time-domain signal sequence. The two original discrete time-domain signals correspond to the same sampling time position at the same sampling point number.

[0019] Optionally, the original discrete-time signals of each acquisition channel are divided into frames according to a preset frame length and frame shift, and the sampling points within each frame are weighted using a window function to obtain the windowed intra-frame sampling value sequence for each acquisition channel, specifically including:

[0020] Within each acquisition channel, the original discrete-time domain signal is divided into frames according to two positive integer parameters: preset frame length and frame shift. This ensures that the number of sampling points in each frame is equal to the frame length, and the sampling point interval between the starting sampling points of two adjacent frames is equal to the frame shift. Based on the relationship between the total number of sampling points, frame length, and frame shift in each acquisition channel, the total number of frames is calculated in the following order: first, subtract the frame length from the total number of sampling points; then, divide the difference by the frame shift; and finally, add one to the quotient. At the same time, it is ensured that the frame shift is not greater than the frame length and the frame length is not greater than the total number of sampling points in each acquisition channel.

[0021] Within each acquisition channel, starting from the initial sampling point, the index of the initial sampling point of each frame is determined sequentially according to the frame shift interval. In each frame, sampling points equal to the frame length are continuously extracted from the corresponding initial sampling point to form the discrete time domain signal sequence of the corresponding acquisition channel in the corresponding frame. The sampling points in each frame sequence are arranged sequentially according to the local index within the frame.

[0022] Configure a window function for each frame, using a rectangular window as the window function form. Perform a windowing operation on each sampling point within each frame. Under the rectangular window condition, set the window function weighting coefficient of each sampling point to one, and multiply the intra-frame sampled value with the corresponding weighting coefficient to obtain the windowed intra-frame sampled value sequence for each acquisition channel.

[0023] Optionally, the step of calculating the temporal energy of each acquisition channel within each frame based on the energy distribution of the windowed intra-frame sampled value sequence of each acquisition channel in the time domain, and constructing a channel weight factor that satisfies the normalization constraint within each frame, specifically includes:

[0024] Within each acquisition channel, the windowed intra-frame sampled values ​​of each frame are squared one by one and accumulated within the frame. The accumulated result is used as the temporal energy of each acquisition channel in each frame, thus obtaining the temporal energy sequence of the first acquisition channel in each frame and the temporal energy sequence of the second acquisition channel in each frame.

[0025] In each frame, the temporal energy of the first acquisition channel and the second acquisition channel are added together to obtain the sum of the temporal energy of the first acquisition channel and the second acquisition channel in each frame. Under the condition that the sum is greater than zero, the temporal energy of each acquisition channel in each frame is divided by the sum of the temporal energy, which is used as the channel weight factor of each acquisition channel in each frame, so that the sum of the channel weight factors of the first acquisition channel and the second acquisition channel in each frame is equal to one.

[0026] In each frame, when the sum of the temporal energy of the first acquisition channel and the second acquisition channel is equal to zero, the channel weight factors of the first acquisition channel and the second acquisition channel are set to equal values ​​respectively, and the sum of the channel weight factors of the first acquisition channel and the second acquisition channel in each frame is equal to one, forming a channel weight factor sequence that satisfies the normalization condition in all frames.

[0027] Optionally, performing a discrete Fourier transform on the windowed intra-frame sampled value sequence of each acquisition channel to obtain the frequency domain complex spectrum of each frame at discrete frequency points for each acquisition channel specifically includes:

[0028] Within each acquisition channel, a discrete Fourier transform with a length equal to the frame length is performed on the windowed intra-frame sampled value sequence of each frame to obtain the frequency domain complex spectrum of each frame at each discrete frequency point. Each frequency domain complex spectrum has a real component and an imaginary component at each discrete frequency point.

[0029] For each acquisition channel, each frame, and each discrete frequency point, the real and imaginary components of the frequency domain complex spectrum are squared respectively. The two squared results are added together to obtain the spectral power of each acquisition channel at each discrete frequency point in each frame, forming a set of spectral power covering all acquisition channels, all frames, and all discrete frequency points.

[0030] Optionally, the step of initializing and recursively updating the spectral power of each discrete frequency point in each frame of each acquisition channel according to a preset number of noise estimation frames, to obtain a set of noise power spectral estimates frame by frame, specifically includes:

[0031] Set the number of frames to be used for noise estimation initialization. The set number of frames is a positive integer that is not less than one and not greater than the total number of frames.

[0032] Within each acquisition channel, for each discrete frequency point, the spectral power of the first few frames is summed and divided by the number of frames to obtain the initial estimate of the noise power spectrum of each acquisition channel at each discrete frequency point, thus forming the initial estimate set of noise power spectrum of each discrete frequency point of each acquisition channel.

[0033] Within a set number of the first few frames, the noise power spectrum estimate for each frame, each acquisition channel, and each discrete frequency point is uniformly set to the corresponding initial noise power spectrum estimate, forming a set of noise power spectrum estimates on the initialization frame.

[0034] Starting from the next frame after the initialization frame until the last frame, at each acquisition channel and at each discrete frequency point, the noise power spectrum estimate of the previous frame and the spectral power of the current frame are weighted and averaged. The weight ratio is set to half for the noise power spectrum estimate of the previous frame and half for the spectral power of the current frame. This yields new noise power spectrum estimates for the current frame, each acquisition channel, and each discrete frequency point. The recursive update process is then performed sequentially on all subsequent frames.

[0035] Optionally, the step of calculating the net spectral power of each acquisition channel at each discrete frequency point in each frame based on the noise power spectrum estimate, and then weighting and fusing the net spectral power of each acquisition channel using a channel weighting factor to generate a dual-channel fused suppressed spectral power distribution, specifically includes:

[0036] At each acquisition channel, each frame, and each discrete frequency point, the corresponding noise power spectrum estimate is subtracted from the spectral power. When the difference is greater than or equal to zero, the difference is used as the net spectral power of each acquisition channel at each discrete frequency point in each frame. When the difference is less than zero, the net spectral power of each acquisition channel at each discrete frequency point in each frame is set to zero, thus forming the net spectral power distribution of the first and second acquisition channels at each discrete frequency point in each frame.

[0037] At each frame and at each discrete frequency point, the net spectral power of the first and second acquisition channels is weighted and fused using the channel weighting factor corresponding to the current frame. The net spectral power of each acquisition channel is multiplied by the corresponding channel weighting factor and summed between the first and second acquisition channels to obtain the value of the dual-channel fusion suppressed spectral power distribution at each discrete frequency point in each frame.

[0038] Optionally, the step of selecting the main channel of each frame based on the channel weighting factor, and combining the dual-channel fusion suppressed spectral power distribution to complete the main channel phase extraction and noise-suppressed frequency domain complex spectrum reconstruction specifically includes:

[0039] In each frame, the channel weight factors of the first acquisition channel and the second acquisition channel are compared. When the channel weight factor of the first acquisition channel is greater than or equal to the channel weight factor of the second acquisition channel, the first acquisition channel is designated as the main channel of the current frame. When the channel weight factor of the first acquisition channel is less than the channel weight factor of the second acquisition channel, the second acquisition channel is designated as the main channel of the current frame, thus forming a sequence of main channel numbers for each frame.

[0040] At each frame and at each discrete frequency point, the phase angle of the corresponding discrete frequency point is read from the complex spectrum of the frequency domain of the main channel to form the phase spectrum of the main channel of the current frame.

[0041] In each frame and at each discrete frequency point, the square root of the power distribution of the dual-channel fused suppression spectrum at the corresponding discrete frequency point is taken as the amplitude of the frequency domain complex spectrum, and the phase angle of the main channel phase spectrum is taken as the phase angle of the frequency domain complex spectrum. The amplitude and phase angle are combined to form the frequency domain complex spectrum reconstructed after noise suppression, covering all frames and all discrete frequency points.

[0042] Optionally, the step of performing a frame-level inverse discrete Fourier transform on the noise-suppressed frequency domain complex spectrum, and then overlapping and splicing the reconstructed time-domain signals of each frame according to the frame shift relationship to obtain a complete noise-suppressed output time-domain signal, specifically includes:

[0043] In each frame, a frame-level inverse discrete Fourier transform is performed on the reconstructed frequency domain complex spectrum after noise suppression. The transform length is the same as the discrete Fourier transform length, resulting in the reconstructed time-domain signal sample sequence at each sampling point of the corresponding frame.

[0044] The total number of discrete sampling points of the final output signal is calculated based on the relationship between the total number of frames, frame length, and frame shift. The product of the total number of frames minus one and the frame shift is added to the frame length to obtain the discrete sampling length of the output signal.

[0045] For each discrete sampling point location, identify all intra-frame sample values ​​that have a coverage relationship at this discrete sampling point location, and sum these intra-frame sample values ​​to obtain the complete output time-domain signal value for the corresponding discrete sampling point location.

[0046] The present invention has the following beneficial effects:

[0047] 1. Microphone sensors are positioned at key locations within the motor cavity and near the cutter head of the electric hair clipper. Simultaneous sampling through at least two channels achieves comprehensive coverage of the spatial distribution of source noise. Compared to traditional single-channel fixed-position acquisition, multi-channel deployment can capture the noise differences within the mechanical cavity and in front of the cutter head, providing rich data for subsequent multi-source information fusion. Simultaneous sampling technology ensures strict temporal alignment of signals from each channel, avoiding fusion errors caused by sampling offsets. This deployment and synchronization strategy exhibits stronger robustness to non-stationary noise in complex mechanical sound environments. On the one hand, it accurately captures noise components from different locations; on the other hand, it provides a reliable temporal alignment basis for multi-channel weighted fusion, ensuring a sufficient foundation for subsequent temporal energy comparison and weight calculation of each channel.

[0048] 2. For massive data under high sampling rates, a frame partitioning mechanism with preset frame length and frame shift is designed, and a rectangular window function is introduced to weight the sampled values ​​within each frame, making energy calculation and frequency domain analysis more stable. Compared with windowless or simple segmented direct transformation, the window function can reduce spectral leakage and improve frequency domain resolution; the frame shift mechanism achieves data overlap while ensuring real-time performance, providing a smooth transition for subsequent overlapping and splicing. It reduces boundary distortion, is compatible with frame-level parallel computing, and improves the accuracy of signal-to-noise ratio estimation; it also allows the length and overlap of each frame signal sequence to be flexibly adjusted to adapt to the computing power and algorithm latency requirements of different devices. Unlike the fixed-length windowless FFT commonly used in existing technologies, this scheme balances real-time performance and spectral analysis quality in the frame partitioning and window weighting stages, laying a solid foundation for subsequent energy calculation, Fourier transform, and inverse transform.

[0049] 3. A method is proposed that utilizes the energy distribution of windowed signals from each acquisition channel in the time domain for normalization processing, thereby dynamically constructing a channel weight factor sequence for each frame. It directly uses time-domain energy as the fusion basis, eliminating the need for prior acoustic models or complex deep networks, and adaptively evaluates the relative contribution of each channel in each frame. When the total channel energy is zero, it automatically assigns equal weights, ensuring the algorithm's stability. It can adapt to scenarios such as channel occlusion and sensor failure, automatically adjusting the weighting ratio. Simultaneously, the normalization condition ensures energy conservation after fusion, avoiding excessive amplification or suppression and reducing speech distortion.

[0050] 4. The windowed intra-frame sampled value sequence is input into the Discrete Fourier Transform to obtain the frequency domain complex spectrum of each channel and each frame. The spectral power is then obtained by summing the squares of the real and imaginary parts. The frequency domain complex spectrum preserves amplitude and phase information, while the spectral power calculation provides the amplitude measure required for noise estimation and net spectrum calculation. This two-stage extraction mechanism is more complete than simple power spectrum calculation, laying the foundation for subsequent phase reconstruction and frequency domain filtering. Furthermore, point-value processing using discrete frequency indexing enables parallel optimization and real-time hardware acceleration. Compared to existing methods that only use power or amplitude spectra, this scheme balances information integrity and computational convenience, providing reliable data support for accurate noise estimation and effective frequency domain suppression.

[0051] 5. A recursive update mechanism is introduced to initialize the spectral power within a preset frame range and recursively update the frame-level noise power spectrum based on a first-order weighted average. This allows for online estimation of the noise power spectrum without the need for silent frame detection, making it suitable for non-stationary noise environments. The simple arithmetic average with a recursive weight of 1 / 2 is easy to implement and exhibits good smoothness, balancing noise response speed and estimation stability. Compared to static estimation in fixed-window-shift statistics or spectral subtraction, the recursive algorithm can quickly track changes in noise power, reducing the risk of residual noise underestimation or over-suppression, and improving environmental adaptability.

[0052] 6. By combining recursively estimated noise power spectrum with channel weighting factors, the net spectral power of each channel is first calculated, and then weighted and fused according to the weighting factors to generate a dual-channel suppressed spectral power distribution. This approach leverages the differences in signal quality among channels while also considering the bandgap suppression effect brought by noise estimation. The final output fused suppressed spectrum is more accurate and has less residual noise than single-channel filtering. This scheme achieves complementary advantages of multiple channels, improving speech intelligibility and naturalness under mechanical noise.

[0053] 7. In each frame, a main channel is selected based on a channel weighting factor, and its phase spectrum is extracted and combined with the power of the fused suppression spectrum for frequency domain complex spectrum reconstruction. This preserves the true phase information of the main channel, and the reconstructed complex spectrum can better restore the detailed structure of the speech, avoiding distortion and metallic sound after pure amplitude spectrum suppression. Simultaneously, dynamically selecting the main channel ensures that the original phase is preserved in channels with better signal quality, improving the phase reconstruction effect. Compared with existing multi-channel schemes that average the phase or use a fixed phase, this dynamic phase selection scheme can more accurately reproduce speech features, improving speech clarity and naturalness.

[0054] 8. At the frame level, an inverse transform of the same length as the discrete Fourier transform is performed on the reconstructed frequency domain complex spectrum, and the results are then overlapped and concatenated according to the frame shift relationship to obtain the complete output time-domain signal. The intra-frame sample indexing and frame shift concatenation process are strictly defined and synchronized to avoid edge distortion and concatenation gaps, while preserving the temporal continuity of the reconstructed signal. The algorithm achieves end-to-end real-time output, ensures controllable algorithm latency, and maximizes the smoothness and coherence of the processed signal. Unlike existing post-noise reduction schemes that often neglect frame shift synchronization at the overlap addition point, this scheme strictly uses a pulse function to control the concatenation, ensuring the integrity and temporal accuracy of the output signal. Attached Figure Description

[0055] Figure 1 This is a schematic diagram of the process of the present invention. Detailed Implementation

[0056] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0057] Example, refer to Figure 1 A method for suppressing electric clipper noise based on multi-source data fusion processing, comprising:

[0058] Multiple microphone sensors with acquisition channels are arranged on the surface of the electric clipper housing to synchronously sample the acoustic signals generated during the operation of the equipment, thereby acquiring multi-channel original discrete time-domain signals covering the space near the motor cavity and the cutter head.

[0059] The original discrete time-domain signals of each acquisition channel are divided into frames according to the preset frame length and frame shift. The sampling points in each frame are weighted by a window function to obtain the windowed intra-frame sampling value sequence of each acquisition channel.

[0060] Based on the energy distribution of the windowed intra-frame sampled value sequence of each acquisition channel in the time domain, the time domain energy of each acquisition channel in each frame is calculated, and a channel weight factor that satisfies the normalization constraint in each frame is constructed.

[0061] Perform a discrete Fourier transform on the windowed intra-frame sampled value sequence of each acquisition channel to obtain the frequency domain complex spectrum of each frame at discrete frequency points;

[0062] Based on the preset number of noise estimation frames, the spectral power of each discrete frequency point in each frame of each acquisition channel is initialized and recursively updated to obtain a set of noise power spectral estimates for each frame.

[0063] The net spectral power of each acquisition channel at each discrete frequency point in each frame is calculated based on the noise power spectrum estimate; the net spectral power of each acquisition channel is weighted and fused using a channel weighting factor to generate a dual-channel fused suppressed spectral power distribution.

[0064] The main channel of each frame is selected based on the channel weighting factor. Combined with the dual-channel fusion suppression spectrum power distribution, the main channel phase extraction and noise suppression frequency domain complex spectrum reconstruction are completed.

[0065] Perform a frame-level inverse discrete Fourier transform on the noise-suppressed frequency domain complex spectrum, and then overlap and splice the reconstructed time-domain signals of each frame according to the frame shift relationship to obtain the complete noise-suppressed output time-domain signal.

[0066] By arranging multiple microphone acquisition channels on the surface of the electric clipper's casing and sampling synchronously, this method overcomes the limitation of traditional single-channel noise suppression, which relies solely on single-point sound source information. After simultaneously acquiring discrete time-domain signals covering the motor cavity and the area in front of the cutter head through multiple channels, frame partitioning and window function weighting techniques are used to divide the continuous raw signal into multiple segments and eliminate edge effects, thereby making subsequent time-domain energy and frequency-domain analysis more stable and accurate. Based on this, the method automatically calculates channel weight factors based on the time-domain energy distribution of each frame, without the need for prior manual setting or calibration. It can adaptively adjust the fusion ratio of each channel according to changes in noise intensity, ensuring that the optimal input signal can be dynamically acquired even in scenarios with uneven noise distribution during mechanical cutting. Next, the discrete Fourier transform is used to extract the frequency domain complex spectrum and further calculate the spectral power. This method retains both phase information and amplitude characteristics, laying a solid foundation for accurate noise estimation. Subsequently, an initialization and recursive update strategy is used to obtain a set of frame-by-frame noise power spectrum estimates. This adaptive recursive algorithm effectively solves the problem that traditional static noise models cannot quickly track mechanical noise spectrum fluctuations over time. Furthermore, this method combines weighting factors to weight and fuse the net spectral power of each channel, generating a dual-channel fused suppressed spectral power distribution, which effectively improves suppression performance and avoids signal residue and speech distortion caused by single-channel filtering. Finally, through main channel phase extraction and frequency domain complex spectrum reconstruction, followed by frame-level inverse transform and overlapping concatenation, a complete time-domain signal is output. This method ensures the continuity of the processed audio and preserves speech details to the greatest extent.

[0067] The microphone sensor with multiple acquisition channels arranged on the surface of the electric hair clipper's outer shell performs synchronous sampling of the acoustic signals generated during the device's operation, acquiring multi-channel raw discrete-time signals covering the space near the motor cavity and the cutter head. Specifically, this includes:

[0068] A first microphone and a second microphone are installed at equal intervals along a preset direction on the outer surface of the electric clipper housing. The first microphone is located on the outer surface of the motor cavity housing, and the second microphone is located on the housing directly in front of the cutter head. The first microphone and the second microphone constitute the first acquisition channel and the second acquisition channel, respectively.

[0069] Under the set system sampling frequency, synchronous sampling is performed on the microphone outputs of the first and second acquisition channels. The continuous sampling duration is set, and the total number of discrete sampling points for each acquisition channel is calculated based on the product of the system sampling frequency and the continuous sampling duration. The time interval between two adjacent discrete sampling points is determined according to the reciprocal of the system sampling frequency.

[0070] The time-domain signal values ​​of each discrete sampling point are acquired sequentially in the first acquisition channel according to the sampling order to form a first original discrete time-domain signal sequence. The time-domain signal values ​​of each discrete sampling point are acquired in the second acquisition channel at the same sampling time as the first acquisition channel to form a second original discrete time-domain signal sequence. The two original discrete time-domain signals correspond to the same sampling time position at the same sampling point number.

[0071] Two microphone sensors, numbered 1, are equidistantly placed on the outer surface of the electric hair clipper. and ;in, For the acquisition channel number, This indicates the first channel microphone. This indicates the second channel microphone;

[0072] Install microphone 1 on the motor cavity housing and microphone 2 in front of the cutter head;

[0073] Two-channel microphones at sampling frequency Synchronous sampling, sampling time is The sampling interval is The number of sampling points is ;in, The sampling frequency of the system; This refers to the duration of continuous sampling. The time interval between two adjacent discrete sampling points; The total number of discrete sampling points collected on each channel;

[0074] The original discrete signals obtained are represented as follows: , , ;in, Index of discrete-time sampling points; For channel 1 in discrete sampling index The discrete-time signal value at the location; For channel 2 in discrete sampling index The discrete-time signal value at that point.

[0075] The original discrete-time domain signals of each acquisition channel are divided into frames according to a preset frame length and frame shift. A window function is applied to the sampling points within each frame to weight them, resulting in a windowed intra-frame sampling value sequence for each acquisition channel. Specifically, this includes:

[0076] Within each acquisition channel, the original discrete-time domain signal is divided into frames according to two positive integer parameters: preset frame length and frame shift. This ensures that the number of sampling points in each frame is equal to the frame length, and the sampling point interval between the starting sampling points of two adjacent frames is equal to the frame shift. Based on the relationship between the total number of sampling points, frame length, and frame shift in each acquisition channel, the total number of frames is calculated in the following order: first, subtract the frame length from the total number of sampling points; then, divide the difference by the frame shift; and finally, add one to the quotient. At the same time, it is ensured that the frame shift is not greater than the frame length and the frame length is not greater than the total number of sampling points in each acquisition channel.

[0077] Within each acquisition channel, starting from the initial sampling point, the index of the initial sampling point of each frame is determined sequentially according to the frame shift interval. In each frame, sampling points equal to the frame length are continuously extracted from the corresponding initial sampling point to form the discrete time domain signal sequence of the corresponding acquisition channel in the corresponding frame. The sampling points in each frame sequence are arranged sequentially according to the local index within the frame.

[0078] Configure a window function for each frame, using a rectangular window as the window function form. Perform a windowing operation on each sampling point within each frame. Under the rectangular window condition, set the window function weighting coefficient of each sampling point to one, and multiply the intra-frame sampled value with the corresponding weighting coefficient to obtain the windowed intra-frame sampled value sequence for each acquisition channel.

[0079] For the signals of the two channels, according to frame length Frame division, frame shifting Frame number is Frame length With frame shift It is a positive integer and satisfies: ;in, The number of sampling points contained in each frame; The sampling interval between the start points of two adjacent frames; Assign frame number; Total number of frames;

[0080] Calculate the first The starting sampling point of the frame is: ;in, For the first The index of the starting sampling point of the frame on the global discrete sampling axis;

[0081] The corresponding frame signal is extracted as follows: , ;in, For local sampling point indices within the frame; For the first Channel 1 In-frame index is The sampled values;

[0082] Introducing window functions Set as a rectangular window The signal after windowing is represented as:

[0083] ;in, For the first frame The window function weighting coefficients for each point; For the first Channel 1 Intra-frame sampled values ​​after frame windowing.

[0084] The method involves calculating the temporal energy of each acquisition channel within each frame based on the energy distribution of the windowed intra-frame sampled value sequence of each acquisition channel in the time domain, and constructing a channel weight factor that satisfies normalization constraints within each frame. Specifically, this includes:

[0085] Within each acquisition channel, the windowed intra-frame sampled values ​​of each frame are squared one by one and accumulated within the frame. The accumulated result is used as the temporal energy of each acquisition channel in each frame, thus obtaining the temporal energy sequence of the first acquisition channel in each frame and the temporal energy sequence of the second acquisition channel in each frame.

[0086] In each frame, the temporal energy of the first acquisition channel and the second acquisition channel are added together to obtain the sum of the temporal energy of the first acquisition channel and the second acquisition channel in each frame. Under the condition that the sum is greater than zero, the temporal energy of each acquisition channel in each frame is divided by the sum of the temporal energy, which is used as the channel weight factor of each acquisition channel in each frame, so that the sum of the channel weight factors of the first acquisition channel and the second acquisition channel in each frame is equal to one.

[0087] In each frame, when the sum of the temporal energy of the first acquisition channel and the second acquisition channel is equal to zero, the channel weight factors of the first acquisition channel and the second acquisition channel are set to equal values ​​respectively, and the sum of the channel weight factors of the first acquisition channel and the second acquisition channel in each frame is equal to one, forming a channel weight factor sequence that satisfies the normalization condition in all frames.

[0088] For the Frame, First The frame energy is calculated from the channel signal as follows:

[0089] ;in, For the first Channel 1 Temporal energy of a frame;

[0090] Calculate the first The sum of the time-domain energy of the two channels of the frame ;

[0091] when At that time, the channel weighting factor is constructed as follows: , ;in, For the first Channel 1 Frame channel weighting factor;

[0092] when At that time, let the channel weighting factor be: , ;

[0093] In both of the above situations: .

[0094] The step of performing a discrete Fourier transform on the windowed intra-frame sampled value sequence of each acquisition channel to obtain the frequency domain complex spectrum of each frame at discrete frequency points specifically includes:

[0095] Within each acquisition channel, a discrete Fourier transform with a length equal to the frame length is performed on the windowed intra-frame sampled value sequence of each frame to obtain the frequency domain complex spectrum of each frame at each discrete frequency point. Each frequency domain complex spectrum has a real component and an imaginary component at each discrete frequency point.

[0096] For each acquisition channel, each frame, and each discrete frequency point, the real and imaginary components of the frequency domain complex spectrum are squared respectively. The two squared results are added together to obtain the spectral power of each acquisition channel at each discrete frequency point in each frame, forming a set of spectral power covering all acquisition channels, all frames, and all discrete frequency points.

[0097] For windowed frame signals implement The point-discrete Fourier transform yields the frequency domain complex spectrum as follows:

[0098] , ;in, For the first Channel 1 Frame in frequency index The discrete Fourier transform result at the location; Frequency index; The imaginary unit; It is a natural exponential function;

[0099] The square of the amplitude of the frequency component is calculated as follows:

[0100] ;in, For the first Channel 1 Frame in frequency index The square of the power spectral amplitude at that point; To take the real part of a complex number; This is a function that takes the imaginary part of a complex number.

[0101] The step involves initializing and recursively updating the spectral power at each discrete frequency point in each frame of each acquisition channel according to a preset number of noise estimation frames, to obtain a set of noise power spectral estimates for each frame. Specifically, this includes:

[0102] Set the number of frames to be used for noise estimation initialization. The set number of frames is a positive integer that is not less than one and not greater than the total number of frames.

[0103] Within each acquisition channel, for each discrete frequency point, the spectral power of the first few frames is summed and divided by the number of frames to obtain the initial estimate of the noise power spectrum of each acquisition channel at each discrete frequency point, thus forming the initial estimate set of noise power spectrum of each discrete frequency point of each acquisition channel.

[0104] Within a set number of the first few frames, the noise power spectrum estimate for each frame, each acquisition channel, and each discrete frequency point is uniformly set to the corresponding initial noise power spectrum estimate, forming a set of noise power spectrum estimates on the initialization frame.

[0105] Starting from the next frame after the initialization frame until the last frame, at each acquisition channel and at each discrete frequency point, the noise power spectrum estimate of the previous frame and the spectral power of the current frame are weighted and averaged. The weight ratio is set to half for the noise power spectrum estimate of the previous frame and half for the spectral power of the current frame. This yields new noise power spectrum estimates for the current frame, each acquisition channel, and each discrete frequency point. The recursive update process is then performed sequentially on all subsequent frames.

[0106] Set the number of frames used to initialize noise estimation. ,satisfy: ;

[0107] Before use The initial noise spectrum value for each channel is calculated for each frame, with each channel numbered... and each frequency index ,make: ;in, For in front On the frame for the first Channel in frequency index The initial estimate of the noise power spectrum obtained from the calculation;

[0108] For all And all and Assignment: ;in, For the first Channel 1 Frame, Frequency Index The noise power spectrum estimate at the location;

[0109] right The current noise spectrum estimate is updated using recursive values ​​for each channel number. and each frequency index ,calculate:

[0110] .

[0111] The net spectral power of each acquisition channel at each discrete frequency point in each frame is calculated based on the noise power spectrum estimate; the net spectral power of each acquisition channel is weighted and fused using a channel weighting factor to generate a dual-channel fused suppressed spectral power distribution, specifically including:

[0112] At each acquisition channel, each frame, and each discrete frequency point, the corresponding noise power spectrum estimate is subtracted from the spectral power. When the difference is greater than or equal to zero, the difference is used as the net spectral power of each acquisition channel at each discrete frequency point in each frame. When the difference is less than zero, the net spectral power of each acquisition channel at each discrete frequency point in each frame is set to zero, thus forming the net spectral power distribution of the first and second acquisition channels at each discrete frequency point in each frame.

[0113] At each frame and at each discrete frequency point, the net spectral power of the first and second acquisition channels is weighted and fused using the channel weighting factor corresponding to the current frame. The net spectral power of each acquisition channel is multiplied by the corresponding channel weighting factor and summed between the first and second acquisition channels to obtain the value of the dual-channel fusion suppressed spectral power distribution at each discrete frequency point in each frame.

[0114] Number each frame Each channel is numbered and each frequency index The net spectral power of each channel is constructed as follows:

[0115] ;

[0116] in, For the first Channel 1 Frame in frequency index Net spectral power at;

[0117] Channel weighting factor and As the fusion factor, the dual-channel fusion suppression spectrum is calculated as follows: ;in, The first after channel weighted fusion Frame rate index The suppression spectral power at that point.

[0118] The process involves selecting the main channel for each frame based on the channel weighting factor, and combining this with the dual-channel fusion suppressed spectral power distribution to complete the main channel phase extraction and noise-suppressed frequency domain complex spectrum reconstruction. Specifically, this includes:

[0119] In each frame, the channel weight factors of the first acquisition channel and the second acquisition channel are compared. When the channel weight factor of the first acquisition channel is greater than or equal to the channel weight factor of the second acquisition channel, the first acquisition channel is designated as the main channel of the current frame. When the channel weight factor of the first acquisition channel is less than the channel weight factor of the second acquisition channel, the second acquisition channel is designated as the main channel of the current frame, thus forming a sequence of main channel numbers for each frame.

[0120] At each frame and at each discrete frequency point, the phase angle of the corresponding discrete frequency point is read from the complex spectrum of the frequency domain of the main channel to form the phase spectrum of the main channel of the current frame.

[0121] In each frame and at each discrete frequency point, the square root of the power distribution of the dual-channel fused suppression spectrum at the corresponding discrete frequency point is taken as the amplitude of the frequency domain complex spectrum, and the phase angle of the main channel phase spectrum is taken as the phase angle of the frequency domain complex spectrum. The amplitude and phase angle are combined to form the frequency domain complex spectrum reconstructed after noise suppression, covering all frames and all discrete frequency points.

[0122] Number each frame The main channel number is determined based on the channel weighting factor. Specifically:

[0123] when season ;in, For the first The selected main channel number in the frame;

[0124] when season ;

[0125] For each frequency index The phase spectrum of the main channel is extracted as follows:

[0126] ;in, For the first Frame rate index Phase at; The spectrum is a complex number, indicating the selection of the main channel. In the Frame, Frequency Index The discrete Fourier transform result at the location;

[0127] For each frequency index The reconstructed complex spectrum is:

[0128] ;

[0129] in, After noise suppression, the first Frame in frequency index The reconstructed complex spectrum at the location.

[0130] The step involves performing a frame-level inverse discrete Fourier transform on the noise-suppressed frequency domain complex spectrum, and then overlapping and concatenating the reconstructed time-domain signals of each frame according to the frame shift relationship to obtain a complete noise-suppressed output time-domain signal. Specifically, this includes:

[0131] In each frame, a frame-level inverse discrete Fourier transform is performed on the reconstructed frequency domain complex spectrum after noise suppression. The transform length is the same as the discrete Fourier transform length, resulting in the reconstructed time-domain signal sample sequence at each sampling point of the corresponding frame.

[0132] The total number of discrete sampling points of the final output signal is calculated based on the relationship between the total number of frames, frame length, and frame shift. The product of the total number of frames minus one and the frame shift is added to the frame length to obtain the discrete sampling length of the output signal.

[0133] For each discrete sampling point location, identify all intra-frame sample values ​​that have a coverage relationship at this discrete sampling point location, and sum these intra-frame sample values ​​to obtain the complete output time-domain signal value for the corresponding discrete sampling point location.

[0134] Number each frame and the index of each intra-frame sampling point Perform an intra-frame inverse discrete Fourier transform to reconstruct the time-domain signal frame as follows:

[0135] ;in, After noise suppression, the first Frame in-frame index Reconstructed time-domain signal samples at the location;

[0136] Calculate the total number of discrete sampling points of the final output signal. ;

[0137] Sample index for each integer The complete time-domain output signal is calculated using the following formula: ;in, For indexing discrete samples The complete output time-domain signal at that point; For a unit impulse function, with respect to an integer independent variable satisfy: when , when .

[0138] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0139] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for suppressing electric hair clipper noise based on multi-source data fusion processing, characterized in that, include: Multiple microphone sensors with acquisition channels are arranged on the surface of the electric clipper housing to synchronously sample the acoustic signals generated during the operation of the equipment, thereby acquiring multi-channel original discrete time-domain signals covering the space near the motor cavity and the cutter head. The original discrete time-domain signals of each acquisition channel are divided into frames according to the preset frame length and frame shift. The sampling points in each frame are weighted by a window function to obtain the windowed intra-frame sampling value sequence of each acquisition channel. Based on the energy distribution of the windowed intra-frame sampled value sequence of each acquisition channel in the time domain, the time domain energy of each acquisition channel in each frame is calculated, and a channel weight factor that satisfies the normalization constraint in each frame is constructed. Perform a discrete Fourier transform on the windowed intra-frame sampled value sequence of each acquisition channel to obtain the frequency domain complex spectrum of each frame at discrete frequency points; Based on the preset number of noise estimation frames, the spectral power of each discrete frequency point in each frame of each acquisition channel is initialized and recursively updated to obtain a set of noise power spectral estimates for each frame. The net spectral power of each acquisition channel at each discrete frequency point in each frame is calculated based on the noise power spectrum estimate; the net spectral power of each acquisition channel is weighted and fused using a channel weighting factor to generate a dual-channel fused suppressed spectral power distribution. The main channel of each frame is selected based on the channel weighting factor. Combined with the dual-channel fusion suppression spectrum power distribution, the main channel phase extraction and noise suppression frequency domain complex spectrum reconstruction are completed. Perform a frame-level inverse discrete Fourier transform on the noise-suppressed frequency domain complex spectrum, and then overlap and splice the reconstructed time-domain signals of each frame according to the frame shift relationship to obtain the complete noise-suppressed output time-domain signal.

2. The method for suppressing electric clipper noise based on multi-source data fusion processing according to claim 1, characterized in that, The microphone sensor with multiple acquisition channels arranged on the surface of the electric hair clipper's outer shell performs synchronous sampling of the acoustic signals generated during the device's operation, acquiring multi-channel raw discrete-time signals covering the space near the motor cavity and the cutter head. Specifically, this includes: A first microphone and a second microphone are installed at equal intervals along a preset direction on the outer surface of the electric clipper housing. The first microphone is located on the outer surface of the motor cavity housing, and the second microphone is located on the housing directly in front of the cutter head. The first microphone and the second microphone constitute the first acquisition channel and the second acquisition channel, respectively. Under the set system sampling frequency, synchronous sampling is performed on the microphone outputs of the first and second acquisition channels. The continuous sampling duration is set, and the total number of discrete sampling points for each acquisition channel is calculated based on the product of the system sampling frequency and the continuous sampling duration. The time interval between two adjacent discrete sampling points is determined according to the reciprocal of the system sampling frequency. The time-domain signal values ​​of each discrete sampling point are acquired sequentially in the first acquisition channel according to the sampling order to form a first original discrete time-domain signal sequence. The time-domain signal values ​​of each discrete sampling point are acquired in the second acquisition channel at the same sampling time as the first acquisition channel to form a second original discrete time-domain signal sequence. The two original discrete time-domain signals correspond to the same sampling time position at the same sampling point number.

3. The method for suppressing electric clipper noise based on multi-source data fusion processing according to claim 2, characterized in that, The original discrete-time domain signals of each acquisition channel are divided into frames according to a preset frame length and frame shift. A window function is applied to the sampling points within each frame to weight them, resulting in a windowed intra-frame sampling value sequence for each acquisition channel. Specifically, this includes: Within each acquisition channel, the original discrete-time domain signal is divided into frames according to two positive integer parameters: preset frame length and frame shift. This ensures that the number of sampling points in each frame is equal to the frame length, and the sampling point interval between the starting sampling points of two adjacent frames is equal to the frame shift. Based on the relationship between the total number of sampling points, frame length, and frame shift in each acquisition channel, the total number of frames is calculated in the following order: first, subtract the frame length from the total number of sampling points; then, divide the difference by the frame shift; and finally, add one to the quotient. At the same time, it is ensured that the frame shift is not greater than the frame length and the frame length is not greater than the total number of sampling points in each acquisition channel. Within each acquisition channel, starting from the initial sampling point, the index of the initial sampling point of each frame is determined sequentially according to the frame shift interval. In each frame, sampling points equal to the frame length are continuously extracted from the corresponding initial sampling point to form the discrete time domain signal sequence of the corresponding acquisition channel in the corresponding frame. The sampling points in each frame sequence are arranged sequentially according to the local index within the frame. Configure a window function for each frame, using a rectangular window as the window function form. Perform a windowing operation on each sampling point within each frame. Under the rectangular window condition, set the window function weighting coefficient of each sampling point to one, and multiply the intra-frame sampled value with the corresponding weighting coefficient to obtain the windowed intra-frame sampled value sequence for each acquisition channel.

4. The method for suppressing electric clipper noise based on multi-source data fusion processing according to claim 3, characterized in that, The method involves calculating the temporal energy of each acquisition channel within each frame based on the energy distribution of the windowed intra-frame sampled value sequence of each acquisition channel in the time domain, and constructing a channel weight factor that satisfies normalization constraints within each frame. Specifically, this includes: Within each acquisition channel, the windowed intra-frame sampled values ​​of each frame are squared one by one and accumulated within the frame. The accumulated result is used as the temporal energy of each acquisition channel in each frame, thus obtaining the temporal energy sequence of the first acquisition channel in each frame and the temporal energy sequence of the second acquisition channel in each frame. In each frame, the temporal energy of the first acquisition channel and the second acquisition channel are added together to obtain the sum of the temporal energy of the first acquisition channel and the second acquisition channel in each frame. Under the condition that the sum is greater than zero, the temporal energy of each acquisition channel in each frame is divided by the sum of the temporal energy, which is used as the channel weight factor of each acquisition channel in each frame, so that the sum of the channel weight factors of the first acquisition channel and the second acquisition channel in each frame is equal to one. In each frame, when the sum of the temporal energy of the first acquisition channel and the second acquisition channel is equal to zero, the channel weight factors of the first acquisition channel and the second acquisition channel are set to equal values ​​respectively, and the sum of the channel weight factors of the first acquisition channel and the second acquisition channel in each frame is equal to one, forming a channel weight factor sequence that satisfies the normalization condition in all frames.

5. The method for suppressing electric clipper noise based on multi-source data fusion processing according to claim 4, characterized in that, The step of performing a discrete Fourier transform on the windowed intra-frame sampled value sequence of each acquisition channel to obtain the frequency domain complex spectrum of each frame at discrete frequency points specifically includes: Within each acquisition channel, a discrete Fourier transform with a length equal to the frame length is performed on the windowed intra-frame sampled value sequence of each frame to obtain the frequency domain complex spectrum of each frame at each discrete frequency point. Each frequency domain complex spectrum has a real component and an imaginary component at each discrete frequency point. For each acquisition channel, each frame, and each discrete frequency point, the real and imaginary components of the frequency domain complex spectrum are squared respectively. The two squared results are added together to obtain the spectral power of each acquisition channel at each discrete frequency point in each frame, forming a set of spectral power covering all acquisition channels, all frames, and all discrete frequency points.

6. The method for suppressing electric clipper noise based on multi-source data fusion processing according to claim 5, characterized in that, The step involves initializing and recursively updating the spectral power at each discrete frequency point in each frame of each acquisition channel according to a preset number of noise estimation frames, to obtain a set of noise power spectral estimates for each frame. Specifically, this includes: Set the number of frames to be used for noise estimation initialization. The set number of frames is a positive integer that is not less than one and not greater than the total number of frames. Within each acquisition channel, for each discrete frequency point, the spectral power of the first few frames is summed and divided by the number of frames to obtain the initial estimate of the noise power spectrum of each acquisition channel at each discrete frequency point, thus forming the initial estimate set of noise power spectrum of each discrete frequency point of each acquisition channel. Within a set number of the first few frames, the noise power spectrum estimate for each frame, each acquisition channel, and each discrete frequency point is uniformly set to the corresponding initial noise power spectrum estimate, forming a set of noise power spectrum estimates on the initialization frame. Starting from the next frame after the initialization frame until the last frame, at each acquisition channel and at each discrete frequency point, the noise power spectrum estimate of the previous frame and the spectral power of the current frame are weighted and averaged. The weight ratio is set to half for the noise power spectrum estimate of the previous frame and half for the spectral power of the current frame. This yields new noise power spectrum estimates for the current frame, each acquisition channel, and each discrete frequency point. The recursive update process is then performed sequentially on all subsequent frames.

7. The method for suppressing electric clipper noise based on multi-source data fusion processing according to claim 6, characterized in that, The net spectral power of each acquisition channel at each discrete frequency point in each frame is calculated based on the noise power spectrum estimation value. The net spectral power of each acquisition channel is weighted and fused using a channel weighting factor to generate a dual-channel fused suppressed spectral power distribution, specifically including: At each acquisition channel, each frame, and each discrete frequency point, the corresponding noise power spectrum estimate is subtracted from the spectral power. When the difference is greater than or equal to zero, the difference is used as the net spectral power of each acquisition channel at each discrete frequency point in each frame. When the difference is less than zero, the net spectral power of each acquisition channel at each discrete frequency point in each frame is set to zero, thus forming the net spectral power distribution of the first and second acquisition channels at each discrete frequency point in each frame. At each frame and at each discrete frequency point, the net spectral power of the first and second acquisition channels is weighted and fused using the channel weighting factor corresponding to the current frame. The net spectral power of each acquisition channel is multiplied by the corresponding channel weighting factor and summed between the first and second acquisition channels to obtain the value of the dual-channel fusion suppressed spectral power distribution at each discrete frequency point in each frame.

8. The method for suppressing electric hair clipper noise based on multi-source data fusion processing according to claim 7, characterized in that, The process involves selecting the main channel for each frame based on the channel weighting factor, and combining this with the dual-channel fusion suppressed spectral power distribution to complete the main channel phase extraction and noise-suppressed frequency domain complex spectrum reconstruction. Specifically, this includes: In each frame, the channel weight factors of the first acquisition channel and the second acquisition channel are compared. When the channel weight factor of the first acquisition channel is greater than or equal to the channel weight factor of the second acquisition channel, the first acquisition channel is designated as the main channel of the current frame. When the channel weight factor of the first acquisition channel is less than the channel weight factor of the second acquisition channel, the second acquisition channel is designated as the main channel of the current frame, thus forming a sequence of main channel numbers for each frame. At each frame and at each discrete frequency point, the phase angle of the corresponding discrete frequency point is read from the complex spectrum of the frequency domain of the main channel to form the phase spectrum of the main channel of the current frame. In each frame and at each discrete frequency point, the square root of the power distribution of the dual-channel fused suppression spectrum at the corresponding discrete frequency point is taken as the amplitude of the frequency domain complex spectrum, and the phase angle of the main channel phase spectrum is taken as the phase angle of the frequency domain complex spectrum. The amplitude and phase angle are combined to form the frequency domain complex spectrum reconstructed after noise suppression, covering all frames and all discrete frequency points.

9. The method for suppressing electric clipper noise based on multi-source data fusion processing according to claim 8, characterized in that, The step involves performing a frame-level inverse discrete Fourier transform on the noise-suppressed frequency domain complex spectrum, and then overlapping and concatenating the reconstructed time-domain signals of each frame according to the frame shift relationship to obtain a complete noise-suppressed output time-domain signal. Specifically, this includes: In each frame, a frame-level inverse discrete Fourier transform is performed on the reconstructed frequency domain complex spectrum after noise suppression. The transform length is the same as the discrete Fourier transform length, resulting in the reconstructed time-domain signal sample sequence at each sampling point of the corresponding frame. The total number of discrete sampling points of the final output signal is calculated based on the relationship between the total number of frames, frame length, and frame shift. The product of the total number of frames minus one and the frame shift is added to the frame length to obtain the discrete sampling length of the output signal. For each discrete sampling point location, identify all intra-frame sample values ​​that have a coverage relationship at this discrete sampling point location, and sum these intra-frame sample values ​​to obtain the complete output time-domain signal value for the corresponding discrete sampling point location.

Images