A method for constructing reference signals of fan duct ANC system based on microphone array

Abstract

A method for constructing a reference signal for a fan duct ANC system based on a microphone array. A microphone array device is used as a reference signal sensor for a feedforward ANC system. The algorithm estimates the rotation noise frequency in the duct fan noise based on FFT and quadratic interpolation. After completing array data preprocessing such as rotation noise component extraction, a circular harmonic transform is first performed for a uniform annular array. For a uniform linear array or a virtual linear array obtained by the transform, subarray division is performed and forward and backward spatial smoothing is performed. The smoothed array data covariance matrix is ​​diagonally loaded, and then MPDR or LCMP beamforming is performed. Finally, the amplitude and phase of the data sequence obtained by beamforming are estimated based on the least squares method, and a high-quality fan rotation noise reference signal is constructed in conjunction with a timer. The method of the present invention improves the stability of feedforward ANC algorithms such as FxLMS and the suppression effect of fan rotation noise.

Description

Technical Field

[0001] The present invention belongs to the field of active noise control (ANC), and relates to a method for constructing a reference signal of a fan duct ANC system based on a microphone array. Technical Background

[0002] In noise control technology, there are two main methods: "passive noise control" and "active noise control". Passive noise control mainly uses special sound-absorbing materials to wrap or fill the noise propagation path to complete noise control. This method is mainly used to suppress high-frequency noise. For low-frequency noise, due to the long wavelength of sound waves, passive noise control has high requirements for material properties and dimensions, which is often difficult to achieve, and in many application scenarios, the installation of passive noise control equipment will be restricted. In comparison, active noise control has the characteristics of simple equipment, convenient installation, and good suppression effect on low-frequency noise, and is gradually becoming the mainstream technology for noise control.

[0003] With the progress of social economy, people are increasingly pursuing good auditory experience in learning and working environments. It is not difficult to imagine that in offices, teaching buildings, libraries and other environments, if the fan equipment such as central air conditioners can effectively control the noise while meeting the air output, the environment at this time will be more conducive to people's focused work and study. The classic topic of fan duct noise control is gaining more and more attention in this context.

[0004] Some literature on fan noise, such as the journal article "Fan Blade Noise Mechanism and Noise Reduction" published by Wu Xianjun et al. of Wuhan University of Technology in 2001 and "Fan Noise Control" published by Zheng Dongming in 1998, jointly discussed the mechanism and characteristics of fan noise, and pointed out that fan noise is mainly composed of medium and low frequency rotation noise (periodic noise) and medium and high frequency turbulent noise. The former constitutes the main element of fan noise under the condition of high speed and low load of the fan. For turbulent noise, passive noise control can be used to suppress it, while for rotation noise, ANC technology needs to be introduced for effective suppression.

[0005] In ANC technology, common control systems are divided into feedforward and feedback structures. The main difference between the two is that in the feedback ANC system, only the error signal is involved in the adjustment of the adaptive filter coefficients, while in the feedforward ANC system, in addition to the error feedback signal, the feedforward reference signal of the primary noise source is also involved in the adaptive filtering process. Therefore, the feedforward system usually has better system stability and noise reduction performance. Feedforward ANC mainly uses classic algorithms such as FxLMS, ​​FxNLMS and FxRLS. In order to take into account the strict real-time requirements of point-by-point adaptation of the algorithm, the ANC controller usually uses DSP or FPGA devices dedicated to real-time audio signal processing, and generally requires two audio input interfaces and one audio output interface.

[0006] In the feedforward ANC system, the reference signal mainly contains two highly coherent interferences: acoustic feedback from the secondary speaker to the reference microphone and acoustic reflection of the multipath propagation of sound waves, as well as incoherent interferences in other directions such as human voice, which will affect the stability and noise reduction performance of the system. Therefore, how to obtain high-quality reference signals has become the research focus of the feedforward ANC system.

[0007] One method for avoiding reference signal interference is described in the Chinese invention patent application, publication number CN113257214A, which uses a tachometer, a non-acoustic sensor, to replace the reference microphone to obtain the fundamental frequency of the fan rotation noise to construct a reference signal.

[0008] Compared with such systems, the original reference signal of the present invention is collected by using a microphone as an acoustic sensor, and combined with the corresponding algorithm, the constructed signal can retain the amplitude and phase information of the rotational noise, which in fact constitute the main adjustment objects of ANC algorithms such as FxLMS.

[0009] When an acoustic sensor is used as the reference sensor, the common related hardware parts are described in the master's thesis "DSP Implementation of Active Control System for Pipeline Low-frequency Noise" published by Yu Mengjiao of Harbin Institute of Technology in 2013. The reference microphone is often composed of only a single microphone, and the primary noise is often generated inside the ANC controller. Parameters such as the frequency of the reference signal are regarded as known quantities.

[0010] Compared with such systems, the present invention replaces a single microphone with a reference microphone array device and combines it with a corresponding estimation algorithm without the need to obtain a priori information on signal frequency and phase.

[0011] Regarding the acoustic feedback interference problem of the reference signal, Kuan-Chun Chen et al. from Chung Yuan Christian University proposed a more direct acoustic feedback suppression solution in the paper "Active noise control in a duct to cancel broadband noise" published in 2017. This solution requires modeling of the acoustic feedback channel, and the effect of acoustic feedback suppression is closely related to the modeling accuracy.

[0012] Compared with such systems, the present invention does not need to introduce modeling of the acoustic feedback channel and the acoustic feedback suppression algorithm, and the reference microphone array device and the reference signal construction algorithm are independent of the ANC controller device. The reference microphone array device can be implemented using a low-cost Linux device, reducing the computing pressure of the ANC controller.

[0013] Another solution to avoid acoustic feedback is to use a cardioid microphone, which uses the spatial directivity brought by its physical properties to directionally extract the signal in the direction of the primary noise source, while suppressing the signal in the direction of the secondary speaker. However, this type of microphone is generally expensive and is often found in professional recording studios and stages. In addition, its amplitude-frequency response characteristics are generally not as flat as those of an omnidirectional microphone.

[0014] Compared with such solutions, the reference microphone of the present invention adopts a common omnidirectional microphone, combined with the beamforming algorithm in the array signal processing technology, so that the array obtains similar spatial directivity.

[0015] The technology of introducing array into ANC system was first seen in the conference article "Application of MVDR beamforming to reject turbulence noise in a duct" published by Daniel J. Sebald in 1996. The main purpose of using array beamforming technology in this article is to reduce the influence of turbulence noise in the fan duct on the reference signal.

[0016] Compared with such systems, the present invention fully considers various major coherent and incoherent interference factors that pollute the fan rotation noise in the application scenario of the fan duct system, and considers the delay compensation problem of the reference signal to improve the quality of the constructed reference signal. Summary of the invention

[0017] The technical problem to be solved by the present invention is to provide a high-quality reference signal for a fan duct feedforward ANC system. The reference signal construction algorithm of the present invention can use the spatial directivity of the array to directionally extract the rotational noise generated by the fan, while suppressing the acoustic feedback of the secondary speaker and interference signals in other directions, and fully retaining the amplitude and phase information of the rotational noise, thereby improving the stability and noise reduction performance of the ANC system.

[0018] The technical solution of the present invention is as follows:

[0019] A method for constructing an ANC reference signal for a fan duct system based on a microphone array is provided. A feedforward ANC system is arranged in the fan duct. After the ANC controller completes the amplitude calibration of two-channel audio inputs and the offline identification of the secondary channel, the microphone array device and the ANC controller are turned on in sequence. The specific process is as follows:

[0020] Step 1: Initialize the parameters such as array element position vector, diagonal loading, FFT window, and the container for storing data such as algorithm output sequence.

[0021] Step 2: Set the maximum time T for a cycle ps , and start the timer.

[0022] Step 3: If it is the first startup, that is, processing the first frame of data, then calculate the FFT-based rotation noise frequency estimate of each channel, and use the average of the frequency estimates of each channel as the rotation noise prior frequency information of the next frame of data; if it is not processing the first frame of data, proceed to step 4. In the FFT-based rotation noise frequency estimation process, a windowing operation is required before the data sequence is transformed by FFT, and then the quadratic interpolation method is used to estimate the frequency of the fan rotation noise (periodic noise) in combination with the maximum value or maximum value of the FFT modulus sequence. After the frequency estimation of the first frame of data is completed, the timer calculates the elapsed time t of the algorithm process at this time. elapse , process sleep (T ps -t elapse ) and then returns to wait for the next call. Since the first frame estimation does not perform actual reference signal construction, the construction signal output by this step is the initialization value of the container described in step 1 (generally initialized to a sequence of all 0s).

[0023] Step 4: The next time it is called (one frame of data is collected), the process starts from step 2 to preprocess the array data. In this step, the frequency estimate obtained in the previous frame is used as the center frequency to design the Butterworth notch filter parameters and the array parameters such as the array manifold vector in the direction of the primary noise source are obtained by combining the array coordinate system. The original data sequence is filtered forward and backward to obtain the bottom noise, and the rotational noise component in the fan noise is obtained by subtracting the bottom noise from the original array data. Then, the rotational noise components of each channel are homogenized to eliminate the amplitude difference, and the Hilbert transform of each channel data after homogenization is calculated so that the data has instantaneous phase information.

[0024] Step 5: Perform FBSS-based MPDR or LCMP beamforming on the array data obtained in step 4. If the array is configured as a uniform annular array, first perform circular harmonic transformation on the array to obtain a virtual linear array. Then, the uniform linear array or virtual linear array is sub-divided and forward and backward spatial smoothing (FBSS) is performed, and the smoothed data covariance matrix is ​​diagonally loaded to reduce the adverse effects of array mismatch. Then, MPDR or LCMP beamforming is performed, and the real part of the sequence is taken as the output of the beamformer.

[0025] Step 6: Perform FFT-based rotation noise frequency estimation on the output of the beamformer in step 5.

[0026] Step 7: Construct a matrix equation based on the frequency estimate of the rotating noise in step 6, the expression of the multi-frequency signal, and the output of the beamformer in step 5. Use the least squares method to obtain the solution vector of the matrix equation, and then solve the amplitude and phase information of the signal encapsulated in the solution vector. So far, the process has obtained four pieces of information: the frequency, amplitude, phase, and inherent duration of a frame of data of the rotating noise. Based on this information, in conjunction with the timer and the algorithm duration setting value, the fan rotating noise reference signal required by the ANC controller can be constructed. In the future, the frequency estimate and amplitude estimate obtained in the current frame will be used as the prior information of the signal for the next frame. The amplitude estimate described in this step and the FFT-based rotating noise frequency estimate described in step 3 can both add a sliding average operation to make the parameters more stable.

[0027] Step 8: The timer calculates the elapsed time t of the algorithm process at this time elapse , process sleep (T ps -t elapse ) and then output the fan rotation noise reference signal obtained in step 7 and return, waiting for the next call. The next call starts from step 2 to enter the process.

[0028] Beneficial effects of the present invention: For the fan duct ANC system, the reference signal construction algorithm of the present invention can prevent the rotation noise from being contaminated by factors such as acoustic feedback from the secondary speaker, and retain the amplitude and phase information of the reference signal, greatly improving the stability and noise reduction effect of the ANC system. The reference microphone array subsystem can be implemented with a low-cost Linux device, independent of the ANC controller. Due to the high quality of the constructed reference signal, the ANC controller can be freed from the computational pressure of reference signal preprocessing, and execute more complex feedforward ANC algorithms such as FxRLS and perform more accurate secondary channel modeling. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 It is the overall system structure block diagram.

[0030] Figure 2(a) is a schematic diagram of the array coordinate system.

[0031] Figure 2(b) is a schematic diagram of the ULA array element layout.

[0032] Figure 2(c) is a schematic diagram of the UCA array element layout.

[0033] Figure 3 It is the main flow chart of the reference signal construction algorithm of the present invention.

[0034] Figure 4 This is a flow chart of the rotation noise frequency estimation based on FFT of the present invention.

[0035] FIG5( a) is a flowchart of array data preprocessing of the present invention;

[0036] FIG5( b ) is a flowchart of MPDR or LCMP beamforming based on FBSS of the present invention;

[0037] FIG5( c ) is a flow chart of amplitude and phase estimation and signal construction based on LS of the present invention;

[0038] Figure 6 This is an example of a single-frequency rotating noise sampled from a real wind turbine; (a) is a time history diagram; (b) is a partial zoomed-in time history diagram; (c) is a power spectrum diagram;

[0039] Figure 7 This is an example of a harmonic rotation noise sampled from a real wind turbine; (a) is a time history diagram; (b) is a partial zoomed-in time history diagram; (c) is a power spectrum diagram;

[0040] Figure 8 The following is an example of the amplitude difference phenomenon of the original data channels of a small-sized array; (a) is a time course difference comparison diagram; (b) is a power spectrum difference comparison diagram. DETAILED DESCRIPTION

[0041] The specific implementation of the present invention is described in detail below in conjunction with the content of the invention and the accompanying drawings.

[0042] Figure 1The figure is a block diagram of the overall system structure, indicating the position of the system in the fan duct and its working principle. In the figure, the reference microphone array device, ANC controller (including three audio interfaces and the parts below the interfaces), secondary speakers and error microphones together constitute a feedforward ANC system. The reference microphone array is a Linux device, and the ANC controller is a DSP or FPGA device. The reference microphone array element is composed of omnidirectional microphones, and the error microphone is an omnidirectional electret microphone. Under the action of adaptive ANC algorithms such as FxLMS, ​​the ANC controller generates secondary noise with the same amplitude but opposite phase as the primary noise (from the duct fan) through the secondary speaker. The two are coherently superimposed to achieve the purpose of fan duct noise suppression. The ANC controller needs to complete the amplitude calibration of the two audio inputs in advance (that is, under the same signal and the same sound pressure excitation, the two audio input channels need to obtain the same sampling value) and offline identification of the secondary channel.

[0043] When arranging the array, the reference microphone array can be arranged as a uniform linear array (ULA) or a uniform circular array (UCA). The array should not be too close to the primary noise source to ensure that the wavefront is a plane wave model when the sound wave reaches the array. At the same time, before constructing the specific implementation of the algorithm, first clarify the array coordinate system for the algorithm. As shown in Figure 2, whether it is ULA or UCA, the center of the array (center of the circle) is usually used as the origin of the array coordinate system, and the 0th array element is arranged at the first position in the positive direction of the x-axis of the three-dimensional rectangular coordinate system. When installing the reference microphone array, align the positive direction of the x-axis with the direction of the primary noise source (that is, Figure 1 The negative direction of the x-axis is aligned with the direction of the secondary speaker, that is, in the ULA case, the array elements are evenly distributed on the x-axis, and the primary noise source direction signal and the acoustic feedback interference signal form an "end-fire". For narrow-band plane waves, after selecting the array reference system, the frequency domain data received by the array elements is expressed as follows:

[0044]

[0045] Where F(ω) is the frequency domain form of the plane wave signal waveform, is called the array manifold vector. is the position vector of the array element in the three-dimensional rectangular coordinate system, is the wave number vector defined by the wave equation, is the direction cosine of the standard spherical coordinate system relative to the three-dimensional rectangular coordinate system, and c is the speed of sound.

[0046] In narrowband beamforming technology, the beamformer is often expressed as a series of complex weights From the frequency domain product corresponding to the time domain convolution of the linear system, we can know that the frequency domain representation of the beamformer output is:

[0047]

[0048] From (2), we can see that the array manifold vector Encapsulates the space-time-frequency information of the array. This parameter will be directly applied to the complex weight value. At the same time, by bringing the ULA case into play, it can be found that this parameter has a Vandermonde form, which is a prerequisite for performing forward and backward spatial smoothing decorrelation. The UCA can be transformed into a virtual linear array.

[0049] Figure 3 This is the main flow chart of the reference signal construction algorithm of the present invention. The algorithm first initializes the parameters such as the array element position vector, diagonal loading and FFT window, and the container for storing the algorithm output sequence and other data. When processing the data frame, the algorithm first enables the timer to start timing, and at the same time sets the maximum time T of an algorithm flow. ps This setting value should be less than the maximum sampling delay of a frame of audio data. Suppose the length of a frame of audio data is N sample , the sampling frequency is f s , then ideally the first point of a frame of data has a delay of 0, and the last point corresponds to the maximum delay For example, if the sampling rate is 8KHz and one frame of audio data is set to 4096 points, the corresponding T maxdelay =512ms, at this time T ps After completing the processing of a frame of data, the timer is enabled again and the time taken by the algorithm at this moment is calculated. elapse , algorithm process sleep T ps -t elapse Then it returns and waits for the next call. The frequency estimate and amplitude estimate obtained in one algorithm flow will be used as the prior information of the rotation noise signal in the next algorithm flow. When the reference signal construction algorithm is called for the first time (when estimating the first frame), the constructed signal output by the algorithm is the initialization sequence of the corresponding container, and the valid output will be obtained in subsequent calls.

[0050] Depend on Figure 3 It can be seen that the reference signal construction algorithm of the present invention mainly consists of four parts: "rotation noise frequency estimation based on FFT", "array data preprocessing", "MPDR or LCMP beamforming based on FBSS" and "amplitude, phase estimation and signal construction based on LS". The flowcharts of each part are shown in Figure 4 and Figure 5.

[0051] The details are as follows:

[0052] 1. Rotational noise frequency estimation based on FFT, the flowchart is shown in Figure 4, its purpose is to obtain Figure 6 (a)-(b) are the frequencies corresponding to the peaks of the rotation noise spectrum. They are applicable when the rotation noise is single-frequency noise or multi-frequency noise:

[0053] In order to avoid the adverse effects of spectrum leakage and fence effect, the present invention is based on the quadratic function vertex formula (quadratic interpolation) for efficient estimation. This method requires the maximum point of the FFT amplitude spectrum or the selected extreme point and one data point before and after it to complete the construction of the equation, that is, three points that are not on a straight line determine a parabola. Among them, the maximum point method determines the fundamental frequency of the rotating noise, and its integer harmonics can be directly multiplied by the harmonic sequence number to obtain an estimate; the maximum point method can be implemented by the difference method, which can estimate the fundamental frequency of the rotating noise and directly estimate the harmonic frequency. It is still applicable when the harmonic frequency is offset due to the acoustic characteristics of the pipeline, and the error of the maximum method in harmonic estimation can be avoided. However, the selection of the maximum point must meet the height threshold condition (the amplitude of the selected extreme point should be greater than a certain height) and the distance threshold condition (the interval between two maximum points should be greater than a certain frequency). Since the FFT sequence of the real sequence has the characteristics of conjugate symmetry, when looking for the maximum value or maximum value, only the positive frequency part of the amplitude spectrum is used. Before FFT transformation, the original data sequence should be windowed to improve the frequency estimation accuracy. Suppose that the data groups of a certain maximum point or a maximum point are (x f ,v f )、(x c ,v c ) and (x b ,v b ), where x c is the maximum or maxima point (index converted to frequency according to frequency resolution), v c Represents the amplitude of the point, then the linear equation system can be constructed:

[0054]

[0055] After solving the quadratic parabola parameters a, b, and c, the frequency estimate can be obtained according to the abscissa formula of the parabola vertex. To avoid the inversion of the Vandermonde matrix, the abscissa of the parabola vertex can also be directly analyzed. The specific formula is:

[0056]

[0057] When the added window is a Gaussian window, v in formula (4) f 、v c and v bThe natural logarithm should be taken. For the estimated value of an algorithm process, both the frequency and the amplitude mentioned later can be smoothed with a sliding average window, so that in a steady-state environment (assuming that the frequency of the rotational noise is constant or only changes slowly within a small range), the change of the estimated parameters is reduced, and the connection between the algorithm output data frames is smoother.

[0058] 2. Array data preprocessing, the flow chart is shown in the corresponding part of Figure 5, which mainly completes the extraction of the rotation noise component in the fan noise, the array data channel homogenization and the Hilbert transform:

[0059] Step 1: Estimated frequency f from the previous frame of data est Design the parameters of the Butterworth notch filter, the acoustic wave number, and the array element position vector in the sense of wavelength for the center frequency, etc., which are needed for the MPDR / LCMP beamforming algorithm. Suppose the original data matrix received by the array is The Butterworth filter is used to filter each row of Rx (equivalent to each channel of the array) to obtain the bottom noise data matrix Rn, and the rotation noise data matrix is ​​obtained using Rs=Rx-Rn. The order of the Butterworth notch filter cannot be too high, and the stopband width cannot be too narrow. For example, a frame of data with 4096 points can be used as the example. The order of the filter can be 6, and the cutoff frequency can be set to f est ±2.5Hz. The filtering should be performed in the form of forward and backward filtering, the purpose of which is to offset the signal delay caused by the digital filter according to the linear system theory. At the same time, in order to avoid serious distortion of the head and tail of the filtered data, part of the original data of the previous frame (such as the last 2048 points of the original data of the previous frame) can be retained to participate in the filtering.

[0060] Step 2: Homogenize each channel of Rs, that is, only consider the phase difference between different array elements, and ignore the amplitude difference. Figure 7 An example of the amplitude difference phenomenon of the raw data channel of a small-sized array is shown, which is equivalent to adding an uncertain filter window to the array element, affecting the performance of the beamformer. This design uses the modulus value of each point of the FFT sequence of Rs to be normalized to the average value to achieve uniformity, which can fully retain the amplitude information of the rotation noise.

[0061] Step 3: Perform Hilbert transform on each channel of Rs to obtain the analytical form data matrix R of the signal a , so that the data sequence carries instantaneous phase information, so that the complex weight vector obtained by the beamformer can be used to perform gain-phase shift operations.

[0062] 3. FBSS-based MPDR or LCMP beamforming, the flow chart is shown in the corresponding part of Figure 5, the purpose is to make the array only extract the rotational noise in the direction of the primary noise source (i.e., the duct fan) and suppress interference in other directions (such as the direction of the secondary speaker):

[0063] Step 1: If the array device is configured as a uniform linear array, perform MPDR or LCMP beamforming based on forward-backward spatial smoothing (FBSS); if it is configured as a uniform annular array, it must first be transformed into a virtual linear array by circular harmonic transformation so that its array manifold vector has a Vandermonde form. According to the array coordinate system determined by the system, the circular harmonic transformation matrix T is defined as:

[0064]

[0065] Where n = -N, -N+1, ..., 0, ... N-1, N is the circular harmonic number, N is the maximum circular harmonic order selected; m = 0, 1, ..., M-1, M is the number of array elements, is the azimuth angle of the mth array element of UCA in the array coordinate system. The circular modal response elimination matrix J is defined as:

[0066]

[0067] Among them J n (kr) is the Bessel function of the first kind, k is the wave number vector The modulus value of r is the radius of the uniform circular array. For UCA, the process of converting the array data domain to the circular harmonic data domain is:

[0068]

[0069] Step 2: To avoid the adverse effects of coherent interference on the performance of the beamformer, forward and backward spatial smoothing is required. First, the uniform linear array or the virtual linear array formed by the circular harmonic transformation needs to be divided into sub-arrays. Taking the uniform linear array with M = 5 array elements as an example, if its array elements are represented as ARRAY = {e0, e1, e2, e3, e4}, it can be divided into 3 sub-arrays with three array elements as ARRAY sub0 ={e0,e1,e2}, ARRAY sub1 ={e1,e2,e3} and ARRAY sub2 = {e2, e3, e4}, corresponding to the analytical form data submatrix R obtained after the above steps a0 , R a1 and R a2 . The sampling data covariance matrix of the i-th sub-matrix is:

[0070]

[0071] The spatial smoothing process is:

[0072]

[0073] Where K+1 is the number of sub-arrays, and the final forward and backward spatial smoothing sampling data covariance matrix is ​​obtained by introducing forward and backward averaging:

[0074]

[0075] Among them J s is a switching matrix with the same order as the number of sub-array elements, with 1 on the sub-diagonal and 0 on the rest of the positions, C x,ss * Indicates C x,ss Elements take conjugate, J s C x,ss * J s Also known as the backward sampling data covariance matrix.

[0076] Step 3: C x,fbss Diagonal loading (DL) is performed to reduce the negative impact on the performance of the beamformer when problems such as array element position mismatch or wave direction mismatch occur. Its essence is to artificially inject a relatively low-power white noise into the sampled data covariance matrix, that is, the loading method is:

[0077]

[0078] in It is called diagonal loading, I is the same as C x,fbss The optimal diagonal loading has no unified analytical expression under different circumstances. It can usually be selected based on the signal-to-noise ratio of the original signal or the power of the data sequence obtained through the above steps, which is highly flexible.

[0079] Step 4: Design the beamformer weight vector, where the optimal MPDR weight vector is:

[0080]

[0081] in represents the array manifold vector in the desired direction (if the array coordinate system is configured according to step 1, it corresponds to the positive x-axis, that is, the direction of the primary noise source), and L represents the number of sub-array elements.

[0082] The optimal weight vector of LCMP is:

[0083]

[0084] in is called the constraint value vector, and F is called the constraint matrix. The constraint setting is relatively flexible, and common ones include distortion constraint and zero point constraint. In the application scenario of the present invention, the constraint can be set as:

[0085]

[0086] in are array manifold vectors of the desired direction and the null direction of the subarray symmetric about the origin of the array coordinate system. As mentioned above, in the application scenario of the present invention, Corresponding to the direction of the primary noise source, Corresponds to the direction of the secondary speaker.

[0087] The present invention recommends using the LCMP optimal weight vector for beamforming. After the above steps, the data matrix of the subarray symmetric about the origin of the array coordinate system is assumed to be R a,sym , then the final beamforming output corresponding to a certain frequency point n is:

[0088]

[0089] 4. Amplitude and phase estimation and signal construction based on LS. The flowchart is shown in the corresponding part of Figure 5. The purpose is to extract the amplitude and phase of the rotation noise and combine the audio delay of one frame of data and the maximum time T of the set algorithm process. ps Delay compensation is completed so that the signal output by the reference microphone array to the ANC controller is transformed from the signal that has already passed to the signal that is about to occur:

[0090] A periodic signal with N harmonics or a multi-frequency signal with N+1 frequency points can be expressed as:

[0091]

[0092] From formula (16), it can be seen that after the frequency, amplitude, phase and action time of each component, the reference signal that meets the above requirements and is sent to the ANC controller for calculation can be constructed. The action time is constructed as a time vector:

[0093]

[0094] Based on the least squares method, the parameter matrix is ​​constructed by equation (16):

[0095]

[0096] Solving Matrix Equations Here is the final signal synthesized at all frequency points in equation (15), where Obviously, the amplitude a n for u n and vn The 2-norm of and The sign and value of the inverse function can be solved to obtain the unambiguous phase

[0097] So far, the algorithm has obtained the four structural elements of rotation noise: frequency, amplitude, phase and action time. Among them, the frequency and amplitude can be added with sliding average operation to make the parameter estimation more stable.

[0098] The final reference signal is constructed as:

[0099]

[0100] Among them, "+T maxdelay "One item means to shift the sine sequence to the left to compensate for the delay caused by audio sampling, "+T ps "One term represents a left shift of the sequence to compensate for the time consumption of the reference signal construction algorithm.

Claims

1. A method for constructing a reference signal of a fan duct ANC system based on a microphone array, characterized in that: Here are the steps: Step 1: Initialize parameters and data containers; Step 2: Set a maximum algorithm cycle time T ps , and start the timer to start timing; Step 3: When starting for the first time, process the first frame of data, calculate the estimated value of the rotation noise frequency of each channel based on FFT, and use the average value of the estimated value of the noise frequency of each channel as the prior frequency information of the rotation noise of the next frame of data; If it is not the first time to start, go directly to step 4; Step 4: preprocess the array data; design the Butterworth notch filter parameters as the center frequency using the frequency estimate obtained in the previous frame, and obtain the array manifold vector in the direction of the primary noise source in combination with the array coordinate system; perform forward and backward filtering operations on the original data sequence to obtain the bottom noise, subtract the obtained bottom noise from the original data sequence, and obtain the rotational noise component in the fan noise; then homogenize the rotational noise components of each channel to eliminate the amplitude difference, and perform Hilbert transform on the homogenized data of each channel so that the data has instantaneous phase information; Step 5: Perform MPDR or LCMP beamforming based on the forward and backward spatial smoothing operation FBSS on the array data preprocessed in step 4; the array device configuration is divided into a uniform annular array and a uniform linear array; when the array device is configured as a uniform annular array, firstly, a circular harmonic transformation is performed on the array to obtain a virtual linear array; then, the array device configured as a uniform linear array or a virtual linear array is divided into sub-arrays and a forward and backward spatial smoothing operation is performed, and the smoothed data covariance matrix is ​​diagonally loaded; then, MPDR or LCMP beamforming is performed, and the real part of the sequence is taken as the output of the beamformer; in is the beamforming weight vector, R a,sym is the data matrix of the sub-array symmetric about the origin of the array coordinate system, real(·) represents the real part operation, (·) T represents the transpose operation, N sample Indicates the number of points in a frame of data; Step 6: Perform FFT-based rotation noise frequency estimation on the output of the beamformer in step 5; Step 7: Output the constructed matrix equation according to the noise frequency estimation value, the multi-frequency signal expression and the beamformer in step 6; Multi-frequency signal expression: where a n and The frequencies are f n The amplitude and phase of the signal component corresponding to the fan rotation noise frequency f n It is estimated from step 6, where The action time is constructed as a moment vector: Based on the least squares method, construct the parameter matrix equation: Solving Matrix Equations Here is the final signal synthesized from all frequency points of formula (1), where Amplitude a n for u n and v n The 2-norm of and The inverse function of the symbol and numerical solution of the unambiguous phase Based on the frequency, amplitude, phase and time vector of the rotating noise, the reference signal of the fan rotating noise is constructed in combination with the timer and the algorithm duration setting value. Where "+T maxdelay " means to shift the sine sequence to the left to compensate for the delay caused by audio sampling, "+T ps "One item represents the left shift of the sequence to compensate for the time consumption of the reference signal construction algorithm; Step 8: The timer calculates the elapsed time t of the algorithm process at this time elapse After the process sleeps for the first time period, it outputs the average value of the noise frequency estimation value of each channel in step 3 as the initialization value of the data container in step 1 and waits for the next call, wherein the first time period is T ps -t elapse ; Output the fan rotation noise reference signal obtained in step 7 and return, waiting for the next call, and the next call starts from step 2.

2. The method for constructing a reference signal of a fan duct ANC system based on a microphone array according to claim 1, characterized in that: The specific steps of the FBSS-based MPDR or LCMP beamforming in step 5 are as follows: Step 5.1: The array device is configured as a uniform linear array, and MPDR or LCMP beamforming based on forward and backward spatial smoothing FBSS is performed; it is configured as a uniform annular array, and circular harmonic transformation is first performed to transform it into a virtual linear array; the circular harmonic transformation matrix T is defined as: Where n = -N, -N+1, ..., 0, ... N-1, N is the circular harmonic number, N is the maximum circular harmonic order selected; m = 0, 1, ..., M-1, M is the number of array elements; is the azimuth angle of the mth element of the uniform annular array in the array coordinate system; the circular modal response elimination matrix J is defined as: Among them J n (kr) is the Bessel function of the first kind, k is the wave number vector The modulus value, r is the radius of the uniform annular array, and j is the imaginary unit; The process of converting the array data domain to the circular harmonic data domain is: Where (·)H represents the conjugate transpose operation, is the array manifold vector under the array data domain, After array data preprocessing, the data sequence of each channel exists in the data matrix in analytical form. and R CH They are the array manifold vector and data matrix of the uniform annular array in the circular harmonic domain; Step 5.2: forward and backward spatial smoothing: First, the uniform linear array or the virtual linear array formed by the circular harmonic transformation is divided into sub-arrays, where the sampling data covariance matrix of the i-th sub-array is: Where R ai is the data matrix of the i-th sub-matrix; The spatial smoothing process is: Where K+1 is the number of sub-arrays, and the final forward and backward spatial smoothing sampling data covariance matrix is ​​obtained by introducing forward and backward averaging: Among them J s is a switching matrix with the same order as the number of sub-array elements, with 1 on the secondary diagonal and 0 on the rest of the positions; C x,ss * Indicates C x,ss Elements take conjugate, J s C x,ss * J s Also known as the backward sampling data covariance matrix; Step 5.3: Load Cx, fbss diagonally; the loading method is: in It is called diagonal loading, I is the same as C x,fbss Identity matrix of the same order; Step 5.4: Design the beamformer weight vector, where the MPDR optimal weight vector is: Where represents the array manifold vector in the desired direction, L represents the number of sub-array elements; The optimal weight vector of LCMP is: where is called the constraint value vector and F is called the constraint matrix.

3. The method for constructing a reference signal of a fan duct ANC system based on a microphone array according to claim 1 or 2, characterized in that: In the FFT-based rotation noise frequency estimation process in step 6, the original data sequence is windowed for FFT transformation; the modulus of the first half of the FFT sequence is taken and the sequence index is converted into the actual frequency according to the FFT frequency resolution; the maximum value point of the modulus sequence or the maximum value point that satisfies both the height threshold condition and the distance threshold condition and one point before and after it are taken for secondary interpolation; when the window added by the FFT transformation is a Gaussian window, the natural logarithm of the modulus value of the taken point is first taken; then the quadratic interpolation method is used to calculate the fan rotation noise frequency estimation value in combination with the taken points; the fan rotation noise frequency estimation value is directly output as the frequency estimation value obtained by quadratic interpolation or is output after sliding averaging; after the frequency estimation of the first frame of data is completed, the timer obtains the elapsed time t of the algorithm process at this time elapse , the process sleeps for the first period of time and then returns to wait for the next call.

4. The method for constructing a reference signal of a fan duct ANC system based on a microphone array according to claim 1, characterized in that: The maximum time T ps The set value is smaller than the maximum sampling delay of one frame of audio data.

Images