Active noise control method and system of self-adaptive combined weight

Abstract

The invention discloses an active noise control method and system of self-adaptive combined weights. The method comprises the following steps: pre-training a control filter; inputting a signal acquired by the reference microphone into each control filter to obtain an output signal; weighted summation is carried out on the combined weight and the output of all the control filters to generate a control signal, so that a loudspeaker is driven to send out an anti-noise signal to cancel interference, and residual noise is collected and fed back by an error microphone; and the instantaneous square error is minimized through an LMS algorithm, the combined weight is iteratively updated, and a quiet area is formed near an error microphone. The method has the characteristic of rapid convergence, achieves good balance between calculation complexity and noise reduction performance, can rapidly adapt to a changing noise environment, and remarkably improves the noise reduction effect and application flexibility of an active noise control system.

Description

Technical Field

[0001] The invention relates to the technical field of noise control, and in particular to an active noise control method and system of adaptive combined weights. Background Art

[0002] The adverse effects of noise pollution on people's physical and mental health and environmental quality are becoming increasingly serious, so noise control is particularly important in various fields. Noise control technology is mainly divided into two categories: passive noise control (PNC) and active noise control (ANC). Passive noise reduction relies on the sound absorption performance of materials and structural design to attenuate noise, especially in the suppression of medium and high frequency noise, but it is still insufficient in the treatment of low-frequency noise. In contrast, active noise reduction is based on the principle of "destructive interference". It generates reverse sound waves with opposite phases and equal amplitudes to the noise to offset the noise waves and achieve noise reduction effects. Due to its excellent performance in low-frequency noise control, ANC technology has been widely used in automotive cockpits, noise-cancelling headphones, smart headrests and other fields, becoming a key technology and research hotspot in modern noise management.

[0003] In traditional ANC systems, the Filters-x Least Mean Square (FxLMS) algorithm is widely used due to its high computational efficiency, and its derivative methods are also often used in engineering practice. However, the convergence speed of this type of adaptive algorithm is slow, which limits the ability of the ANC system to handle dynamic or non-stationary noise. In order to improve the responsiveness and robustness of the system, the ANC method of fixed control filters has been widely used in practice. However, fixed control filters are usually targeted at a specific type of noise, and the effect is greatly reduced when dealing with other types of noise. To solve this problem, the prior art proposes a generative fixed-filter active noise control (GFANC) method. This method uses a convolutional neural network (CNN) to generate corresponding control filters for different noises, and the parameters can be automatically learned from a synthetic noise dataset without manual intervention. However, the GFANC method requires the use of a labeled noise dataset to supervise the training of the CNN, and the collection of labeled noise data often consumes a lot of resources and has the risk of labeling errors. Therefore, the existing technology applies unsupervised learning to the GFANC method. The proposed unsupervised-GFANC method can train CNN using an unlabeled noise dataset, eliminating the labeling process and further improving the noise reduction performance. Although the above methods show excellent noise reduction capabilities, they still rely on CNN training. Neural network training requires a lot of data and computing resources, and requires careful design of the network structure and initial parameter values ​​to ensure model convergence and stability. In addition, GFANC and unsupervised-GFANC cannot handle noise within the first second because these methods only update the control filter once per second. Summary of the invention

[0004] In order to solve the deficiencies in the prior art, the present invention provides an active noise control method and system with adaptive combined weights, which is particularly suitable for scenarios facing rapidly changing or non-stationary noise. In a dynamic noise control environment, the method has good tracking ability and robustness, overcoming the problem of slow convergence speed of the traditional FxLMS algorithm; in addition, the method has low computational complexity, which enhances the practicality of the technology.

[0005] The present invention provides an active noise control method and system with adaptive combined weights, which is mainly divided into two stages:

[0006] 1) Pre-training stage

[0007] The pre-training phase is completed offline, so there is no additional computational cost. Specifically, the FxLMS algorithm is used to pre-train the control filter using Gaussian white noise of different frequency bands as reference signals to generate control filters suitable for various noise types. The parameters of these filters are automatically learned through noise data without manual intervention, and are stored in the control filter library after training, providing prior information support for real-time control.

[0008] 2) Real-time control stage

[0009] In the real-time control stage, the system inputs the signal collected by the reference microphone into the pre-trained control filter, and uses the combined weights to weight the sum of all filter outputs to generate a control signal to drive the speaker to emit an anti-noise signal, thereby achieving effective noise reduction. At the same time, the instantaneous square error is minimized by the LMS algorithm, and the combined weights are iteratively updated to adapt to different noise environments. In a rapidly changing noise environment, this method can dynamically adjust the combined weights to achieve a stable noise reduction effect, thereby forming a quiet area around the error microphone.

[0010] Compared with the prior art, the beneficial effects of the present invention include at least:

[0011] Through the optimized LMS algorithm, the system can update the combined weights at every moment, and use the weights to perform linear correction on the output of the pre-trained control filter, thereby generating a control signal for noise control. The system structure of the present invention is complete, the algorithm is efficient, and it not only ensures low iteration time and computational complexity, but also achieves fast convergence. Especially in a dynamic noise environment, the system can converge quickly and show excellent noise reduction performance, providing a strong guarantee for the reliability of the ANC system.

[0012] In addition, the method of the present invention does not require training of a neural network and only requires a small amount of prior data (pre-training control filter) to process rapidly changing or non-stationary noise, and is suitable for a wide range of ANC application scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments recorded in the embodiments of the present invention. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0014] Figure 1 A flow chart of the active noise control method of the adaptive combined weights provided by the present invention;

[0015] Figure 2A principle block diagram of the adaptive combined weight method in the online control stage provided by the present invention;

[0016] Figure 3 The algorithm block diagram of FxLMS used in the offline training stage provided by the present invention;

[0017] Figure 4 The performance comparison diagram of controlling street noise by the method provided by the present invention, unsupervised-GFANC and FxLMS methods under simulation conditions; (a)-(c) are waveform diagrams before and after noise reduction by the three methods; (d) is a comparison diagram of noise attenuation levels per second;

[0018] Figure 5 This is a comparison of the power spectral density of street noise simulated using different methods. DETAILED DESCRIPTION

[0019] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0020] The embodiment of the present invention provides an active noise control method with an adaptive combination weight, the principle of which is as follows: Figure 2 As shown in the figure, this method generates a control signal by weighting the combined weights and the output of the pre-trained control filter to achieve real-time noise control; at the same time, the error microphone collects residual noise to update the combined weights. Where P(z) is the transfer function of the primary path from the reference microphone to the error microphone, and S(z) is the transfer function of the primary path from the secondary sound source to the error microphone, which is estimated as W m and g m denote the mth pre-trained control filter and the mth combination weight respectively, x(n) is the reference signal obtained by the reference microphone, d(n) is the interference signal obtained after x(n) passes through the primary path, and y m is the output signal of the mth pre-trained control filter, y(n) is the control signal used to drive the speaker, y′(n) is the anti-noise signal obtained after y(n) passes through the secondary path, and e(n) is the error signal obtained by superimposing the interference signal d(n) and the anti-noise signal y′(n).

[0021] Specific as Figure 1-Figure 3 As shown, the method comprises the following steps:

[0022] S1: Pre-trained control filter.

[0023] S1 is the offline training phase, which includes:

[0024] S11: setting the number M of pre-trained control filters, and then artificially synthesizing M Gaussian white noise signals in different frequency ranges;

[0025] S12: Set the length L of the pre-trained control filter c , the synthesized Gaussian white noise signal is used as the reference input, and the Filters-x Least Mean Square (FxLMS) algorithm is used for training to obtain the optimal control filter corresponding to each noise; finally, all pre-trained control filters will be stored in the control filter library.

[0026] S1 is the key step in the adaptive combined weighted active noise control method of the present invention, based on Figure 3 The FxLMS algorithm shown in the figure controls the update equation of the filter coefficients as follows:

[0027] w(n+1)=w(n)+μ w e(n)x′(n)

[0028] Among them, w(n+1) and w(n) represent the control filters at the n+1th iteration and the nth iteration respectively, and their lengths are L c , e(n) represents the error signal collected by the error microphone at the nth iteration, u w Represents the step size, which is used to control the update rate of the filter coefficients and should meet the following conditions:

[0029]

[0030] Among them, λ max is the autocorrelation matrix The largest eigenvalue.

[0031] The expression of the filtered reference signal vector is:

[0032]

[0033] After the algorithm converges, multiple control filters for different noise types can be obtained.

[0034] S2: Input the signal collected by the reference microphone into each control filter to obtain an output signal;

[0035] At the nth iteration, the reference signal is expressed as:

[0036] x(n)=[x(n),x(n-1),…,x(nL c +1)] T

[0037] Where x(n) represents the reference signal vector at the nth iteration, L c represents the length of the pre-trained control filter, (·) T Represents a transpose operation.

[0038] The control filter obtained by S1 pre-training processes the reference signal to obtain the respective output signals. The output signal expression is:

[0039] y m (n) = x T (n)W m m∈[1,M]

[0040] Among them, y m (n) represents the output signal of the mth pre-trained control filter, W m represents the mth pre-trained control filter, and M represents the number of pre-trained control filters.

[0041] S3: The combined weights are weighted and summed with the outputs of all control filters to generate a control signal, thereby driving the speaker to emit an anti-noise signal to cancel out the interference, and the residual noise is collected and fed back by the error microphone.

[0042] The control signal is the inner product of the output signal vector and the combination weight vector:

[0043] y(n)=y T (n)g(n)

[0044] The output signal vector is y(n)=[y1(n),y2(n),…,y M (n)] T , the combined weight vector is g(n)=[g1(n),g2(n),…,g M (n)] T , y(n) represents the control signal at the nth iteration;

[0045] The interference signal and the anti-noise signal obtained after the control signal passes through the secondary path are superimposed to obtain an error signal, which can be expressed as:

[0046] e(n)=d(n)-s(n)*y(n)

[0047] Where d(n) and s(n) represent the impulse responses of the interference signal and the secondary path, respectively; * is the linear convolution.

[0048] S4: Minimize the instantaneous square error through the LMS algorithm, iteratively update the combination weights, and form a quiet area near the error microphone.

[0049] The instantaneous square error, i.e., the loss function, can be expressed as:

[0050] J(n)=e 2 (n)

[0051] Using the gradient descent method, update the combined weight vector with a step size μ in the negative gradient direction:

[0052]

[0053] Among them, ▽J(n) is the instantaneous estimate of the mean square error gradient at time n, which can be expressed as:

[0054] ▽J(n)=-2y′(n)e(n)

[0055] The expression of the filtered output signal vector is:

[0056]

[0057] In practical ANC applications, the secondary path is unknown and needs to be estimated through an additional filter. Therefore, the filtered output signal vector is generated by estimating the output signal vector through the secondary path, where is the estimated value of s(n);

[0058] Finally, the update equation of the combined weight is:

[0059] g(n+1)=g(n)+μe(n)y′(n)

[0060] Among them, g(n+1) is the combined weight vector at time n+1, g(n) is the combined weight vector at time n, e(n) is the error signal at time n, y′(n) is the filtered output signal vector at time n, and μ is the convergence step size, which affects the convergence and steady-state performance of the algorithm. If the step size is too small, the convergence speed of the algorithm will slow down, resulting in a decrease in noise reduction performance; if the step size is too large, the algorithm will diverge. Therefore, μ should meet the following conditions:

[0061]

[0062] Where M is the number of pre-trained control filters, λ k and λ max Represents the autocorrelation matrix The kth eigenvalue and the largest eigenvalue of .

[0063] It is worth noting that the adaptive combination weight algorithm uses the filtered output signal y′(n) to update the combination weight, while the FxLMS algorithm uses the filtered reference signal x′(n) to update the control filter coefficients; since the dimension of y′(n) is usually significantly smaller than that of x′(n), the algorithm can significantly reduce the amount of calculation in the adaptive process;

[0064] The computational complexity is a key issue that needs to be considered in practical applications of adaptive algorithms. For real-time platforms, such as digital signal processors (DSPs), microcontrollers (MCUs), or field programmable gate arrays (FPGAs), the main computational burden of algorithm execution comes from multiplication, division, and addition or subtraction operations. The computational complexity of the two algorithms is shown in Table 1. a and L s denote the length of the adaptive filter and the estimated secondary path, L c represents the length of the pre-trained control filter, and M represents the number of pre-trained control filters.

[0065] Table 1 Computational complexity of the two algorithms

[0066]

[0067] It should be noted that the adaptive combined weight algorithm of the present invention usually only requires a small number of pre-trained control filters to achieve a good noise reduction effect. Therefore, compared with the traditional FxLMS algorithm, the algorithm achieves a better balance between computational cost and noise reduction performance. For example, when the estimated secondary path length is 256, the lengths of the adaptive filter and the control filter are 1024 and 256 respectively, and when the number of control filters does not exceed 8, the computational complexity of the method of the present invention is lower than that of the FxLMS algorithm based on the adaptive filter.

[0068] The present invention also provides an embodiment of an active noise control system of adaptive combined weights, which is used to implement the above-mentioned active noise control method of adaptive combined weights, and comprises:

[0069] Offline training module: used to pre-train control filters. Gaussian white noise in different frequency ranges is used as reference signals. Multiple control filters are trained using the FxLMS algorithm and saved.

[0070] Online control module: used to weightedly sum the combined weights with the output signal of the pre-trained control filter to generate a control signal, thereby achieving noise reduction; at the same time, this module uses the LMS algorithm to update the combined weights through the error signal and the filtered output signal to achieve adaptive adjustment to different noise environments to obtain the optimal noise reduction effect.

[0071] The application effect of the present invention is further described below in conjunction with simulation:

[0072] 1) Simulation conditions

[0073] The sampling rate of the system is 16kHz, and the frequency ranges of the noise in the training phase are 20-4000Hz, 4000-7980Hz, and 20-7980Hz. The primary and secondary paths are both 20-7980Hz bandpass filters with lengths of 512 and 256, respectively. The lengths of the adaptive filter and the pre-trained control filter are 1024 and 256, respectively, and the step sizes of the FxLMS algorithm and the adaptive combination weight algorithm are 0.0001 and 0.01, respectively. The noise signal to be controlled is street noise collected in the real world.

[0074] 2) Simulation results

[0075] Figure 4 (a) Figure 4 (b) and Figure 4 (c) shows the noise waveforms before and after controlling street noise using three methods: adaptive combined weights, unsupervised generation of fixed filters, and filtered x least mean squares; Figure 4 (d) shows the comparison of noise attenuation levels per second for the three methods.

[0076] from Figure 4 It can be seen that compared with the filtering x-minimum mean square and the unsupervised generation of fixed filter methods, the adaptive combination weight method of the present invention has significantly improved convergence speed and noise reduction performance. In addition, as an adaptive algorithm, the present method can update the combination weight at every moment, thereby overcoming the limitation that the generation of fixed filter methods cannot process the first second noise.

[0077] Figure 5 The comparison of the noise power spectrum density before and after the three methods are used to control street noise is shown. The results show that the proposed method has a good noise reduction effect in the entire frequency range and performs better than the other two methods.

[0078] A person skilled in the art can understand that all or part of the steps in the various methods of the above embodiments can be completed by instructing the relevant hardware through a program, and the program can be stored in a computer-readable storage medium, which can include: ROM, RAM, disk or CD, etc.

[0079] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. An active noise control method with adaptive combined weights, characterized in that: include: S1: pre-trained control filter; S2: Input the signal collected by the reference microphone into each control filter to obtain an output signal; S3: The combined weights are weighted and summed with the outputs of all control filters to generate a control signal, thereby driving the speaker to emit an anti-noise signal to cancel out the interference, and the residual noise is collected and fed back by the error microphone; S4: Minimize the instantaneous square error through the LMS algorithm, iteratively update the combination weights, and form a quiet area near the error microphone.

2. The active noise control method of the adaptive combined weight according to claim 1, characterized in that: In S1, the pre-training control filter includes: S11: setting the number M of pre-trained control filters, and then artificially synthesizing M Gaussian white noise signals in different frequency ranges; S12: Set the length L of the pre-trained control filter c , the synthesized Gaussian white noise signal is used as the reference input, and the Filters-x Least Mean Square (FxLMS) algorithm is used for training to obtain the optimal control filter corresponding to each noise; finally, all pre-trained control filters will be stored in the control filter library.

3. The active noise control method of the adaptive combined weight according to claim 1, characterized in that: In S2, the signal collected by the reference microphone, i.e., the reference signal, is expressed as: x(n)=[x(n),x(n-1),,x(n-L c +1)] T Where x(n) represents the reference signal vector at the nth iteration, (·) T Represents the transpose operation; The expression of the output signal of the control filter is: y m (n)=x T (n)W m m∈[1,M] Among them, y m (n) represents the output signal of the Mth pre-trained control filter, W m represents the mth pre-trained control filter.

4. The active noise control method of adaptive combined weights according to claim 1, characterized in that: In S3, the control signal is the inner product of the output signal vector and the combination weight vector: y(n)=y T (n)g(n) The output signal vector is y(n)=[y1(n),y2(n),,y M (n)] T , the combined weight vector is g(n)=[g1(n),g2(n),,g M (n)] T , y(n) represents the control signal at the nth iteration; The error signal can be expressed as: e(n)=d(n)-s(n)*y(n) Where d(n) and s(n) represent the impulse responses of the interference signal and the secondary path, respectively; * is the linear convolution.

5. The active noise control method of adaptive combined weights according to claim 1, characterized in that: In S4, the instantaneous square error, i.e., the loss function, can be expressed as: J(n)=e 2 (n) Using the gradient descent method, update the combined weight vector with a step size μ in the negative gradient direction: Among them, ▽J(n) is the instantaneous estimate of the mean square error gradient at time n, which can be expressed as: The expression of the filtered output signal vector is: In practical ANC applications, the secondary path is unknown and needs to be estimated through an additional filter. Therefore, the filtered output signal vector is generated by estimating the output signal vector through the secondary path, where is the estimated value of s(n); Finally, the update equation of the algorithm is: g(n+1)=g(n)+μe(n)y′(n) Among them, g(n+1) is the combined weight vector at time n+1, g(n) is the combined weight vector at time n, e(n) is the error signal at time n, y′(n) is the filtered output signal vector at time n, and μ is the convergence step size, which affects the convergence and steady-state performance of the algorithm. If the step size is too small, the convergence speed of the algorithm will slow down, resulting in a decrease in noise reduction performance; if the step size is too large, the algorithm will diverge. Therefore, μ should meet the following conditions: Where M is the number of pre-trained control filters, λ k and λ max Represents the autocorrelation matrix The kth eigenvalue and the largest eigenvalue of .

6. The active noise control method of adaptive combined weights according to claim 1 or 5, characterized in that: The proposed algorithm, the adaptive combined weight algorithm, is based on the LMS algorithm. Although it is similar to the traditional FxLMS algorithm, it has two significant differences: 1) The adaptive combination weight algorithm uses the filtered output signal y′(n) to update the combination weights, while the FxLMS algorithm uses the filtered reference signal x′(n) to update the control filter coefficients; it is worth noting that the dimension of y′(n) is usually significantly smaller than that of x′(n), so the algorithm can significantly save computational effort in the adaptive process; 2) The algorithm of the present invention linearly corrects the output of the control filter by combining weights to generate a control signal; while the FxLMS algorithm directly uses the output of the control filter as the control signal.

7. An active noise control system with adaptive combined weights, used to implement the adaptive combined weight method according to any one of claims 1 to 6, characterized in that: include: Offline training module: used to pre-train control filters. Gaussian white noise in different frequency ranges is used as reference signals. Multiple control filters are trained using the FxLMS algorithm and saved. Online control module: used to weightedly sum the combined weights with the output signal of the pre-trained control filter to generate a control signal, thereby achieving noise reduction; at the same time, this module uses the LMS algorithm to update the combined weights through the error signal and the filtered output signal to achieve adaptive adjustment to different noise environments to obtain the optimal noise reduction effect.

Images