An improved LMS-based variable order filter and sparse system identification method

Abstract

The application relates to a variable order length filter and a sparse system identification method based on an improved LMS, and belongs to the technical field of communication. The variable order length filter and the sparse system identification method based on the improved LMS algorithm, an improved algorithm in which two identification strategies are coordinated with each other is provided. On the basis of the FT-LMS variable order length adaptive filter algorithm with variable error width, in the application scene of the sparse channel environment in which a signal actually propagates, the coefficients of an estimated filter are segmented for zero attraction, a large penalty is adopted for the coefficients close to zero, the attraction strength is gradually decreased from the small coefficient to the large coefficient, a larger sparse coefficient change range is covered, and better performance is achieved for the application of the sparse channel. Therefore, under the premise of having the order length identification function, the application can be more targetedly applied to an echo suppression scene.

Description

Technical Field

[0001] This invention belongs to the field of communication technology and relates to a variable-length filter and sparse system identification method based on an improved LMS. Background Technology

[0002] In practical applications, sound reinforcement systems often experience acoustic feedback, which significantly limits their performance. Acoustic feedback occurs when an audio signal, after being played by a loudspeaker, is reflected by objects in the room and then picked up again by a microphone, converted into an electrical signal, processed, and finally played back by the loudspeaker. In a real sound field, this creates a positive feedback loop of "loudspeaker-microphone-loudspeaker," resulting in the simultaneous amplification of both the audio signal and the echo. Furthermore, when self-oscillation occurs at certain frequencies, a howling phenomenon occurs. Howling not only affects the quality of the output signal, causing distortion, but can also damage the sound reinforcement system due to excessive output power, even burning out the power amplifier. Adaptive acoustic feedback cancellation uses an adaptive filter that adjusts its weights based on the input signal to model and identify the acoustic feedback path in the sound field. This allows for the estimation of the acoustic feedback signal picked up by the microphone, which is then removed from the input signal, thus suppressing the echo. Theoretically, if the acoustic feedback path estimation is accurate enough, adaptive acoustic feedback cancellation can completely eliminate the echo.

[0003] For a communication channel with an unknown channel order, choosing a large tap length reduces convergence speed at the cost of computational complexity. Conversely, choosing a small tap length results in a large steady-state mean square error. Therefore, it is necessary to adaptively adjust the tap length to its optimal value. YuGong et al. proposed a variable tap length algorithm using the concept of pseudo-fractional tap length. This algorithm updates the tap length using instantaneous errors and has been widely applied. Most subsequent improvements to the algorithm are based on the FT-LMS algorithm. LiNing et al. proposed an improved fractional tap length LMS (FT-LMS) algorithm, which achieved better tap length convergence speed and steady-state tap length deviation. In 2020, Pauline SH et al. proposed a nonparametric variable step size NLMS algorithm with variable tap length. This combination of variable step size and variable tap length provides faster convergence and reduced steady-state mean square error, while overcoming the shortcomings of the traditional FT-LMS algorithm, which has many parameters to be adjusted and limited practical applicability. However, in most wireless communication channels, echo channels exhibit sparsity, with most coefficients being equal to or close to zero, and a small portion having very large coefficients. To address the problem of adapting to sparse channels, researchers have proposed many zero-attraction algorithms. Considering area, power consumption, and data processing speed, l0-norm type algorithms are more suitable for hardware implementation.

[0004] When implementing adaptive filtering algorithms in hardware, hardware resource consumption and complexity significantly impact processing speed. For echo suppression systems, fast convergence is crucial to prevent self-excited howling. Therefore, based on the FT-LMS variable-order-length adaptive filtering algorithm with variable error width, a zero-attractivity sparse variable-order-length algorithm is proposed, which achieves better convergence for sparse channels. This reduces the filter's resource consumption for unknown channel order lengths while achieving better performance. Summary of the Invention

[0005] In view of this, the purpose of the present invention is to provide a variable order length filter and sparse system identification method based on improved LMS.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A variable-order-length filter and sparse system identification method based on improved LMS, the method includes the following steps:

[0008] Step 1: Obtain the power input from the loudspeaker system to the pickup device. The initial signal x(n) forms a zero-mean Gaussian input signal vector X(n) = [x(n) x(n-1)...x(n-L+1)] T , where n represents the number of signal sequences, and L is a filter length that does not match the sparse channel order length;

[0009] Step 2: Input the initial clean speech signal x(n) into the estimation filter, process it to obtain the output signal y(n) of the estimation filter, and input the input signal x(n) into the sparse communication channel to obtain the echo signal H. L T The echo signal is synthesized with zero-mean Gaussian white noise in the sparse communication channel to obtain the desired output signal d(n), where d(n) = H L T (n)X(n)+n(n), where H(n) is the coefficient of the sparse communication channel, L is the length of the sparse channel in the simulated echo environment setting, where most of the sparse channel has the characteristic of being equal to or close to zero, and n(n) is the power of Zero-mean Gaussian white noise;

[0010] Step 3: Process the obtained expected output signal d(n) and the output signal y(n) obtained by passing the input signal through the estimation filter to obtain the error signal e(n);

[0011] Step 4: Based on the error signal value, the length of the estimated filter is updated iteratively by the pseudo-fractional tap length, so that the estimated filter can identify the length of the echo sparse channel and the filter order length is dynamically adjusted, wherein the error width and the adaptive filter step size are also dynamically adjusted accordingly.

[0012] Step 5: The adaptively adjusted dynamic filter order length is further applied to sparse channel identification. The error signal e(n) guides the iterative update of the coefficients of the estimation filter. The sparse multi-coefficient switching method is used to enable the estimation filter to identify the coefficients of the echo communication channel. Finally, the echo channel passing through the sparse channel is canceled to obtain a single output signal e(n). This output signal is used as an amplified signal and transmitted to the signal amplification end for amplified output.

[0013] The iterative update process of the adaptive filter in the improved MCSFT-LMS algorithm is as follows:

[0014]

[0015]

[0016]

[0017]

[0018]

[0019]

[0020]

[0021] Where, the zero attraction function f CS1 (W L(n) (n))=[f CS1 (W0(n))f CS1 (W1(n))…f CS1 (W L(n)-1 (n))] T A single function f in CS1 (W i (n) is defined as:

[0022]

[0023] Where y(n)=X L(n) T (n)W L(n) (n), e(n)=d(n)-y(n), W(n)=[w0 w1...w L(n)-1 ] TTo estimate the coefficients of the adaptive filter, β is the tuning parameter used to adjust the zero-attraction range of small coefficients in the sparse channel, φ is used to adjust the zero-attraction range of small and large coefficients, λ is used to adjust the zero-attraction range of even larger coefficients, μ(n) is the dynamic step size factor, μ' is a step size constant, and Δ(n) is the dynamic error width.

[0024] Optionally, in step 4, the dynamic error width is calculated using the following method:

[0025]

[0026]

[0027]

[0028] Where Δ max is a predetermined constant, is the upper bound of the error width, Δ(∞) is the final steady-state error width, the value of Δ(∞) is 1 / 10 of the optimal tap length, and P is a constant value calculated in advance;

[0029] The calculation method for step size change is as follows:

[0030]

[0031] Where 0 < μ' < 2, the step size parameter controls the convergence speed of the LMS algorithm; the stability boundary of the step size will decrease as the filter length increases; the constraint method of dynamic step size adjustment is used to make the step size dynamically adjusted according to the filter length and the input signal power.

[0032] Optionally, in step 5, the zero attraction function f CS1 The calculation characteristics are as follows:

[0033] In the attraction function, ρ is the strength of the attraction, β is the threshold parameter that distinguishes between small coefficients and zero coefficients, φ is the threshold parameter that distinguishes between small coefficients and large coefficients, and λ is the threshold parameter that distinguishes between even larger coefficients. Large attraction is applied to near-zero coefficients, and small attraction is applied to large coefficients in progressively decreasing order to maintain the large coefficient characteristics of the sparse channel. The different attraction methods of the three segments of the function are dynamically adjusted.

[0034] The beneficial effects of this invention are as follows: Based on an improved LMS algorithm and a variable-length filter and sparse system identification method, this invention proposes an improved algorithm in which the two identification strategies work synergistically. Building upon the FT-LMS variable-length adaptive filtering algorithm with variable error width, and in the application scenario of sparse channel environments where signals actually propagate, the coefficients of the estimated filter are subjected to piecewise zero attraction. A large penalty is applied to coefficients close to zero, and the attraction strength is gradually reduced from small to large coefficients, covering a wider range of sparse coefficient variations and resulting in better performance for sparse channel applications. Therefore, this invention, while possessing the function of length identification, can be more specifically applied to echo suppression scenarios.

[0035] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0036] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein:

[0037] Figure 1 This is the identification model for the adaptive filter system of the present invention;

[0038] Figure 2 This invention relates to the application of the present invention in an echo suppression system;

[0039] Figure 3 Steady-state MSD plots for adaptive filters of different orders;

[0040] Figure 4 To simulate the strain order and MSE curve convergence plot of an unknown 128th order sparse channel;

[0041] Figure 5 For LMS, the convergence plot of the MSE curve of the CS-LMS algorithm when the fixed order length is smaller than the real channel;

[0042] Figure 6 The simulation convergence plot shows the MCSFT-LMS algorithm inputting an 8kHz real speech signal. Detailed Implementation

[0043] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0044] The accompanying drawings are for illustrative purposes only and are schematic diagrams, not actual pictures. They should not be construed as limiting the invention. To better illustrate the embodiments of the invention, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0045] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components. In the description of the present invention, it should be understood that if terms such as "upper," "lower," "left," "right," "front," and "rear" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the terms used to describe positional relationships in the drawings are only for illustrative purposes and should not be construed as limiting the present invention. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.

[0046] like Figure 1 As shown, this embodiment of the invention discloses a variable-length filter and a sparse system identification method for an improved LMS algorithm. Figure 2 As shown, the speech signal v(n) emitted by the speaker passes through the echo path to generate an echo signal, which, together with the echo speech signal and noise signal b(n) from the microphone, constitutes the desired signal d(n). Furthermore, X(n) and the weights W(n) of the adaptive filter generate an estimated echo signal y(n). Subtracting this from d(n) yields the error signal e(n), which is then transmitted to the far room. If W(n) achieves an exact match with the impulse response H(n) of the echo path, y(n) is an accurate estimate of the actual echo signal; in this case, e(n) will not contain the echo signal. The essence of echo cancellation is to use an adaptive algorithm to adjust the filter coefficients to estimate the actual acoustic echo path, simulate the echo signal, and subtract it from the speech input from the near-end microphone, thereby canceling the echo.

[0047] In acoustic echo channels, the channel length through which the echo travels is unknown and fluctuates, and the echo path model often exhibits a long impulse response, requiring the use of high-order adaptive filters to model the echo path. This leads to a high computational cost for adaptive filtering algorithms and slows down convergence. For the LMS algorithm, for a communication channel with an unknown channel order, choosing a large tap length reduces convergence speed at the cost of computational complexity. Conversely, choosing a small tap length results in a larger steady-state mean square error, such as... Figure 3 As can be seen, when the channel order is set to 128, the steady-state MSD performance of the basic LMS algorithm with different filter orders was simulated. The filtering effect is significant as long as the filter order is greater than the channel order. However, as the filter order increases, the steady-state MSD performance decreases. Therefore, the filter order has a significant impact on the algorithm. Consequently, we study a variable-order adaptive filtering (FT-LMS) algorithm and combine it with a sparse channel to form the MCSFT-LMS algorithm.

[0048] In the prior art, the iterative update equation for the estimated filter coefficients of the FT-LMS algorithm is:

[0049]

[0050]

[0051]

[0052] The iterative update equation for the estimated filter coefficients of the FT-LMS algorithm with variable error width is:

[0053]

[0054]

[0055]

[0056] The iterative update equation for the estimated filter coefficients of the MCSFT-LMS algorithm proposed in this invention is as follows:

[0057]

[0058]

[0059]

[0060]

[0061]

[0062]

[0063]

[0064] in,

[0065]

[0066] In hardware implementation, the zero-attractivity subterm added in this invention, when combined with the variable-order algorithm, does not require additional real-time division and exponentiation operations compared to the LMS algorithm structure. This results in minimal increase in hardware resource consumption while improving MSE performance. High hardware resource consumption can slow down data processing, potentially leading to self-excited howling in echo suppression systems. Algorithm complexity often affects hardware data processing time; the iterative method of the zero-attractivity algorithm coefficient equation proposed in this invention is easy to implement in hardware and offers fast processing speed.

[0067] To better illustrate the advantages of this invention, the algorithm of this invention is used to identify the order of an unknown channel and simulate a sparse channel. In the simulation experiment, the minimum mean square error MSE = E[e 2 [n] is used as the criterion for channel estimation. The following experiments used 100 Monte Carlo runs to obtain the simulated values ​​for each data point. This invention selects the input signal power... The noise power is 1. The value is 0.01. Because the coefficients of real-world wireless sparse channels are random numbers, in the simulation experiment, the unknown sparse channel H = [0.3; zeros(L / 4-2,1); 0.6; -0.5; zeros(L / 2-2,1); zeros(L / 4-0,1); 0.8]. Considering that the length of the echo channel in a real sound reinforcement system is usually quite large, the length L of the unknown channel is set to 128. The length of the system estimation filter for FT-ILMS and MCSFT-LMS is set to 150, and the order length of the LMS filter with no variable order length is set to 128. Stationary and non-stationary simulations of the random generation of sparse channel coefficients are performed. The initial step size of the above algorithm is μ' = 0.5, ρ = 0.99, and for the variable order parameters α = 0.005, γ = 3, δ = 5, Δ... max =50, for the sparse attraction range parameter β=100.

[0068] from Figure 4 As can be seen from the graph, in a sparse channel with an unknown order length, the left side shows that the MSE performance of the MCSFT-LMS algorithm is superior to that of the FT-ILMS algorithm. The right side shows that in the filter order length convergence part, the performance is basically the same for the filter converging from the initial order length of 150 to the actual channel order length of 128, and the order length fluctuation of the MCSFT-LMS algorithm is more stable. From... Figure 5It can be seen that setting a fixed filter order length is unreasonable. When it is smaller than the actual order length, it cannot effectively identify unknown sparse channels, and the filter performance degrades significantly. Therefore, the initial filter order length should be set greater than the echo channel, and a variable order length method should be used for convergence. For real speech signals, such as... Figure 6 As shown, we establish a simulation model, using an 8kHz real speech signal as the input of X(n) and the MCSFT-LMS algorithm as the update criterion for the estimation filter. It can be seen that the estimation filter can identify the echo signal passing through the unknown sparse channel very well. The estimated signal cancels out the echo part of the reference signal d(n). Therefore, the error signal e(n) will be almost equal to the noise, making the error signal tend to converge.

[0069] This invention addresses the self-oscillation problem in sound reinforcement systems with large-order echo channels and the echo phenomenon generated when a microphone receives amplified signals from a power amplifier. It can further approximate the unknown channel order, reducing hardware resource consumption and enabling lossless amplification of sound signals.

[0070] Therefore, it is understandable that the method of this invention can effectively identify the coefficients of the echo sparse channel and solve the problem of unknown channel order length in unknown systems. By adding a zero attractor to the variable order adaptive filter, the real echo sparse channel can be better identified, resulting in lower resource consumption, higher hardware stability, and lower risk when the algorithm processes data in hardware, which is beneficial for the implementation of the algorithm in hardware.

[0071] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A variable-order-length filter and sparse system identification method based on improved LMS, characterized in that: The method includes the following steps: Step 1: Obtain the power input from the loudspeaker system to the pickup device. The initial signal x(n) forms a zero-mean Gaussian input signal vector X(n) = [x(n) x(n-1) ... x(n-L+1)] T , where n represents the number of signal sequences, and L is a filter length that does not match the sparse channel order length; Step 2: Input the initial clean speech signal x(n) into the estimation filter, process it to obtain the output signal y(n) of the estimation filter, and input the input signal x(n) into the sparse communication channel to obtain the echo signal H. L T The echo signal is synthesized with zero-mean Gaussian white noise in the sparse communication channel to obtain the desired output signal d(n), where d(n) = H L T (n)X(n)+n(n), where H(n) is the coefficient of the sparse communication channel, L is the length of the sparse channel in the simulated echo environment setting, where most of the sparse channel has the characteristic of being equal to or close to zero, and n(n) is the power of Zero-mean Gaussian white noise; Step 3: Process the obtained expected output signal d(n) and the output signal y(n) obtained by passing the input signal through the estimation filter to obtain the error signal e(n); Step 4: Based on the error signal value, the length of the estimated filter is updated iteratively by the pseudo-fractional tap length, so that the estimated filter can identify the length of the echo sparse channel and the filter order length is dynamically adjusted, wherein the error width and the adaptive filter step size are also dynamically adjusted accordingly. Step 5: The adaptively adjusted dynamic filter order length is further applied to sparse channel identification. The error signal e(n) guides the iterative update of the coefficients of the estimation filter. The sparse multi-coefficient switching method is used to enable the estimation filter to identify the coefficients of the echo communication channel. Finally, the echo channel passing through the sparse channel is canceled to obtain a single output signal e(n). This output signal is used as an amplified signal and transmitted to the signal amplification end for amplified output. The iterative update process of the adaptive filter in the improved MCSFT-LMS algorithm is as follows: Where, the zero attraction function f CS1 (W L(n) (n))=[f CS1 (W0(n))f CS1 (W1(n))…f CS1 (W L(n)-1 (n))] T A single function f in CS1 (W i (n) is defined as: Where y(n)=X L(n) T (n)W L(n) (n), e(n)=d(n)-y(n), W(n)=[w0 w1 ... w L(n)-1 ] T To estimate the coefficients of the adaptive filter, β is the tuning parameter used to adjust the zero-attraction range of small coefficients in the sparse channel, φ is used to adjust the zero-attraction range of small and large coefficients, λ is used to adjust the zero-attraction range of even larger coefficients, μ(n) is the dynamic step size factor, μ' is a step size constant, and Δ(n) is the dynamic error width.

2. The method for identifying sparse systems based on an improved LMS-based variable-order-length filter according to claim 1, characterized in that: In step 4, the dynamic error width is calculated as follows: Where Δ max is a predetermined constant, is the upper bound of the error width, Δ(∞) is the final steady-state error width, the value of Δ(∞) is 1 / 10 of the optimal tap length, and P is a constant value calculated in advance; The calculation method for step size change is as follows: Where 0 < μ' < 2, the step size parameter controls the convergence speed of the LMS algorithm; the stability boundary of the step size will decrease as the filter length increases; the constraint method of dynamic step size adjustment is used to make the step size dynamically adjusted according to the filter length and the input signal power.

3. The method for identifying sparse systems based on an improved LMS-based variable-order-length filter according to claim 1, characterized in that: In step 5, the zero attraction function f CS1 The calculation characteristics are as follows: In the attraction function, ρ is the strength of the attraction, β is the threshold parameter that distinguishes between small coefficients and zero coefficients, φ is the threshold parameter that distinguishes between small coefficients and large coefficients, and λ is the threshold parameter that distinguishes between even larger coefficients. Large attraction is applied to near-zero coefficients, and small attraction is applied to large coefficients in progressively decreasing order to maintain the large coefficient characteristics of the sparse channel. The different attraction methods of the three segments of the function are dynamically adjusted.

Images