Distributed Active Noise Control Method Based on Frequency Domain Augmented Gradient Diffusion Algorithm

By using the frequency domain augmented gradient diffusion filter reference minimum mean square algorithm, the problem of high computational and communication resources in the existing technology is solved, realizing distributed active noise control with low resource requirements and fast convergence, while maintaining the multi-node collaborative noise reduction performance.

CN123072364APending Publication Date: 2026-11-10NATIONAL INSTITUTE OF METROLOGY CHINA
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611308536.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-27
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing time-domain augmented diffusion type distributed active noise control algorithms have high computational and communication resource requirements in sample-by-sample implementation, and reducing the communication frequency can easily slow down the convergence speed.

Method used

We employ a frequency-domain augmented gradient diffusion filter reference minimum mean square algorithm. By using fast Fourier transform block processing, we convert linear convolution and correlation operations into frequency-domain point-by-point operations. We also reconstruct the traditional two-round control filter coefficient communication into a single-round augmented gradient communication. By combining frequency-domain block processing and normalization techniques, we optimize the computation and communication burden.

Benefits of technology

While maintaining the asymmetric acoustic path, we achieved low computational and communication overhead, preserved multi-node collaborative noise reduction performance, and improved convergence speed.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123072364A_ABST
    Figure CN123072364A_ABST
Patent Text Reader

Abstract

This invention discloses a distributed active noise control method based on a frequency-domain augmented gradient diffusion algorithm, comprising the following steps: acoustic nodes perform block-based and frequency-domain transformations on the reference signal, error signal, and pre-identified secondary path model to construct a frequency-domain augmented filter reference signal matrix and a frequency-domain error signal matrix; calculate the frequency-domain augmented gradient spectrum matrix and the frequency-point-by-frequency normalized vector, and combine them with the node step size and combination coefficients to form a normalized weighted frequency-domain augmented gradient vector; receive the vectors from neighboring acoustic nodes through a single neighborhood communication and sum and combine them in the frequency domain; perform a frequency-domain to time-domain transformation on the combination result, extract the time-domain augmented gradient component, update the time-domain control filter coefficients, and generate a control acoustic signal. This method is used for real-time distributed active noise control on resource-constrained node processors, which can reduce computational load, communication data volume, and communication rounds, and improve convergence speed.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the fields of active noise control and distributed adaptive signal processing technology, specifically to a distributed active noise control method based on the frequency domain augmented gradient diffusion type filter reference minimum mean square algorithm. Background Technology

[0002] Multi-channel active noise control (ANC) systems utilize multiple error sensors to collect residual noise signals within the control area and drive multiple secondary sound sources to radiate control sound. This control sound is then superimposed on the primary noise to suppress low-frequency noise. Centralized control algorithms typically update all control filters using a global error signal, fully utilizing multi-channel information. However, the computational burden on the central processing unit increases significantly with the number of channels, the length of the control filters, and the length of the secondary path model. Distributed control algorithms, where each local processor updates its corresponding control filter based only on its local error signal, reduce the computational load on a single processor. However, by ignoring the coupling between acoustic local systems, this can lead to decreased noise reduction performance or even algorithm divergence.

[0003] Distributed active noise control (ANC) algorithms organize multi-channel ANC systems into multiple cooperating acoustic nodes. Each acoustic node performs local estimation and exchanges information with neighboring nodes through a communication network to achieve distributed processing of computational tasks and collaborative control among nodes. The diffused distributed algorithm employs a parallel communication topology based on neighboring node information exchange, featuring flexible communication link configuration, low inter-node waiting latency, and ease of parallel implementation. It is suitable as a framework for implementing collaborative updates of control filters among acoustic nodes in a distributed ANC system.

[0004] Classical multi-task diffusion-type filtering reference least mean square (MMS) algorithms improve system stability through neighborhood space regularization. However, this regularization can cause the control filters of different acoustic nodes to tend towards similar steady-state solutions. When the primary path, secondary path, or multi-channel acoustic coupling is asymmetrically distributed, the optimal control filters required by each node are usually different. Directly smoothing the neighborhood control filters spatially can introduce convergence bias. Augmented diffusion-type filtering reference least mean square (ANC) algorithms, through a neighborhood-based adaptive process and a node-based combination process, enable each acoustic node to estimate an augmented vector containing the neighborhood control filters and combine the estimation results of the control filters of the same node only. This reduces the aforementioned convergence bias and improves the steady-state noise reduction performance of distributed ANC algorithms under asymmetric acoustic path conditions.

[0005] Existing augmented-diffusion distributed ANC algorithms primarily employ a time-domain sample-by-sample iterative approach. Each acoustic node needs to complete the calculation of the filtered reference signal based on linear convolution, the correlation operation between the error signal and the filtered reference signal, and the adaptive update of the augmented control filter within a single sampling interval, exchanging control filter coefficients through forward and backward communication. As the number of acoustic nodes, the length of the control filter, and the length of the secondary path model increase, the sample-by-sample computation and the amount of control filter coefficient transmission by each node increase accordingly. Furthermore, the need for forward and backward communication in each iteration further burdens the real-time communication between nodes, thus limiting the real-time scaling of this algorithm on resource-constrained node processors.

[0006] To reduce the aforementioned communication burden, existing technologies employ intermittent or periodic communication methods, exchanging node information at preset time intervals to reduce the number of communications and the amount of communication data per unit time. However, reducing the information exchange frequency prolongs the update interval of neighboring node information, thereby slowing down the initial convergence speed of the algorithm and the reconvergence speed after changes in the acoustic path. Some real-time implementations also require offline-designed compensation filters or external coprocessors to complete data combination (J. Ji, D. Shi, X. Shen, W.-S. Gan, J. Chen, and J. Yang, “Implementation of Distributed Multichannel Active Noise Control With Intermittent Communication and Coprocessor Assisted Data Combination,” Mechanical Systems and SignalProcessing, vol. 248, Art. no. 114024, 2026). Therefore, simply increasing the communication interval to reduce communication overhead is insufficient to simultaneously achieve low resource consumption, fast convergence speed, and low hardware implementation complexity.

[0007] To address the computational burden of time-domain sample-by-sample updates, a Fast Fourier Transform (FFT) block processing approach can be employed. This transforms the linear convolution required to generate the filter reference signal and the related operations needed to calculate the control filter update into frequency-domain point-by-point operations, thereby reducing the average computational cost per sample point. On the other hand, regarding the communication burden caused by forward and backward communication, further analysis of the control filter update relationship in existing augmented-diffusion algorithms reveals that although bidirectional communication directly exchanges control filter coefficients, the information directly related to the current control filter update can be equivalently represented by the weighted augmented gradient calculated by each neighboring acoustic node based on the local error signal and the filter reference signal. Based on this equivalence, the traditional two-round control filter coefficient communication can be reconstructed into a single augmented gradient communication while maintaining the control filter update result unchanged. However, combining this communication reconstruction with frequency-domain block processing requires addressing issues such as the frequency-domain construction of the augmented filter reference signal, frequency-by-frequency normalization of the augmented gradient, the correspondence between the neighboring combination coefficients and the augmented gradient components, and the consistency between the multi-node frequency-domain augmented gradient combination and the time-domain control filter update. Summary of the Invention

[0008] The purpose of this invention is to provide a distributed active noise control method based on the frequency domain augmented gradient diffusion type filter reference least mean square algorithm, in order to solve the problems of high computational resource requirements in the sample-by-sample implementation of the existing time domain augmented diffusion type distributed ANC algorithm, high communication resource requirements caused by communication of the two rounds of control filter coefficients, and slow convergence when reducing communication frequency to reduce communication overhead. Thus, while maintaining the multi-node collaborative noise reduction performance under asymmetric acoustic path conditions, it achieves real-time distributed active noise control with low computational and communication overhead and fast convergence.

[0009] To achieve the above objectives, the technical solution adopted by the present invention is as follows: The distributed active noise control method based on the frequency domain augmented gradient diffusion algorithm is executed by the acoustic nodes in the distributed active noise control system, and includes: S1, acquire the reference signal, error signal, pre-identified secondary path model, and neighborhood node set corresponding to the target acoustic node of the target acoustic node; S2, perform block division and frequency domain transformation on the reference signal, error signal and pre-identified secondary path model to construct a frequency domain augmented filter reference signal matrix and a frequency domain error signal matrix; S3, multiply the conjugate matrix of the frequency domain augmented filter reference signal matrix element by element with the frequency domain error signal matrix to obtain the frequency domain augmented gradient spectrum matrix, and construct a normalized vector based on the frequency domain augmented filter reference signal matrix for each frequency point. S4. Based on the normalized vector, node step size, and combination coefficient, the corresponding gradient components in the frequency domain augmented gradient spectrum matrix are weighted to form a normalized weighted frequency domain augmented gradient vector for neighborhood communication. S5, the target acoustic node receives the normalized weighted frequency domain augmented gradient vector sent by the neighboring acoustic nodes through a neighborhood communication, and sums and combines the normalized weighted frequency domain augmented gradient vectors corresponding to the target acoustic node in the frequency domain of the target acoustic node and the neighboring acoustic nodes to obtain the combined frequency domain augmented gradient vector. S6, perform a frequency-domain to time-domain transformation on the combined frequency-domain augmented gradient vector, extract the time-domain augmented gradient component from the transformation result, update the time-domain control filter coefficients of the target acoustic node according to the time-domain augmented gradient component, and generate a control acoustic signal for driving the secondary sound source according to the updated time-domain control filter coefficients.

[0010] Furthermore, the distributed active noise control system includes multiple acoustic nodes, each acoustic node including a reference signal input terminal, an error sensor, a secondary sound source, and a node processor; the neighborhood node set includes the target acoustic node itself and acoustic nodes that have established communication relationships with the target acoustic node.

[0011] Furthermore, the target acoustic node performs local computation and neighborhood communication according to signal blocks, and adopts a block processing delay: during the current signal block, the current reference signal sample and the current error signal sample are cached, and the data corresponding to the previous signal block are used to complete the block and frequency domain transformation, frequency domain augmented gradient estimation, normalized weighted frequency domain augmented gradient vector construction, one neighborhood communication, frequency domain summation and combination, and time domain control filter update, so as to provide a complete signal block processing interval for block computation and neighborhood communication.

[0012] Furthermore, the segmentation and frequency domain transformation employ overlap-preserving segmentation processing: the reference signal is divided into overlapping blocks according to a preset block length and a preset overlap rate, and the pre-identified secondary path model is zero-padded, so that the overlapping blocks of the reference signal and the zero-padded pre-identified secondary path model have the same length. In a preferred embodiment, the block length is... L Time-domain controlled filter length H and the length of the pre-identified secondary path model R Equal, the frequency domain transformation is 2 L Point Fast Fourier Transform.

[0013] Further, the reference signal spectrum is element-wise multiplied with the corresponding pre-identified secondary path model spectrum and an inverse fast Fourier transform is performed to obtain a circular convolution sequence. Samples in the circular convolution sequence that do not belong to the linear convolution result are removed, and samples corresponding to the linear convolution result are retained. The linear convolution results of two adjacent signal blocks are combined in chronological order and a fast Fourier transform is performed to obtain the frequency domain augmented filter reference signal matrix corresponding to the control filter of the target acoustic node and its neighboring acoustic nodes. The error signal is padded before the block division. L 1 zero sample and perform 2 once L The frequency domain error signals are then stacked according to the number of nodes in the neighborhood node set to obtain the frequency domain error signal matrix.

[0014] Furthermore, the normalization vector includes a frequency power term and a low-energy frequency constraint term; the frequency power term is obtained by summing the squared magnitudes of the frequency-domain augmented filter reference signal matrix at each frequency along the augmented filter dimension, and the low-energy frequency constraint term is generated based on the average frequency power and regularization constant of the frequency-domain augmented filter reference signal matrix; the normalization vector is used to adjust the update amplitude of the frequency-domain augmented gradient at different frequencies and to limit the update amplitude of the low-energy frequencies.

[0015] Furthermore, the combination coefficients are applied to the corresponding gradient components of the frequency-domain augmented gradient spectrum matrix according to the neighboring acoustic nodes to be transmitted; each normalized weighted frequency-domain augmented gradient vector corresponds to a receiving acoustic node and serves as the data object for the first neighboring communication. The target acoustic node receives the normalized weighted frequency-domain augmented gradient vector corresponding to its time-domain control filter from each neighboring acoustic node, and sums and combines the normalized weighted frequency-domain augmented gradient vectors before performing the frequency-domain to time-domain transformation.

[0016] Furthermore, in the block length L Time-domain controlled filter length H In a preferred embodiment where the length R of the pre-identified secondary path model is equal, the frequency domain to time domain transformation is a single 2-1 transformation. L Point-wise inverse fast Fourier transform, the target acoustic node extracts the previous value corresponding to the length of the time-domain control filter from the transform result. L A time-domain augmented gradient component. The updated time-domain control filter filters the reference signal to generate a control acoustic signal, which drives the secondary sound source of the target acoustic node and is superimposed on the primary noise signal via the secondary path at the location of the error sensor.

[0017] This invention provides a distributed active noise control method suitable for real-time implementation on resource-constrained node processors. Compared with the prior art, this invention has at least the following advantages: (1) By using Fast Fourier Transform block processing, the linear convolution required to generate the filter reference signal and the correlation operation required to calculate the update amount of the control filter are converted into frequency domain point-by-point operation, reducing the average computation amount per sample point of the node processor; (2) By augmented gradient equivalent reconstruction, the traditional two-round control filter coefficient communication is rewritten into a normalized weighted frequency domain augmented gradient communication, and a frequency domain to time domain transformation is performed after frequency domain summation and combination, reducing the number of communication rounds, communication data volume and redundant inverse transformation; (3) By normalizing each frequency point, the update amplitude of each frequency point is adapted to the energy of the corresponding filtered reference signal, thereby improving the convergence speed; (4) By retaining the neighborhood control filter estimation relationship of the augmented diffusion algorithm, the multi-node collaborative noise reduction performance is maintained under the asymmetric acoustic path condition.

[0018] Therefore, this invention can reduce the computational and communication resource requirements of node processors while maintaining a relatively fast convergence speed without reducing communication overhead by extending the neighborhood communication interval. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the distributed active noise control method of the present invention.

[0020] Figure 2 This is a schematic diagram of the algorithm data flow from acoustic signal segmentation to control acoustic signal in the method of the present invention.

[0021] Figure 3 These are the time-domain denoising curve and steady-state error signal spectrum under band-limited white noise excitation in the embodiment, where (a) is the time-domain denoising curve and (b) is the steady-state error signal spectrum.

[0022] Figure 4 These are the time-domain noise reduction curve and steady-state error signal spectrum under factory noise excitation in the embodiment, where (a) is the time-domain noise reduction curve and (b) is the steady-state error signal spectrum.

[0023] Figure 5 These are the time-domain noise reduction curve and steady-state error signal spectrum under ship engine room noise excitation in the embodiment, where (a) is the time-domain noise reduction curve and (b) is the steady-state error signal spectrum.

[0024] Figure 6 These are the primary noise signal spectra and Wiener despectral spectra observed before and after speech interference in a ship engine room noise experiment. Detailed Implementation

[0025] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. It should be noted that the following embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the scope of protection of the present invention. Equivalent substitutions or transformations made by those skilled in the art, without departing from the concept of the present invention, regarding the number of acoustic nodes, neighborhood topology, block length, combination coefficients, the method of obtaining the pre-identified secondary path model, or the node processor platform, should all be included within the scope of protection of the present invention.

[0026] like Figure 1 As shown, the method of this invention is executed by each acoustic node in a distributed active noise control system according to signal blocks. Each acoustic node constructs a frequency-domain augmented filter reference signal matrix and a frequency-domain error signal matrix based on the reference signal, local error signal, and pre-identified secondary path model. It then calculates the frequency-domain augmented gradient spectrum matrix and performs frequency-point normalization to form a normalized weighted frequency-domain augmented gradient vector for each neighboring acoustic node. The target acoustic node receives the vector through a neighborhood communication, performs a summation and combination in the frequency domain, executes an inverse fast Fourier transform, extracts the time-domain augmented gradient component to update the time-domain control filter coefficients, and generates a control acoustic signal based on the updated coefficients.

[0027] Specifically, the method of the present invention includes the following steps: S1, obtain the reference signal, error signal, pre-identified secondary path model, and the set of neighboring nodes corresponding to the target acoustic node. S2, perform block-based and frequency-domain transformation on the reference signal, error signal and pre-identified secondary path model to construct the frequency-domain augmented filter reference signal matrix and the frequency-domain error signal matrix; S3, multiply the conjugate matrix of the frequency domain augmented filter reference signal matrix element by element with the frequency domain error signal matrix to obtain the frequency domain augmented gradient spectrum matrix, and construct a normalized vector based on the frequency domain augmented filter reference signal matrix for each frequency point. S4. Based on the normalized vector, node step size, and combination coefficient, the corresponding gradient components in the frequency domain augmented gradient spectrum matrix are weighted to form a normalized weighted frequency domain augmented gradient vector for neighborhood communication. S5, the target acoustic node receives the normalized weighted frequency domain augmented gradient vector sent by the neighboring acoustic nodes through a neighborhood communication, and sums and combines the normalized weighted frequency domain augmented gradient vectors corresponding to the target acoustic node in the frequency domain of the target acoustic node and the neighboring acoustic nodes to obtain the combined frequency domain augmented gradient vector. S6 performs a frequency-domain to time-domain transformation on the combined frequency-domain augmented gradient vector, extracts the time-domain augmented gradient component from the transformation result, updates the time-domain control filter coefficients of the target acoustic node based on the time-domain augmented gradient component, and generates a control acoustic signal for driving the secondary sound source based on the updated time-domain control filter coefficients.

[0028] The theoretical aspects involved in the above steps are detailed below.

[0029] Assume a distributed active noise control system includes J There are 5 acoustic nodes, each exchanging information with its neighboring acoustic nodes via a communication network. Each acoustic node includes a reference signal input, a secondary sound source, an error sensor, and a node processor; the reference signal is provided by a common reference sensor, a local reference input, or a signal generator. j Each node processor performs local signal processing and exchanges information with neighboring acoustic node processors. The neighboring node set includes the target acoustic node itself and the acoustic nodes that have established communication relationships with the target acoustic node.

[0030] At sampling time n , No. j The local error signal measured by each error sensor is formed by the superposition of the primary noise signal propagating to that location and the control sound components of all secondary sound sources propagating to that location through their respective secondary paths, and can be expressed as:

[0031] in, Indicates the first j Error signals measured by an error sensor This represents the primary noise signal propagating to the error sensor. Indicates the first k The secondary sound source propagates through the secondary path to the first... j The control acoustic component at each error sensor. J This represents the total number of acoustic nodes.

[0032] No. k The control sound signal output from the secondary sound source is transmitted to the first secondary sound source. j After secondary path propagation by an error sensor, it can be represented as:

[0033] in, Indicates by the first k Secondary sound sources to the first j One error sensor R Tap secondary path impulse response vector Indicates containing the first kThe nearest secondary sound source R The vector of output samples, with superscript T This indicates transpose.

[0034] Used to drive the k The control sound signal of the secondary sound source is obtained from the reference signal via the first... k The result is obtained by filtering with time-domain controlled filters, and can be expressed as:

[0035] in, This indicates that the current reference signal sample and its preceding samples are used to represent the signal. H -1 sample constitutes H 3D reference signal vector, Indicates the driver's first k Secondary sound sources H Tapped time-domain controlled filter coefficient vector.

[0036] Under the condition that the coefficients of the time-domain control filter remain approximately constant during the duration of the secondary path impulse response, the first k The secondary sound source propagates to the first j The control acoustic component at each error sensor can be rewritten as:

[0037] Substituting the above control acoustic components into the local error signal expression, we obtain the local error signal represented by the filtered reference signal vector and the coefficients of the time-domain control filter:

[0038] in, Indicates that the reference signal passes through the first k Secondary sound sources to the first j The pre-identification secondary path model of each error sensor is formed after filtering. H 3D filtered reference signal vector. The pre-identified secondary path model can be obtained through offline identification before control operation or online identification during control operation.

[0039] In the existing time-domain augmented diffusion filter reference minimum mean square algorithm, the first... j The cost function for each acoustic node includes a local error signal minimization term and a neighborhood space regularization term, and can be expressed as:

[0040] in, Indicates the first j The cost function for each acoustic node, where E{·} represents the mathematical expectation, Indicates the first j The set of neighboring nodes of the acoustic node, including the . jEach acoustic node itself; This represents the vector of augmented control filter coefficients formed by stacking neighboring node control filters sequentially. Indicates by the first k The first acoustic node inferred the j Each control filter is estimated. and The non-negative regularization coefficients are used to select the matrix. and A control filter used to extract the corresponding node from the augmented vector.

[0041] An adaptive post-combination strategy is adopted, and the local error signal minimization term is updated according to stochastic gradient descent to obtain the time-domain augmented adaptive relation:

[0042] in, Indicates the first j The intermediate estimates obtained by local augmentation and adaptation of acoustic nodes This represents the step size of the time-domain augmentation-diffusion algorithm. Indicates by the first j The augmented filter reference signal vector is formed by the filter reference signal vectors corresponding to each control filter in the neighborhood of each node.

[0043] To limit the update amplitude of the control filter and improve numerical stability, the above time-domain augmented adaptive relationship is written in a normalized form:

[0044] in, This is a regularization constant used to avoid the normalization denominator being too small.

[0045] In the traditional forward communication phase, the first... j The intermediate estimates of the augmented control filter formed by each acoustic node from its neighboring acoustic nodes are combined to update the _th ... j One time-domain controlled filter:

[0046] in, This represents the combination coefficient matrix extended to the dimension of the control filter. express Dimensional unit matrix express Dimensional combination coefficient vector.

[0047] In the traditional backward communication phase, each acoustic node receives the updated time-domain control filter coefficients from neighboring acoustic nodes to reconstruct the augmented control filter coefficient vector for the next sampling time.

[0048] in, This represents the global control filter coefficient vector composed of all node-controlled filters. This indicates the extraction of the first [value] from the global control filter coefficient vector. j The reconstruction matrix of the neighborhood control filter for each node. The augmented adaptation, forward communication, and backward communication described above are all performed on a single sampling point basis, resulting in a high per-sample computational burden and a two-round communication burden for the control filter coefficients.

[0049] To rewrite the above sample-by-sample iteration as block iteration, let the signal block length be... L The block number is m , No. m The first signal block in the i Each sample corresponds to a sampling time mL + i For a signal block L The time-domain augmented adaptive update is averaged to obtain:

[0050] The first j An acoustic node within a signal block L The local error signal samples are arranged in columns to construct the time-domain error signal matrix:

[0051] The first j The filter reference signals corresponding to each time-domain control filter in the neighboring acoustic nodes of each acoustic node are arranged in columns to construct a time-domain augmented filter reference signal matrix:

[0052] In the time-domain augmented filter reference signal matrix, with the first k The column vector corresponding to the time-domain control filter of each neighboring acoustic node is determined by the signal block. L Composed of several filtered reference signal vectors:

[0053] Using the aforementioned time-domain error signal matrix and time-domain augmented filter reference signal matrix, the block-domain time-domain augmented adaptive relationship can be written in the following compact matrix form:

[0054] in, and They represent the first jThe intermediate estimation matrix and coefficient matrix of the augmented control filter for each acoustic node, with the block step size already including the values ​​of... L Calculate the average scaling factor generated from each sample.

[0055] If in the 1st m The block update and neighborhood communication are only performed after each signal block is completed. Therefore, the corresponding calculations and communication need to be completed in the [number]th [number]th [number]. m The last sampling point of the first signal block and the first m The processing is completed between the first sampling point of +1 signal blocks. To provide a real-time processing interval for a complete signal block, this implementation adopts a block processing delay: in the first... m During each signal block, the current sample (including the current reference signal sample and the current error signal sample) is buffered, and the first sample is used... m -1 data blocks of signal are used to update the control filter, resulting in:

[0056] Under the block-based iteration condition, the traditional forward communication relationship can be expressed as:

[0057] The traditional backward communication relationship corresponding to the aforementioned forward communication relationship can be expressed as:

[0058] To reconstruct the two rounds of control filter coefficient communication into a single augmented gradient communication, the block-domain augmented adaptive relation is substituted into the traditional forward communication relation. Furthermore, the information directly related to the current control filter update is equivalently represented as the weighted augmented gradient calculated by each neighboring acoustic node based on the local error signal and the filter reference signal. Thus:

[0059] The above formula shows that the first j The update of each control filter can be achieved by subtracting the sum of the weighted augmented gradient vectors provided by each neighboring acoustic node from the coefficients of its previous block. Indicates by the first k The acoustic node is calculated and used to update the [number]th [node]. j The weighted augmented gradient vectors of each control filter. The combination coefficients are applied to the corresponding gradient components of the frequency domain augmented gradient spectrum matrix according to the neighboring acoustic nodes to be transmitted; each normalized weighted frequency domain augmented gradient vector corresponds to a receiving acoustic node and serves as the data object for one neighborhood communication.

[0060] To generate a weighted augmented gradient vector for all neighboring receiving nodes in a single neighborhood communication, the th... jThe augmented gradient matrix of each acoustic node is weighted column-by-column according to its corresponding combination coefficients to obtain the time-domain weighted augmented gradient matrix:

[0061] in, Indicates the first j The time-domain weighted augmented gradient matrix of each acoustic node This represents a matrix of combined coefficients, denoted by [symbol]. This represents element-wise multiplication. The above augmented gradient equivalent reconstruction rewrites the traditional forward and backward communication into a single augmented gradient communication.

[0062] In the frequency domain implementation, fast Fourier transform block processing and overlap-preserving block processing are used to complete the linear convolution required to generate the augmented filter reference signal, and the control filter update amount is calculated based on the linearly correlated frequency domain equivalence relation. In one embodiment, the signal block length... L With time-domain control of filter length H and the length of the pre-identified secondary path model R Equal, and retain 50% overlap. Used for processing the first... m -1 signal block of 2 L The point reference signal overlap is divided into blocks as follows:

[0063] Will be by the first k Secondary sound sources to the first j The pre-identification secondary path model of each error sensor is zero-padding to 2. L Points are used to ensure that the overlapping blocks of the reference signal have the same length as the pre-identified secondary path model:

[0064] Perform 2 on the reference signal overlap block L Point Fast Fourier Transform yields the spectrum of the reference signal:

[0065] Perform 2 on the pre-identified secondary path model with zero padding L Point Fast Fourier Transform yields the spectrum of the pre-identified secondary path:

[0066] in, Indicates the reference signal block in the th... l Spectral components at each frequency point Indicates by the first k Secondary sound sources to the first j The pre-identification secondary path model of the error sensor in the first... lThe spectral components at each frequency point.

[0067] The reference signal spectrum is multiplied element-wise with the corresponding pre-identified secondary path spectrum, and the resulting product is subjected to 2... L Perform a pointwise inverse fast Fourier transform to obtain a circular convolution sequence; remove samples from the circular convolution sequence that do not belong to the linear convolution result, and retain the samples from the circular convolution sequence that correspond to the linear convolution result. L From each sample, we obtain the linear convolution output of the current signal block:

[0068] in, Indicates from 2 L After retaining the point sequence L The selection matrix of each sample This indicates element-wise multiplication. Performing this process on two adjacent blocks yields 2-1 for linear correlation. L Point augmented filter reference signal sequence.

[0069] The two neighboring acoustic nodes of the j-th acoustic node are associated with each time-domain control filter. L Arrange the point augmented filter reference signal sequence column-wise to obtain the time-domain augmented filter reference signal block matrix:

[0070] Among them, with the first k The column vectors corresponding to the neighborhood control filters are formed by concatenating the linear convolution outputs of the previous block and the block to be processed in chronological order:

[0071] In containing L Supplement before dividing the error signal of each sample into blocks L Using zero samples, and copying the resulting column vectors according to the number of nodes in the neighborhood node set, construct the time-domain error signal block matrix:

[0072] Perform 2 on each column of the time-domain augmented filter reference signal block matrix. L Point Fast Fourier Transform yields the frequency domain augmented filter reference signal matrix:

[0073] Perform a 2-step process on one error signal column of the time-domain error signal block matrix. L Perform a point-wise fast Fourier transform and copy the resulting spectrum column. Stack the frequency domain error signals according to the number of nodes in the neighborhood node set to obtain the frequency domain error signal matrix:

[0074] Since all columns of the frequency domain error signal matrix are identical, each acoustic node only needs to perform a 2x2 operation on the error signal once in each signal block. L The point fast Fourier transform is then performed, and the resulting spectrum is copied according to the number of nodes in the neighborhood node set.

[0075] Based on the frequency domain equivalence relation of linear correlation, the conjugate matrix of the frequency domain augmented filter reference signal matrix is ​​multiplied element-wise with the frequency domain error signal matrix to obtain the frequency domain augmented gradient spectrum matrix:

[0076] in, Indicates the first j The frequency domain augmented gradient spectrum matrix of each acoustic node, with the superscript * indicating element-wise complex conjugation.

[0077] To ensure that the update amplitude at each frequency point is adapted to the energy of the corresponding augmented filter reference signal, and to limit the update amplitude at low-energy frequencies, a normalized vector is constructed:

[0078] Wherein, the normalized vector The first term is the summation of the squared magnitudes of the frequency-domain augmented filter reference signal matrix at each frequency point along the direction of the neighborhood control filter, i.e., the frequency-point power term; the second term is determined by the regularization constant. Transformation length 2 L It is constructed using the square Euclidean norm of the frequency domain augmented filter reference signal matrix, and serves as the average frequency power regularization term shared by all frequency points, i.e., the low-energy frequency constraint term.

[0079] The frequency-domain augmented gradient spectrum matrix is ​​normalized frequency-by-frequency using the normalized vector, and the node step size and combination coefficient matrix are applied to the corresponding columns of the normalized frequency-domain augmented gradient spectrum matrix to obtain the normalized weighted frequency-domain augmented gradient spectrum matrix:

[0080] in, Indicates the first j Normalized weighted frequency domain augmented gradient spectrum matrix of each acoustic node This indicates the step size for frequency domain augmented gradient estimation. Represents the frequency domain combination coefficient matrix, symbol This indicates element-wise division. The matrix's [number]th [value] represents the [number]th [value]. i Listed as the number j The acoustic node sends to its first... i The normalized weighted frequency domain augmented gradient vector of the neighboring nodes.

[0081] In a normalized weighted frequency domain augmented gradient communication phase, each acoustic node only sends the corresponding column of the normalized weighted frequency domain augmented gradient spectrum matrix to its neighboring acoustic nodes. j The acoustic node receives signals from each neighboring acoustic node, which are related to the first acoustic node. j The normalized weighted frequency-domain augmented gradient vectors corresponding to each time-domain controlled filter are first summed and combined in the frequency domain, and then the process is repeated once. L Point-wise inverse fast Fourier transform, and extract the previous matrix by selection matrix. L The time-domain augmented gradient components are updated accordingly. j One time-domain controlled filter:

[0082] in, Indicates from 2 L Extracting the first part of the result from the inverse fast Fourier transform of the point L The selection matrix for each time-domain component; the summation condition is used to select the acoustic components transmitted by each neighboring acoustic node and that are consistent with the first time-domain component. j The normalized weighted frequency domain augmented gradient vectors corresponding to each control filter. Since both the inverse fast Fourier transform and sample selection are linear operations, performing an inverse fast Fourier transform after summing and combining in the frequency domain yields the same control filter update result as performing inverse fast Fourier transforms separately and then summing and combining in the time domain, but avoids redundant inverse transforms performed separately for multiple neighborhood gradients.

[0083] like Figure 2 As shown, the method of this invention uses a normalized weighted frequency-domain augmented gradient vector as the object for local node estimation and inter-node communication. During the current signal block segmentation, the target acoustic node buffers reference signal samples and error signal samples. It constructs a frequency-domain augmented filter reference signal matrix, a frequency-domain error signal matrix, and a normalized weighted frequency-domain augmented gradient vector using the signal block to be processed. After one neighborhood communication and frequency-domain summation combination, an inverse fast Fourier transform is performed to extract the time-domain augmented gradient component and update the time-domain control filter coefficients. The updated time-domain control filter is used to filter the reference signal, and the resulting control acoustic signal is output to the secondary sound source of the target acoustic node. This secondary sound signal is superimposed on the primary noise signal along the secondary path at the location of the error sensor, forming a secondary sound field to cancel the primary noise.

[0084] In one hardware implementation, multiple acoustic nodes are implemented by multiple node processing cores in a multi-core processor. Each node processing core stores its own acoustic signal block, pre-identified secondary path spectrum, frequency-domain augmented gradient spectrum matrix, normalized vector, and time-domain control filter coefficients. An on-chip communication module transmits the normalized weighted frequency-domain augmented gradient vector between the node processing cores based on the neighboring node set. The multi-core processor and on-chip communication module are only used to illustrate one hardware implementation of the method of the present invention and do not constitute a limitation on the processor model or communication structure. Example

[0085] To verify the noise reduction performance, convergence speed, computational complexity, and communication overhead of the method of this invention in a real-time distributed active noise control system, this embodiment employs an eight-channel distributed active noise control system. The system is arranged in a reverberation chamber approximately 7.7 m long, 5.7 m wide, and 3.3 m high, with a reverberation time of approximately 0.7 s. Eight secondary sound sources and eight error sensors are installed, and each secondary sound source and its corresponding error sensor is designated as a distributed acoustic node. The pre-identified secondary path model length, the time-domain control filter length, and the fast Fourier transform block length are all set to 256, and 50% overlap retention is applied.

[0086] The real-time experimental platform uses the TMS320C6678 multi-core digital signal processor as its core, and utilizes eight C66xCorePacs to execute signal processing tasks for eight acoustic nodes. The L2 local memory of each CorePac is used to store acoustic signals, algorithm variables, neighborhood exchange data, and program instructions. The Multicore Navigator and TeraNet on-chip communication structure are used to transmit normalized weighted frequency domain augmented gradient vectors between CorePacs. This experimental platform is only used to illustrate that the method of this invention can be implemented in real time on a multi-core processor and on-chip communication module, and does not constitute a limitation on the processor model or on-chip communication structure.

[0087] In this embodiment, the method of the present invention employing frequency-point normalization is denoted as FD-AGDF. x LMS, the time-domain augmented diffusion filter reference least mean square algorithm is denoted as TD-ADF. x LMS; meanwhile, the distributed filter reference minimum mean square algorithm in the time domain and frequency domain are denoted as TD-DCFxLMS and FD-DCFxLMS, ​​respectively, and the centralized filter reference minimum mean square algorithm in the time domain and frequency domain are denoted as TD-CF. x LMS and FD-CF x LMS was used, and the above algorithm was used as a comparison algorithm. Experimental evaluation metrics included noise reduction, convergence time required to reach steady-state noise reduction level, steady-state error signal spectrum, computational complexity, and communication overhead. Noise reduction was calculated using the following formula:

[0088] in, Indicates the first j Primary noise signal at each error sensor This represents the error signal after active noise control. Computational complexity is measured using the TIMESTAMP register of the TMS320C6678 processor, with processing latency measured in clock cycles. Communication overhead is measured based on the amount of data transmitted in each iteration of the TeraNet on-chip communication structure and its usage in the CorePac's internal memory. The sampling frequency for each experimental group is 2^6kbps. k Hz.

[0089] When band-limited white noise of 100Hz to 900Hz is used as the primary noise signal, the experimental results are as follows: Figure 3 As shown. By Figure 3 As can be seen in (a), the frequency domain algorithm, through frequency-by-frequency normalization, adapts the update amplitude of different frequencies to the energy of the corresponding filtered reference signal, thus achieving a faster overall convergence speed than the corresponding time domain algorithm. FD-DCF x LMS reaches its maximum noise reduction of approximately 24.9 dB at around 28 seconds and then diverges, TD-DCF x The LMS reaches its maximum noise reduction of approximately 24.2 dB at about 37 seconds before diverging; the method of this invention, FD-AGDF x LMS reaches the steady-state noise reduction level corresponding to the Wiener solution at approximately 44 seconds, while TD-ADF x LMS reaches the same level at approximately 71 seconds. In the 75-80 second interval, FD-CF... x LMS, TD-CF x LMS, FD-AGDF x LMS and TD-ADF x The noise reduction of LMS is approximately 26.2dB, 24.2dB, 27.9dB and 27.8dB, respectively. Figure 3 (b) gives the steady-state error signal spectrum within this interval; the error signal spectrum of the method of the present invention in the main frequency band from 100Hz to 900Hz is basically consistent with the ideal Wiener solution, indicating that the frequency domain augmented gradient update maintains the steady-state noise reduction performance of the augmented diffusion algorithm while shortening the convergence time.

[0090] When factory noise from the NOISEX-92 database is used as the primary noise signal, the experimental results are as follows: Figure 4 As shown. By Figure 4 As can be seen in (a), during the period of approximately 30 to 50 seconds, the non-stationary impact disturbances in the factory noise cause significant fluctuations in the noise reduction curves of each algorithm. Distributed Algorithm TD-DCF xLMS and FD-DCF x LMS diverges under the perturbation, while augmented-diffusion and centralized algorithms remain stable and continue to converge. The method of this invention reaches a steady-state noise reduction level at approximately 53 seconds, with a steady-state noise reduction of approximately 24.7 dB in the 85-90 second range; TD-ADF x LMS reaches a steady-state noise reduction level at approximately 80 seconds, with a steady-state noise reduction of approximately 24.8 dB. Figure 4 (b) gives the steady-state error signal spectrum in the 85s to 90s interval. The distributed algorithm's error signal spectrum in the approximately 180Hz to 250Hz frequency band is higher than the primary noise signal spectrum, and... Figure 4 Corresponding to the divergence phenomenon shown in (a); FD-AGDF x LMS and TD-ADF x The error signal spectrum of LMS is basically consistent with that of the ideal Wiener solution, while the error signal spectrum of the centralized algorithm in the frequency band of about 650Hz to 950Hz is still higher than that of the Wiener solution, indicating that it has not yet fully converged within the given experimental time.

[0091] When using ship engine room noise from the NOISEX-92 database as the primary noise signal, the experimental results are as follows: Figure 5 As shown. Short-duration speech interference superimposed on the experimental signal causes pulse-like components to appear in the Wiener solution curve during the interference period, but it does not change the overall convergence trend of each algorithm. Figure 5 As can be seen in (a), the method of the present invention reaches the steady-state noise reduction level corresponding to the Wiener solution in about 55s, TD-ADF x LMS reaches the same level at approximately 84s, with both achieving a steady-state noise reduction of approximately 28.4dB; FD-CF x LMS and TD-CF x The noise reduction achieved by LMS within the same experimental duration was approximately 23.0 dB and 25.8 dB, respectively. Figure 5 As can be seen in (b), FD-AGDF x LMS and TD-ADF x The steady-state error signal spectrum of LMS is basically consistent with that of Wiener's solution, while the centralized algorithm still has a high residual spectral component in multiple frequency bands, indicating that frequency-point normalized frequency domain augmentation gradient update can shorten the time required to reach the steady-state noise reduction level.

[0092] Figure 6Further comparisons were made of the primary noise signal spectrum and Wiener solution spectrum before and after the introduction of speech interference at approximately 10 seconds in a ship engine room noise experiment. The speech interference formed identifiable spectral components around 256Hz, 487Hz, 634Hz, and 728Hz. Specifically, the Wiener solution spectrum amplitude at 728Hz increased from approximately -96.7dB without interference to approximately -92.2dB with interference. The primary noise signal, after being interfered with at 256Hz, had a spectral amplitude of approximately -70.3dB, still lower than its inherent spectral peak of approximately -52.8dB at 238Hz. Therefore, the speech interference... Figure 5 The overall noise reduction curve in Figure (a) is not significant, but it can be distinguished in the spectral results. These results indicate that the spectral results can serve as a supplement to the time-domain noise reduction curves, characterizing the impact of non-stationary interference on specific frequency points.

[0093] The above three types of noise experiments demonstrate that the method of this invention can maintain the steady-state noise reduction performance of the augmented-diffusion distributed active noise control algorithm under asymmetric sound path conditions, under band-limited white noise, factory noise containing non-stationary impact disturbances, and ship engine room noise containing speech interference. Compared with the corresponding time-domain augmented-diffusion algorithm, the method of this invention shortens the time required to reach the steady-state noise reduction level through fast Fourier transform block processing, frequency-point normalization, and frequency-domain augmented gradient update.

[0094] To compare the real-time computational complexity under different control structures and operational domains, the clock cycles required for each algorithm to complete one sampling update were measured using the timestamp of the TMS320C6678 multi-core digital signal processor. Each algorithm used the aforementioned eight-channel system configuration, and the measured values ​​of the eight CorePacs were recorded and summed to obtain the total measured computational cost per sampling point. The measurement results for the six algorithms are shown in the table below; the measured range for a single processing core is from the minimum to the maximum value among the eight processing cores.

[0095] Table 1 Measurement results of six algorithms As shown in Table 1, under the same eight-channel real-time experimental conditions, the total measured clock cycles for the eight processing steps of the time-domain centralized algorithm, the time-domain distributed algorithm, and the time-domain augmented diffusion algorithm are 273,203, 36,394, and 132,060, respectively; the corresponding frequency-domain algorithms are 23,093, 3,587, and 10,120, which are approximately 8.5%, 9.9%, and 7.7% of the time-domain algorithms of the same type, respectively. The centralized algorithm has the highest measured computational cost due to its use of global error signal estimation; the distributed algorithm uses only local error signals and has the lowest measured computational cost, but combined with... Figures 3 to 5The results show that it diverges under asymmetric acoustic paths and non-stationary perturbation conditions; the augmented-diffusion algorithm maintains stable convergence through neighborhood cooperation, and its measured computational cost is between that of the centralized and decentralized algorithms. The method of this invention, while maintaining the steady-state noise reduction performance of the augmented-diffusion algorithm, converts the linear convolution required to generate the filter reference signal and the correlation operations required to calculate the update amount of the control filter into point-by-point operations in the frequency domain, and performs a frequency-to-time domain transformation after summing and combining in the frequency domain, reducing the measured computational cost to approximately 7.7% of that of the time-domain augmented-diffusion algorithm.

[0096] To compare the communication overhead before and after the communication architecture refactoring, under the same augmented diffusion neighborhood cooperative structure, TD-ADF was used. x LMS is compared with the method of this invention. The control structures, communication objects, and communication topologies of the centralized and decentralized algorithms differ from those of the augmentation-diffusion algorithm, and therefore are not directly compared in this respect. The measurement results are shown in the table below.

[0097] Table 2 Comparison of Measurement Results As shown in Table 2, the method of this invention uses the normalized weighted frequency domain augmented gradient vector as the neighborhood communication object, completes the exchange of the vector through one neighborhood communication, and performs a frequency domain to time domain transformation after frequency domain summation and combination. Its comprehensive measured communication overhead is approximately TD-ADF. x The LMS is 0.78%. Therefore, the method of the present invention reduces the real-time computing burden and the amount of neighborhood communication data while maintaining the multi-node collaborative noise reduction performance under asymmetric acoustic path conditions.

Claims

1. A distributed active noise control method based on a frequency-domain augmented gradient diffusion algorithm, characterized in that, Performed by the acoustic nodes in a distributed active noise control system, including: S1, acquire the reference signal, error signal, pre-identified secondary path model, and the set of neighboring nodes corresponding to the target acoustic node. S2, perform block division and frequency domain transformation on the reference signal, error signal and pre-identified secondary path model to construct a frequency domain augmented filter reference signal matrix and a frequency domain error signal matrix; S3, multiply the conjugate matrix of the frequency domain augmented filter reference signal matrix element by element with the frequency domain error signal matrix to obtain the frequency domain augmented gradient spectrum matrix, and construct a normalized vector based on the frequency domain augmented filter reference signal matrix for each frequency point. S4. Based on the normalized vector, node step size, and combination coefficient, the corresponding gradient components in the frequency domain augmented gradient spectrum matrix are weighted to form a normalized weighted frequency domain augmented gradient vector for neighborhood communication. S5, the target acoustic node receives the normalized weighted frequency domain augmented gradient vector sent by the neighboring acoustic nodes through a neighborhood communication, and sums and combines the normalized weighted frequency domain augmented gradient vectors corresponding to the target acoustic node in the frequency domain of the target acoustic node and the neighboring acoustic nodes to obtain the combined frequency domain augmented gradient vector. S6, perform a frequency-domain to time-domain transformation on the combined frequency-domain augmented gradient vector, extract the time-domain augmented gradient component from the transformation result, update the time-domain control filter coefficients of the target acoustic node according to the time-domain augmented gradient component, and generate a control acoustic signal for driving the secondary sound source according to the updated time-domain control filter coefficients.

2. The method according to claim 1, characterized in that, The distributed active noise control system includes multiple acoustic nodes, each acoustic node including a reference signal input terminal, an error sensor, a secondary sound source, and a node processor; the neighborhood node set includes the target acoustic node itself and acoustic nodes that have established communication relationships with the target acoustic node.

3. The method according to claim 1, characterized in that, The target acoustic node employs a block processing delay, buffering the current reference signal sample and the current error signal sample during the current signal block, and simultaneously using the data corresponding to the previous signal block to execute steps S2 to S6, thereby providing a complete signal block processing interval for block calculation and neighborhood communication.

4. The method according to claim 1, characterized in that, Step S2 employs overlapping block processing: the reference signal is divided into reference signal overlapping blocks according to a preset block length and a preset overlap rate, and the pre-identified secondary path model is zero-filled so that the reference signal overlapping blocks and the pre-identified secondary path model have the same length.

5. The method according to claim 4, characterized in that, The block length L Time-domain controlled filter length H and the length of the pre-identified secondary path model R Equal, the frequency domain transformation is 2 L Point Fast Fourier Transform.

6. The method according to claim 5, characterized in that, The construction of the frequency domain augmented filter reference signal matrix includes: multiplying the reference signal spectrum element-wise with the spectrum of the corresponding pre-identified secondary path model and performing an inverse fast Fourier transform to obtain a circular convolution sequence; removing samples in the circular convolution sequence that do not belong to the linear convolution result and retaining the samples corresponding to the linear convolution result; combining the linear convolution results of two adjacent signal blocks in chronological order and performing a fast Fourier transform to obtain the frequency domain augmented filter reference signal matrix corresponding to the control filter of the target acoustic node and its neighboring nodes.

7. The method according to claim 5, characterized in that, The frequency domain error signal matrix is ​​padded before the error signal is divided into blocks. L 1 zero sample and perform 2 once L The frequency domain error signal is obtained by constructing a point fast Fourier transform and stacking the signals according to the number of nodes in the neighborhood node set.

8. The method according to claim 1, characterized in that, In step S3, the normalization vector includes: a frequency power term obtained by summing the squared magnitudes of the frequency domain augmented filter reference signal matrix at each frequency point along the augmented filter dimension, and a low-energy frequency constraint term generated based on the average frequency power and regularization constant of the frequency domain augmented filter reference signal matrix.

9. The method according to claim 1, characterized in that, In step S4, the combination coefficients are applied to the corresponding gradient components of the frequency domain augmented gradient spectrum matrix according to the neighboring acoustic nodes to be transmitted; each normalized weighted frequency domain augmented gradient vector corresponds to a receiving acoustic node and serves as the data object of the first neighboring communication.

10. The method according to claim 1, characterized in that, In step S5, the target acoustic node receives the normalized weighted frequency domain augmented gradient vector corresponding to the time domain control filter of the target acoustic node from each neighboring acoustic node, and sums and combines the normalized weighted frequency domain augmented gradient vector before performing the frequency domain to time domain transformation.