Design of delayless multi-sampled subband functional link neural network with application to active noise control☆
Introduction
Neural networks (NNs), such as radial basis function NN, echo state network, and reaction-diffusion NN, have had a great deal of success in the study of many complex applications [1], [2], [3], [4], [5], for example, nonlinear system identification, communication channel equalization, and image encryption. Recently, functional link NNs (FLNNs) as a single-layer feedforward neural network with low computational complexity have attracted considerable interest in active noise control (ANC) [6], [7], [8], digital predistortion [9], and many other applications. According to different ways of functional expansion, the commonly used FLNN includes trigonometric-FLNN (TFLNN) [10], Legendre-FLNN (LFLNN) [11], Chebyshev-FLNN (CFLNN) [12] and so on.
In the past decades, to enhance the nonlinear identification capability of the FLNN, several different filter structures and methods have been proposed, such as pipelined filter [13], sparse representations [9], set-membership method [14] and collaborative TFLNNs [15], [16]. By means of an exponential sinusoid, an adaptive exponential TFLNN (AETFLNN) was reported in [8]. Using the infinite impulse response (IIR) scheme, a recursive adaptive sparse exponential TFLNN (RASETFLNN) was proposed in [17]. Furthermore, a generalized exponential TFLNN structure was developed in [18]. Recently, many improved variants of the FLNNs were proposed in the literature [19], [20], [21]. For example, a novel CFLNN based on correntropy was introduced in [19], which can achieve excellent performance in impulsive noise environments. However, because all these FLNNs expand the original input signal to a higher dimensional expansion space, it will bring an expanded input covariance matrix with larger eigenvalue spread. Its disadvantage is that using the well-known least mean square (LMS) strategy will lead to a slow rate of convergence.
It is known that the subband adaptive filter (SAF) structures increase the algorithmic convergence speed when the input signal involves the covariance matrix with large eigenvalue spread [22]. For the linear SAF, several subband structures had been designed, including conventional SAF [23], [24], [25], [26] and multiband-structured SAF (MSAF) mechanism [27], [28]. Compared with the conventional SAF method, the MSAF mechanism updates the fullband taps based on the subband error signals. Recently, several modified MSAFs have been presented, such as variable step size MSAF [29], [30], robust MSAF [31], [32], multi-sampled MSAF [33], etc. To increase the nonlinear processing capability of subband filtering, subband second-order Volterra filter (SBVF) and its proportionate version were developed [34], [35]. Because of using the conventional SAF structure, the SBVF updates the taps of each subfilter in the first- and second-order subband channels independently.
By generating an anti-noise with equal amplitude and opposite phase, the ANC system can cancel the undesired noise signal, such as in the headset, car and infant incubators. The linear subband ANC systems were developed in [36], [37], [38]. However, in real-time implementation, significant nonlinearities are frequently generated by the loudspeaker of the ANC system. The performance of linear ANC systems deteriorates when applied for such nonlinear ANC (NANC) scenario [6], [39]. To overcome this situation, the FLNN-based ANC has been increasingly employed, such as [8], [40]. Unfortunately, these FLNN-based ANC systems using the LMS method will be subjected to slow convergence speed for correlated noise.
In order to speed up the convergence of the TFLNN, in this work we will design a delayless multi-sampled multiband-structured subband FLNN (DMSFLNN) structure. The key ingredient is to employ the multi-sampled multiband-structured SAF scheme proposed in [33] in the TFLNN. In the meantime, to avoid the delay caused by subband filters, a delayless channel is introduced, which copies the multi-sampled MSAF taps. With the use of the principle of minimum disturbance over the DMSFLNN, an effective normalized subband adaptive algorithm is designed. In order to study the performance, stability conditions, optimal step size and computational complexity of the proposed algorithm are deduced. Additionally, for application to the NANC, a delayless multi-sampled multiband-structured filtered-s normalized LMS (NLMS) algorithm, called DMSFsNLMS, is designed. To assess the performance of the DMSFLNN, extensive simulation experiments are conducted in nonlinear system identification and NANC applications. Finally, to clarify the contribution in this work, the main contributions have been summarized as follows:
- i)
This paper firstly proposes the DMSFLNN, which employs the multi-sampled MSAF scheme to accelerate the convergence speed of the TFLNN.
- ii)
For the proposed algorithmic formulation over the DMSFLNN, the mean, mean-square stabilities and optimal step size are derived.
- iii)
Based on the DMSFLNN structure, the DMSFsNLMS algorithm is devised to handle NANC systems.
The remainder of the paper is arranged as follows. The TFLNN is briefly reviewed in Section 2. The DMSFLNN is presented in Section 3. Section 4 provides an analysis of the stability conditions, optimal step size and computational complexity. Section 5 develops the DMSFLNN-based NANC algorithm. Section 6 provides Monte-Carlo simulations and the conclusion is drawn in Section 7. The symbols and abbreviations are listed as follows:
Section snippets
Overview of traditional TFLNN
Consider a nonlinear system identification problem. At time instant , the observed output signal and known input signal are related bywhere denotes the additive background noise, is an unknown nonlinear system, and with being the delayed number.
To model and approximate the nonlinear system , the signal in the TFLNN is expanded to a higher dimensional vector :
Proposed DMSFLNN and its weight update
In this section, to speed up the convergence of the TFLNN, we aim to design the DMSFLNN filter, which integrates the multi-sampled multiband-structured subband method into the TFLNN. In the DMSFLNN, a delayless channel is also introduced to avoid the delay caused by subband filters. Moreover, we provide an algorithmic formulation over the DMSFLNN.
Analysis of stability conditions, optimal step size and computational complexity over the DMSFLNN
In the section, we will analyze the stability conditions, optimal and computational complexity for the proposed subband adaptive algorithm (11).
To begin with, we introduce a model as followswhere denotes the deterministic coefficients vector of the linear MSE estimator of based on the trigonometric extended input vector . Some commonly used assumptions on the signal and the estimation error are introduced as below [33]. Assumption 1 The input sequence is a
DMSFLNN-based ANC application
In this section, the DMSFLNN-based NANC system is designed, as depicted in Fig. 2. In this figure, and denote the transfer functions of the primary path and secondary path, respectively, is the reference noise measured by a reference microphone, is the primary noise to be cancelled, is the output signal of the secondary path at the cancellation point, is an estimate of , and denotes the weight of the adaptive controller.
At time , the error signal is
Monte-Carlo simulation study
In this section, we provide Monte-Carlo simulation results to validate the usefulness of the DMSFLNN. The cosine-modulated filter banks are adopted in the subband structure[33], where the lengths of the prototype filter are , 128, and 256 for , respectively. The initial weight vector is a zero vector. All curves are generated by averaging the results of 100 Monte-Carlo runs.
Conclusion
In this paper, the DMSFLNN filter has been designed. By employing the multi-sampled MSAF method, the DMSFLNN filter can increase the convergence speed of the TFLNN in the case of colored input signals. The principle of minimum disturbance has been used over the DMSFLNN and then a subband algorithm has been constructed. The analysis of stability conditions, optimal step size and computational complexity has been undertaken for the proposed DMSFLNN filter. Moreover, the DMSFLNN-based NANC
Clarification of Relationship to Conference Paper
A part of the initial work in this SIGPRO submission was presented at the 52nd IEEE International Symposium on Circuits and Systems (ISCAS’2019) in Sapporo, Hokkaido, Japan in May 2019, which is also cited as Reference [1] in this SIGPRO submission.
In this SIGPRO submission, we have significantly expanded the ISCAS’2019 conference paper by including a lot of substantive new material. In more details, in this SIGPRO submission,
- •
we show the mean stability and mean-square stability of the proposed
CRediT authorship contribution statement
Sheng Zhang: Conceptualization, Formal analysis, Methodology, Writing – original draft. Wei Xing Zheng: Methodology, Writing – review & editing. Hongyu Han: Validation, Investigation.
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
References (41)
- et al.
Orthogonalized power filters for nonlinear acoustic echo cancellation
Signal Process.
(2006) - et al.
A study about Chebyshev nonlinear filters
Signal Process.
(2016) - et al.
Advances in active noise control: a survey, with emphasis on recent nonlinear techniques
Signal Process.
(2013) - et al.
Nonlinear channel equalization for wireless communication systems using Legendre neural networks
Signal Process.
(2009) - et al.
A novel nonlinear adaptive filter using a pipelined second-order Volterra recurrent neural network
Neural Netw.
(2009) - et al.
Pipelined set-membership approach to adaptive Volterra filtering
Signal Process.
(2016) - et al.
A generalized exponential functional link artificial neural networks filter with channel-reduced diagonal structure for nonlinear active noise control
Appl. Acoust.
(2018) - et al.
Hierarchical partial update generalized functional link artificial neural network filter for nonlinear active noise control
Digit. Signal Process.
(2019) - et al.
A robust band-dependent variable step size NSAF algorithm against impulsive noises
Signal Process.
(2016) - et al.
A variable step-size matrix normalized subband adaptive filter
IEEE Trans. Audio Speech Lang. Process.
(2010)
Diffusion sign subband adaptive filtering algorithm with enlarged cooperation and its variant, circuits
Syst. Signal Process.
Subband adaptive filtering algorithm over functional link neural network
Proc. 52nd IEEE Int. Symp. Circuits Syst.
Polynomial Signal Processing
Impulsive synchronization of reaction diffusion neural networks with mixed delays and its application to image encryption
IEEE Trans. Neural Netw. Learn. Syst.
Frequency domain exponential functional link network filter: design and implementation
Signal Process.
Design of adaptive exponential functional link network-based nonlinear filters
IEEE Trans. Circuits Syst. I Reg. Pap.
Adaptive Pattern Recognition and Neural Networks
Identification of nonlinear dynamic systems using functional link artificial neural networks
IEEE Trans. Syst. Man Cybern. B Cybern.
Nonlinear dynamic system identification using Chebyshev functional link artificial neural networks
IEEE Trans. Syst. Man Cybern. B Cybern.
Functional link adaptive filters for nonlinear acoustic echo cancellation
IEEE Trans. Audio Speech Lang. Process.
Cited by (6)
An efficient nonlinear adaptive filter algorithm based on the rectified linear unit
2024, Digital Signal Processing: A Review JournalFrequency-domain Volterra kernel-based adaptation: Formulations and algorithms
2024, Signal ProcessingDelayless partial subband update algorithm for feed-forward active road noise control system in pure electric vehicles
2023, Mechanical Systems and Signal ProcessingError Reused Filtered-X Least Mean Square Algorithm for Active Noise Control
2024, IEEE/ACM Transactions on Audio Speech and Language ProcessingPerformance Analysis of Robust Subband Hammerstein Spline Adaptive Filter
2024, Circuits, Systems, and Signal ProcessingOptimal Subband Adaptive Filtering Algorithm Over Functional Link Neural Network
2023, IEEE International Workshop on Machine Learning for Signal Processing, MLSP