Elsevier

Signal Processing

Volume 190, January 2022, 108318
Signal Processing

Censored regression distributed functional link adaptive filtering algorithm over nonlinear networks

https://doi.org/10.1016/j.sigpro.2021.108318Get rights and content

Highlights

  • A class of DFLMS algorithm is derived for distributed nonlinear networks, in which four types of DFLNs are respectively used.

  • Based on the probit regression, the proposed CRDFLAF algorithm corrects the bias of censored measurements.

  • The stability condition of the learning rate is derived, and the corresponding computational complexity is provided.

  • Simulation results of two types of distributed nonlinear networks verify the effectiveness of the proposed algorithms.

Abstract

Wireless sensor network (WSN) is an important part of the Internet of Things (IoT) and has emerged in various new forms, such as smart home, smart city, and intelligent manufacturing system. Due to its high reliability, distributed estimation over nonlinear WSNs is one of the most active fields in recent years. In this paper, a novel distributed functional link least mean square (DFLMS) algorithm based on rblackthe diffusion strategy is proposed, in which the diffusion functional link network (DFLN) is used to model the nonlinear dynamic behavior of the distributed system. In particular, by using different orthogonal polynomials, we develop four types of DFLNs, i.e., trigonometric DFLN (TDFLN), Legendre DFLN (LDFLN), Chebyshev DFLN (CDFLN), and Hermite DFLN (HDFLN). However, the censored measurement caused by the range of sensors brings great challenges to the traditional distributed nonlinear estimation. To tackle this problem, a censored regression-distributed functional link adaptive filtering (CR-DFLAF) algorithm is further proposed. Compared with the DFLMS algorithm, the CR-DFLAF algorithm can compensate the estimated bias in the CR scenario at the price of slightly increased computational complexity. Simulations involving two distributed nonlinear networks verify the effectiveness of the proposed algorithms.

Introduction

The wireless sensor network (WSN) is a multi-hop self-organizing network system formed by a large number of sensor nodes deployed in the monitoring area to communicate with each other [1], [2], [3]. It is an important technology in the underlying network of the Internet of Things (IoT). As the perception of external information, WSN has become a development trend of the IoT technology. Distributed estimation over WSNs extracts data collected by nodes distributed in geographic areas, thereby estimating system parameters of interest from noise measurements [4], [5]. Centralized estimation is performed by a central processor that fuses all nodes together for estimation [6], [7], while distributed estimation, after obtaining the estimated value on each node, is used for data exchange with direct neighbors in the network. Distributed estimation performs estimation tasks on each node and communicates with nodes from the same neighborhood to exchange information, which reduces the demand for communication resources and enhances the reliability and robustness of the network [8], [9]. For different application scenarios, a number of distributed algorithms have been developed. For example, the distributed multi-error filtered-x least mean square (DMEFxLMS) algorithm was developed for distributed active noise control (ANC) networks [10], and the diffusion augmented complex LMS (DACLMS) algorithm was proposed for distributed complex networks [11], etc. The α-stable distribution is considered to be an effective modeling method for impulsive noise signals that are often encountered in real-life [12]. By minimizing the real-time fractional powers of an error measure, the authors in [13] developed a class of distributed adaptive filtering solutions for α-stable signals. More distributed processing methods of α-stable noise can be found in [14], [15].

The above algorithms assume that the input-output of each node is a linear regression model, and do not consider the nonlinear behavior of the system. However, when the system exhibits nonlinearity, the reduced performance of these algorithms cannot meet the requirements of distributed estimation. To the best of our knowledge, the literature is limited on distributed nonlinear estimation, including kernel space-based [16], [17], Volterra filter (VF)-based [18], [19], and spline filter (SF)-based [20] distributed estimation algorithms. The principle of the kernel method is to map the input samples into the high-dimensional reproducing kernel Hilbert space of linear modeling, and then use the corresponding kernel function for nonlinear approximation [16]. However, conventional kernel adaptive methods suffer from the high computational load due to the increase of dictionary dimension [21]. To alleviate this problem, several methods have been proposed [22], [23]. Based on the principles of set-membership filtering, two computationally efficient kernel adaptive filters were developed in [23]. The VF is derived from the Taylor expansion with memory, which belongs to the linear-in-parameters (LIP) nonlinear filter, i.e., its output depends linearly on coefficients [24]. The main bottleneck of VF is that the computational complexity increases exponentially as the order of polynomials increases [25]. The SF characterizes a nonlinear system by using a series of local interpolation polynomials, which adapts a small subset of the parameters at each time step [20].

In many engineering distributed applications, the measurement data of a nonlinear system may be partially censored due to the saturation of sensors. For example, distributed acoustically coupled sensor networks are usually used to deal with nonlinear active noise problems, where the primary or secondary path may exhibit nonlinear distortions [26]. When the measurement of the primary or secondary disturbance exceeds the range of acoustic sensors, part of the statistical information of the recorded signal will be lost or censored. Algorithms based on the censored regression (CR) process were proposed to eliminate the estimation bias caused by censored observations [27], [28]. Note that due to incomplete observations, the CR problem is an intentional nonlinear process independent of system nonlinearity. Since current CR-based distributed algorithms only focus on solving the censored measurement of distributed linear estimation, the bias compensation of CR models in distributed nonlinear networks is still a promising research gap.

On the other hand, to model the nonlinearity of the system, the functional link network (FLN) has been considered as a good alternative to VF in recent years due to its fewer parameters, simple structure, and efficient calculation. The FLN resorts to point-wise orthogonal polynomial expansions of the current input signal, which are then linearly combined by an adaptive filter to obtain the output. Therefore, an FLN is also a LIP filter with the same flexibility as the VF. Inspired by the attractive properties of the FLN, it is natural to consider developing a new distributed nonlinear algorithm. However, expansion functions based on different orthogonal polynomials lead to different diffusion FLNs (DFLNs), and the modeling capabilities of various DFLNs for nonlinear distributed networks need to be further compared.

In this work, we focus on solving distributed nonlinear estimation problems. By developing different DFLNs, a class of distributed function link LMS (DFLMS) algorithm is proposed, in which the input signals expanded by DFLNs are represented in a unified fashion. Moreover, to compensate the bias caused by CR measurement, the CR-distributed functional link adaptive filtering (CR-DFLAF) algorithm is further proposed. Specifically, the main contributions are outlined as follows.

1) A class of DFLMS algorithm is derived for distributed nonlinear networks, in which four types of DFLNs including Chebyshev DFLN (CDFLN), Hermite DFLN (HDFLN), Legendre DFLN (LDFLN), and trigonometric DFLN (TDFLN), are respectively used as controllers to model the nonlinearity of the system.

2) Consider censored measurements at distributed nodes that are corrupted by additional noise with unknown standard deviation. Based on the probit regression, the proposed CR-DFLAF algorithm corrects the bias of censored measurements by constructing an adaptive estimator and adopts the least square criterion and distributed collaboration strategy to obtain the estimate of the parameters of interest.

3) For the proposed approaches, the stability condition of the learning rate is derived, and the corresponding computational complexity is provided.

4) Simulation results of two types of distributed nonlinear networks verify the effectiveness of the proposed algorithms.

The paper is organized as follows. Section 2 reviews the distributed nonlinear network model and censored regression. In Section 3, the DFLMS and CR-DFLAF algorithms are derived. In Section 4, the performance of the proposed algorithms is analyzed. In Section 5, simulations are conducted to verify the performance of the proposed algorithms. Conclusions are provided in Section 6.

Notation: Normal font stands for scalars. Uppercase and Lowercase boldface letters denote matrices and vectors, respectively. Operators (·)T and E(·) are transpose and expectation. The symbol R is the set of real numbers.

Section snippets

Distributed nonlinear network model

The considered model is a spatial network, composed of K nodes. At each time instant n, each node k{1,2,,K} receives the random data {xk(n),dk(n)}, where xk(n)=[xk(n),xk(n1),,xk(nM+1)]T is the node input, M is the order of the input, and dk(n) is the scalar measurement. The nonlinear relationship between xk(n) and dk(n) is as follows [16], [29]:dk(n)=f(xkT(n))wo+vk(n)where f(·) denotes the nonlinear behavior of nodes, wo is the unknown vector of the system parameters, and vk(n) represents

DFLN model

By developing the DFLN, the general form of the expanded input signal f(xk(n)) is given byf(xk(n))=[f0(xk(n)),f1(xk(n)),,fp(xk(n)),,fP(xk(n)),f0(xk(n1)),f1(xk(n1)),,fp(xk(n1)),,fP(xk(n1)),,f0(xk(nM+1)),f1(xk(nM+1)),,fp(xk(nM+1))],,fP(xk(nM+1))]Twhere fp(·), p=0,1,,P is the expansion function of DFLN, and P is the nonlinear order. A generic set of functional link for the P-order trigonometric DFLN (TDFLN) can be expressed byfp(xk(ni))={1,ifp=0xk(ni),ifp=1sin(pπxk(ni)),ifp=2mcos

Performance analysis

In this section, we only focus on the analysis of the CR-DFLAF algorithm based on the P-order TDFLN, since the analysis of the DFLMS algorithm is similar to that of the DLMS algorithm. The weight vector wk(n) and intermediate estimate vector ψk(n) are vectors with dimension (2P+1)M.

Numerical simulations

In this section, we perform experimental studies to evaluate the effectiveness of the proposed algorithms, where the distributed network is considered to be composed of K=20 nodes, as shown in Fig. 1. The variance of the measurement noise σv2 and the SNRs of each node are shown in Fig. 2. The original input signal xk(n) is a uniform noise over the interval [1,+1].

Conclusions and discussions

In this work, we focused on the distributed nonlinear network and used DFLN to model the nonlinear characteristics of the system. Based on the diffusion strategy, the DFLMS algorithm was proposed for distributed nonlinear estimation. In particular, four DFLNs were developed to model the nonlinearity of the system. In the case of censored distributed nonlinear networks, the CR-DFLAF algorithm was further proposed by constructing an adaptive estimator and combining with the least square

CRediT authorship contribution statement

Kai-Li Yin: Data curation, Formal analysis, Writing – original draft, Methodology. Yi-Fei Pu: Writing – original draft, Funding acquisition, Writing – review & editing. Lu Lu: Investigation, Writing – original draft, Writing – review & editing.

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.

Acknowledgment

This work was supported by the National Key Research and Development Program Foundation of PR China under Grant 2018YFC0830300, China South Industries Group Corporation (Chengdu) Fire Control Technology Center Project (non-secret) under Grant HK20-03, National Natural Science Foundation of PR China under Grant 62171303 and Grant 61901285, China Postdoctoral Science Foundation under Grant 2020T130453, the Fundamental Research Funds for the Central Universities under Grant 2021SCU12063, and

References (39)

  • M.M. Rana et al.

    IoT-based state estimation for microgrids

    IEEE IoT J.

    (2018)
  • M. Ozay et al.

    Sparse attack construction and state estimation in the smart grid: centralized and distributed models

    IEEE J. Sel. Areas Commun.

    (2013)
  • S. Kar et al.

    Distributed parameter estimation in sensor networks: nonlinear observation models and imperfect communication

    IEEE Trans. Inf. Theory

    (2012)
  • Y. Xia et al.

    An adaptive diffusion augmented CLMS algorithm for distributed filtering of noncircular complex signals

    IEEE Signal Process. Lett.

    (2011)
  • S.P. Talebi et al.

    Complex-valued nonlinear adaptive filters with applications in α-stable environments

    IEEE Signal Process. Lett.

    (2019)
  • S.P. Talebi et al.

    Distributed adaptive filtering of α-stable signals

    IEEE Signal Process. Lett.

    (2018)
  • V.C. Gogineni et al.

    Fractional-order correntropy adaptive filters for distributed processing of α-stable signals

    IEEE Signal Process. Lett.

    (2020)
  • S.P. Talebi et al.

    Distributed particle filtering of α-stable signals

    IEEE Signal Process. Lett.

    (2017)
  • B.-S. Shin et al.

    Distributed adaptive learning with multiple kernels in diffusion networks

    IEEE Trans. Signal Process.

    (2018)
  • Cited by (12)

    • A class of robust censored regression adaptive filtering algorithms

      2024, Digital Signal Processing: A Review Journal
    • Distributed functional link adaptive filtering for nonlinear graph signal processing

      2022, Digital Signal Processing: A Review Journal
      Citation Excerpt :

      So the FLAF can perform a nonlinear expansion of the input signal and carry out a linear filtering in a higher dimensional space. Some existing nonlinear extension functions include trigonometric function [19,25], Lagrangian function [26], and exponential function [27], and so on. In addition, owing to the low computational burden, FLAF has been widely used in the field of acoustic echo cancellation [28,29] and active noise control [30].

    • Low-Complexity Normalized Subband Adaptive Filter for the Censored Regression Model

      2024, IEEE Transactions on Circuits and Systems II: Express Briefs
    • A method of grounding fault location in power system based on adaptive filtering

      2023, International Journal of Power and Energy Conversion
    View all citing articles on Scopus
    View full text