\title{A new method of designing complementary filters for sensor fusion using \(\mathcal{H}_\infty\) synthesis}
\begin{document}
\hypersetup{allcolors=teal}
\begin{frontmatter}
\begin{abstract}
Sensor have limited bandwidth and are accurate only in a certain frequency band.
In many applications, the signals of different sensor are fused together in order to either enhance the stability or improve the operational bandwidth of the system.
The sensor signals can be fused using complementary filters.
The tuning of complementary filters is a complex task and is the subject of this paper.
The filters needs to meet design specifications while satisfying the complementary property.
This paper presents a framework to shape the norm of complementary filters using the \(\mathcal{H}_\infty\) norm minimization.
The design specifications are imposed as constraints in the optimization problem by appropriate selection of weighting functions.
The proposed method is quite general and easily extendable to cases where more than two sensors are fused.
Finally, the proposed method is applied to the design of complementary filter design for active vibration isolation of the Laser Interferometer Gravitation-wave Observatory (LIGO).
\item\cite{pascoal99_navig_system_desig_using_time} use LMI to generate complementary filters
\item\cite{hua05_low_ligo,hua04_polyp_fir_compl_filter_contr_system}: FIR + convex optimization
\item Similar to feedback system:
\begin{itemize}
\item\cite{plummer06_optim_compl_filter_their_applic_motion_measur} use H-Infinity to optimize complementary filters (flatten the super sensor noise spectral density)
\item Robustness problems: \cite{zimmermann92_high_bandw_orien_measur_contr} change of phase near the merging frequency
\item Trial and error
\item Although many design methods of complementary filters have been proposed in the literature, no simple method that allows to shape the norm of the complementary filters is available.
A general sensor fusion architecture using complementary filters is shown in Figure \ref{fig:sensor_fusion_overview} where several sensors (here two) are measuring the same physical quantity \(x\).
The resulting sensor, termed as ``super sensor'', can have larger bandwidth and better noise characteristics in comparison to the individual sensor.
This means that the super sensor provides an estimate \(\hat{x}\) of \(x\) which can be more accurate over a larger frequency band than the outputs of the individual sensors.
Such model is shown in Figure \ref{fig:sensor_model} and consists of a linear time invariant (LTI) system \(G_i(s)\) representing the dynamics of the sensor and an additive noise input \(n_i\) representing its noise.
The model input \(x\) is the measured physical quantity and its output \(\tilde{x}_i\) is the ``raw'' output of the sensor.
Before filtering the sensor outputs \(\tilde{x}_i\) by the complementary filters, the sensors are usually normalized to simplify the fusion.
This normalization consists of first obtaining an estimate \(\hat{G}_i(s)\) of the sensor dynamics \(G_i(s)\).
It is supposed that the estimate of the sensor dynamics \(\hat{G}_i(s)\) can be inverted and that its inverse \(\hat{G}_i^{-1}(s)\) is proper and stable.
The raw output of the sensor \(\tilde{x}_i\) is then passed through \(\hat{G}_i^{-1}(s)\) as shown in Figure \ref{fig:sensor_model_calibrated}.
The two sensors are measuring the same physical quantity \(x\) with dynamics \(G_1(s)\) and \(G_2(s)\), and with \emph{uncorrelated} noises \(n_1\) and \(n_2\).
The normalized signals from both calibrated sensors are fed into two complementary filters \(H_1(s)\) and \(H_2(s)\) and then combined to yield an estimate \(\hat{x}\) of \(x\) as shown in Fig. \ref{fig:fusion_super_sensor}.
Thus, this sensor fusion architecture permits to filter the noise of both sensors without introducing any distortion in the physical quantity to be measured.
The estimation error \(\delta x\), defined as the difference between the sensor output \(\hat{x}\) and the measured quantity \(x\), is computed for the super sensor \eqref{eq:estimate_error}.
As shown in \eqref{eq:noise_filtering_psd}, the Power Spectral Density (PSD) of the estimation error \(\Phi_{\delta x}\) depends both on the norm of the two complementary filters and on the PSD of the noise sources \(\Phi_{n_1}\) and \(\Phi_{n_2}\).
If the two sensors have identical noise characteristics (\(\Phi_{n_1}(\omega)=\Phi_{n_2}(\omega)\)), a simple averaging (\(H_1(s)= H_2(s)=0.5\)) is what would minimize the super sensor noise.
This the simplest form of sensor fusion with complementary filters.
However, the two sensors have usually high noise levels over distinct frequency regions.
In such case, to lower the noise of the super sensor, the value of the norm \(|H_1|\) has to be lowered when \(\Phi_{n_1}\) is larger than \(\Phi_{n_2}\) and that of \(|H_2|\) lowered when \(\Phi_{n_2}\) is larger than \(\Phi_{n_1}\).
Therefore, by properly shaping the norm of the complementary filters, it is possible to minimize the noise of the super sensor noise.
In practical systems the sensor normalization is not perfect and condition \eqref{eq:perfect_dynamics} is not verified.
In order to study such imperfection, a multiplicative input uncertainty is added to the sensor dynamics (Figure \ref{fig:sensor_model_uncertainty}), where the nominal model is taken as the estimated model for the normalization \(\hat{G}_i(s)\), \(\Delta_i\) is any stable transfer function satisfying \(|\Delta_i(j\omega)| \le1,\ \forall\omega\), and \(w_i(s)\) is a weight representing the magnitude of the uncertainty.
The weight \(w_i(s)\) is chosen such that the real sensor dynamics is always contained in the uncertain region represented by a circle centered on \(1\) and with a radius equal to \(|w_i(j\omega)|\).
As the nominal sensor dynamics is taken as the normalized filter, the normalized sensor can be further simplified as shown in Figure \ref{fig:sensor_model_uncertainty_simplified}.
\caption{\label{fig:sensor_fusion_dynamic_uncertainty}Sensor fusion architecture with sensor dynamics uncertainty}
\end{figure}
The super sensor dynamics \eqref{eq:super_sensor_dyn_uncertainty} is no longer equal to \(1\) and now depends on the sensor dynamics uncertainty weights \(w_i(s)\) as well as on the complementary filters \(H_i(s)\).
The dynamical uncertainty of the super sensor can be graphically represented in the complex plane by a circle centered on \(1\) with a radius equal to \(|w_1(j\omega) H_1(j\omega)| + |w_2(j\omega) H_2(j\omega)|\) as shown in Figure \ref{fig:uncertainty_set_super_sensor}.
\caption{\label{fig:uncertainty_set_super_sensor}Uncertainty region of the super sensor dynamics in the complex plane (solid circle). The contribution of both sensors 1 and 2 to the uncertainty are represented respectively by a blue circle and a red circle. The frequency dependency \(\omega\) is here omitted.}
The super sensor dynamical uncertainty (i.e. the robustness of the fusion) clearly depends on the complementary filters norms.
For instance, the phase uncertainty \(\Delta\phi(\omega)\) added by the super sensor dynamics at frequency \(\omega\) can be found by drawing a tangent from the origin to the uncertainty circle of super sensor (Figure \ref{fig:uncertainty_set_super_sensor}) and is bounded by \eqref{eq:max_phase_uncertainty}.
As it is generally desired to limit the maximum phase added by the super sensor, \(H_1(s)\) and \(H_2(s)\) should be designed such that \(\Delta\phi\) is bounded to acceptable values.
Typically, the norm of the complementary filter \(|H_i(j\omega)|\) should be made small when \(|w_i(j\omega)|\) is large, i.e., at frequencies where the sensor dynamics is uncertain.
The synthesis objective is to shape the norm of two filters \(H_1(s)\) and \(H_2(s)\) while ensuring their complementary property \eqref{eq:comp_filter}.
This is equivalent as to finding proper and stable transfer functions \(H_1(s)\) and \(H_2(s)\) such that conditions \eqref{eq:comp_filter_problem_form} are satisfied.
where \(W_1(s)\) and \(W_2(s)\) are two weighting transfer functions that are chosen to specify the maximum wanted norms of the complementary filters during the synthesis.
In this section, it is shown that the synthesis objective can be easily expressed as a standard \(\mathcal{H}_\infty\) optimal control problem and therefore solved using convenient tools readily available.
Consider the generalized plant \(P(s)\) shown in Figure \ref{fig:h_infinity_robust_fusion} and mathematically described by \eqref{eq:generalized_plant}.
Applying the standard \(\mathcal{H}_\infty\) synthesis on the generalized plant \(P(s)\) is then equivalent as finding a stable filter \(H_2(s)\) which based on \(v\), generates a signal \(u\) such that the \(\mathcal{H}_\infty\) norm from \(w\) to \([z_1, \ z_2]\) is less than one \eqref{eq:hinf_syn_obj}.
By then defining \(H_1(s)\) to be the complementary of \(H_2(s)\)\eqref{eq:definition_H1}, the \(\mathcal{H}_\infty\) synthesis objective becomes equivalent to \eqref{eq:hinf_problem} which ensure that \eqref{eq:hinf_cond_h1} and \eqref{eq:hinf_cond_h2} are satisfied.
Therefore, applying the \(\mathcal{H}_\infty\) synthesis on the standard plant \(P(s)\)\eqref{eq:generalized_plant} will generate two filters \(H_2(s)\) and \(H_1(s)\triangleq1- H_2(s)\) that are complementary \eqref{eq:comp_filter_problem_form} and such that there norms are bellow specified bounds \eqref{eq:hinf_cond_h1},\eqref{eq:hinf_cond_h2}.
The above optimization problem can be efficiently solved in Matlab \cite{matlab20} using the Robust Control Toolbox.
Weighting functions are used during the synthesis to specify what is the maximum allowed norms of the complementary filters.
The proper design of these weighting functions is of primary importance for the success of the presented complementary filters \(\mathcal{H}_\infty\) synthesis.
First, only proper and stable transfer functions should be used.
Second, the order of the weighting functions should stay reasonably small in order to reduce the computational costs associated with the solving of the optimization problem and for the physical implementation of the filters (the order of the synthesized filters being equal to the sum of the weighting functions order).
This implies for instance that \(|H_1(j\omega)|\) and \(|H_2(j\omega)|\) cannot be made small at the same frequency.
When designing complementary filters, it is usually desired to specify its slope, its crossover frequency and its maximum gain at low and high frequency.
To help with the design of the weighting functions such that the above specification can be easily expressed, the formula \eqref{eq:weight_formula} is proposed.
\caption{\label{fig:weight_formula}Magnitude of a weighting function generated using the proposed formula \eqref{eq:weight_formula}, \(G_0=1e^{-3}\), \(G_\infty=10\), \(\omega_c =\SI{10}{Hz}\), \(G_c =2\), \(n =3\)}
\end{figure}
\subsection{Validation of the proposed synthesis method}
The proposed methodology for the design of complementary filters is now applied on a simple example where two complementary filters \(H_1(s)\) and \(H_2(s)\) have to be designed such that:
The \(\mathcal{H}_\infty\) synthesis is applied on the generalized plant of Figure \ref{fig:h_infinity_robust_fusion} using the Matlab \texttt{hinfsyn} command.
The synthesized filter \(H_2(s)\) is such that \(\mathcal{H}_\infty\) norm between \(w\) and \([z_1,\ z_2]^T\) is minimized and here found close to one \eqref{eq:hinf_synthesis_result}.
The bode plots of the obtained complementary filters are shown by solid lines in Figure \ref{fig:hinf_filters_results} and their transfer functions in the Laplace domain are given in \eqref{eq:hinf_synthesis_result_tf}.
The obtained transfer functions are of order \(5\) as expected (sum of the weighting functions orders), and their magnitudes are bellow the maximum specified ones as ensured by \eqref{eq:hinf_synthesis_result}.
\caption{\label{fig:hinf_filters_results}Frequency response of the weighting functions and complementary filters obtained using \(\mathcal{H}_\infty\) synthesis}
Sensor fusion using complementary filters are widely used in active vibration isolation systems in gravitational wave detectors such at the LIGO \cite{matichard15_seism_isolat_advan_ligo,hua05_low_ligo}, the VIRGO \cite{lucia18_low_frequen_optim_perfor_advan,heijningen18_low} and the KAGRA \cite{akutsu21_vibrat_isolat_system_beam_split}.
\caption{\label{fig:ligo_super_sensor_architecture}Simplified block diagram of the sensor blending strategy for the first stage at the LIGO \cite{hua04_low_ligo}}
\end{figure}
The fusion of the position sensor at low frequency with the ``inertial super sensor'' at high frequency using the complementary filters (\(L_1,H_1\)) is done for several reasons, first of which is to give the super sensor a DC sensibility and therefore allow the feedback loop to have authority at zero frequency.
The requirements on those filters are very tight and thus their design is complex and should be expressed as an optimization problem.
The approach used in \cite{hua05_low_ligo} is to use FIR complementary filters and to write the synthesis as a convex optimization problem.
After synthesis, the obtained FIR filters were found to be compliant with the requirements.
However they are of very high order so their implementation is quite complex.
In this section, the effectiveness of the proposed complementary filter synthesis strategy is demonstrated on the same set of requirements.
The specifications for the set of complementary filters (\(L_1,H_1\)) used at the LIGO are summarized below (for further details, refer to \cite{hua04_polyp_fir_compl_filter_contr_system}):
The weighting functions should be designed such that their inverse magnitude is as close as possible to the specifications in order to not over-constrain the synthesis problem.
However, the order of each weight should stay reasonably small in order to reduce the computational costs of the optimization problem as well as for the physical implementation of the filters.
A Type I Chebyshev filter of order \(20\) is used as the weighting transfer function \(w_L(s)\) corresponding to the low pass filter.
For the one corresponding to the high pass filter \(w_H(s)\), a \(7^{\text{th}}\) order transfer function is designed.
The magnitudes of the weighting functions are shown in Fig. \ref{fig:ligo_weights}.
\caption{\label{fig:comp_fir_ligo_hinf}Comparison of the FIR filters (solid) designed in \cite{hua05_low_ligo} with the filters obtained with \(\mathcal{H}_\infty\) synthesis (dashed)}
&\left| H_i(j\omega) \right| < \frac{1}{\left| W_i(j\omega) \right|}, \quad\forall\omega,\ i = 1 \dots n \label{eq:hinf_cond_perf_gen}
\end{align}
\end{subequations}
The synthesis method is generalized here for the synthesis of three complementary filters using the architecture shown in Fig. \ref{fig:comp_filter_three_hinf}.
The \(\mathcal{H}_\infty\) synthesis objective applied on \(P(s)\) is to design two stable filters \(H_2(s)\) and \(H_3(s)\) such that the \(\mathcal{H}_\infty\) norm of the transfer function from \(w\) to \([z_1,\ z_2, \ z_3]\) is less than one \eqref{eq:hinf_syn_obj_three}.
\caption{\label{fig:comp_filter_three_hinf}Architecture for \(\mathcal{H}_\infty\) synthesis of three complementary filters}
\end{figure}
By choosing \(H_1(s)\triangleq1- H_2(s)- H_3(s)\), the proposed \(\mathcal{H}_\infty\) synthesis solves the design problem \eqref{eq:hinf_problem_gen}. \par
An example is given to validate the method where three sensors are used in different frequency bands (up to \(\SI{1}{Hz}\), from \(1\) to \(\SI{10}{Hz}\) and above \(\SI{10}{Hz}\) respectively).
Three weighting functions are designed using \eqref{eq:weight_formula} and shown by dashed curves in Fig. \ref{fig:three_complementary_filters_results}.
The bode plots of the obtained complementary filters are shown in Fig. \ref{fig:three_complementary_filters_results}.
\caption{\label{fig:three_complementary_filters_results}Frequency response of the weighting functions and three complementary filters obtained using \(\mathcal{H}_\infty\) synthesis}
This paper has shown how complementary filters can be used to combine multiple sensors in order to obtain a super sensor.
Typical specification on the super sensor noise and on the robustness of the sensor fusion has been shown to be linked to the norm of the complementary filters.
Therefore, a synthesis method that permits the shaping of the complementary filters norms has been proposed and has been successfully applied for the design of complex filters.
Future work will aim at further developing this synthesis method for the robust and optimal synthesis of complementary filters used in sensor fusion.
