% Controller design for a ducted fan VTOL micro-UAV. % % Copyright (c) 2024, Naoki Sean Pross, ETH Zürich % This work is distributed under a permissive license, see LICENSE.txt % ------------------------------------------------------------------------ % Clear environment and generate parameters clear; clc; close all; s = tf('s'); do_plots = true; % runs faster without do_hinf = true; % midterm do_musyn = true; % endterm fprintf('Controller synthesis for ducted fan VTOL micro-UAV\n') fprintf('Will do:\n') if do_plots fprintf(' - Produce plots\n') end if do_hinf fprintf(' - H-infinity synthesis\n') end if do_musyn fprintf(' - Mu synthesis\n') end % Synthesized controllers will be stored here ctrl = struct(); %% ------------------------------------------------------------------------ % Define system parameters fprintf('Generating system parameters...\n') params = uav_params(); %% ------------------------------------------------------------------------ % Define performance requirements fprintf('Generating performance requirements...\n') perf = uav_performance_musyn(params, do_plots); %% ------------------------------------------------------------------------ % Define stability requirements fprintf('Generating stability requirements...\n') uncert = uav_uncertainty(params, do_plots); %% ------------------------------------------------------------------------ % Create UAV model fprintf('Generating system model...\n'); model = uav_model(params, perf, uncert); %% ------------------------------------------------------------------------ % Perform H-infinity design if do_hinf fprintf('Performing H-infinty controller design...\n') idx = model.uncertain.index; P = model.uncertain.StateSpace; % Get nominal system without uncertainty (for lower LFT) P_nom = minreal(P([idx.OutputError; idx.OutputNominal], ... [idx.InputExogenous; idx.InputNominal]), [], false); nmeas = model.uncertain.Ny; nctrl = model.uncertain.Nu; hinfopt = hinfsynOptions('Display', 'off', 'Method', 'RIC', ... 'AutoScale', 'off', 'RelTol', 1e-3); [K_inf, ~, gamma, info] = hinfsyn(P_nom, nmeas, nctrl, hinfopt); fprintf(' - H-infinity synthesis gamma: %g\n', gamma); ctrl.hinf = struct('Name', '$\mathcal{H}_{\infty}$', 'K', K_inf); if gamma >= 1 error('Failed to syntesize controller (closed loop is unstable).') end %% ------------------------------------------------------------------------ % Measure Performance of H-infinity design fprintf(' - Simulating closed loop...\n'); T = 60; nsamples = 500; do_noise = true; simout = uav_sim_step(params, model, ctrl.hinf, nsamples, T, do_plots, do_noise); fprintf(' - Writing simulation results...\n'); cols = [ simout.StepX(:, simout.index.Position), ... simout.StepX(:, simout.index.Velocity), ... simout.StepX(:, simout.index.FlapAngles) * 180 / pi, ... simout.StepX(:, simout.index.Angles) * 180 / pi]; writematrix([simout.Time', cols], 'fig/stepsim.dat', 'Delimiter', 'tab') end %% ------------------------------------------------------------------------ % Perform mu-Analysis & DK iteration if do_musyn drawnow; fprintf('Performing mu-synthesis controller design...\n') % Get complete system (without debugging outputs for plots) idx = model.uncertain.index; P = minreal(model.uncertain.StateSpace(... [idx.OutputUncertain; idx.OutputError; idx.OutputNominal], ... [idx.InputUncertain; idx.InputExogenous; idx.InputNominal]), ... [], false); % Options for H-infinity nmeas = model.uncertain.Ny; nctrl = model.uncertain.Nu; hinfopt = hinfsynOptions('Display', 'off', 'Method', 'RIC', ... 'AutoScale', 'on', 'RelTol', 1e-3); % Frequency raster resolution to fit D scales nsamples = 600; omega_max = 3; omega_min = -3; omega = logspace(omega_min, omega_max, nsamples); omega_range = {10^omega_min, 10^omega_max}; % Initial values for D-K iteration are identity matrices D_left = tf(eye(model.uncertain.Nz + model.uncertain.Ne + model.uncertain.Ny)); D_right = tf(eye(model.uncertain.Nv + model.uncertain.Nw + model.uncertain.Nu)); % Maximum number of D-K iterations niters = 5; fprintf(' - Will do at most %d iterations.\n', niters); % Maximum degree of D-scales and error d_scales_max_degree = 3; d_scales_max_err_p = .4; % in percentage d_scales_improvement_p = .15; % in percentage % for plotting later mu_plot_legend = {}; % Start DK-iteration dkstart = tic; for it = 1:niters fprintf(' * Running D-K iteration %d / %d...\n', it, niters); itstart = tic(); % Find controller using H-infinity P_scaled = minreal(D_left * P * inv(D_right), [], false); [P_scaled, ~] = prescale(P_scaled, omega_range); [K, ~, gamma, ~] = hinfsyn(P_scaled, nmeas, nctrl, hinfopt); fprintf(' H-infinity synthesis gamma: %g\n', gamma); if gamma == inf error('Failed to synethesize H-infinity controller'); end % Calculate frequency response of closed loop N = minreal(lft(P, K), [], false); M = minreal(N(idx.OutputUncertain, idx.InputUncertain), [], false); [N, ~] = prescale(N, omega_range); [M, ~] = prescale(M, omega_range); N_frd = frd(N, omega); M_frd = frd(M, omega); % Calculate upper bound D scaling fprintf(' Computing Performance SSV... ') [mu_bounds_rp, mu_info_rp] = mussv(N_frd, model.uncertain.BlockStructurePerf, 'U'); fprintf(' Computing Stability SSV... ') [mu_bounds_rs, mu_info_rs] = mussv(M_frd, model.uncertain.BlockStructure, 'U'); mu_rp = norm(mu_bounds_rp(1,1), inf, 1e-6); mu_rs = norm(mu_bounds_rs(1,1), inf, 1e-6); fprintf(' SSV for Performance: %g, for Stability: %g\n', mu_rp, mu_rs); if do_plots fprintf(' Plotting SSV mu\n'); figure(100); hold on; bodemag(mu_bounds_rp(1,1)); mu_plot_legend = {mu_plot_legend{:}, sprintf('$\\mu_{P,%d}$', it)}; bodemag(mu_bounds_rs(1,1), 'k:'); mu_plot_legend = {mu_plot_legend{:}, sprintf('$\\mu_{S,%d}$', it)}; title('\bfseries $\mu_\Delta(\omega)$ for both Stability and Performance', ... 'interpreter', 'latex'); legend(mu_plot_legend, 'interpreter', 'latex'); grid on; drawnow; end % Are we done yet? if mu_rp < 1 fprintf(' - Found robust controller that meets performance.\n'); break; end if mu_rs < 1 fprintf(' Found robust controller that is stable.\n') ctrl.musyn = struct('Name', '$\mu$-Synthesis', 'K', K, ... 'mu_rp', mu_rp, 'mu_rs', mu_rs); end % Fit D-scales [D_left_frd, D_right_frd] = mussvunwrap(mu_info_rp); fprintf(' Fitting D-scales\n'); % There are three complex, square, full block uncertainties and % a non-square full complex block for performance i_alpha = [1, 1]; i_omega = model.uncertain.BlockStructure(1, :) + 1; % after first block i_state = sum(model.uncertain.BlockStructure(1:2, :)) + 1; % after second block i_perf = sum(model.uncertain.BlockStructurePerf(1:3, :)) + 1; % after third block D_frd = { D_left_frd(i_alpha(1), i_alpha(1)); D_left_frd(i_omega(1), i_omega(1)); D_left_frd(i_state(1), i_state(1)); D_left_frd(i_perf(1), i_perf(1)); }; D_max_sv = { max(max(sigma(D_frd{1, 1}))); max(max(sigma(D_frd{2, 1}))); max(max(sigma(D_frd{3, 1}))); max(max(sigma(D_frd{4, 1}))); }; D_names = {'alpha', 'omega', 'state', 'perf'}; D_fitted = {}; % for each block for j = 1:4 % for each in left and right fprintf(' %s', D_names{j}); best_fit_deg = inf; best_fit_err = inf; for deg = 0:d_scales_max_degree % Fit D-scale D_fit = fitmagfrd(D_frd{j}, deg); % D_fit = fitfrd(genphase(D_frd{j}), deg); % Check if it is a good fit max_sv = max(max(sigma(D_fit, omega))); fit_err = abs(D_max_sv{j} - max_sv); if fit_err < best_fit_err % Choose higher degree only if we improve by at least a specified % percentage over the previous best fit (or we are at the first % iteration). This is a heuristic to avoid adding too many states % to the controller as it depends on the order of the D-scales. if abs(best_fit_err - fit_err) / best_fit_err > d_scales_improvement_p || best_fit_err == inf best_fit_deg = deg; best_fit_err = fit_err; D_fitted{j} = D_fit; end end if (fit_err / D_max_sv{j} < d_scales_max_err_p) break; end fprintf('.'); end fprintf(' degree %d, error %g (%g %%)\n', ... best_fit_deg, best_fit_err, 100 * best_fit_err / D_max_sv{j}); end % Construct full matrices D_left = blkdiag(D_fitted{1} * eye(4), ... D_fitted{2} * eye(1), ... D_fitted{3} * eye(12), ... D_fitted{4} * eye(14), ... eye(12)); D_right = blkdiag(D_fitted{1} * eye(4), ... D_fitted{2} * eye(1), ... D_fitted{3} * eye(12), ... D_fitted{4} * eye(10), ... eye(5)); % Compute peak of singular values for to estimate how good is the % approximation of the D-scales sv_left_frd = sigma(D_left_frd); max_sv_left_frd = max(max(sv_left_frd)); sv_left = sigma(D_left, omega); max_sv_left = max(max(sv_left)); fprintf(' Max SVD of D: %g, Dhat: %g\n', max_sv_left_frd, max_sv_left); fprintf(' D scales fit rel. error: %g %%\n', ... 100 * abs(max_sv_left_frd - max_sv_left) / max_sv_left_frd); % Plot fitted D-scales if do_plots fprintf(' Plotting D-scales'); f = figure(101); clf(f); hold on; bodemag(D_frd{1}, omega, 'r-'); bodemag(D_fitted{1}, omega, 'b'); fprintf('.'); bodemag(D_frd{2}, omega, 'r--'); bodemag(D_fitted{2}, omega, 'b--'); fprintf('.'); bodemag(D_frd{3}, omega, 'c-'); bodemag(D_fitted{3}, omega, 'm-'); fprintf('.'); bodemag(D_frd{4}, omega, 'c--'); bodemag(D_fitted{4}, omega, 'm--'); fprintf('.'); fprintf('\n'); title(sprintf('\\bfseries $D(\\omega)$ Scales Approximations at Iteration %d', it), ... 'interpreter', 'latex') legend(... '$D_{\alpha}$', '$\hat{D}_{\alpha}$', ... '$D_{\omega}$', '$\hat{D}_{\omega}$', ... '$D_{\mathbf{x}}$', '$\hat{D}_{\mathbf{x}}$', ... '$D_{\Delta}$', '$\hat{D}_{\Delta}$', ... 'interpreter', 'latex' ... ); grid on; drawnow; end itend = toc(itstart); fprintf(' Iteration took %.1f seconds\n', itend); end dkend = toc(dkstart); fprintf(' - D-K iteration took %.1f seconds\n', dkend); if mu_rp > 1 && mu_rs > 1 error(' - Failed to synthesize robust controller that meets the desired performance.\n'); end %% Fit worst-case perturbation fprintf(' - Computing worst case perturbation.\n') % Find peak of mu [mu_upper_bound_rp, ~] = frdata(mu_bounds_rp(1,1)); max_mu_rp_idx = find(mu_rp == mu_upper_bound_rp, 1); Delta = mussvunwrap(mu_info_rp); % TODO: finish here % Save controller ctrl.musyn = struct('Name', '$\mu$-Synthesis', ... 'K', K, 'Delta', Delta, ... 'mu_rp', mu_rp, 'mu_rs', mu_rs); if mu_rp >= 1 fprintf(' - Synthetized robust stable controller does not meet the desired performance.\n'); end %% ------------------------------------------------------------------------ % Measure Performance of mu synthesis design fprintf('Simulating nominal closed loop...\n'); T = 60; nsamples = 5000; do_noise = true; simout = uav_sim_step(params, model, ctrl.musyn, nsamples, T, do_plots, do_noise); fprintf('Simulating worst-case closed loop...\n'); end