"""M3 实验:对数线性 vs 非线性最小二乘(三档 SNR + 已知/估基线)。 运行:python -m fiber_delay_sim.experiments.exp04_m3_algorithms """ import argparse import numpy as np from ..core import fitting from ..export import record from ._plot import setup_style def run(seeds=8, runs_dir="runs"): plt = setup_style() t = np.arange(0.0, 400.0, 10.0) true_tau = 52.8 snrs = (30.0, 50.0, 70.0) # 三档 SNR × 多种子:三种方法的平均相对误差 summary = [] for snr in snrs: errs = {"nls": [], "ll_known": [], "ll_est": []} for s in range(seeds): rng = np.random.default_rng(s) y = np.exp(-t / true_tau) sigma = np.max(y) * 10 ** (-snr / 20) yn = y + rng.normal(0, sigma, len(t)) r1 = fitting.fit_single_exp_nls(t, yn) errs["nls"].append(abs(r1["tau"] - true_tau) / true_tau) r2 = fitting.fit_log_linear(t, yn, b=0.0) # 已知基线(理想) if r2["tau"]: errs["ll_known"].append(abs(r2["tau"] - true_tau) / true_tau) r3 = fitting.fit_log_linear(t, yn) # 默认尾段估基线 if r3["tau"]: errs["ll_est"].append(abs(r3["tau"] - true_tau) / true_tau) summary.append({"snr": snr, "nls": float(np.mean(errs["nls"])), "ll_known": float(np.mean(errs["ll_known"])) if errs["ll_known"] else None, "ll_est": float(np.mean(errs["ll_est"])) if errs["ll_est"] else None}) print(f"\n=== M3 算法对比(真值 τ = {true_tau} ns,{seeds} 个种子平均)===") print(f"{'SNR(dB)':>8} {'NLS':>10} {'对数(已知基线)':>14} {'对数(估基线)':>14}") for r in summary: print(f"{r['snr']:>8.0f} {r['nls'] * 100:>9.2f}% " f"{(r['ll_known'] * 100 if r['ll_known'] is not None else float('nan')):>13.2f}% " f"{(r['ll_est'] * 100 if r['ll_est'] is not None else float('nan')):>13.2f}%") # ---------- 图 1:SNR 50 示例,两方法拟合对比(对数坐标展示尾部偏离) ---------- fig, ax = plt.subplots(figsize=(7.5, 4.4)) rng = np.random.default_rng(3) y = np.exp(-t / true_tau) sigma = np.max(y) * 10 ** (-50 / 20) yn = y + rng.normal(0, sigma, len(t)) ax.semilogy(t, yn, "o", ms=4, color="#9aa5b1", label="采样数据(SNR 50 dB)") ax.semilogy(t, np.exp(-t / true_tau), "k-", lw=1.2, label="真值") r_nls = fitting.fit_single_exp_nls(t, yn) r_ll = fitting.fit_log_linear(t, yn) # 估基线 ax.semilogy(t, r_nls["A"] * np.exp(-t / r_nls["tau"]) + r_nls["B"], "-", color="#2D5F8A", label=f"NLS:τ̂ = {r_nls['tau']:.2f} ns") if r_ll["tau"]: ax.semilogy(t, r_ll["A"] * np.exp(-t / r_ll["tau"]), "--", color="#C25B2A", label=f"对数线性(估基线):τ̂ = {r_ll['tau']:.2f} ns") ax.set_xlabel("时间 (ns)") ax.set_ylabel("归一化强度(对数坐标)") ax.set_title("M3:对数线性 vs NLS——低幅值尾部经对数变换后噪声权重失真") ax.legend(fontsize=9) ax.grid(alpha=0.3, which="both") # ---------- 图 2:平均误差 vs SNR 折线 ---------- fig2, ax2 = plt.subplots(figsize=(7, 4)) xs = [r["snr"] for r in summary] ax2.plot(xs, [r["nls"] * 100 for r in summary], "o-", color="#2D5F8A", label="NLS") ax2.plot(xs, [r["ll_known"] * 100 for r in summary], "s--", color="#6B2A35", label="对数线性(已知基线)") ax2.plot(xs, [r["ll_est"] * 100 for r in summary], "^--", color="#C25B2A", label="对数线性(尾段估基线)") ax2.set_xlabel("SNR (dB)") ax2.set_ylabel("平均相对误差 |τ̂−τ|/τ (%)") ax2.set_title("M3:三档信噪比下算法平均误差") ax2.legend() ax2.grid(alpha=0.3) run_dir = record.save_run( runs_dir, "exp04_m3_algorithms", records=[{"meta": {"experiment": "m3_loglinear_vs_nls", "data_origin": "software-simulated"}, "summary": summary, "true_tau": true_tau}], figures=[("fig1_fit_example", fig), ("fig2_snr_curve", fig2)], notes=("M3 算法对比:NLS 稳健;对数线性在已知基线下数学等价、" "在默认尾段估基线下引入系统偏差(对数变换放大低幅值尾部噪声)。"), ) print(f"\n[M3] 记录包:{run_dir}") return run_dir if __name__ == "__main__": ap = argparse.ArgumentParser() ap.add_argument("--seeds", type=int, default=8) ap.add_argument("--runs-dir", default="runs") a = ap.parse_args() run(seeds=a.seeds, runs_dir=a.runs_dir)