153 lines
6.5 KiB
Python
153 lines
6.5 KiB
Python
import torch
|
|
import torch.nn as nn
|
|
import torch.optim as optim
|
|
import numpy as np
|
|
import matplotlib.pyplot as plt
|
|
import time
|
|
|
|
# ==========================================
|
|
# 0. 全局设置 (对标你的 -35° 强杂波挑战)
|
|
# ==========================================
|
|
plt.rcParams['font.sans-serif'] = ['Microsoft YaHei']
|
|
plt.rcParams['axes.unicode_minus'] = False
|
|
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
|
print(f"========== 启动 TSC-DL 扇区聚焦引擎 ==========\n使用设备: {device}")
|
|
|
|
N_ant, N_RF, d_lambda = 24, 4, 0.5
|
|
noise_power = 1.0
|
|
INR_linear = 10**(30/10) # 30dB 强杂波
|
|
|
|
#核心测试场景
|
|
test_t_ang, test_c_ang = 15.0, -25.0
|
|
|
|
def gen_a_batch(theta_deg_tensor):
|
|
theta_rad = torch.deg2rad(theta_deg_tensor)
|
|
n_idx = torch.arange(-(N_ant-1)/2, (N_ant-1)/2 + 0.1, device=device).unsqueeze(0)
|
|
phases = 2 * torch.pi * d_lambda * n_idx * torch.sin(theta_rad.unsqueeze(1))
|
|
return torch.exp(1j * phases).unsqueeze(2)
|
|
|
|
# ==========================================
|
|
# 1. 复数域正交网络 (彻底消灭相位梯度死锁)
|
|
# ==========================================
|
|
class SectorHBFNet(nn.Module):
|
|
def __init__(self):
|
|
super().__init__()
|
|
self.net = nn.Sequential(
|
|
nn.Linear(48, 512), nn.GELU(),
|
|
nn.Linear(512, 512), nn.GELU(),
|
|
# 【核心架构升级】：不直接输出相位！输出复数的实部和虚部 (2个通道)
|
|
nn.Linear(512, N_ant * N_RF * 2)
|
|
)
|
|
|
|
def forward(self, w_opt):
|
|
w_flat = w_opt.squeeze(2)
|
|
x = torch.cat([w_flat.real, w_flat.imag], dim=1)
|
|
out = self.net(x).view(-1, N_ant, N_RF, 2)
|
|
|
|
complex_out = out[..., 0] + 1j * out[..., 1]
|
|
# 【物理强制约束】：将复数除以自身的模长，绝对精准地满足恒模约束，且梯度无比丝滑！
|
|
F_RF = complex_out / (torch.abs(complex_out) + 1e-8) / np.sqrt(N_ant)
|
|
return F_RF
|
|
|
|
# ==========================================
|
|
# 2. TSC-DL 扇区聚焦训练
|
|
# ==========================================
|
|
model = SectorHBFNet().to(device)
|
|
optimizer = optim.Adam(model.parameters(), lr=0.002)
|
|
|
|
Batch_Size, Epochs = 500, 1500 #训练次数
|
|
loss_history = []
|
|
|
|
print("启动 TSC-DL 训练")
|
|
start_time = time.time()
|
|
for epoch in range(Epochs):
|
|
model.train()
|
|
optimizer.zero_grad()
|
|
|
|
# 【与 TSC-OMP 呼应的核心逻辑】：扇区聚焦训练！
|
|
# 不再全空域瞎猜，而是在目标和杂波的局部战术扇区（±3°）内生成海量数据进行特训
|
|
theta_t = test_t_ang + (torch.rand(Batch_Size, device=device) - 0.5) * 6.0
|
|
theta_c = test_c_ang + (torch.rand(Batch_Size, device=device) - 0.5) * 6.0
|
|
|
|
a_t, a_c = gen_a_batch(theta_t), gen_a_batch(theta_c)
|
|
|
|
# 构建纯净的基准协方差矩阵
|
|
R_interf = noise_power * torch.eye(N_ant, device=device).unsqueeze(0) + INR_linear * (a_c @ a_c.mH)
|
|
w_opt = torch.linalg.pinv(R_interf) @ a_t
|
|
w_opt = w_opt / torch.norm(w_opt, dim=1, keepdim=True)
|
|
|
|
F_RF = model(w_opt)
|
|
F_BB = torch.linalg.pinv(F_RF.mH @ F_RF + 1e-3*torch.eye(N_RF, device=device).unsqueeze(0)) @ (F_RF.mH @ w_opt)
|
|
w_ai = F_RF @ F_BB
|
|
w_ai = w_ai / torch.norm(w_ai, dim=1, keepdim=True)
|
|
|
|
# 波束相似度
|
|
sim = torch.abs(torch.sum(w_opt.conj() * w_ai, dim=(1,2)))**2
|
|
loss_sim = torch.mean(1.0 - sim)
|
|
# 杂波方向绝对抑制
|
|
loss_null = torch.mean(torch.abs(torch.sum(a_c.conj() * w_ai, dim=(1,2)))**2)
|
|
|
|
# 【重拳出击】：给予杂波方向 150 倍的极限死亡惩罚！逼迫 AI 挖出深坑！
|
|
loss = loss_sim + 150.0 * loss_null
|
|
loss.backward()
|
|
optimizer.step()
|
|
loss_history.append(loss.item())
|
|
|
|
print(f"训练完成！耗时: {time.time() - start_time:.2f} 秒\n")
|
|
|
|
# ==========================================
|
|
# 3. 静态成果检验：生成完美的波束方向图
|
|
# ==========================================
|
|
print(f"正在生成目标 {test_t_ang}°，强杂波 {test_c_ang}° 的终极抗干扰方向图...")
|
|
theta_scan = torch.arange(-90, 90.1, 0.1, device=device)
|
|
A_scan = gen_a_batch(theta_scan).squeeze(2).T
|
|
|
|
a_t_stat = gen_a_batch(torch.tensor([test_t_ang], device=device))
|
|
a_c_stat = gen_a_batch(torch.tensor([test_c_ang], device=device))
|
|
R_interf_stat = noise_power * torch.eye(N_ant, device=device).unsqueeze(0) + INR_linear * (a_c_stat @ a_c_stat.mH)
|
|
|
|
with torch.no_grad():
|
|
model.eval()
|
|
w_opt_stat = torch.linalg.pinv(R_interf_stat) @ a_t_stat
|
|
w_opt_stat /= torch.norm(w_opt_stat, dim=1, keepdim=True)
|
|
|
|
F_RF_stat = model(w_opt_stat)
|
|
F_BB_stat = torch.linalg.pinv(F_RF_stat.mH @ F_RF_stat + 1e-3*torch.eye(N_RF, device=device).unsqueeze(0)) @ (F_RF_stat.mH @ w_opt_stat)
|
|
w_ai_stat = F_RF_stat @ F_BB_stat
|
|
w_ai_stat /= torch.norm(w_ai_stat, dim=1, keepdim=True)
|
|
|
|
# 极简且精准的点乘测向
|
|
gain_dig = torch.abs(torch.sum(w_opt_stat.conj() * A_scan.unsqueeze(0), dim=1))**2
|
|
BP_dig = 10 * torch.log10(gain_dig).cpu().numpy().flatten()
|
|
BP_dig -= np.max(BP_dig)
|
|
|
|
gain_ai = torch.abs(torch.sum(w_ai_stat.conj() * A_scan.unsqueeze(0), dim=1))**2
|
|
BP_ai = 10 * torch.log10(gain_ai).cpu().numpy().flatten()
|
|
BP_ai -= np.max(BP_ai)
|
|
|
|
# ==========================================
|
|
# 4. 绘制终极展示双图
|
|
# ==========================================
|
|
fig, (ax1, ax2) = plt.subplots(1, 2, figsize=(14, 5))
|
|
|
|
ax1.plot(loss_history, color='#0072BD', linewidth=2)
|
|
ax1.set_title('扇区聚焦网络 (TSC-DL) 极限优化损失曲线', fontweight='bold', fontsize=12)
|
|
ax1.set_xlabel('迭代轮数 (Epochs)')
|
|
ax1.set_ylabel('联合 Loss')
|
|
ax1.grid(True, ls='--')
|
|
|
|
ax2.plot(theta_scan.cpu().numpy(), BP_dig, color='#000000', linestyle='-', linewidth=2, alpha=0.5, label='纯数字理想波束')
|
|
ax2.plot(theta_scan.cpu().numpy(), BP_ai, color='#D95319', linestyle='--', linewidth=2.5, label='扇区聚焦 AI 混合波束')
|
|
ax2.axvline(test_t_ang, color='g', linestyle='-.', label=f'机动目标 ({test_t_ang}°)')
|
|
ax2.axvline(test_c_ang, color='r', linestyle='-.', label=f'强杂波 ({test_c_ang}°)')
|
|
ax2.set_ylim([-60, 0])
|
|
ax2.set_xlim([-60, 60])
|
|
ax2.set_title(f'深度学习端到端波束响应 (目标 {test_t_ang}° 杂波 {test_c_ang}° )', fontweight='bold', fontsize=12)
|
|
ax2.set_xlabel('角度 θ (°)')
|
|
ax2.set_ylabel('归一化增益 (dB)')
|
|
ax2.grid(True, ls='--')
|
|
ax2.legend(loc='lower center', fontsize=10)
|
|
|
|
plt.tight_layout()
|
|
plt.savefig('DL_Sector_Focus_Final.png', dpi=300)
|
|
plt.show() |