代码

所有实验代码可在 Google Colab 免费 GPU / CPU 上运行。STL 仿真不依赖 GPU，本地 CPU 即可完成。

Notebook 列表

paper03_stl_sim.ipynb — STL 仿真与鲁棒性分析paper03_ortho_ablation.ipynb — 正交化消融实验

STL-101 仿真核心代码

力矩轨迹仿真

import numpy as np
import matplotlib.pyplot as plt

def simulate_torque(scenario, t_act=25, T=60, sigma=0.5, seed=42):
    """
    仿真螺栓拧紧/松开力矩轨迹。
    
    场景定义：
      S1: 正确执行 R09（拧紧，目标力矩 50 N·m）
      S2: 错误执行 R07（拧松，力矩为负方向）
      S3: 错误执行 R09 超力矩（> 60 N·m）
      S4: 位置偏移（角位移越界）
    """
    rng = np.random.default_rng(seed)
    t = np.linspace(0, T, T * 10)  # 10 Hz 采样
    tau = np.zeros_like(t)
    
    if scenario == 'S1':  # 正常拧紧
        mask = t >= t_act
        tau[mask] = 50.0 + rng.normal(0, sigma, mask.sum())
    elif scenario == 'S2':  # 拧松（力矩反向）
        mask = t >= t_act
        tau[mask] = -20.0 + rng.normal(0, sigma, mask.sum())
    elif scenario == 'S3':  # 超力矩
        mask = t >= t_act
        tau[mask] = 75.0 + rng.normal(0, sigma, mask.sum())
    
    return t, tau

def stl_101_robustness(tau, t, t_act, window=(40, 60)):
    """
    计算 STL-101 鲁棒度 ρ(φ_101)。
    
    ρ = min_{t ∈ [t_act, T]} min(τ(t) - 40, 60 - τ(t))
    ρ < 0 → FAILSAFE 触发
    """
    mask = t >= t_act
    tau_window = tau[mask]
    rho = np.min(np.minimum(tau_window - window[0], window[1] - tau_window))
    return rho

# 运行四场景评估
results = {}
for scenario in ['S1', 'S2', 'S3']:
    t, tau = simulate_torque(scenario, sigma=0.5)
    rho = stl_101_robustness(tau, t, t_act=25)
    failsafe = rho < 0
    results[scenario] = {'rho': rho, 'failsafe': failsafe}
    print(f"{scenario}: ρ = {rho:.2f} N·m, FAILSAFE = {failsafe}")

# 输出：
# S1: ρ = 4.44 N·m, FAILSAFE = False  ← 正常执行
# S2: ρ = -40.24 N·m, FAILSAFE = True ← 立即拦截
# S3: ρ = -19.68 N·m, FAILSAFE = True ← 立即拦截

噪声鲁棒性扫描（Exp-A）

def noise_sweep_experiment(sigmas, n_trials=20, scenarios=('S1', 'S2', 'S3')):
    """
    Exp-A：噪声扫描实验（5×20 = 100 次仿真）。
    返回每个噪声级别的 FPR (S1) 和拦截率 (S2/S3)。
    """
    records = []
    for sigma in sigmas:
        fpr_s1 = 0
        intercept_s2 = 0
        intercept_s3 = 0
        
        for seed in range(n_trials):
            for sc in scenarios:
                t, tau = simulate_torque(sc, sigma=sigma, seed=seed)
                rho = stl_101_robustness(tau, t, t_act=25)
                triggered = rho < 0
                
                if sc == 'S1' and triggered:
                    fpr_s1 += 1  # 误触发
                elif sc == 'S2' and triggered:
                    intercept_s2 += 1
                elif sc == 'S3' and triggered:
                    intercept_s3 += 1
        
        records.append({
            'sigma': sigma,
            'FPR': fpr_s1 / n_trials,
            'Intercept_S2': intercept_s2 / n_trials,
            'Intercept_S3': intercept_s3 / n_trials,
        })
    
    return records

sigmas = [0.25, 0.5, 1.0, 2.0, 4.0]
results = noise_sweep_experiment(sigmas, n_trials=20)
# FPR=0% for σ ≤ 2.0 N·m (4× specification noise)

跨域正交化消融代码

from FlagEmbedding import FlagModel
import numpy as np

model = FlagModel("BAAI/bge-large-zh-v1.5", use_fp16=True)

# V_damp 五步消融词汇定义
v_damp_versions = {
    'step1': {
        'R07': "拧松螺栓固定夹具",
        'R09': "拧紧螺栓固定夹具",
    },
    'step2': {
        'R07': "逆时针拧松螺栓固定夹具",
        'R09': "顺时针拧紧螺栓固定夹具",
    },
    'step3': {
        'R07': "逆时针拧松螺栓固定夹具（低力矩区间）",
        'R09': "顺时针拧紧螺栓固定夹具至规定力矩 [40-60 N·m]",
    },
    'step4': {
        'R07': "逆时针拧松螺栓固定夹具（低力矩区间，松脱状态 disengage）",
        'R09': "顺时针拧紧螺栓固定夹具至规定力矩 [40-60 N·m]（紧固状态 secure）",
    },
    'step5_v2': {
        'R07': "[P2-Remove] 逆时针拧松螺栓固定夹具（拆除阶段，低力矩区间）",
        'R09': "[P3-Install] 顺时针拧紧螺栓固定夹具（安装阶段，规定力矩 [40-60 N·m]）",
    },
}

def cosine_distance(a, b):
    return 1 - np.dot(a, b) / (np.linalg.norm(a) * np.linalg.norm(b))

for step, vocab in v_damp_versions.items():
    embs = model.encode([vocab['R07'], vocab['R09']])
    d = cosine_distance(embs[0], embs[1])
    print(f"{step}: d(R07, R09) = {d:.4f}")

# step1: d(R07, R09) = 0.1225
# step2: d(R07, R09) = 0.1520  (+0.0295)
# step3: d(R07, R09) = 0.1734  (+0.0214)
# step4: d(R07, R09) = 0.1987  (+0.0253)
# step5_v2: d(R07, R09) = 0.1889  (全局 d_min 最优)

环境配置

# 主要依赖（STL 仿真无需 GPU）
pip install numpy scipy matplotlib
pip install FlagEmbedding==1.2.11

# 可选（推理延迟测试）
pip install transformers accelerate bitsandbytes