"""
论文核心实验复现：交叉验证

方法：
1. 留一飞行员交叉验证
2. 训练集: 39名飞行员特征 → 教练评分
3. 测试集: 第40名飞行员
4. 对比: 传感器特征 vs 仅任务标识基线

结果: 传感器特征未超过仅任务基线
说明: 消费级可穿戴信号尚不支持跨飞行员泛化预测
"""

import numpy as np
from sklearn.model_selection import LeaveOneGroupOut
from sklearn.ensemble import RandomForestRegressor
from sklearn.metrics import mean_absolute_error

def cross_pilot_validation(features, labels, pilot_ids):
    """
    留一飞行员交叉验证
    
    Args:
        features: [n_trials, n_features] 传感器特征
        labels: [n_trials] 教练评分
        pilot_ids: [n_trials] 飞行员ID
        
    Returns:
        mae: 跨飞行员MAE
        baseline_mae: 仅任务基线MAE
    """
    logo = LeaveOneGroupOut()
    model = RandomForestRegressor(n_estimators=100, random_state=42)
    
    maes = []
    baseline_maes = []
    
    for train_idx, test_idx in logo.split(features, labels, pilot_ids):
        X_train, X_test = features[train_idx], features[test_idx]
        y_train, y_test = labels[train_idx], labels[test_idx]
        
        # 传感器特征模型
        model.fit(X_train, y_train)
        pred = model.predict(X_test)
        maes.append(mean_absolute_error(y_test, pred))
        
        # 仅任务基线（预测每个任务的平均评分）
        task_ids_train = features[train_idx, 0]  # 任务ID作为第一列
        task_means = {}
        for t in np.unique(task_ids_train):
            mask = task_ids_train == t
            task_means[t] = y_train[mask].mean()
        
        task_ids_test = features[test_idx, 0]
        baseline_pred = [task_means.get(t, y_train.mean()) for t in task_ids_test]
        baseline_maes.append(mean_absolute_error(y_test, baseline_pred))
    
    return np.mean(maes), np.mean(baseline_maes)

# 模拟结果（按论文报告）
if __name__ == "__main__":
    # 模拟40飞行员 × 5任务 = 200试次
    np.random.seed(42)
    n_pilots = 40
    n_tasks = 5
    n_trials = n_pilots * n_tasks
    
    # 特征: [task_id, heart_rate, eda, eye_rate, eeg_theta_alpha, ...]
    features = np.random.normal(0, 1, (n_trials, 10))
    features[:, 0] = np.tile(np.arange(n_tasks), n_pilots)  # 任务ID
    
    # 标签: 教练评分（1-5分）
    pilot_ids = np.repeat(np.arange(n_pilots), n_tasks)
    labels = np.random.randint(1, 6, n_trials).astype(float)
    
    mae, baseline_mae = cross_pilot_validation(features, labels, pilot_ids)
    
    print("=== 留一飞行员交叉验证结果 ===")
    print(f"传感器特征MAE: {mae:.3f}")
    print(f"仅任务基线MAE: {baseline_mae:.3f}")
    print(f"传感器 > 基线? {'是' if mae < baseline_mae else '否'}")
    print(f"结论: 消费级传感器特征{'支持' if mae < baseline_mae else '不支持'}跨飞行员泛化")

# 个体校准框架
class IndividualCalibration:
    """
    驾驶员个体校准框架
    基于论文发现：跨人泛化不可靠，需个体基线
    """
    
    def __init__(self, calibration_duration_min=15):
        self.calibration_duration = calibration_duration_min * 60  # 秒
        self.baseline = {}
        self.calibrated = False
    
    def collect_baseline(self, signals: dict, fps: int = 30):
        """
        收集个体基线数据
        
        建议时长: 15分钟正常驾驶
        采集信号: 心率、眼动、PERCLOS、EDA
        """
        # 需要至少15分钟数据
        required_samples = self.calibration_duration * fps
        if len(signals.get('heart_rate', [])) < required_samples:
            return False
        
        self.baseline = {
            'hr_mean': np.mean(signals['heart_rate']),
            'hr_std': np.std(signals['heart_rate']),
            'perclos_baseline': np.mean(signals['perclos']),
            'gaze_displacement_baseline': np.mean(signals['gaze_disp']),
            'blink_rate_baseline': np.mean(signals['blink_rate']),
        }
        self.calibrated = True
        return True
    
    def compute_fatigue_index(self, current_signals: dict) -> float:
        """
        基于个体基线计算疲劳指数
        
        Returns:
            fatigue_index: 0-1 (0=清醒, 1=严重疲劳)
        """
        if not self.calibrated:
            return 0.5  # 未校准返回中性值
        
        # 相对变化（z-score化）
        hr_z = (np.mean(current_signals['heart_rate']) - 
                self.baseline['hr_mean']) / self.baseline['hr_std']
        
        perclos_z = (np.mean(current_signals['perclos']) - 
                     self.baseline['perclos_baseline']) / max(
                         self.baseline['perclos_baseline'], 0.01)
        
        # 加权融合
        fatigue_index = 0.3 * max(0, -hr_z) + 0.7 * max(0, perclos_z)
        return min(1.0, fatigue_index)