Anomaly Detection Foundation Model — 异常检测基础模型：零样本时序异常感知

"""
Anomaly Detection Foundation Model (ChronosAD-style)
时序基础模型驱动的零样本异常检测
"""
import numpy as np
from scipy import stats


class FoundationAnomalyDetector:
    """
    ChronosAD 风格的异常检测器
    核心：用预测误差量化异常程度
    生产环境推荐: pip install chronos-forecasting 后替换预测模块
    """

    def __init__(self, context_len: int = 30, sensitivity: float = 2.5):
        self.context_len = context_len
        self.sensitivity = sensitivity  # 异常阈值（σ倍数）

    def _foundation_predict(self, history: np.ndarray, steps: int = 1) -> tuple:
        """
        模拟基础模型预测（生产替换为 ChronosPipeline.predict()）
        返回 (预测中位数, 预测标准差)
        """
        if len(history) < 7:
            return history.mean(), history.std() + 1e-8

        # 简化预测：局部趋势 + 季节性
        recent = history[-min(14, len(history)):]
        trend = np.polyfit(range(len(recent)), recent, 1)[0]

        # 季节性（7天周期）
        if len(history) >= 14:
            weekly = np.array([history[i::7].mean() for i in range(7)])
            seasonal_factor = weekly[len(history) % 7] / (weekly.mean() + 1e-8)
        else:
            seasonal_factor = 1.0

        pred_mean = (recent[-1] + trend) * seasonal_factor
        # 不确定性：历史方差 + 预测步数（越远越不确定）
        pred_std = recent.std() * (1 + 0.1 * steps) + 1e-8

        return float(pred_mean), float(pred_std)

    def compute_anomaly_scores(self, timeseries: np.ndarray) -> np.ndarray:
        """
        计算每个时间点的异常分数
        使用滑动窗口：用历史预测当前值
        """
        n = len(timeseries)
        scores = np.zeros(n)

        for t in range(min(self.context_len, 7), n):
            history = timeseries[max(0, t - self.context_len):t]
            pred_mean, pred_std = self._foundation_predict(history)
            actual = timeseries[t]
            # 标准化异常分
            scores[t] = abs(actual - pred_mean) / pred_std

        return scores