HC900-Crawler/scripts/analysis/c6111_operator_assist.py

"""
Operator-assist 패키징 (작업3).

사용법:
  python3 c6111_operator_assist.py --data c61_data.pkl --prefix c61
  python3 c6111_operator_assist.py --data c61_data.pkl --prefix c61 --live '{"feed":500,"product":300,"T_C":84.7}'
"""
import argparse
import json
import numpy as np
import pandas as pd
from sklearn.ensemble import IsolationForest

BASE = "/home/windpacer/projects/hc900_ax/scripts/analysis/"
FEATURES = ["feed", "product", "T_C"]
PROD_SMOOTH = 40


class OperatorAssist:
    def __init__(self, df):
        self.df = df
        self.mode = "UNKNOWN"
        self.model = None
        self.inv = None
        self.ood = None
        self.env_lo = None
        self.env_hi = None
        self._train()

    def _train(self):
        prod = self.df[self.df["mode"] == "PROD"].copy()
        prod = prod[(prod["feed"] > 50) & (prod["steam_flow"] > 10) & (prod["steam_op"] > 1)]
        prod = prod.dropna(subset=FEATURES + ["steam_op", "steam_flow"])
        if len(prod) < 100:
            print("  [WARN] PROD 데이터 부족 — advisory 신뢰도 낮음")
        points = (prod.set_index("dtat").resample("6h").median(numeric_only=True)
                  .dropna(subset=["steam_flow", "feed"]))
        points = points[points["feed"] > 50]
        from sklearn.ensemble import GradientBoostingRegressor
        self.model = GradientBoostingRegressor(n_estimators=200, max_depth=2,
                                               learning_rate=0.05, random_state=0)
        self.model.fit(points[FEATURES].values, points["steam_flow"].values)
        self.inv = np.polyfit(prod["steam_flow"], prod["steam_op"], 3)
        self.env_lo = points[FEATURES].quantile(0.01)
        self.env_hi = points[FEATURES].quantile(0.99)
        self.ood = IsolationForest(contamination=0.05, random_state=0).fit(points[FEATURES].values)
        print(f"  학습 운전점: {len(points)}개  envelope:")
        for c in FEATURES:
            print(f"    {c}: [{self.env_lo[c]:.0f}, {self.env_hi[c]:.1f}]")

    def classify_mode(self, tags):
        """tags dict → mode 추정 (classify_phases 단순 replica).
        
        steam_op 없으면 feed/product로 판단 (live advisory용).
        """
        prod = tags.get("product", 0)
        feed = tags.get("feed", 0)
        steam = tags.get("steam_op", None)
        reb = tags.get("reb_temp", 60)
        if prod > 100:
            if steam is None or steam > 10:
                return "PROD"
        if steam is not None:
            if steam > 10 and reb > 60:
                return "LINEOUT"
            if steam > 10 and feed < 50:
                return "STARTUP"
        if feed > 50:
            return "PROD"  # fallback: steam_op 없이 feed>50 + product>100는 PROD
        return "STOPPED"

    def in_envelope(self, tags):
        x = np.array([[tags[c] for c in FEATURES]])
        return ((x >= self.env_lo.values) & (x <= self.env_hi.values)).all()

    def ood_score(self, tags):
        return self.ood.decision_function(np.array([[tags[c] for c in FEATURES]]))[0]

    def predict(self, tags, smooth_history=None):
        """live_tags dict → advisory dict.
        
        tags: {"feed": float, "product": float, "T_C": float}
        smooth_history: optional list of prior tag dicts for causal smoothing
        
        Returns:
          {"rec_OP": float or None, "rec_steam": float, "confidence": str,
           "mode": str, "ood": bool, "in_env": bool, "message": str}
        """
        mode = self.classify_mode(tags)
        self.mode = mode
        env = self.in_envelope(tags)
        ood = self.ood_score(tags) < 0
        raw = np.array([[[tags[c] for c in FEATURES]]])

        if mode != "PROD":
            msg = f"운전모드={mode} — advisory는 PROD에서만 제공 (STARTUP/LINEOUT은 레시피 참조)"
            return {"rec_OP": None, "rec_steam": None, "confidence": "N/A",
                    "mode": mode, "ood": ood, "in_env": env, "message": msg}

        # smooth: causal trailing median over recent history
        if smooth_history and len(smooth_history) >= PROD_SMOOTH:
            buf = pd.DataFrame(smooth_history[-PROD_SMOOTH:])[FEATURES].median()
            x = np.array([[buf[c] for c in FEATURES]])
        else:
            x = raw[0]

        sf = self.model.predict(x)[0]
        op = np.clip(np.polyval(self.inv, sf), 0, 100)

        if not env:
            confidence = "LOW_OOD"
            msg = (f"⚠ 범위밖 입력 — 권장 OP={op:.1f}% (외삽, 신뢰도 낮음). "
                   "오퍼레이터 판단 우선")
        elif ood:
            confidence = "MEDIUM"
            msg = f"권장 OP={op:.1f}% (신뢰: 구간내, IForest 이상감지 — 주의)"
        else:
            confidence = "HIGH"
            msg = f"권장 OP={op:.1f}% (신뢰: 구간내)"

        return {"rec_OP": round(op, 1), "rec_steam": round(sf, 1),
                "confidence": confidence, "mode": mode, "ood": bool(ood),
                "in_env": bool(env), "feed": float(x[0][0]),
                "product": float(x[0][1]), "T_C": float(x[0][2]),
                "message": msg}


def main():
    parser = argparse.ArgumentParser()
    parser.add_argument("--data", default=BASE + "C-6111_data.pkl")
    parser.add_argument("--prefix", default="C-6111")
    parser.add_argument("--live", help='JSON live_tags for single predict test')
    args = parser.parse_args()
    df = pd.read_pickle(args.data)
    assist = OperatorAssist(df)

    if args.live:
        tags = json.loads(args.live)
        res = assist.predict(tags)
        print(f"\n=== Operator Advisory ({args.prefix}) ===")
        for k, v in res.items():
            print(f"  {k:15s}: {v}")
        return

    # 전체 shadow 리플레이: PROD 행 벡터화 처리
    prod = df[df["mode"] == "PROD"].sort_values("dtat").copy()
    prod = prod[(prod["feed"] > 50) & (prod["steam_flow"] > 10) & (prod["steam_op"] > 1)
                & prod[FEATURES + ["steam_op"]].notna().all(axis=1)]
    if len(prod) == 0:
        print("  PROD 없음 — advisory 불가")
        return

    X = prod[FEATURES].values
    sf = assist.model.predict(X)
    op = np.clip(np.polyval(assist.inv, sf), 0, 100)
    env_mask = ((X >= assist.env_lo.values) & (X <= assist.env_hi.values)).all(axis=1)
    ood_mask = assist.ood.decision_function(X) < 0
    errors = op - prod["steam_op"].values

    ood_rate = np.mean(ood_mask) * 100
    within_2 = np.mean(np.abs(errors) <= 2.0) * 100
    print(f"\n=== Shadow Advisory Report ({args.prefix}) ===")
    print(f"  PROD 행수   : {len(prod)}")
    print(f"  OOD 비율    : {ood_rate:.1f}%")
    print(f"  OP MAE      : {np.abs(errors).mean():.2f}%")
    print(f"  |Δ|≤2%      : {within_2:.1f}%  (검증기준: 90%+ in-envelope)")
    env_only = errors[~ood_mask[:len(errors)]]
    if len(env_only):
        print(f"  in-env MAE : {np.abs(env_only).mean():.2f}%  "
              f"|Δ|≤2%={np.mean(np.abs(env_only)<=2)*100:.1f}%")

    # 권장 OP vs 실제 OP 시계열 플롯
    import matplotlib
    matplotlib.use("Agg")
    import matplotlib.pyplot as plt
    fig, ax = plt.subplots(2, 1, figsize=(14, 8))
    s = prod.iloc[::10]
    ax[0].plot(s["dtat"], s["steam_op"], lw=.6, label="actual OP")
    ax[0].plot(s["dtat"], op[::10], lw=.6, c="r", label="advisory OP")
    ax[0].set_ylabel("OP %"); ax[0].legend(fontsize=8)
    ax[0].set_title(f"Operator Advisory vs Actual OP ({args.prefix})")
    ax[1].hist(errors, bins=60)
    ax[1].axvline(0, c="k", lw=.5)
    ax[1].set_title(f"Advisory error (rec-actual): median {np.median(errors):+.2f}%, "
                    f"within 2%={within_2:.1f}%")
    fig.tight_layout()
    path = BASE + f"{args.prefix}_advisory.png"
    fig.savefig(path, dpi=95)
    print(f"\n  플롯 저장: {path}")


if __name__ == "__main__":
    main()