Thibaut Forest
Claude Fable 5
Phase 34-35: retire cross-fleet CV, re-screen 72 configs, champion unchanged
7d93fb8 | """ | |
| Per-(device, origin) Wiener-process parameters for first-hitting-time RUL. | |
| Models temp-corrected daily voltage as Brownian motion with drift; EOL is the | |
| first passage of the validated 2.40 V threshold. For each (device, origin) we | |
| estimate, from data <= origin only: | |
| wnr_m - drift toward the threshold (V/day, positive = declining) | |
| wnr_s2 - diffusion variance (V^2/day) from drift-adjusted increments | |
| wnr_n - number of increments behind the estimate (for shrinkage) | |
| Estimation window: last FIT_DAYS days of in-band daily medians. Cached to | |
| dataset/train/fht_features.parquet, keyed like features.parquet. | |
| """ | |
| from pathlib import Path | |
| import os | |
| import sys | |
| import numpy as np | |
| import pandas as pd | |
| MIN_TEMP, MAX_TEMP = 10, 30 | |
| BETA = 0.005 # V/degC temperature correction (validated in features.py) | |
| FIT_DAYS = 180 | |
| FHT_FEATURES = ['wnr_m', 'wnr_s2', 'wnr_n'] | |
| def daily_corrected(timeseries: pd.DataFrame) -> pd.DataFrame: | |
| df = timeseries.reset_index() | |
| df['device_id'] = df['device_id'].astype(str) | |
| df = df[(df['temperature'] > MIN_TEMP) & (df['temperature'] < MAX_TEMP)] | |
| df['day'] = df['end_time'].dt.normalize() | |
| g = df.groupby(['device_id', 'day']) | |
| out = pd.DataFrame({'v': g['voltage'].median(), 't': g['temperature'].median()}) | |
| out['vc'] = out['v'] - BETA * (out['t'] - 20.0) | |
| return out.reset_index().sort_values('day') | |
| def wiener_at_origin(daily: pd.DataFrame, origin: pd.Timestamp) -> pd.DataFrame: | |
| h = daily[(daily['day'] <= origin) & | |
| (daily['day'] > origin - pd.Timedelta(days=FIT_DAYS))] | |
| def fit(g): | |
| if len(g) < 4: | |
| return pd.Series({'wnr_m': np.nan, 'wnr_s2': np.nan, | |
| 'wnr_n': float(len(g)) - 1}) | |
| tt = (g['day'] - g['day'].iloc[0]).dt.days.values.astype(float) | |
| vv = g['vc'].values | |
| span = tt[-1] - tt[0] | |
| if span < 7: | |
| return pd.Series({'wnr_m': np.nan, 'wnr_s2': np.nan, | |
| 'wnr_n': float(len(g)) - 1}) | |
| # drift: endpoints estimator (MLE for BM); positive = declining voltage | |
| m = (vv[0] - vv[-1]) / span | |
| dt = np.diff(tt) | |
| dv = -np.diff(vv) # positive = decline | |
| s2 = np.mean((dv - m * dt) ** 2 / dt) | |
| return pd.Series({'wnr_m': m, 'wnr_s2': s2, 'wnr_n': float(len(dt))}) | |
| return h.groupby('device_id').apply(fit, include_groups=False) | |
| def build(dataset_path: Path) -> pd.DataFrame: | |
| from batteryswap_public.utils import load_dataset | |
| locations, timeseries, eol_times, scenarios = load_dataset(dataset_path / 'train') | |
| table = pd.read_parquet(dataset_path / 'train' / 'features.parquet') | |
| origins = (table[['origin_idx', 'origin']].drop_duplicates() | |
| .sort_values('origin_idx')) | |
| daily = daily_corrected(timeseries) | |
| rows = [] | |
| for _, r in origins.iterrows(): | |
| f = wiener_at_origin(daily, r['origin']) | |
| f['origin_idx'] = r['origin_idx'] | |
| rows.append(f.reset_index().rename(columns={'index': 'device_id'})) | |
| print(f"fht origin {r['origin_idx']} {r['origin'].date()}: {len(f)}", | |
| file=sys.stderr) | |
| out = pd.concat(rows, ignore_index=True) | |
| out.to_parquet(dataset_path / 'train' / 'fht_features.parquet') | |
| return out | |
| def load(dataset_path: Path = Path('dataset'), force: bool = False) -> pd.DataFrame: | |
| p = dataset_path / 'train' / 'fht_features.parquet' | |
| if p.exists() and not force: | |
| return pd.read_parquet(p) | |
| return build(dataset_path) | |
| if __name__ == '__main__': | |
| dataset_path = Path(os.environ.get('BATTERYSWAP_DATASET_PATH', 'dataset')) | |
| t = load(dataset_path, force='--force' in sys.argv) | |
| print(t[FHT_FEATURES].describe().round(5).to_string()) | |