eseguiti fix e introdotta versione wavelets

shared_utils.py

@@ -9,6 +9,10 @@ from typing import Dict, List, Optional, Sequence, Tuple
 import numpy as np
 import pandas as pd
 import pyodbc
+try:
+    import pywt
+except ImportError:  # pragma: no cover - optional dependency
+    pywt = None
 
 DEFAULT_CONFIG_PATH = Path("config/pattern_knn_config.json")
 
@@ -87,6 +91,58 @@ def z_norm(arr: np.ndarray) -> Optional[np.ndarray]:
     return (arr - mu) / (sd + 1e-12)
 
 
+def wavelet_denoise(
+    series: pd.Series,
+    wavelet: str = "db3",
+    level: int = 3,
+    mode: str = "symmetric",
+    threshold_mode: str = "soft",
+) -> Optional[pd.Series]:
+    """Denoise/reshape the series with a wavelet decomposition.
+
+    Keeps the original index length; if PyWavelets is missing the function
+    returns None so callers can gracefully fall back to the raw signal.
+    """
+    if pywt is None:
+        print("[WARN] pywt non installato: salto il filtraggio wavelet.")
+        return None
+    s = pd.to_numeric(series, errors="coerce")
+    if s.dropna().empty:
+        return None
+
+    w = pywt.Wavelet(wavelet)
+    max_level = pywt.dwt_max_level(len(s.dropna()), w.dec_len)
+    lvl = max(1, min(level, max_level)) if max_level > 0 else 1
+
+    valid = s.dropna()
+    coeffs = pywt.wavedec(valid.values, w, mode=mode, level=lvl)
+    # Universal threshold (Donoho-Johnstone)
+    sigma = np.median(np.abs(coeffs[-1])) / 0.6745 if len(coeffs[-1]) > 0 else 0.0
+    thresh = sigma * np.sqrt(2 * np.log(len(valid))) if sigma > 0 else 0.0
+    if thresh <= 0:
+        coeffs_f = coeffs
+    else:
+        def _safe_thresh(c: np.ndarray) -> np.ndarray:
+            if c is None or c.size == 0:
+                return c
+            if threshold_mode == "hard":
+                return pywt.threshold(c, value=thresh, mode="hard")
+            # soft threshold without divide-by-zero warnings
+            mag = np.abs(c)
+            mask = mag > thresh
+            out = np.zeros_like(c)
+            out[mask] = np.sign(c[mask]) * (mag[mask] - thresh)
+            return out
+
+        coeffs_f = [coeffs[0]] + [_safe_thresh(c) for c in coeffs[1:]]
+
+    rec = pywt.waverec(coeffs_f, w, mode=mode)
+    rec = rec[: len(valid)]
+    filt = pd.Series(rec, index=valid.index)
+    # Re-allineamento all'indice originale
+    return filt.reindex(s.index).interpolate(limit_direction="both")
+
+
 def build_pattern_library(
     ret_series: pd.Series,
     wp: int,