fix: MNFTransformer 数值稳定性 — 防 inf/NaN 导致 SVR 拒绝输入

- fit: 正则化按协方差总迹的1%自适应 + 截断噪声主导特征值
- fit: 输入先做 nan_to_num 清洗
- transform: 输出 clip 到 float32 安全范围，防止极端值
- 噪声协方差正则化从固定 1e-9 改为 trace*0.01，对 113 维特征更稳健

src/preprocessing/spectral_Preprocessing.py

@@ -378,37 +378,48 @@ class MNFTransformer(TransformerMixin, BaseEstimator, _ArrayAsFloat64):
 
     def fit(self, X, y=None):
         X = self._to_ndarray(X)
+        X = np.nan_to_num(X, nan=0.0, posinf=0.0, neginf=0.0)
         n_samples, n_features = X.shape
 
-        # 1) 噪声矩阵：相邻行差分（假设输入按空间/时间连续排列）
+        # 1) 噪声矩阵：相邻行差分
         noise = X[1:] - X[:-1]                          # (n-1, p)
         Cn = np.cov(noise, rowvar=False)                # (p, p)
-        Cn += np.eye(n_features) * 1e-9                 # 防止奇异
+        # 正则化：按总方差的 1% 加对角项，强于固定 1e-9
+        _reg = np.trace(Cn) / n_features * 0.01
+        Cn += np.eye(n_features) * max(_reg, 1e-6)
 
-        # 2) 噪声白化：Cn = U diag(S) Vt — 对称矩阵用 eigh
+        # 2) 噪声白化：仅对显著特征值做白化，忽略噪声主导的成分
         eigvals_n, eigvecs_n = np.linalg.eigh(Cn)
-        # eigh 返回升序排列；白化矩阵 = V @ diag(1/sqrt(s)) @ V.T
-        inv_sqrt = 1.0 / np.sqrt(np.maximum(eigvals_n, 1e-12))
-        Wn = eigvecs_n @ np.diag(inv_sqrt) @ eigvecs_n.T   # (p, p)
+        # 截断：特征值 < max * 1e-4 的视为噪声，不参与白化（置单位矩阵）
+        _keep_mask = eigvals_n > eigvals_n.max() * 1e-4
+        if _keep_mask.sum() < 2:
+            _keep_mask[:2] = True  # 至少保留 2 个成分
+        _kept_vals = np.where(_keep_mask, eigvals_n, 1.0)
+        inv_sqrt = np.where(_keep_mask, 1.0 / np.sqrt(np.maximum(_kept_vals, 1e-12)), 1.0)
+        Wn = eigvecs_n @ np.diag(inv_sqrt) @ eigvecs_n.T
 
         # 3) 均值
-        self.mean_ = X.mean(axis=0, keepdims=True)         # (1, p)
+        self.mean_ = X.mean(axis=0, keepdims=True)
 
-        # 4) 白化后的数据协方差 → PCA
-        X_w = (X - self.mean_) @ Wn                        # (n, p)
-        Cw = np.cov(X_w, rowvar=False)                     # (p, p)
+        # 4) 白化后 PCA
+        X_w = (X - self.mean_) @ Wn
+        Cw = np.cov(X_w, rowvar=False)
         eigvals_w, eigvecs_w = np.linalg.eigh(Cw)
-        # 降序排列（eigh 升序，翻转）
-        Wp = eigvecs_w[:, ::-1]                             # (p, p)
+        Wp = eigvecs_w[:, ::-1]
 
         # 5) 最终变换矩阵
-        self.W_mnf_ = Wn @ Wp                               # (p, p)
+        self.W_mnf_ = Wn @ Wp
 
         return self
 
     def transform(self, X):
         X = self._to_ndarray(X)
-        return (X - self.mean_) @ self.W_mnf_[:, :self.n_components]
+        X = np.nan_to_num(X, nan=0.0, posinf=0.0, neginf=0.0)
+        result = (X - self.mean_) @ self.W_mnf_[:, :self.n_components]
+        # 防 SVR 拒绝 inf：clip 到 float32 安全范围
+        result = np.nan_to_num(result, nan=0.0, posinf=0.0, neginf=0.0)
+        result = np.clip(result, -1e15, 1e15)
+        return result
 
 
 # ============================================================================