diff --git a/src/core/modeling/modeling_batch.py b/src/core/modeling/modeling_batch.py
index 10e4cc9..9d26ded 100644
--- a/src/core/modeling/modeling_batch.py
+++ b/src/core/modeling/modeling_batch.py
@@ -22,6 +22,8 @@ from sklearn.tree import DecisionTreeRegressor
 from sklearn.neural_network import MLPRegressor
 from sklearn.pipeline import Pipeline
 from sklearn.impute import SimpleImputer
+from sklearn.compose import TransformedTargetRegressor
+from sklearn.preprocessing import StandardScaler
 from joblib import parallel_backend
 # 第三方模型导入
 # try:
@@ -655,6 +657,17 @@ class WaterQualityModelingBatch:
         elif model_name == 'LightGBM':
             base_model.set_params(verbose=-1)
 
+        # ★ 致命修复 1：SVR 对 y 尺度极度敏感。若 y 未缩放（如浊度 100~1000），
+        #    epsilon=0.1 相对 y 量级几乎为零 → 所有样本都变成支持向量 → 过拟合/躺平。
+        #    TransformedTargetRegressor 在 fit 时自动 StandardScaler(y)，
+        #    predict 时自动 inverse_transform，对调用方完全透明。
+        _is_svr = (config['model'] == SVR)
+        if _is_svr:
+            base_model = TransformedTargetRegressor(
+                regressor=base_model,
+                transformer=StandardScaler()
+            )
+
         # ============ 关键：把预处理器塞进 Pipeline ============
         # DualStream_MNF / Physical_Only 需要波长列表，供 PhysicalFeatureExtractor 定位波段
         _wl_list = None
@@ -677,9 +690,16 @@ class WaterQualityModelingBatch:
 
         # 以「步骤名__参数名」的格式索引参数网格；
         # config['params'] 是模型层的（无 __），统一加 model__ 前缀。
-        prefixed_params = {
-            f"model__{k}": v for k, v in config['params'].items()
-        }
+        # ★ SVR 被 TransformedTargetRegressor 包裹后，实际模型在 model.regressor_ 下，
+        #    GridSearchCV 参数路径变为 model__regressor__{param}
+        if _is_svr:
+            prefixed_params = {
+                f"model__regressor__{k}": v for k, v in config['params'].items()
+            }
+        else:
+            prefixed_params = {
+                f"model__{k}": v for k, v in config['params'].items()
+            }
 
         # 全量网格搜索：SVR 超参组合仅 216 种（4×6×3×3），
         # 穷举远优于 RandomizedSearchCV(n_iter=10) 的随机抽样
@@ -921,8 +941,16 @@ class WaterQualityModelingBatch:
     @staticmethod
     def _find_svr_in_pipeline(pipeline):
         model_step = pipeline.named_steps.get('model')
-        if model_step is not None and model_step.__class__.__name__ == 'SVR':
+        if model_step is None:
+            return None
+        # 直接就是 SVR（旧模型 / 非 SVR 的其他回归器）
+        if model_step.__class__.__name__ == 'SVR':
             return model_step
+        # ★ TransformedTargetRegressor 包裹的 SVR（修复 1 引入）
+        if hasattr(model_step, 'regressor_'):
+            inner = model_step.regressor_
+            if inner.__class__.__name__ == 'SVR':
+                return inner
         return None
 
     @staticmethod
diff --git a/src/preprocessing/spectral_Preprocessing.py b/src/preprocessing/spectral_Preprocessing.py
index 7fc696f..1198fd9 100644
--- a/src/preprocessing/spectral_Preprocessing.py
+++ b/src/preprocessing/spectral_Preprocessing.py
@@ -381,8 +381,28 @@ class MNFTransformer(TransformerMixin, BaseEstimator, _ArrayAsFloat64):
         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
-        noise = X[1:] - X[:-1]
+        # 1) ★ 致命修复 3：噪声矩阵估计。
+        #   旧方案 noise = X[1:] - X[:-1]（相邻行差分）假设数据是空间连续的图像，
+        #   相邻像素差异 = 传感器噪声。但进入 Pipeline.fit() 的 X_train 已被
+        #   split_data() 随机抽取 + KFold(shuffle=True) 打乱，相邻行可能是空间上
+        #   相隔十万八千里的不同水质样本。此时差分结果是自然方差而非噪声，
+        #   白化操作会把真实水质信号当噪声除掉。
+        #   新方案：Savitzky-Golay 滤波提取光谱维平滑曲线，残差 = 原始 - 平滑 ≈ 纯传感器噪声。
+        from scipy.signal import savgol_filter
+        try:
+            # window_length 必须是奇数，取 7（或波段数的一半向下取奇数）
+            _wl = min(7, n_features - 1)
+            if _wl % 2 == 0:
+                _wl -= 1
+            if _wl >= 5:
+                X_smooth = savgol_filter(X, window_length=_wl, polyorder=2, axis=1)
+                noise = X - X_smooth
+            else:
+                # 波段太少，无法滤波，回退到去均值残差
+                noise = X - np.mean(X, axis=1, keepdims=True)
+        except Exception:
+            # 兜底：去均值残差
+            noise = X - np.mean(X, axis=1, keepdims=True)
         Cn = np.cov(noise, rowvar=False)
         _reg = np.trace(Cn) / n_features * 0.01
         Cn += np.eye(n_features) * max(_reg, 1e-6)
@@ -611,10 +631,18 @@ def get_preprocessing_transformer(method: str, wavelengths=None,
         return IdentityTransformer()
     if method == "DualStream_MNF":
         from sklearn.pipeline import FeatureUnion
-        return FeatureUnion([
+        from sklearn.pipeline import Pipeline as _Pipeline
+        # ★ 致命修复 2：MNF 主成分（量级 ~±10）与 Physical 指数（量级 ~0.001）
+        #   尺度差异可达 10^4 倍，SVR 的 RBF 距离会被大值特征主导。
+        #   FeatureUnion 之后必须接 StandardScaler 统一量纲。
+        mnf_union = FeatureUnion([
             ('physical', PhysicalFeatureExtractor(wavelengths=wavelengths)),
             ('mnf',      MNFTransformer(n_components=mnf_n_components)),
         ])
+        return _Pipeline([
+            ('features', mnf_union),
+            ('scaler',   StandardScaler()),
+        ])
     if method == "Physical_Only":
         from sklearn.pipeline import Pipeline as _Pipeline
         return _Pipeline([