refactor: DualStream 纯光谱输入 — 消除 WQI 冗余双向断层

训练端:
- DualStream_MNF 时只取纯光谱列(50列)传给 Pipeline
- PhysicalFeatureExtractor + MNFTransformer 都只收纯光谱
- 不再从 CSV 预读 WQI 列混入输入

推理端:
- 只做 308→50 光谱重采样，不补 WQI
- pipeline.predict() 自动完成 Physical 指数计算 + MNF 降维
- 删除 30+ 行 WQI 补齐代码

训练/推理完全对称: 纯光谱入 → FeatureUnion → SVR 出

src/core/modeling/modeling_batch.py

@@ -606,7 +606,16 @@ class WaterQualityModelingBatch:
 
         print(f"开始训练模型: {model_name} (预处理: {preprocess_method})")
 
-        # 使用指定方法分割训练集和测试集（用原始 X_raw，Pipeline 内置 transform 处理）
+        # ═══════════════════════════════════════════════════════════
+        # ★ DualStream_MNF：只取纯光谱列，WQI 由 Pipeline 内 PhysicalExtractor 动态计算
+        # ═══════════════════════════════════════════════════════════
+        if preprocess_method == "DualStream_MNF":
+            _spec_cols = [c for c in X_raw.columns if self._is_wavelength_column(c)]
+            X_raw = X_raw[_spec_cols]
+            print(f"[DualStream_MNF] 精简为纯光谱: {X_raw.shape[1]} 列 "
+                  f"({X_raw.columns[0]} ~ {X_raw.columns[-1]} nm)")
+
+        # 使用指定方法分割训练集和测试集
         X_train, X_test, y_train, y_test = self.split_data(
             X_raw, y, method=split_method, test_size=test_size, random_state=random_state
         )

src/core/prediction/inference_batch.py

@@ -477,15 +477,14 @@ class WaterQualityInference:
         metadata = self.loaded_model_data.get('metadata', {})
 
         # ═══════════════════════════════════════════════════════════
-        # ★ DualStream_MNF：重建训练特征空间（308→113列），
+        # ★ DualStream_MNF：纯光谱 → 重采样到训练波长 → pipeline.predict()
-        #   让 Pipeline.predict() 完整走 imputer→preproc→cleaner→SVR
+        #   Pipeline 内置的 FeatureUnion[PhysicalExtractor + MNF]
+        #   自动计算物理指数和降维，无需外部补 WQI
         # ═══════════════════════════════════════════════════════════
         if actual_preprocess_method == "DualStream_MNF" and isinstance(model, Pipeline):
-            print("[DualStream_MNF] 推理：重建训练特征集，交给 Pipeline 完整处理")
+            print("[DualStream_MNF] 推理：纯光谱重采样 → Pipeline 全自动处理")
             train_wl = metadata.get('train_wavelengths', None)
-            train_cols = metadata.get('train_columns', None)
+            if train_wl is not None and len(train_wl) > 0:
-            if train_wl is not None and len(train_wl) > 0 and train_cols is not None:
-                # 1) 重采样 308 → 训练波长
                 spec_cols = []
                 for c in spectra.columns:
                     try:
@@ -500,26 +499,11 @@ class WaterQualityInference:
                     resampled[i] = np.interp(dst_wl, src_wl, spec_data[i],
                                              left=np.nan, right=np.nan)
                 resampled = np.nan_to_num(resampled, nan=0.0)
-                wl_col_names = [f'{wl:.6f}' for wl in train_wl]
+                spectra = pd.DataFrame(resampled,
-                spectra = pd.DataFrame(resampled, columns=wl_col_names)
+                    columns=[f'{wl:.6f}' for wl in train_wl])
-                # 2) 计算 WQI + 补零到训练列数
+            print(f"[DualStream_MNF] 纯光谱重采样完成: {spectra.shape[1]} 列 "
-                n_need = len(train_cols)
+                  f"→ pipeline.predict() 将自动完成 Physical+MNF 变换")
-                if spectra.shape[1] < n_need:
-                    print(f"[DualStream_MNF] 计算 WQI: {spectra.shape[1]} → {n_need}")
-                    try:
-                        from src.utils.water_index import WaterQualityIndexCalculator
-                        wqi = WaterQualityIndexCalculator().calculate_many(
-                            WaterQualityIndexCalculator().list_available(),
-                            spectra, fast=True)
-                        spectra = pd.concat([spectra, wqi], axis=1)
-                    except Exception as e:
-                        print(f"[DualStream_MNF] WQI 失败: {e}")
-                    while spectra.shape[1] < n_need:
-                        spectra[f'_pad_{spectra.shape[1]}'] = 0.0
-                    spectra = spectra.iloc[:, :n_need]
-            print(f"[DualStream_MNF] 特征重建完成: {spectra.shape}")
             print(f"[特征对齐] 最终输入维度: {spectra.shape}")
-            # 返回 DataFrame，后续 model.predict() 走完整 Pipeline
             return spectra.values
 
         train_wavelengths = metadata.get('train_wavelengths', None)