From 3edb73b6521bd3dced67b9fc379147258b96cc9f Mon Sep 17 00:00:00 2001
From: duxin <duxin@example.com>
Date: Wed, 29 Jul 2026 15:41:25 +0800
Subject: [PATCH] =?UTF-8?q?fix:=20DualStream=20Pipeline=20=E4=B8=89?=
 =?UTF-8?q?=E9=A1=B9=E4=BF=AE=E5=A4=8D=20=E2=80=94=20StandardScaler=20+=20?=
 =?UTF-8?q?=E9=83=A8=E7=BD=B2=E5=AF=BC=E5=87=BA=20+=20=E6=8E=A8=E7=90=86?=
 =?UTF-8?q?=E6=8B=86=E5=BC=B9?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

任务1: Pipeline 插入 StandardScaler
  imputer → FeatureUnion → cleaner → scaler(StandardScaler) → SVR
  MNF(≈1e4量级)与物理指数(≈1e0量级)统一尺度后入SVR

任务2: .npz 导出 scaler_mean / scaler_scale
  C++端可复现 (X - mean) / scale

任务3: 推理端 train_wavelengths 缺失时直接抛异常
  不再静默裸退308列导致 SimpleImputer 维度不匹配崩溃
---
 src/core/modeling/modeling_batch.py    | 8 +++++++-
 src/core/prediction/inference_batch.py | 7 +++++--
 2 files changed, 12 insertions(+), 3 deletions(-)

diff --git a/src/core/modeling/modeling_batch.py b/src/core/modeling/modeling_batch.py
index 0b9b5d0..8f7f79a 100644
--- a/src/core/modeling/modeling_batch.py
+++ b/src/core/modeling/modeling_batch.py
@@ -646,7 +646,8 @@ class WaterQualityModelingBatch:
             ('imputer', SimpleImputer(strategy='median')),
             ('preproc', preproc),
             ('cleaner', _SafeFiniteTransformer()),
-            ('model', base_model),
+            ('scaler',  StandardScaler()),
+            ('model',   base_model),
         ])
 
         # ★ 入口清洗：inf → NaN，后续 SimpleImputer/cleaner 接力处理
@@ -785,6 +786,11 @@ class WaterQualityModelingBatch:
                 'mnf_W':    np.asarray(_mnf.W_mnf_),
                 'mnf_n_components': int(getattr(_mnf, 'n_components_', _mnf.n_components)),
             }
+            # StandardScaler 参数（C++ 端需复现 (X - mean) / scale）
+            _scaler = model.named_steps.get('scaler')
+            if _scaler is not None and hasattr(_scaler, 'mean_'):
+                _deploy['scaler_mean']  = np.asarray(_scaler.mean_, dtype=np.float64)
+                _deploy['scaler_scale'] = np.asarray(_scaler.scale_, dtype=np.float64)
             # SVR 部署矩阵（rbf kernel 需要 support_vectors_）
             _deploy['svr_dual_coef'] = np.asarray(_svr.dual_coef_)
             _deploy['svr_intercept'] = np.asarray(_svr.intercept_)
diff --git a/src/core/prediction/inference_batch.py b/src/core/prediction/inference_batch.py
index e778178..9489a2d 100644
--- a/src/core/prediction/inference_batch.py
+++ b/src/core/prediction/inference_batch.py
@@ -446,8 +446,11 @@ class WaterQualityInference:
         """
         train_wl = metadata.get('train_wavelengths', None)
         if train_wl is None or len(train_wl) == 0:
-            print("[DualStream_MNF] ⚠ 模型无 train_wavelengths，fallback 原样输入")
-            return spectra.values
+            raise ValueError(
+                "推理失败：metadata 中缺失 train_wavelengths。"
+                "DualStream_MNF 必须对齐训练波段以完成光谱重采样。"
+                "请使用包含 train_wavelengths 元数据的模型文件。"
+            )
 
         # 提取纯光谱列
         spec_cols = []