Upload folder using huggingface_hub

.gitignore

@@ -0,0 +1,35 @@
+# Python
+__pycache__/
+*.py[cod]
+*$py.class
+*.so
+.Python
+
+# Environment
+.env
+.venv
+env/
+venv/
+
+# IDE
+.vscode/
+.idea/
+*.swp
+*.swo
+
+# OS
+.DS_Store
+Thumbs.db
+
+# Gradio
+flagged/
+gradio_cached_examples/
+
+# Large files (if using Git LFS)
+*.pt
+*.pth
+*.ckpt
+
+# Logs
+*.log
+

README.md

@@ -4,7 +4,7 @@ emoji: 📟
 colorFrom: indigo
 colorTo: blue
 sdk: gradio
-sdk_version: "4.12.0"
+sdk_version: "4.44.0"
 app_file: gradio_app.py
 pinned: false
 ---

configs/case_builder_config.json

@@ -0,0 +1,11 @@
+{
+  "main_model_confirmation": {
+    "enabled": true,
+    "mode": "auto",
+    "model_dir": "checkpoints/phase2/exp_factor_balanced",
+    "device": "cpu",
+    "threshold": null,
+    "min_duration_days": 3
+  }
+}
+

configs/case_builder_config_README.md

@@ -0,0 +1,151 @@
+# CaseBuilder 配置说明
+
+## 配置文件位置
+
+`configs/case_builder_config.json`
+
+## 配置项说明
+
+```json
+{
+  "main_model_confirmation": {
+    "enabled": false,           // 是否启用主时序模型确认
+    "mode": "auto",             // 模式：auto（自动）或 manual（手工）
+    "model_dir": "checkpoints/phase2/exp_factor_balanced",  // 主时序模型路径
+    "device": "cpu",            // 设备：cpu 或 cuda
+    "threshold": null,          // 异常阈值（null 表示使用模型默认值）
+    "min_duration_days": 3      // 最小持续天数（用于 detect_pattern）
+  }
+}
+```
+
+## 使用场景
+
+### 场景1：自动模式（推荐）
+
+**配置：**
+```json
+{
+  "main_model_confirmation": {
+    "enabled": true,
+    "mode": "auto",
+    "model_dir": "checkpoints/phase2/exp_factor_balanced"
+  }
+}
+```
+
+**代码：**
+```python
+from utils.case_builder import CaseBuilder
+
+builder = CaseBuilder()  # 自动从配置文件加载
+result = builder.build_case(payload)
+
+# 检查主时序模型确认结果
+if result.get("risk_confirmed") is False:
+    print("主时序模型确认无风险，终止流程")
+    return
+
+# 继续生成 LLM 输入
+llm_input = result["llm_input"]
+```
+
+**行为：**
+- CaseBuilder 内部自动调用主时序模型确认风险
+- 返回 `risk_confirmed: true/false`
+- 外部可以根据结果决定是否继续
+
+### 场景2：手工模式
+
+**配置：**
+```json
+{
+  "main_model_confirmation": {
+    "enabled": false
+  }
+}
+```
+
+**代码：**
+```python
+from utils.case_builder import CaseBuilder
+from wearable_anomaly_detector import WearableAnomalyDetector
+
+builder = CaseBuilder()  # 只用 PatchAD，不调用主时序模型
+result = builder.build_case(payload)
+
+# 外部自己调用主时序模型确认
+detector = WearableAnomalyDetector(model_dir="...")
+all_data_points = []
+for item in payload["history_windows"]:
+    all_data_points.extend(item["window"])
+
+pattern_result = detector.detect_pattern(all_data_points, days=len(payload["history_windows"]))
+risk_confirmed = pattern_result.get('anomaly_pattern', {}).get('has_pattern', False)
+
+if not risk_confirmed:
+    print("主时序模型确认无风险，终止流程")
+    return
+```
+
+**行为：**
+- CaseBuilder 只用 PatchAD 分数，不调用主时序模型
+- 外部可以自己调用 `WearableAnomalyDetector.detect_pattern()` 来确认风险
+
+### 场景3：外部传入检测器（复用已加载的模型）
+
+**代码：**
+```python
+from utils.case_builder import CaseBuilder
+from wearable_anomaly_detector import WearableAnomalyDetector
+
+# 外部已加载的检测器
+detector = WearableAnomalyDetector(model_dir="...")
+
+# 传入检测器实例，避免重复加载
+builder = CaseBuilder(detector=detector)
+result = builder.build_case(payload)
+
+# 检查确认结果
+if result.get("risk_confirmed") is False:
+    print("主时序模型确认无风险，终止流程")
+    return
+```
+
+**行为：**
+- CaseBuilder 使用外部传入的检测器
+- 如果配置 `enabled=true`，会自动调用确认
+- 避免重复加载模型，提高性能
+
+## 返回值说明
+
+`build_case()` 返回的字典包含以下字段：
+
+```python
+{
+    "case": {...},                    # 标准 case JSON
+    "llm_input": "...",               # LLM 输入 Markdown
+    "validation": {...},              # 验证结果
+    "risk_confirmed": True/False/None, # 主时序模型确认结果（如果启用）
+    "main_model_result": {...}        # 主时序模型完整结果（如果启用）
+}
+```
+
+**`risk_confirmed` 说明：**
+- `True`: 主时序模型确认有风险
+- `False`: 主时序模型确认无风险
+- `None`: 未启用主时序模型确认，或确认失败
+
+## 向后兼容
+
+- 如果配置文件不存在，默认 `enabled=false`（只用 PatchAD）
+- 如果 `enabled=false`，行为与现有代码完全一致
+- 现有代码无需修改即可正常工作
+
+## 注意事项
+
+1. **性能**：自动模式需要加载模型，可能增加初始化时间（约 1-2 秒）
+2. **内存**：自动模式会占用更多内存（模型加载）
+3. **错误处理**：如果模型加载失败，会自动降级为只用 PatchAD，不会抛出异常
+4. **路径**：`model_dir` 可以是相对路径（相对于 `hf_release` 目录）或绝对路径
+

demo_patchad_cases/demo_anomaly_official.json

@@ -2,18 +2,18 @@
   "mode": "official",
   "sample": "demo_anomaly",
   "precheck": {
-    "event_id": "demo_user-cf1207f8",
-    "user_id": "demo_user",
+    "event_id": "demo_anomaly-4ba26c5e",
+    "user_id": "demo_anomaly",
     "status": "abnormal",
     "score": 0.3484,
     "threshold": 0.18,
-    "created_at": "2025-11-27T08:33:52.547782+00:00"
+    "created_at": "2025-11-28T02:50:34.582609+00:00"
   },
   "case_bundle": {
     "case": {
-      "case_id": "demo_user-cf1207f8",
-      "user_id": "demo_user",
-      "generated_at": "2025-11-27T08:33:52.548064+00:00",
+      "case_id": "demo_anomaly-4ba26c5e",
+      "user_id": "demo_anomaly",
+      "generated_at": "2025-11-28T02:50:34.603862+00:00",
       "anomaly_pattern": {
         "type": "continuous_anomaly",
         "duration_days": 3,
@@ -87,6 +87,41 @@
       "passed": true,
       "errors": [],
       "warnings": []
+    },
+    "risk_confirmed": false,
+    "main_model_result": {
+      "anomaly_pattern": {
+        "has_pattern": false,
+        "duration_days": 0,
+        "anomaly_dates": [],
+        "anomaly_scores": [],
+        "pattern_description": "异常仅持续0天，未达到最小持续天数3天"
+      },
+      "baseline_info": null,
+      "related_indicators": null,
+      "daily_results": [
+        {
+          "date": "2025-11-27T13:47:23.434200",
+          "anomaly_score": 0.5196528434753418,
+          "is_anomaly": false,
+          "hrv_rmssd": 68.38103114993835,
+          "hr": 73.06798211937824
+        },
+        {
+          "date": "2025-11-27T13:47:23.434200",
+          "anomaly_score": 0.5196528434753418,
+          "is_anomaly": false,
+          "hrv_rmssd": 68.38103114993835,
+          "hr": 73.06798211937824
+        },
+        {
+          "date": "2025-11-27T13:47:23.434200",
+          "anomaly_score": 0.5196528434753418,
+          "is_anomaly": false,
+          "hrv_rmssd": 68.38103114993835,
+          "hr": 73.06798211937824
+        }
+      ]
     }
   }
 }

demo_patchad_cases/demo_anomaly_platform.json

@@ -10,7 +10,7 @@
     "case": {
       "case_id": "platform-demo-evt",
       "user_id": "demo_anomaly",
-      "generated_at": "2025-11-27T08:33:52.547100+00:00",
+      "generated_at": "2025-11-28T02:50:34.467699+00:00",
       "anomaly_pattern": {
         "type": "continuous_anomaly",
         "duration_days": 3,
@@ -83,6 +83,41 @@
       "passed": true,
       "errors": [],
       "warnings": []
+    },
+    "risk_confirmed": false,
+    "main_model_result": {
+      "anomaly_pattern": {
+        "has_pattern": false,
+        "duration_days": 0,
+        "anomaly_dates": [],
+        "anomaly_scores": [],
+        "pattern_description": "异常仅持续0天，未达到最小持续天数3天"
+      },
+      "baseline_info": null,
+      "related_indicators": null,
+      "daily_results": [
+        {
+          "date": "2025-11-27T13:47:23.434200",
+          "anomaly_score": 0.5196528434753418,
+          "is_anomaly": false,
+          "hrv_rmssd": 68.38103114993835,
+          "hr": 73.06798211937824
+        },
+        {
+          "date": "2025-11-27T13:47:23.434200",
+          "anomaly_score": 0.5196528434753418,
+          "is_anomaly": false,
+          "hrv_rmssd": 68.38103114993835,
+          "hr": 73.06798211937824
+        },
+        {
+          "date": "2025-11-27T13:47:23.434200",
+          "anomaly_score": 0.5196528434753418,
+          "is_anomaly": false,
+          "hrv_rmssd": 68.38103114993835,
+          "hr": 73.06798211937824
+        }
+      ]
     }
   }
 }

demo_patchad_cases/demo_pattern_official.json

@@ -2,18 +2,18 @@
   "mode": "official",
   "sample": "demo_pattern",
   "precheck": {
-    "event_id": "demo_user-7f34fd17",
-    "user_id": "demo_user",
+    "event_id": "demo_pattern-e89db848",
+    "user_id": "demo_pattern",
     "status": "abnormal",
     "score": 0.1969,
     "threshold": 0.18,
-    "created_at": "2025-11-27T08:32:52.564723+00:00"
+    "created_at": "2025-11-28T02:50:31.275982+00:00"
   },
   "case_bundle": {
     "case": {
-      "case_id": "demo_user-7f34fd17",
-      "user_id": "demo_user",
-      "generated_at": "2025-11-27T08:32:52.565035+00:00",
+      "case_id": "demo_pattern-e89db848",
+      "user_id": "demo_pattern",
+      "generated_at": "2025-11-28T02:50:31.298789+00:00",
       "anomaly_pattern": {
         "type": "continuous_anomaly",
         "duration_days": 3,
@@ -87,6 +87,41 @@
       "passed": true,
       "errors": [],
       "warnings": []
+    },
+    "risk_confirmed": false,
+    "main_model_result": {
+      "anomaly_pattern": {
+        "has_pattern": false,
+        "duration_days": 0,
+        "anomaly_dates": [],
+        "anomaly_scores": [],
+        "pattern_description": "异常仅持续0天，未达到最小持续天数3天"
+      },
+      "baseline_info": null,
+      "related_indicators": null,
+      "daily_results": [
+        {
+          "date": "2025-11-20T08:00:00",
+          "anomaly_score": 0.5221930146217346,
+          "is_anomaly": false,
+          "hrv_rmssd": 76.75195878346763,
+          "hr": 69.55658937643992
+        },
+        {
+          "date": "2025-11-21T08:00:00",
+          "anomaly_score": 0.522602379322052,
+          "is_anomaly": false,
+          "hrv_rmssd": 74.44976175238928,
+          "hr": 69.4180874776572
+        },
+        {
+          "date": "2025-11-22T08:00:00",
+          "anomaly_score": 0.5226739645004272,
+          "is_anomaly": false,
+          "hrv_rmssd": 75.61461435432079,
+          "hr": 69.15957575082764
+        }
+      ]
     }
   }
 }

demo_patchad_cases/demo_pattern_platform.json

@@ -10,7 +10,7 @@
     "case": {
       "case_id": "platform-demo-evt",
       "user_id": "demo_pattern",
-      "generated_at": "2025-11-27T08:32:52.563962+00:00",
+      "generated_at": "2025-11-28T02:50:31.157741+00:00",
       "anomaly_pattern": {
         "type": "continuous_anomaly",
         "duration_days": 3,
@@ -83,6 +83,41 @@
       "passed": true,
       "errors": [],
       "warnings": []
+    },
+    "risk_confirmed": false,
+    "main_model_result": {
+      "anomaly_pattern": {
+        "has_pattern": false,
+        "duration_days": 0,
+        "anomaly_dates": [],
+        "anomaly_scores": [],
+        "pattern_description": "异常仅持续0天，未达到最小持续天数3天"
+      },
+      "baseline_info": null,
+      "related_indicators": null,
+      "daily_results": [
+        {
+          "date": "2025-11-20T08:00:00",
+          "anomaly_score": 0.5221930146217346,
+          "is_anomaly": false,
+          "hrv_rmssd": 76.75195878346763,
+          "hr": 69.55658937643992
+        },
+        {
+          "date": "2025-11-21T08:00:00",
+          "anomaly_score": 0.522602379322052,
+          "is_anomaly": false,
+          "hrv_rmssd": 74.44976175238928,
+          "hr": 69.4180874776572
+        },
+        {
+          "date": "2025-11-22T08:00:00",
+          "anomaly_score": 0.5226739645004272,
+          "is_anomaly": false,
+          "hrv_rmssd": 75.61461435432079,
+          "hr": 69.15957575082764
+        }
+      ]
     }
   }
 }

demo_patchad_cases/demo_pattern_test_official.json

@@ -0,0 +1,127 @@
+{
+  "mode": "official",
+  "sample": "demo_pattern_test",
+  "precheck": {
+    "event_id": "demo_pattern-f92fc28c",
+    "user_id": "demo_pattern",
+    "status": "abnormal",
+    "score": 0.1969,
+    "threshold": 0.18,
+    "created_at": "2025-11-28T02:45:33.134958+00:00"
+  },
+  "case_bundle": {
+    "case": {
+      "case_id": "demo_pattern-f92fc28c",
+      "user_id": "demo_pattern",
+      "generated_at": "2025-11-28T02:45:33.158854+00:00",
+      "anomaly_pattern": {
+        "type": "continuous_anomaly",
+        "duration_days": 3,
+        "trend": "improving",
+        "min_score": 0.1969,
+        "max_score": 0.2275,
+        "avg_score": 0.2153,
+        "threshold": 0.18
+      },
+      "baseline_info": {
+        "personal_mean": 75.0,
+        "personal_std": 0.0,
+        "group_mean": 67.5,
+        "record_count": 36,
+        "baseline_type": "personal",
+        "baseline_reliability": "high"
+      },
+      "related_indicators": {
+        "activity_level": {
+          "level": "低",
+          "avg_steps": 0.0,
+          "trend": "stable"
+        },
+        "sleep_quality": {
+          "available": false,
+          "quality": "数据不可用"
+        },
+        "stress_indicators": {
+          "level": "unknown"
+        }
+      },
+      "daily_results": [
+        {
+          "date": "2025-11-20",
+          "hrv_rmssd": 76.75,
+          "hr": 69.56,
+          "anomaly_score": 0.2275
+        },
+        {
+          "date": "2025-11-21",
+          "hrv_rmssd": 74.45,
+          "hr": 69.42,
+          "anomaly_score": 0.2215
+        },
+        {
+          "date": "2025-11-22",
+          "hrv_rmssd": 75.61,
+          "hr": 69.16,
+          "anomaly_score": 0.1969
+        }
+      ],
+      "user_profile": {
+        "age_group": "30-35岁",
+        "estimated_age": 33,
+        "sex": "男性",
+        "exercise": "每周5次以上",
+        "coffee": "偶尔",
+        "smoking": "不吸烟",
+        "drinking": "经常饮酒",
+        "MEQ": 62.0,
+        "MEQ_type": "晨型"
+      },
+      "metadata": {
+        "detector": "official_patchad",
+        "patchad_score": 0.1969,
+        "threshold": 0.18
+      }
+    },
+    "llm_input": "# 健康异常检测分析报告\n\n## 异常概览\n**异常类型**：continuous_anomaly  \n**持续时长**：3天  \n**异常趋势**：improving  \n\n## 异常评分分析\n- **异常分数范围**：0.1969 - 0.2275\n- **平均异常分数**：0.2153\n- **检测阈值**：0.1800\n\n## 当前生理状态评估\n| 指标 | 当前值 | 基线值 | 偏离基线 |\n|------|--------|--------|----------|\n| HRV RMSSD | 75.61 ms | 75.00 ms | +0.8% |\n\n## 历史趋势分析\n| 日期 | HRV (ms) | 心率 (bpm) | 异常分数 |\n|------|----------|------------|----------|\n| 2025-11-20 | 76.75 | 69.56 | 0.2275 |\n| 2025-11-21 | 74.45 | 69.42 | 0.2215 |\n| 2025-11-22 | 75.61 | 69.16 | 0.1969 |\n\n## 相关健康指标分析\n- 活动水平：低（平均步数=0.0，趋势=stable）\n- 睡眠质量：数据不可用\n- 压力指标：暂无显著异常\n\n## 用户背景信息\n- 年龄：33岁（30-35岁）\n- 性别：男性\n- 运动习惯：每周5次以上\n- 咖啡：偶尔 / 饮酒：经常饮酒\n- 生物节律：MEQ=62.0 (晨型)",
+    "validation": {
+      "passed": true,
+      "errors": [],
+      "warnings": []
+    },
+    "risk_confirmed": false,
+    "main_model_result": {
+      "anomaly_pattern": {
+        "has_pattern": false,
+        "duration_days": 0,
+        "anomaly_dates": [],
+        "anomaly_scores": [],
+        "pattern_description": "异常仅持续0天，未达到最小持续天数3天"
+      },
+      "baseline_info": null,
+      "related_indicators": null,
+      "daily_results": [
+        {
+          "date": "2025-11-20T08:00:00",
+          "anomaly_score": 0.5221930146217346,
+          "is_anomaly": false,
+          "hrv_rmssd": 76.75195878346763,
+          "hr": 69.55658937643992
+        },
+        {
+          "date": "2025-11-21T08:00:00",
+          "anomaly_score": 0.522602379322052,
+          "is_anomaly": false,
+          "hrv_rmssd": 74.44976175238928,
+          "hr": 69.4180874776572
+        },
+        {
+          "date": "2025-11-22T08:00:00",
+          "anomaly_score": 0.5226739645004272,
+          "is_anomaly": false,
+          "hrv_rmssd": 75.61461435432079,
+          "hr": 69.15957575082764
+        }
+      ]
+    }
+  }
+}

demo_patchad_cases/manifest.json

@@ -1,26 +1,16 @@
 [
   {
     "sample": "demo_pattern",
-    "mode": "platform",
-    "title": "demo_pattern · 模式A：平台自带",
-    "file": "demo_pattern_platform.json"
-  },
-  {
-    "sample": "demo_pattern",
-    "mode": "official",
-    "title": "demo_pattern · 模式B：官方预筛",
-    "file": "demo_pattern_official.json"
-  },
-  {
-    "sample": "demo_anomaly",
-    "mode": "platform",
-    "title": "demo_anomaly · 模式A：平台自带",
-    "file": "demo_anomaly_platform.json"
+    "title": "正常模式案例",
+    "file_platform": "demo_pattern_platform.json",
+    "file_official": "demo_pattern_official.json",
+    "description": "正常心率变异性模式，展示正常情况下的处理流程"
   },
   {
     "sample": "demo_anomaly",
-    "mode": "official",
-    "title": "demo_anomaly · 模式B：官方预筛",
-    "file": "demo_anomaly_official.json"
+    "title": "异常模式案例",
+    "file_platform": "demo_anomaly_platform.json",
+    "file_official": "demo_anomaly_official.json",
+    "description": "异常心率变异性模式，展示异常情况下的处理流程"
   }
-]
+]

requirements.txt

@@ -1,3 +1,4 @@
+gradio>=4.44.0
 torch>=2.1.0
 numpy>=1.24
 pandas>=2.0

simulate_patchad_case_pipeline.py

@@ -139,33 +139,38 @@ def run_mode_platform(records: List[Dict[str, Any]], sample_name: str, save_dir:
 
 def run_mode_official(records: List[Dict[str, Any]], sample_name: str, save_dir: Path | None) -> Dict[str, Any] | None:
     print("\n=== 模式B：官方 precheck + build_case ===")
-    windows = split_windows(records, days=3)
-    if len(windows) < 2:
-        raise RuntimeError("样例数据不足以生成历史窗口")
-    current_window = windows[-1]["window"]
+    print("步骤1：平台传入当前窗口到 precheck 接口")
+    current_window = records[-12:] if len(records) >= 12 else records
+    user_id = current_window[0].get("deviceId", "demo_user") if current_window else "demo_user"
 
     server = PrecheckServer(score_threshold=0.18)
-    pre_result = server.precheck("demo_user", current_window)
+    pre_result = server.precheck(user_id, current_window)
     print(f"预筛接口返回: {pre_result}")
     if pre_result["status"] != "abnormal":
         print("预筛未触发异常，示例终止。")
         return None
 
-    pending = server.pop_event(pre_result["event_id"])
+    print(f"\n步骤2：平台收到 event_id={pre_result['event_id']}，平台内部合成历史窗口数据")
+    # 模拟平台自己合成历史窗口（从历史数据平台获取）
+    windows = split_windows(records, days=3)
+    if len(windows) < 2:
+        raise RuntimeError("样例数据不足以生成历史窗口")
+    print(f"平台合成历史窗口: {len(windows)} 天")
+
+    print(f"\n步骤3：平台调用 build_case，传入 event_id 和合成的历史窗口（不传 window_data，系统从缓存获取）")
     payload = {
-        "event_id": pre_result["event_id"],
-        "user_id": pending["user_id"],
-        "window_data": pending["window_data"],
+        "event_id": pre_result["event_id"],  # 关键：只传 event_id，不传 window_data
+        "user_id": user_id,
         "user_profile": sample_profile(),
-        "history_windows": windows,
+        "history_windows": windows,  # 平台自己合成的历史窗口
         "metadata": {
             "detector": "official_patchad",
-            "patchad_score": pending["metadata"]["patchad_score"],
-            "threshold": pending["metadata"]["threshold"],
+            "patchad_score": pre_result["score"],
+            "threshold": pre_result["threshold"],
         },
     }
 
-    builder = CaseBuilder()
+    builder = CaseBuilder(precheck_server=server)  # 传入 PrecheckServer 实例
     result = builder.build_case(payload)
     print(f"验证结果: {result['validation']}")
     print(f"Case ID: {result['case']['case_id']}")

utils/case_builder.py

@@ -5,11 +5,21 @@ import statistics
 from dataclasses import dataclass
 from datetime import datetime, timezone
 from pathlib import Path
-from typing import Any, Dict, List, Tuple
+from typing import Any, Dict, List, Optional, Tuple
 from uuid import uuid4
 
 from utils.patchad_filter import PatchADFilter
 
+# 可选导入主时序模型（如果不存在则使用 None）
+try:
+    import sys
+    base_dir = Path(__file__).parent.parent
+    if str(base_dir) not in sys.path:
+        sys.path.insert(0, str(base_dir))
+    from wearable_anomaly_detector import WearableAnomalyDetector
+except ImportError:
+    WearableAnomalyDetector = None
+
 
 @dataclass
 class ValidationResult:
@@ -28,10 +38,80 @@ class ValidationResult:
 class CaseBuilder:
     """
     根据规范化 payload 生成 case JSON + Markdown 输入描述。
+    
+    支持配置主时序模型确认（自动/手工模式）：
+    - enabled=true, mode=auto: 自动调用主时序模型确认风险
+    - enabled=false: 只用 PatchAD 分数，不调用主时序模型
+    
+    支持两种模式：
+    - 模式A：平台有PatchAD，直接传入所有数据（window_data + history_windows）
+    - 模式B：平台没有PatchAD，通过 event_id 从 PrecheckServer 缓存获取 window_data
     """
 
-    def __init__(self) -> None:
+    def __init__(
+        self,
+        config_path: Optional[Path] = None,
+        detector: Optional[Any] = None,
+        precheck_server: Optional[Any] = None
+    ) -> None:
+        """
+        初始化 CaseBuilder
+        
+        参数:
+            config_path: 配置文件路径，如果为None则使用默认配置
+            detector: 可选的主时序模型检测器实例（如果外部已加载）
+            precheck_server: 可选的 PrecheckServer 实例，用于模式B从缓存获取 window_data
+        """
         self.patchad = PatchADFilter()
+        self.detector = detector  # 外部传入的检测器（可选）
+        self.precheck_server = precheck_server  # PrecheckServer 实例（用于模式B）
+        self.config = self._load_config(config_path)
+        
+        # 如果配置为自动且未传入 detector，则自动加载
+        if self.config.get("main_model_confirmation", {}).get("enabled", False):
+            if self.detector is None and WearableAnomalyDetector is not None:
+                model_dir = self.config["main_model_confirmation"].get("model_dir")
+                if model_dir:
+                    # 转换为绝对路径（相对于 hf_release 目录）
+                    base_dir = Path(__file__).parent.parent
+                    model_path = base_dir / model_dir if not Path(model_dir).is_absolute() else Path(model_dir)
+                    device = self.config["main_model_confirmation"].get("device", "cpu")
+                    threshold = self.config["main_model_confirmation"].get("threshold")
+                    
+                    try:
+                        self.detector = WearableAnomalyDetector(
+                            model_dir=model_path,
+                            device=device,
+                            threshold=threshold
+                        )
+                        print(f"✅ 已自动加载主时序模型: {model_path}")
+                    except Exception as e:
+                        print(f"⚠️  自动加载主时序模型失败: {e}，将只使用 PatchAD 分数")
+                        self.detector = None
+    
+    def _load_config(self, config_path: Optional[Path]) -> Dict[str, Any]:
+        """加载配置文件"""
+        if config_path is None:
+            config_path = Path(__file__).parent.parent / "configs" / "case_builder_config.json"
+        
+        if config_path.exists():
+            try:
+                with open(config_path, 'r', encoding='utf-8') as f:
+                    return json.load(f)
+            except Exception as e:
+                print(f"⚠️  加载 CaseBuilder 配置失败: {e}，使用默认配置")
+        
+        # 返回默认配置（向后兼容）
+        return {
+            "main_model_confirmation": {
+                "enabled": False,
+                "mode": "auto",
+                "model_dir": "checkpoints/phase2/exp_factor_balanced",
+                "device": "cpu",
+                "threshold": None,
+                "min_duration_days": 3
+            }
+        }
 
     def validate_payload(self, payload: Dict[str, Any]) -> ValidationResult:
         errors: List[str] = []
@@ -40,9 +120,31 @@ class CaseBuilder:
         if not payload.get("user_id"):
             errors.append("缺少 user_id")
 
-        window = payload.get("window_data", [])
+        # 模式B：通过 event_id 从缓存获取 window_data
+        event_id = payload.get("event_id")
+        window = payload.get("window_data")
+        
+        if event_id and self.precheck_server and not window:
+            # 模式B：从 PrecheckServer 缓存获取 window_data（不删除，因为可能还需要）
+            if self.precheck_server.has_pending(event_id):
+                # 使用 pop_event 获取并删除（因为已经使用过了）
+                pending = self.precheck_server.pop_event(event_id)
+                window = pending.get("window_data")
+                payload["window_data"] = window
+                if pending.get("user_id") and not payload.get("user_id"):
+                    payload["user_id"] = pending.get("user_id")
+                # 补充 metadata 中的 patchad 信息
+                if not payload.get("metadata", {}).get("patchad_score"):
+                    metadata = payload.get("metadata", {})
+                    if "patchad_score" not in metadata:
+                        metadata["patchad_score"] = pending.get("metadata", {}).get("patchad_score", 0.0)
+                    if "threshold" not in metadata:
+                        metadata["threshold"] = pending.get("metadata", {}).get("threshold", 0.0)
+                    payload["metadata"] = metadata
+        
+        # 模式A：直接传入 window_data
         if not window or len(window) < 12:
-            errors.append("window_data 不足 12 条")
+            errors.append("window_data 不足 12 条（模式A需直接传入，模式B需通过 event_id 从缓存获取）")
 
         profile = payload.get("user_profile")
         if not profile:
@@ -50,7 +152,7 @@ class CaseBuilder:
 
         history = payload.get("history_windows", [])
         if not history:
-            errors.append("缺少 history_windows")
+            errors.append("缺少 history_windows（平台需自己合成历史窗口数据）")
 
         metadata = payload.get("metadata", {})
         if not metadata.get("detector"):
@@ -242,10 +344,40 @@ class CaseBuilder:
 
         llm_input = self._format_llm_input(case)
 
+        # 【新增】主时序模型确认（如果配置启用）
+        risk_confirmed = None
+        main_model_result = None
+        
+        if self.config.get("main_model_confirmation", {}).get("enabled", False):
+            if self.detector:
+                try:
+                    # 使用主时序模型对历史窗口进行确认
+                    # 将数据组织成按天的格式：[[day1_data], [day2_data], ...]
+                    daily_data_points = []
+                    for item in normalized_history:
+                        daily_data_points.append(item["window"])
+                    
+                    if daily_data_points:
+                        pattern_result = self.detector.detect_pattern(
+                            daily_data_points,
+                            days=len(normalized_history),
+                            min_duration_days=self.config["main_model_confirmation"].get("min_duration_days", 3)
+                        )
+                        
+                        # 判断是否确认有风险
+                        risk_confirmed = pattern_result.get('anomaly_pattern', {}).get('has_pattern', False)
+                        main_model_result = pattern_result
+                except Exception as e:
+                    print(f"⚠️  主时序模型确认失败: {e}，将只使用 PatchAD 分数")
+                    risk_confirmed = None
+                    main_model_result = None
+
         return {
             "case": case,
             "llm_input": llm_input,
             "validation": validation.to_dict(),
+            "risk_confirmed": risk_confirmed,  # 【新增】主时序模型确认结果
+            "main_model_result": main_model_result,  # 【新增】主时序模型完整结果
         }
 
 

流程说明_修正版.md

@@ -0,0 +1,102 @@
+# CaseBuilder 流程说明（修正版）
+
+## 两种流程模式
+
+### 模式A：平台有 PatchAD
+
+**流程**：
+1. 平台自己有 PatchAD，自己检测到异常
+2. 平台自己合成数据：
+   - `window_data`：当前窗口（12条）
+   - `history_windows`：历史窗口（3-7天，每天12条）
+   - `user_profile`：用户档案
+3. 平台直接调用 `build_case`，传入所有数据
+
+**代码示例**：
+```python
+payload = {
+    "event_id": "platform-evt-123",
+    "user_id": "user_001",
+    "window_data": current_window,  # 平台自己传入
+    "user_profile": user_profile,    # 平台自己传入
+    "history_windows": history_windows,  # 平台自己合成的
+    "metadata": {"detector": "platform_patchad", "score": 0.82},
+}
+builder = CaseBuilder()
+result = builder.build_case(payload)
+```
+
+---
+
+### 模式B：平台没有 PatchAD
+
+**流程**：
+1. **步骤1**：平台传入当前窗口（12条）到系统的 `/precheck` 接口
+   - 系统返回 `event_id` 和 `status`
+   - 系统内部缓存 `window_data`
+
+2. **步骤2**：平台收到 `event_id` 后，**在平台内部合成历史窗口数据**
+   - 平台从历史数据平台获取最近几天的数据
+   - 平台自己按日期分割，生成 `history_windows`
+   - 平台自己获取 `user_profile`
+
+3. **步骤3**：平台调用 `build_case`，传入：
+   - `event_id`：系统根据此从缓存获取 `window_data`
+   - `history_windows`：平台自己合成的
+   - `user_profile`：平台自己获取的
+   - **不传 `window_data`**（系统从缓存获取）
+
+**代码示例**：
+```python
+# 步骤1：precheck
+from utils.patchad_filter import PrecheckServer
+server = PrecheckServer()
+pre_result = server.precheck(user_id, current_window)
+# 返回: {"event_id": "evt-123", "status": "abnormal", ...}
+
+# 步骤2：平台自己合成历史窗口（平台内部操作）
+history_windows = platform_synthesize_history_windows(user_id, days=3)
+user_profile = platform_get_user_profile(user_id)
+
+# 步骤3：build_case（只传 event_id，不传 window_data）
+from utils.case_builder import CaseBuilder
+builder = CaseBuilder(precheck_server=server)  # 传入 PrecheckServer 实例
+payload = {
+    "event_id": pre_result["event_id"],  # 关键：只传 event_id
+    "user_id": user_id,
+    "user_profile": user_profile,  # 平台自己获取
+    "history_windows": history_windows,  # 平台自己合成的
+    "metadata": {"detector": "official_patchad"},
+}
+result = builder.build_case(payload)  # 系统从缓存获取 window_data
+```
+
+---
+
+## 关键区别
+
+| 项目 | 模式A | 模式B |
+|------|-------|-------|
+| **预筛位置** | 平台内部 | 系统 `/precheck` 接口 |
+| **window_data 来源** | 平台直接传入 | 系统从缓存获取（通过 event_id） |
+| **history_windows 来源** | 平台自己合成 | 平台自己合成 |
+| **user_profile 来源** | 平台自己获取 | 平台自己获取 |
+| **交互次数** | 1次（build_case） | 2次（precheck + build_case） |
+
+---
+
+## 重要说明
+
+1. **模式B中，平台需要自己合成 `history_windows`**
+   - 这不是系统的工作
+   - 平台从历史数据平台获取数据，按日期分割
+   - 系统只负责根据 `event_id` 从缓存获取 `window_data`
+
+2. **模式B中，`build_case` 不需要传入 `window_data`**
+   - 系统会根据 `event_id` 从 `PrecheckServer` 的缓存中获取
+   - 如果传入了 `window_data`，则优先使用传入的（兼容模式A）
+
+3. **CaseBuilder 需要传入 `precheck_server` 参数（模式B）**
+   - 这样系统才能从缓存中获取 `window_data`
+   - 模式A不需要传入
+