update pay ui

ml/detector.py

@@ -243,9 +243,12 @@ class MLScorer:
     ):
         self.checkpoint_dir = Path(checkpoint_dir)
 
-        # 设备
+        # 设备检测
         if device == 'auto':
             self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
+        elif device == 'cuda' and not torch.cuda.is_available():
+            print("警告: CUDA 不可用，使用 CPU")
+            self.device = torch.device('cpu')
         else:
             self.device = torch.device(device)
 
@@ -276,8 +279,8 @@ class MLScorer:
                 with open(config_path, 'r') as f:
                     self.config = json.load(f)
 
-            # 加载模型
-            checkpoint = torch.load(model_path, map_location=self.device)
+            # 先用 CPU 加载模型（避免 CUDA 不可用问题），再移动到目标设备
+            checkpoint = torch.load(model_path, map_location='cpu')
 
             model_config = self.config.get('model', {}) if self.config else {}
             self.model = create_model(model_config)
@@ -294,6 +297,8 @@ class MLScorer:
 
         except Exception as e:
             print(f"警告: 模型加载失败 - {e}")
+            import traceback
+            traceback.print_exc()
             self.model = None
 
     def is_ready(self) -> bool:
@@ -551,7 +556,8 @@ if __name__ == "__main__":
         },
     ]
 
-    print("\n测试结果:")
+    print("\n" + "-" * 60)
+    print("测试1: 只用规则（无序列数据）")
     print("-" * 60)
 
     for case in test_cases:
@@ -567,5 +573,63 @@ if __name__ == "__main__":
             print(f"  异动类型: {result.anomaly_type}")
             print(f"  触发规则: {list(result.rule_details.keys())}")
 
+    # 测试2: 带序列数据的融合检测
+    print("\n" + "-" * 60)
+    print("测试2: 融合检测（规则 + ML）")
+    print("-" * 60)
+
+    # 生成模拟序列数据
+    seq_len = 30
+    n_features = 6
+
+    # 正常序列：小幅波动
+    normal_sequence = np.random.randn(seq_len, n_features) * 0.3
+    normal_sequence[:, 0] = np.linspace(0, 0.5, seq_len)  # alpha 缓慢上升
+    normal_sequence[:, 2] = np.abs(normal_sequence[:, 2]) + 1  # amt_ratio > 0
+
+    # 异常序列：最后几个时间步突然变化
+    anomaly_sequence = np.random.randn(seq_len, n_features) * 0.3
+    anomaly_sequence[-5:, 0] = np.linspace(1, 4, 5)  # alpha 突然飙升
+    anomaly_sequence[-5:, 1] = np.linspace(0.2, 1.5, 5)  # alpha_delta 加速
+    anomaly_sequence[-5:, 2] = np.linspace(2, 6, 5)  # amt_ratio 放量
+    anomaly_sequence[:, 2] = np.abs(anomaly_sequence[:, 2]) + 1
+
+    # 测试正常序列
+    normal_features = {
+        'alpha': float(normal_sequence[-1, 0]),
+        'alpha_delta': float(normal_sequence[-1, 1]),
+        'amt_ratio': float(normal_sequence[-1, 2]),
+        'amt_delta': float(normal_sequence[-1, 3]),
+        'rank_pct': 0.5,
+        'limit_up_ratio': 0.02
+    }
+
+    result = detector.detect(normal_features, normal_sequence)
+    print(f"\n正常序列:")
+    print(f"  异动: {'是' if result.is_anomaly else '否'}")
+    print(f"  最终得分: {result.final_score:.1f}")
+    print(f"  规则得分: {result.rule_score:.1f}")
+    print(f"  ML得分: {result.ml_score:.1f}")
+
+    # 测试异常序列
+    anomaly_features = {
+        'alpha': float(anomaly_sequence[-1, 0]),
+        'alpha_delta': float(anomaly_sequence[-1, 1]),
+        'amt_ratio': float(anomaly_sequence[-1, 2]),
+        'amt_delta': float(anomaly_sequence[-1, 3]),
+        'rank_pct': 0.95,
+        'limit_up_ratio': 0.15
+    }
+
+    result = detector.detect(anomaly_features, anomaly_sequence)
+    print(f"\n异常序列:")
+    print(f"  异动: {'是' if result.is_anomaly else '否'}")
+    print(f"  最终得分: {result.final_score:.1f}")
+    print(f"  规则得分: {result.rule_score:.1f}")
+    print(f"  ML得分: {result.ml_score:.1f}")
+    if result.is_anomaly:
+        print(f"  触发原因: {result.trigger_reason}")
+        print(f"  异动类型: {result.anomaly_type}")
+
     print("\n" + "=" * 60)
     print("测试完成!")