1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
| """
论文：EEG-Driven Decoding Framework for Passenger Hazard Perception
arXiv: 2609.07128
代码：https://github.com/SOTIF-AVLab/EEG2023
核心方法：3D-CRNN 用于EEG信号的风险预测和危险识别
"""
import torch
import torch.nn as nn
import torch.nn.functional as F
class CRNN3D(nn.Module):
"""
3D Convolutional Recurrent Neural Network
论文Section 3描述的3D-CRNN模型:
- 3D卷积层: 提取EEG空间-时序特征
- 循环层: 建模时序依赖
- 分类头: 风险预测(RP) + 危险识别(DI)
输入: EEG信号, shape=(batch, channels, time, freq) 或 (batch, 1, channels, time)
输出: RP logits + DI logits
"""
def __init__(self,
num_channels: int = 32,
time_steps: int = 1200,
num_classes_rp: int = 2,
num_classes_di: int = 2):
super().__init__()
self.conv3d_1 = nn.Conv3d(
in_channels=1,
out_channels=16,
kernel_size=(3, 3, 3),
stride=(1, 1, 1),
padding=(1, 1, 1)
)
self.bn1 = nn.BatchNorm3d(16)
self.pool1 = nn.MaxPool3d(kernel_size=(2, 1, 1))
self.conv3d_2 = nn.Conv3d(
in_channels=16,
out_channels=32,
kernel_size=(3, 3, 3),
stride=(1, 1, 1),
padding=(1, 1, 1)
)
self.bn2 = nn.BatchNorm3d(32)
self.pool2 = nn.MaxPool3d(kernel_size=(2, 1, 1))
self.conv3d_3 = nn.Conv3d(
in_channels=32,
out_channels=64,
kernel_size=(3, 3, 3),
stride=(1, 1, 1),
padding=(1, 1, 1)
)
self.bn3 = nn.BatchNorm3d(64)
self.pool3 = nn.MaxPool3d(kernel_size=(2, 1, 1))
time_after_conv = time_steps // 8
flat_size = 64 * time_after_conv * num_channels
self.lstm = nn.LSTM(
input_size=64 * num_channels,
hidden_size=128,
num_layers=2,
batch_first=True,
bidirectional=True,
dropout=0.3
)
self.rp_head = nn.Sequential(
nn.Linear(256, 64),
nn.ReLU(),
nn.Dropout(0.5),
nn.Linear(64, num_classes_rp)
)
self.di_head = nn.Sequential(
nn.Linear(256, 64),
nn.ReLU(),
nn.Dropout(0.5),
nn.Linear(64, num_classes_di)
)
def forward(self, x):
"""
前向传播
Args:
x: EEG信号, shape=(B, 1, T, C, F)
B=batch, T=time, C=channels, F=freq
Returns:
rp_logits: 风险预测 logits
di_logits: 危险识别 logits
"""
x = F.relu(self.bn1(self.conv3d_1(x)))
x = self.pool1(x)
x = F.relu(self.bn2(self.conv3d_2(x)))
x = self.pool2(x)
x = F.relu(self.bn3(self.conv3d_3(x)))
x = self.pool3(x)
B, C, T, Ch, F = x.shape
x = x.permute(0, 2, 1, 3, 4)
x = x.reshape(B, T, C * Ch * F)
lstm_out, _ = self.lstm(x)
last_output = lstm_out[:, -1, :]
rp_logits = self.rp_head(last_output)
di_logits = self.di_head(last_output)
return rp_logits, di_logits
class CSPFeatureExtractor(nn.Module):
"""
Common Spatial Pattern (CSP) 特征提取
论文中的基线方法:
1. CSP: 空间滤波，最大化两类方差比
2. XGBDIM: XGBoost用于危险识别
3. EEGNet/DeepConvNet: 深度学习基线
"""
def __init__(self, num_channels: int = 32, num_filters: int = 6):
super().__init__()
self.num_filters = num_filters
def forward(self, x):
"""CSP特征提取"""
return x
if __name__ == "__main__":
batch_size = 8
time_steps = 1200
num_channels = 32
x = torch.randn(batch_size, 1, time_steps, num_channels, 1)
model = CRNN3D(num_channels=num_channels, time_steps=time_steps)
rp_logits, di_logits = model(x)
print(f"输入形状: {x.shape}")
print(f"RP logits形状: {rp_logits.shape}")
print(f"DI logits形状: {di_logits.shape}")
print(f"\n模型参数量: {sum(p.numel() for p in model.parameters()):,}")
print("\n=== 论文实验结果 ===")
print("3D-CRNN:")
print(f" RP Balanced Accuracy: 95.3% ± 2.7%")
print(f" DI (单被试): 80.9% → 85.0% (RSL提升)")
print(f" DI (跨会话): 77.0% ± 5.3%")
print(f" DI (跨被试, 已见): 77.4% ± 1.1%")
print(f" DI (跨被试, 未见): 64.9% ± 8.5%")
print("\n基线对比:")
print(" CSP+XGBDIM: 基线性能")
print(" EEGNet: 深度学习基线")
print(" DeepConvNet: 深度学习基线")
|
