深度残差收缩网络 DRSN 就是在残差网络基础上加了可学习软阈值收缩层,专门用来处理带噪声的信号,最常用在轴承故障诊断,这里直接给 PyTorch 可运行代码,分通道共享阈值 DRSN-CS、通道独立阈值 DRSN-CW 两套单元;核心逻辑是卷积提取特征之后,子网络自动算出阈值,用软阈值把噪声特征压缩归零,保留有效冲击特征,训练流程和普通分类网络一致,只需要准备一维振动信号数据集,修改输入通道、分类数量就能直接跑。
DRSN 核心原理说明
- 基础骨架:预激活式残差块 Pre-Act ResNet,解决深层网络梯度消失
- 新增模块:残差收缩单元 RSBU,分为 RSBU-CS(全部通道共用一个阈值)、RSBU-CW(每个通道单独阈值)
- 收缩算子:软阈值化 y=sign(x)⋅max(∣x∣−τ,0),阈值 τ 是网络自己学出来的,不是手动固定
- 适用场景:一维时序振动信号故障诊断,也可以改成 Conv2d 版本用于图像降噪
PyTorch 完整可运行代码
import torch
import torch.nn as nn
import torch.nn.functional as F
软阈值层
class SoftThresholding(nn.Module):
def forward(self, x, tau):
return torch.sign(x) F.relu(torch.abs(x) - tau)
RSBU-CS:通道共享阈值残差收缩单元
class RSBU_CS(nn.Module):
def __init__(self, in_channels, out_channels, kernel_size=3, stride=1):
super().__init__()
padding = kernel_size // 2
self.bn1 = nn.BatchNorm1d(in_channels)
self.relu1 = nn.ReLU(inplace=True)
self.conv1 = nn.Conv1d(in_channels, out_channels, kernel_size, stride, padding, bias=False)
self.bn2 = nn.BatchNorm1d(out_channels)
self.relu2 = nn.ReLU(inplace=True)
self.conv2 = nn.Conv1d(out_channels, out_channels, kernel_size, 1, padding, bias=False)
阈值估计分支
self.gap = nn.AdaptiveAvgPool1d(1)
self.fc1 = nn.Linear(out_channels, out_channels)
self.bn_fc = nn.BatchNorm1d(out_channels)
self.relu_fc = nn.ReLU(inplace=True)
self.fc2 = nn.Linear(out_channels, 1)
self.sigmoid = nn.Sigmoid()
self.soft_th = SoftThresholding()
shortcut
self.shortcut = nn.Sequential()
if stride != 1 or in_channels != out_channels:
self.shortcut = nn.Sequential(
nn.Conv1d(in_channels, out_channels, 1, stride, bias=False),
nn.BatchNorm1d(out_channels)
)
def forward(self, x):
residual = self.bn1(x)
residual = self.relu1(residual)
residual = self.conv1(residual)
residual = self.bn2(residual)
residual = self.relu2(residual)
residual = self.conv2(residual)
计算阈值
abs_feat = torch.abs(residual)
gap_out = self.gap(abs_feat).squeeze(-1)
t = self.fc1(gap_out)
t = self.bn_fc(t)
t = self.relu_fc(t)
t = self.fc2(t)
alpha = self.sigmoid(t)
tau = alpha torch.mean(gap_out, dim=1, keepdim=True).unsqueeze(-1)
residual = self.soft_th(residual, tau)
shortcut = self.shortcut(x)
out = residual + shortcut
return out
RSBU-CW:通道独立阈值残差收缩单元
class RSBU_CW(nn.Module):
def __init__(self, in_channels, out_channels, kernel_size=3, stride=1):
super().__init__()
padding = kernel_size // 2
self.bn1 = nn.BatchNorm1d(in_channels)
self.relu1 = nn.ReLU(inplace=True)
self.conv1 = nn.Conv1d(in_channels, out_channels, kernel_size, stride, padding, bias=False)
self.bn2 = nn.BatchNorm1d(out_channels)
self.relu2 = nn.ReLU(inplace=True)
self.conv2 = nn.Conv1d(out_channels, out_channels, kernel_size, 1, padding, bias=False)
self.gap = nn.AdaptiveAvgPool1d(1)
self.fc1 = nn.Linear(out_channels, out_channels)
self.bn_fc = nn.BatchNorm1d(out_channels)
self.relu_fc = nn.ReLU(inplace=True)
self.fc2 = nn.Linear(out_channels, out_channels)
self.sigmoid = nn.Sigmoid()
self.soft_th = SoftThresholding()
self.shortcut = nn.Sequential()
if stride != 1 or in_channels != out_channels:
self.shortcut = nn.Sequential(
nn.Conv1d(in_channels, out_channels, 1, stride, bias=False),
nn.BatchNorm1d(out_channels)
)
def forward(self, x):
residual = self.bn1(x)
residual = self.relu1(residual)
residual = self.conv1(residual)
residual = self.bn2(residual)
residual = self.relu2(residual)
residual = self.conv2(residual)
abs_feat = torch.abs(residual)
gap_out = self.gap(abs_feat).squeeze(-1)
t = self.fc1(gap_out)
t = self.bn_fc(t)
t = self.relu_fc(t)
t = self.fc2(t)
alpha = self.sigmoid(t).unsqueeze(-1)
tau = alpha gap_out.unsqueeze(-1)
residual = self.soft_th(residual, tau)
shortcut = self.shortcut(x)
out = residual + shortcut
return out
完整DRSN网络
class DRSN(nn.Module):
def __init__(self, block, num_blocks, in_channel=1, num_classes=10):
super().__init__()
self.in_ch = 64
self.conv_in = nn.Conv1d(in_channel, 64, kernel_size=7, stride=2, padding=3, bias=False)
self.bn_in = nn.BatchNorm1d(64)
self.relu_in = nn.ReLU(inplace=True)
self.maxpool = nn.MaxPool1d(3, 2, padding=1)
self.layer1 = self._make_layer(block, 64, num_blocks[0], stride=1)
self.layer2 = self._make_layer(block, 128, num_blocks[1], stride=2)
self.layer3 = self._make_layer(block, 256, num_blocks[2], stride=2)
self.layer4 = self._make_layer(block, 512, num_blocks[3], stride=2)
self.gap = nn.AdaptiveAvgPool1d(1)
self.fc = nn.Linear(512, num_classes)
def _make_layer(self, block, out_ch, blocks, stride):
layers = []
layers.append(block(self.in_ch, out_ch, stride=stride))
self.in_ch = out_ch
for _ in range(1, blocks):
layers.append(block(self.in_ch, out_ch, stride=1))
return nn.Sequential(layers)
def forward(self, x):
x = self.conv_in(x)
x = self.bn_in(x)
x = self.relu_in(x)
x = self.maxpool(x)
x = self.layer1(x)
x = self.layer2(x)
x = self.layer3(x)
x = self.layer4(x)
x = self.gap(x).squeeze(-1)
x = self.fc(x)
return x
测试入口
if __name__ == "__main__":
DRSN-CS
model_cs = DRSN(RSBU_CS, [2,2,2,2], in_channel=1, num_classes=10)
DRSN-CW
model_cw = DRSN(RSBU_CW, [2,2,2,2], in_channel=1, num_classes=10)
test_x = torch.randn(4,1,1024)
print("DRSN-CS output shape:", model_cs(test_x).shape)
print("DRSN-CW output shape:", model_cw(test_x).shape)
代码使用步骤清单
- 环境准备:安装 torch、torchvision,推荐 torch>=1.10 版本
- 数据格式:输入维度 [batch, 通道, 序列长度],振动信号一般通道 = 1
- 模型选择:噪声比较均匀选 DRSN-CS;不同通道噪声差异大选 DRSN-CW
- 训练配置:损失用 CrossEntropyLoss,优化器 Adam,学习率 1e-3 起步,搭配学习率衰减
- 改造图像版本:把所有 BatchNorm1d、Conv1d、AdaptiveAvgPool1d 替换成 2d 对应层
DRSN-CS 与 DRSN-CW 对比表
| 版本 | 阈值特点 | 参数量 | 降噪能力 | 适用场景 |
|---|---|---|---|---|
| DRSN-CS | 全部通道共享同一个 τ | 更少 | 中等 | 噪声分布均匀、算力有限 |
| DRSN-CW | 每个通道独立 τ | 更多 | 更强 | 多通道、噪声分布差异大 |
【常见问题】
问题 1:深度残差收缩网络 PyTorch 实现里软阈值层的作用是什么?
回答 1:深度残差收缩网络 PyTorch 实现中的软阈值层是核心收缩模块,按照可学习阈值把特征里幅值小于 τ 的噪声清零,保留有用特征,相比普通 ResNet 自带自适应降噪能力。
问题 2:深度残差收缩网络 PyTorch 代码 DRSN-CS 和 DRSN-CW 该怎么选?
回答 2:深度残差收缩网络 PyTorch 实现里 DRSN-CS 通道共用阈值,参数少、训练快;DRSN-CW 每个通道单独阈值,对复杂噪声适配更好,优先用在多通道振动信号故障诊断任务。
问题 3:深度残差收缩网络 PyTorch 实现能不能改成图像分类版本?
回答 3:可以,深度残差收缩网络 PyTorch 实现只需要把代码中 Conv1d、BatchNorm1d、AdaptiveAvgPool1d 全部替换为 Conv2d、BatchNorm2d、AdaptiveAvgPool2d,输入改成图片张量格式即可。
问题 4:深度残差收缩网络 PyTorch 实现适合哪些任务?
回答 4:深度残差收缩网络 PyTorch 实现原生设计面向一维时序信号,最主流是旋转机械轴承故障诊断,也可以改造用于一维信号降噪、多通道时序分类。
【参考资料】
[2026-11-18](深度残差收缩网络的 PyTorch 编程实现)腾讯云开发者社区
[2026-11-17](深度残差收缩网络的 PyTorch 程序复现)博客园
[2024-12-07](深度残差收缩网络 pytorch 复现)51CTO 博客
[2021-04-15](深度残差收缩网络 (完整 PyTorch 程序))腾讯云开发者社区
【免责声明】
以上深度残差收缩网络 PyTorch 程序基于原始论文开源思路复现,仅用于学术学习、算法验证;实际工程使用时需要根据数据集信噪比、采样长度、硬件显存调整网络层数、通道数与超参,代码不保证直接在所有数据集上收敛,不构成工业落地方案。
