一、思路:用重构误差刻画正常
自编码器在仅用正常样本训练时,会学会压缩并重建正常模式;当出现异常,重建误差会显著增大。相比固定阈值,这种方法能自适应复杂工况。我们用 LSTM 捕捉时序依赖,对单变量或多变量传感器序列做滑动窗口重建。
二、可复现代码
构造窗口样本:
import numpy as np, torch
def make_windows(x, w=24):
return np.array([x[i:i+w] for i in range(len(x)-w+1)])
x = np.sin(np.linspace(0, 40, 1000)) + np.random.normal(0, 0.05, 1000)
X = make_windows(x, 24).astype('float32')
loader = torch.utils.data.DataLoader(torch.from_numpy(X), batch_size=64, shuffle=True)定义 LSTM 自编码器并训练:
class AE(torch.nn.Module):
def __init__(self, n=1, h=32):
super().__init__()
self.enc = torch.nn.LSTM(n, h, batch_first=True)
self.dec = torch.nn.LSTM(h, n, batch_first=True)
def forward(self, x):
z, _ = self.enc(x)
out, _ = self.dec(z)
return out
model = AE()
opt = torch.optim.Adam(model.parameters(), lr=1e-3)
for epoch in range(20):
for b in loader:
b = b.unsqueeze(-1)
opt.zero_grad()
loss = torch.nn.MSELoss()(model(b), b)
loss.backward(); opt.step()以验证集误差分位数设定阈值,超出即判异常:
model.eval()
with torch.no_grad():
errs = [torch.nn.MSELoss()(model(t.unsqueeze(0).unsqueeze(-1)), t.unsqueeze(0).unsqueeze(-1)).item() for t in torch.from_numpy(X)]
thr = np.quantile(errs, 0.99)
print('anomaly threshold =', round(thr, 6))提醒:须用充足且标注清晰的正常样本训练;工况切换、传感器漂移也会抬升误差,判异常后仍需工程人员结合上下文确认,避免误停机。
参考来源
PyTorch 官方文档(torch.nn.LSTM 与训练循环) - PyTorch;scikit-learn 异常检测与 Novelty 检测指南 - scikit-learn 官方文档;LSTM 自编码器用于时序异常检测的公开研究综述 - 学术文献数据库。