序列标记器

Question

我正在分析时间序列信号。我设置了一个阈值以将噪声与基线噪声分开。为了识别每个信号序列的特性（持续时间，幅度，最大信号...），我构建了一个函数，可将连续的所有信号点聚合为不同的“峰值”。尽管此功能可以实现我想要的功能，但我想知道是否有人可以帮助我提高效率-e。 G。向量化，因为我的目标是在超过1M行的data.table上运行该函数。以下是具有该功能的示例数据：

# Generate dummy data
x <- sin(seq(from = 0, to = 20, length.out = 200)) + rnorm(200, 0,0.1)
x <- zoo(x)
plot(x)


# Label each point as signal (== )1) or noise (0)

y <- ifelse(x > 0.5, 1, 0)

# Function to label each peak

peak_labeler <- function(x) {

  tmp <- NULL

  for (i in seq_along(x)) {

    if (x[i] == 0) { tmp[i] <- 0 } # If baseline, mark as 0

    if (x[i] == 1) {

      # If x[n] belongs to a peak
      if (i == 1) {tmp[i] <- 1} # Label as 1 at t0

      else{

        if (!exists("Peak")) {Peak <- 0}

        if (x[i - 1] == 0) {
          # if previous point is no peak, add as peak

          Peak <- Peak + 1
          tmp[i] <- Peak
        }

        if (x[i - 1] == 1) {
          tmp[i] <- Peak
        }
      }
    }

  }

  return(tmp)

  rm(tmp, Peak, i) # Garbage collection
}

# Label peaks

dummy <- data.frame(t = 1:200, x,y,tmp = peak_labeler(y))

# Show data

ggplot(dummy, aes(x = t, y = x)) +
  geom_point(aes(col = as.factor(tmp), group = 1))

Answer 1

这是使用dplyr的方法。

cross_threshold行中的测试通过评估y与先前y是否在0.5的另一侧而起作用。如果是这样，则两个项y - threshold和lag(y) - threshold的符号将不同，从而导致TRUE，该值乘以1便成为1。如果它们在0.5的同一边，您将得到FALSE和0。 default = 0部分处理第一行，其中lag（y）未定义。然后，我们累加定义tmp组的累计交叉次数。

library(dplyr)

threshold = 0.5 
dummy <- data.frame(t = 1:200, x, y) %>%
mutate(cross_threshold = 1 * (sign(y - threshold) != sign(lag(y, default = 0) - threshold)),
     # Line above now optional, just if we want to label all crossings
     up = 1 * ((y > threshold) & (lag(y) < threshold)),
     tmp = if_else(y > threshold, cumsum(up), 0))

ggplot(dummy, aes(x = t, y = x)) +
  geom_point(aes(col = as.factor(tmp), group = 1)) +
  geom_point(data = filter(dummy, cross_threshold == 1), shape = 21, size = 5)