我发现这个答案对我帮助最大，但发现有一些边缘情况，它似乎不能正确处理，特别是消极的情况，以及极限距离为0的情况(如果我们从最大/最小数据中获取极限，就会发生这种情况)。测试似乎表明，这是一致的

我使用以下代码。这里我假设我们有[x1,x2]我们想把它变换成[y1,y2]。我处理这个问题的方法是将[x1,x2]转换为[0,1](一个足够简单的转换)，然后[0,1]转换为[y1,y2]。

climate <- tibble(
  Month = 1:12,
  Temp = c(-4,-4,0,5,11,15,16,15,11,6,1,-3),
  Precip = c(49,36,47,41,53,65,81,89,90,84,73,55)
)
#Set the limits of each axis manually:

  ylim.prim <- c(0, 180)   # in this example, precipitation
ylim.sec <- c(-4, 18)    # in this example, temperature



  b <- diff(ylim.sec)/diff(ylim.prim)

#If all values are the same this messes up the transformation, so we need to modify it here
if(b==0){
  ylim.sec <- c(ylim.sec[1]-1, ylim.sec[2]+1)
  b <- diff(ylim.sec)/diff(ylim.prim)
}
if (is.na(b)){
  ylim.prim <- c(ylim.prim[1]-1, ylim.prim[2]+1)
  b <- diff(ylim.sec)/diff(ylim.prim)
}


ggplot(climate, aes(Month, Precip)) +
  geom_col() +
  geom_line(aes(y = ylim.prim[1]+(Temp-ylim.sec[1])/b), color = "red") +
  scale_y_continuous("Precipitation", sec.axis = sec_axis(~((.-ylim.prim[1]) *b  + ylim.sec[1]), name = "Temperature"), limits = ylim.prim) +
  scale_x_continuous("Month", breaks = 1:12) +
  ggtitle("Climatogram for Oslo (1961-1990)")  

这里的关键部分是，我们用~((.-ylim.prim[1]) *b + ylim.sec[1])转换次要y轴，然后对实际值y = ylim.prim[1]+(Temp-ylim.sec[1])/b)应用逆。我们还应该确保limits = ylim.prim。

这是我对如何做二次轴变换的两种看法。首先，您希望将主数据和辅助数据的范围耦合起来。这通常是混乱的，因为您不想要的变量污染了全局环境。

为了简化这一点，我们将创建一个生成两个函数的函数工厂，其中scales::rescale()完成所有繁重的工作。因为这些是闭包，所以它们知道创建它们的环境，所以它们“有”创建之前生成的to和from参数的“内存”。

一个函数进行正向转换:将辅助数据转换为主要尺度。 第二个函数进行反向转换:将主要单位中的数据转换为次要单位。

library(ggplot2)
library(scales)

# Function factory for secondary axis transforms
train_sec <- function(primary, secondary, na.rm = TRUE) {
  # Thanks Henry Holm for including the na.rm argument!
  from <- range(secondary, na.rm = na.rm)
  to   <- range(primary, na.rm = na.rm)
  # Forward transform for the data
  forward <- function(x) {
    rescale(x, from = from, to = to)
  }
  # Reverse transform for the secondary axis
  reverse <- function(x) {
    rescale(x, from = to, to = from)
  }
  list(fwd = forward, rev = reverse)
}

这看起来相当复杂，但是创建函数工厂会使其余的一切变得更简单。现在，在绘制图形之前，我们将通过向工厂显示主要和次要数据来生成相关函数。我们将使用经济学数据集，它的失业列和pasavert列的范围非常不同。

sec <- with(economics, train_sec(unemploy, psavert))

然后我们使用y = sec$fwd(psavert)将辅助数据重新缩放到主轴，并指定~ sec$rev(.)作为辅助轴的转换参数。这给了我们一个主要范围和次要范围在图上占据相同空间的图。

ggplot(economics, aes(date)) +
  geom_line(aes(y = unemploy), colour = "blue") +
  geom_line(aes(y = sec$fwd(psavert)), colour = "red") +
  scale_y_continuous(sec.axis = sec_axis(~sec$rev(.), name = "psavert"))

工厂比这稍微灵活一些，因为如果您只是想重新调整最大值，您可以传入下限为0的数据。

# Rescaling the maximum
sec <- with(economics, train_sec(c(0, max(unemploy)),
                                 c(0, max(psavert))))

ggplot(economics, aes(date)) +
  geom_line(aes(y = unemploy), colour = "blue") +
  geom_line(aes(y = sec$fwd(psavert)), colour = "red") +
  scale_y_continuous(sec.axis = sec_axis(~sec$rev(.), name = "psavert"))

由reprex包于2021-02-05创建(v0.3.0)

我承认这个例子中的区别不是很明显，但如果你仔细观察，你会发现最大值是相同的，红线比蓝色的线低。

编辑:

这种方法现在已经在ggh4x包中的help_secondary()函数中被捕获和扩展。声明:我是ggh4x的作者。

Kohske大约在3年前提供了解决这一挑战的技术骨干。在Stackoverflow [id: 18989001, 29235405, 21026598]的几个实例中已经讨论过这个主题及其解决方案的技术细节。因此，我将只提供一个特定的变化和一些解释性演练，使用上述解决方案。

让我们假设我们确实在组G1中有一些数据y1，而组G2中的一些数据y2以某种方式相关，例如范围/比例转换或添加了一些噪声。我们想把数据画在一张图上，左边是y1右边是y2。

  df <- data.frame(item=LETTERS[1:n],  y1=c(-0.8684, 4.2242, -0.3181, 0.5797, -0.4875), y2=c(-5.719, 205.184, 4.781, 41.952, 9.911 )) # made up!

> df
  item      y1         y2
1    A -0.8684 -19.154567
2    B  4.2242 219.092499
3    C -0.3181  18.849686
4    D  0.5797  46.945161
5    E -0.4875  -4.721973

如果我们现在把数据画在一起

ggplot(data=df, aes(label=item)) +
  theme_bw() + 
  geom_segment(aes(x='G1', xend='G2', y=y1, yend=y2), color='grey')+
  geom_text(aes(x='G1', y=y1), color='blue') +
  geom_text(aes(x='G2', y=y2), color='red') +
  theme(legend.position='none', panel.grid=element_blank())

它并没有很好地对齐，因为小尺度y1明显被大尺度y2折叠了。

这里应对挑战的技巧是在技术上根据第一个尺度y1绘制两个数据集，但根据二级轴报告第二个数据集，并使用标签显示原始尺度y2。

因此，我们构建了第一个辅助函数CalcFudgeAxis，它计算并收集要显示的新轴的特征。这个函数可以被修改成任意的形式(这个函数只是将y2映射到y1的范围上)。

CalcFudgeAxis = function( y1, y2=y1) {
  Cast2To1 = function(x) ((ylim1[2]-ylim1[1])/(ylim2[2]-ylim2[1])*x) # x gets mapped to range of ylim2
  ylim1 <- c(min(y1),max(y1))
  ylim2 <- c(min(y2),max(y2))    
  yf <- Cast2To1(y2)
  labelsyf <- pretty(y2)  
  return(list(
    yf=yf,
    labels=labelsyf,
    breaks=Cast2To1(labelsyf)
  ))
}

什么产生了一些:

> FudgeAxis <- CalcFudgeAxis( df$y1, df$y2 )

> FudgeAxis
$yf
[1] -0.4094344  4.6831656  0.4029175  1.0034664 -0.1009335

$labels
[1] -50   0  50 100 150 200 250

$breaks
[1] -1.068764  0.000000  1.068764  2.137529  3.206293  4.275058  5.343822


> cbind(df, FudgeAxis$yf)
  item      y1         y2 FudgeAxis$yf
1    A -0.8684 -19.154567   -0.4094344
2    B  4.2242 219.092499    4.6831656
3    C -0.3181  18.849686    0.4029175
4    D  0.5797  46.945161    1.0034664
5    E -0.4875  -4.721973   -0.1009335

现在我将Kohske的解决方案包装在第二个辅助函数PlotWithFudgeAxis中(我们将ggplot对象和新轴的辅助对象放入其中):

library(gtable)
library(grid)

PlotWithFudgeAxis = function( plot1, FudgeAxis) {
  # based on: https://rpubs.com/kohske/dual_axis_in_ggplot2
  plot2 <- plot1 + with(FudgeAxis, scale_y_continuous( breaks=breaks, labels=labels))

  #extract gtable
  g1<-ggplot_gtable(ggplot_build(plot1))
  g2<-ggplot_gtable(ggplot_build(plot2))

  #overlap the panel of the 2nd plot on that of the 1st plot
  pp<-c(subset(g1$layout, name=="panel", se=t:r))
  g<-gtable_add_grob(g1, g2$grobs[[which(g2$layout$name=="panel")]], pp$t, pp$l, pp$b,pp$l)

  ia <- which(g2$layout$name == "axis-l")
  ga <- g2$grobs[[ia]]
  ax <- ga$children[[2]]
  ax$widths <- rev(ax$widths)
  ax$grobs <- rev(ax$grobs)
  ax$grobs[[1]]$x <- ax$grobs[[1]]$x - unit(1, "npc") + unit(0.15, "cm")
  g <- gtable_add_cols(g, g2$widths[g2$layout[ia, ]$l], length(g$widths) - 1)
  g <- gtable_add_grob(g, ax, pp$t, length(g$widths) - 1, pp$b)

  grid.draw(g)
}

现在可以将所有内容放在一起:下面的代码显示了建议的解决方案如何在日常环境中使用。plot调用现在不再绘制原始数据y2，而是一个克隆版本yf(保存在预先计算的辅助对象FudgeAxis中)，它以y1的规模运行。然后使用Kohske的辅助函数PlotWithFudgeAxis操作原始ggplot对象，以添加第二个轴，保留y2的刻度。它的情节和被操纵的情节一样。

FudgeAxis <- CalcFudgeAxis( df$y1, df$y2 )

tmpPlot <- ggplot(data=df, aes(label=item)) +
      theme_bw() + 
      geom_segment(aes(x='G1', xend='G2', y=y1, yend=FudgeAxis$yf), color='grey')+
      geom_text(aes(x='G1', y=y1), color='blue') +
      geom_text(aes(x='G2', y=FudgeAxis$yf), color='red') +
      theme(legend.position='none', panel.grid=element_blank())

PlotWithFudgeAxis(tmpPlot, FudgeAxis)

现在它有两个轴，左边是y1右边是y2

Above solution is, to put it straight, a limited shaky hack. As it plays with the ggplot kernel it will throw some warnings that we exchange post-the-fact scales, etc. It has to be handled with care and may produce some undesired behaviour in another setting. As well one may need to fiddle around with the helper functions to get the layout as desired. The placement of the legend is such an issue (it would be placed between the panel and the new axis; this is why I droped it). The scaling / alignment of the 2 axis is as well a bit challenging: The code above works nicely when both scales contain the "0", else one axis gets shifted. So definetly with some opportunities to improve...

如果on想要保存图片，就必须将调用包装成设备打开/关闭:

png(...)
PlotWithFudgeAxis(tmpPlot, FudgeAxis)
dev.off()

这在ggplot2中是不可能的，因为我认为具有单独y尺度的图(不是相互转换的y尺度)从根本上是有缺陷的。一些问题:

The are not invertible: given a point on the plot space, you can not uniquely map it back to a point in the data space. They are relatively hard to read correctly compared to other options. See A Study on Dual-Scale Data Charts by Petra Isenberg, Anastasia Bezerianos, Pierre Dragicevic, and Jean-Daniel Fekete for details. They are easily manipulated to mislead: there is no unique way to specify the relative scales of the axes, leaving them open to manipulation. Two examples from the Junkcharts blog: one, two They are arbitrary: why have only 2 scales, not 3, 4 or ten?

你也可能想要阅读Stephen Few关于双缩放轴在图形中的主题的冗长讨论，它们是最好的解决方案吗?

我承认并同意哈德利(和其他人)的观点，即单独的y量表“存在根本缺陷”。说到这里，我经常希望ggplot2有这个特性——特别是当数据是宽格式的，并且我想快速地可视化或检查数据时(即仅供个人使用)。

虽然tidyverse库可以很容易地将数据转换为长格式(这样facet_grid()就可以工作)，但这个过程仍然不是简单的，如下所示:

library(tidyverse)
df.wide %>%
    # Select only the columns you need for the plot.
    select(date, column1, column2, column3) %>%
    # Create an id column – needed in the `gather()` function.
    mutate(id = n()) %>%
    # The `gather()` function converts to long-format. 
    # In which the `type` column will contain three factors (column1, column2, column3),
    # and the `value` column will contain the respective values.
    # All the while we retain the `id` and `date` columns.
    gather(type, value, -id, -date) %>%
    # Create the plot according to your specifications
    ggplot(aes(x = date, y = value)) +
        geom_line() +
        # Create a panel for each `type` (ie. column1, column2, column3).
        # If the types have different scales, you can use the `scales="free"` option.
        facet_grid(type~., scales = "free")

It seemingly appears to be a simple question but it boggles around 2 fundamental questions. A) How to deal with a multi-scalar data while presenting in a comparative chart, and secondly, B) whether this can be done without some thumb rule practices of R programming such as i) melting data, ii) faceting, iii) adding another layer to existing one. The solution given below satisfies both the above conditions as it deals data without having to rescale it and secondly, the techniques mentioned are not used.

这是结果，

如果有兴趣了解更多关于此方法的信息，请点击下面的链接。 如何绘制一个2 y轴图表与条形并排而不重新缩放数据