There are some functions of my code that could not be run using NChannelSet (i.e, Biobase::exprs(raw_data) ). I was just curious if exists some way to transform a NChannelSet object to an expressionSet.

Any idea?

It might depend exactly on your object, but extracting a channel might do the trick. From ?NChannelSet:

>  ## An NChannelSet with two channels and channel-specific phenoData
> R <- matrix(0, 10, 3, dimnames=list(NULL, LETTERS[1:3]))
> G <- matrix(1, 10, 3, dimnames=list(NULL, LETTERS[1:3]))
> assayData <- assayDataNew(R=R, G=G)
> data <- data.frame(ChannelRData=numeric(ncol(R)),
+                    ChannelGData=numeric(ncol(R)),
+                    ChannelRAndG=numeric(ncol(R)))
> varMetadata <- data.frame(labelDescription=c(
+                             "R-specific phenoData",
+                             "G-specific phenoData",
+                             "Both channel phenoData"),
+                           channel=factor(c("R", "G", "_ALL_")))
> phenoData <- AnnotatedDataFrame(data=data, varMetadata=varMetadata)
> obj <- NChannelSet(assayData=assayData, phenoData=phenoData)
> obj
NChannelSet (storageMode: lockedEnvironment)
assayData: 10 features, 3 samples 
  element names: G, R 
protocolData: none
phenoData
  sampleNames: A B C
  varLabels: ChannelRData ChannelGData ChannelRAndG
  varMetadata: labelDescription channel
featureData: none
experimentData: use 'experimentData(object)'
Annotation:  
> channel(obj, "G")
ExpressionSet (storageMode: lockedEnvironment)
assayData: 10 features, 3 samples 
  element names: exprs 
protocolData: none
phenoData
  sampleNames: A B C
  varLabels: ChannelGData ChannelRAndG
  varMetadata: labelDescription channel
featureData: none
experimentData: use 'experimentData(object)'
Annotation:

NChannel objects are almost always two-color microarrays. There is more information in the two-channels than in any single expression summary, so any attempt to convert to an expressionSet will lose information. Two-color microarrays are analysed in terms of log-ratios (also called M-values), so the least bad way convert an NChannel object to an expressionSet is to compute the normalized M-values and to use those as the expression matrix. Simply extracting one channel and ignoring the other will give much noisier and possibly meaningless results.