Watson's Guattari's Diagrammatic Thought addresses this directly, and in particular goes over the evolution of this concept in some detail (from p. 12):
The concept of the diagram appears in A Thousand Plateaus (ATP 141-144, 531 n. 41 / 176-180, 177 n. 38) but the details of its development are found in Guattari's writings of the 1970s. The notion was adapted from Charles Sanders Peirce, who includes the diagram among the icons in his index-icon-symbol model of the sign. Peirce identifies three types of icon: image, metaphor, and diagram. For him, the icon operates through a relation of resemblance between the sign and its referent. Guattari would agree that the image and the metaphor signify through resemblance, which is to say representation, but his version of the diagram functions differently because as he defines it, the diagram does not signify; it is "a-signifying"... Examples of the diagram at work include the algorithms of logic, algebra and topology; as well processes of recording, data storage, and computer processing; all of which are used in mathematics, science, technology and polyphonic music. Neither mathematics nor musical notation are languages -- rather, both bypass signification altogether...
Already in his notes for Anti-Oedipus, Guattari senses that Peirce's diagram is somehow special, that it unleashes "deterritorialized polyvocity," that it must be understood as distinct from the image because the diagram is a site of production (AOP 72, 214, 243-255/97, 308, 346-349). He continues reflecting on the powerful, productive diagram in Revolution moleculaire [Molecular Revolution] and L'inconscient Machinique [The Machinic Unconscious], concluding that diagrams "are no longer, strictly speaking, semiotic entities." Their "purpose is not to denote or to image the morphemes of an already-constituted referent, but to produce them" (IM 223, 224). In other words, diagrams do not represent thought; rather, they generate thought. Diagrams abound in experimental science, he says, because it is "a sphere where signs have a direct effect on things," involving "both material technology and a complex manipulation of sign machines" (MR 166/RM 303). The diagrammatic consists precisely in this conjunction between deterritorialized signs and deterritorialized objects.
Watson proceeds to demonstrate how Guattari on several occasions illuminates the notion of the diagram with the example of theoretical physics. The takeaway for me here was how the diagram functions 'outside of' language but nevertheless 'makes use' of signs. Watson formulates this point as follows (p. 13): 'This "diagrammatic process" makes use of signs, but not language, and therefore uses neither signifiers nor signification.'
Finally, with regard to the drawings, Watson also notes that
Guattari never claims that the drawings which illustrate his books are "diagrams," according to his concept, but his drawings do figure heavily in his analytical writing. His drawings work like diagrams in the sense that they seem at times to generate ideas, as if they were operating on their own, like little machines.
Watson also sheds some light on the relationship between schizoanalysis and cartography, and helps place it within a larger cultural-theoretical context:
[Guattari] characterizes schizoanalysis not only as metamodeling, but also as map-making, a process of building "a map of the unconscious--with its strata, lines of deterritorialization, and black holes." Guattari's emphasis on cartography (as for example in the title Cartographies schizoanalytiques) can be placed within a larger poststructuralist vogue of mapping which presupposes "the unremitting deconstruction of representational thinking" and therefore "excludes a metaphysical definition of mapping the classical mimetic sense." Recognizing this rejection of representation and mimesis is crucial to understanding how Guatarri defines modeling, mapping and the diagram... Metamodeling can be understood as a very special form of map-making. It consists in making maps that are not content to merely illustrate, but which also create and produce.