In physics, good ideas that fail tend to be refined and recycled into correct theories later.
LeSage gravity and Feynman diagrams
One of the best examples is the LeSage theory of gravitation. LeSage imagined that all of space is filled with particles moving at some enormous speed c, and that gravity is caused by the absorption of these particles by matter, casting a shadow in the particles behind them. The shadow leads to inverse-square attraction.
This theory died because it would naively predict friction in the LeSage ether, which would wreck the orbits of the planets (of course, with out modern views, we would make the LeSage ether relativistically invariant, fixing this problem, but this was unavailable at the time). But despite this, it was very popular throughout the 19th century. Preston and DePretto, two 19th century lay-scientists, conjectured based on this theory that all mass contains an amount of energy proportional to $mc^2$ (the speed of the LeSage particles had settled down to c by the end of the 19th century).
In the 20th century, Feynman's particle exchange theory of forces revived LeSagian ideas, but this time for good. Feynman's explanation keeps the idea that the inverse square law comes from particle exchange diluted by geometry as they go from on object to another, but makes it quantum, and makes it relativistically invariant. It also becomes universal to all forces, showing how repulsion works.
Feynman, I suppose in honor of LeSage, devotes a section of the Feynman lectures to the LeSage model, to show how theories can fail.
Kelvin's ether atoms
The idea of "knots in the ether", that atoms are topological vortex lines in a fluid filling all space, was explicitly revived by Tony Skyrme in 1960, when he made a model of Baryons as topological defects in the newly discovered pion condensate. This idea was considered dead for many years, until Rajeev, Nair, and Balachandran, followed closely by Witten, showed that it emerges within the large-N (string) model of QCD.
The idea that Baryon interactions are well modeled by tying knots in the pion ether is roughly successful at predicting the structure of small nuclei.
The original idea of the bootstrap dates back to Heisenberg 1940, but really got going after Pomaranchuk, Gribov, Mandelstam and Gell-Mann started to think about the idea in the late 1950s. The point of bootstrap was to give a description of hadronic interactions based on the exchange of particles which would be formed through the same interaction. Geoffrey Chew proselyzed for this so effectively that it became the dominant paradigm in physics between 1964 and 1974.
During this time, the interaction of Regge trajectories was understood qualitatively, then quantitatively, and a whole host of new ideas emerged from this: the pomeron, conspiracies, dispersion relations, unitarity cuts, etc. These ideas are not as well understood today as they were back then, because almost all those people were lost to physics.
But they developed string theory in the last 6 years, and string theory pulled all their ideas out of the dustbin.
Yang Mills theory (as Yang Mills wanted)
When Yang and Mills proposed their theory of an SU(2) gauge theory, they wanted it to gauge Isospin. The fundamental interactions don't do that, but Sakurai predicted the $\rho$ meson based on this idea. There are two light vector mesons, the $\rho$ and $\omega$, and the idea that they are gauging isospin and hypercharge led to the idea that they mix with the photon at hadronic energies. This predicted photon nucleon scattering well.
The theory predicts coupling constant universality-- all isodoublets should have the same coupling to the rho. But it doesn't explain the mass of the rho, since Isospin is unbroken. These ideas died out, as it was realized that these particles are not fundamental. But in modern AdS/QCD it is completely rehabilitated.
I don't know what the experimental situation is with respect to coupling constant universality. I will ask that as a question.