Tokamak H-mode plasmas frequently exhibit edge-localized modes (ELMs). ELMs allow maintaining sufficient plasma purity and thus enable stationary H-mode. On the other hand in a future device ELMs may cause divertor power flux densities far in excess of tolerable material limits. The size of the energy loss per ELM is determined by saturation effects in the non-linear phase of the ELM, which at present is hardly understood. ASDEX Upgrade is now equipped with a set of fast sampling diagnostics, which is well suited to investigate the chain of events around the ELM crash with appropriate temporal resolution (10 µs). Solitary magnetic perturbations (SMPs) are identified as dominant features in the radial magnetic fluctuations below 100 kHz. They are typically observed close (±100 µs) to the onset of pedestal erosion. SMPs are field aligned structures rotating in the electron diamagnetic drift direction with perpendicular velocities of about 10 km s-1. A comparison of perpendicular velocities suggests that the perturbation evoking SMPs is located at or inside the separatrix. Analysis of very pronounced examples showed that the number of peaks per toroidal turn is 1 or 2, which is clearly lower than the corresponding numbers in linear stability calculations. In combination with strong peaking of the magnetic signals this results in a solitary appearance resembling modes like palm tree modes, edge snakes or outer modes. This behaviour has been quantified as solitariness and correlated with main plasma parameters. SMPs may be considered as a signature of the non-linearELMphase originating at the separatrix or further inside. Thus they provide a handle to investigate the transition from linear to non-linear ELM phase. By comparison with data from gas puff imaging processes in the non-linear phase at or inside the separatrix and in the scrape-off layer (SOL) can be correlated. A connection between the passing of an SMP and the onset of radial filament propagation has been found. Eventually the findings related to SMPs may contribute to a future quantitative understanding of the non-linear ELM evolution.