In this contribution, we introduce a hybrid model for virtual sensing applications which combines a frequency response function model with a Long Short‐Term Memory network. It estimates the behavior of non‐linear dynamic systems with multiple input and output channels by generating predictions on short subsequences of signals and recombining them using a windowing technique. The approach is tested on an experimental dataset composed of measurements from a 3‐component servo hydraulic fatigue test bench. The model is parameterized using noise data, while fatigue serviceloads with variable amplitudes are used for validation and testing.