This work proposes a reference-free damage identification technique based on the statistical modelling of vibration signals. This technique does not require a reference signal to identify damage, but uses the information embedded on the signal at the sensor location. The vibration signal statistical distribution is modelled using Gaussian Mixture Models (GMM) for density estimation, from which indicators such variance, skewers, kurtosis, energy and entropy are computed. Using Principal Component Analysis (PCA), the deviation from the mean value of the first and second principal components is used as indicators of damage. This approach is validated on a simulated Finite Element (FE) model of a simply supported beam, where structural damage is simulated as a reduction of local stiffness. The ability to detect damage efficiently is first investigated by modelling directly the time vibration signal. Furthermore, a frequency approach is considered, where the vibration signal is first decomposed into frequency subbands and then the estimated signal density is used in each subband to compile a space-frequency error map. Simulation results highlight the superiority of the frequency based approach over the time domain one.