Experimental identification of structural dynamics models is usually based on the modal analysis approach. In the classical modal parameter estimation approach, the baseline data which are processed are frequency response functions measured under laboratory conditions. However, in many applications, the real operating conditions may differ significantly from those applied during the modal test. Hence, the need arises to identify a modal model in operational conditions. This issue is even more complicated by the fact that in most cases, only response data are measurable while the actual loading conditions are unknown. Therefore, the system identification process will need to base itself on output-only data. In the last decade, the problem of output-only modal analysis has typically been approached by applying a peak-picking technique to the auto-and cross-powers of the measured responses, resulting in operational deflection shapes and approximate estimates for the resonance frequencies. These shapes were then compared to or even decomposed into the laboratory modal results. Over the past years, several modal parameter estimation techniques have been proposed and studied for modal parameter extraction from output-only data. They include autoregressive moving averaging models, natural excitation technique (NExT) and stochastic subspace methods. In this paper, the capabilities and limitations of the NExT technique and two subspace techniques, referred to as the balanced realization and the canonical variate analysis, are evaluated for its applicability to industrial cases. The theory of the methods is briefly outlined. Subsequently, the performance of the methods is critically evaluated and compared for three industrial cases: the modal characterisation of the rear suspension system of a family car during road tests in order to understand a booming noise problem, flight flutter analysis of a commercial aircraft and the identification of the modes of a three-span concrete bridge under ambient excitation. Practical aspects of the output-only system identification process such as the selection of the model order, handling responses measured in different patches due to a limited number of channels and the validation of the derived structural dynamics models by overlaying synthesised auto-and cross-power spectra with measured ones are discussed.