Electrical generators of offshore wind or tidal current turbines are exposed to harsh marine and operating conditions. Predictive maintenance is therefore a key issue for the competitiveness of these energy generation systems. Generally speaking, the predictive maintenance is based on the monitoring of a Diagnostic Indicator (DI): the interpretation of its value or drift is used for the optimal planning of the corrective maintenance. In this work, we present briefly our new online monitoring technique of electrical machine winding insulation. This model-based approach consists in monitoring the drift of a DI built from the in-situ estimation of high-frequency electrical model parameters. The involved model structures are derived from the RLC network modeling of the winding insulation, with more or less lumped parameters. In the second part of the work, we investigate the effects of temperature changes on the estimated parameters of diagnostic models. A 1.5 kW low power wound stator is exposed to different temperature levels, from 30°C to 160°C, and for each temperature a series of experimental acquisitions is realized. Identification results show that resistance and inductance of a simple HF model structure are almost independent of temperature changes, while insulation capacitance increases with temperature increases: at 160°C it is 8% higher than its initial value at room temperature.