A working wind turbine generates a large amount of multivariate time-series data, which contain abundant operation state information and can predict impending anomalies. The anomaly detection of the wind turbine nacelle that houses all of the generating components in a turbine have been challenging due to its inherent complexities, systematic oscillations and noise. To address these problems, this paper proposes an unsupervised time-series anomaly detection approach, which combines deep learning with multi-parameter relative variability detection. A normal behavior model (NBM) of nacelle vibration is firstly built upon training normal historical data of the supervisory control and data acquisition (SCADA) system in the high-resolution domain. To better capture the temporal characteristics and frequency information of vibration signals, the vibration spectrum vector is integrated with the multivariate time-series data as inputs and the spectrum-embedded temporal convolutional network (SETCN) is then used to extract latent features. The anomalies are detected through a multi-variate coefficient of variation (MCV) based anomaly assessment index (AAI) of relative variability among vibration residuals and environment parameters of the nacelle. The approach considers the time-series characteristics of input data and preserves the spatio-temporal correlation between variables. Validations using data collected from real-world wind farms demonstrate the effectiveness of the proposed approach.