We introduce the usage of multivariate orthogonal space transformations and vectorized time-series models in combination with data-driven system identification models to achieve an enhanced performance of residual-based fault detection in condition monitoring systems equipped with multi-sensor networks. Neither time-consuming annotated samples nor fault patterns/models need to be available, as our approach is solely based on on-line recorded data streams. The system identification step acts as a fusion operation by searching for relations and dependencies between sensor channels measuring the state of system variables. We therefore apply three different vectorized time-series variants: (i) non-linear finite impulse response models (NFIR) relying only on the lagged input variables, (ii) non-linear output error models (NOE), also including the lags of the own predictions and (iii) non-linear Box?Jenkins models (NBJ) which include the lags of the predictions errors as well. The use of multivariate orthogonal space transformations allows to produce more compact and accurate models due to an integrated dimensionality (noise) reduction step. Fault detection is conducted based on finding anomalies (untypical occurrences) in the temporal residual signal in incremental manner. Our experimental results achieved on four real-world condition monitoring scenarios employing multi-sensor network systems demonstrate that the Receiver Operating Characteristic (ROC) curves are improved over those ones achieved with native static models (w/o lags, w/o transformations) by about 20?30%.