Condition monitoring signals provide sufficient information about the health of machines and, therefore, are widely used for fault diagnosis, prognosis, and health management. Existing approaches generally extract one or more degradation features from original signals collected in a time interval and predict the remaining useful life of machinery based on a selected or fused feature under a pre-specified threshold. However, using a single feature is often inadequate in terms of the accuracy of fault prognosis due to the interval-based extraction procedure. To overcome the shortcoming, a new prognosis framework is proposed for machinery based on original condition monitoring signals in this paper. Technically, the Box-Cox transformation is first performed on the original signals point by point to construct a series of degradation features without losing information. Then, the neural basis expansion analysis for time series (N-BEATS) that has robust performance in time-series prediction is utilized to predict the future evolution for each feature with high volatility. By leveraging the similarity of multiple Box-Cox features, a parameter-based transfer learning method is proposed to reduce the computation complexity. Finally, we reconstruct the future original signals of machinery through the inverse Box-Cox transformation. Since the reconstructed original signals are noisy, a new failure criterion suitable for decision-level fault prognosis is defined from an industrial application perspective. An application on high speed train wheels and two follow-up simulations are used to illustrate the performance of our proposed framework in machinery fault prognosis.