Condition monitoring and prognosis is a key issue in ensuring stable and reliable operation of mechanical transmissions. Wear in a mechanical transmission, which leads to the production of wear particles followed by severe wear, is a slow degradation process that can be monitored by spectral analysis of oil, but the actual degree of degradation is often difficult to evaluate in practical applications due to the complexity of multiple oil spectra. To solve this problem, a health index extraction methodology is proposed to better characterize the degree of degradation compared to relying solely on spectral oil data, which leads to an accurate estimation of the failure time when the transmission no longer fulfils its function. The health index is extracted using a weighted average method with selection of degradation data with allocation steps for weight coefficients that lead to a reasonable mechanical transmission degradation model. First, the degradation data used as input are selected based on source entropy which can describe the information volume contained in each set of spectral oil data. Then, the weight coefficient of each set of degradation data is modelled by measuring the relative scale of the permutation entropy from the selected degradation data. Finally, the selected degradation data are fused, and the health index is extracted. The proposed methodology was verified using a case study involving a degradation dataset of multispectral oil data sampled from several power-shift steering transmissions.