Heavy　key　mechanical　devices　relate　to　production　quality　and　quantity　of　complex　industrial　process　directly.　It　is　necessary　to　estimate　some　difficulty-to-measure　process　parameters　inside　these　devices.　Multi-component　and　non-stationary　mechanical　signals,　such　as　vibration　and　acoustic　ones,　are　always　employed　to　model　these　process　parameters　indirectly.　How　to　effective　extract　and　select　interesting　information　from　these　signals　is　the　key　step　to　build　effective　soft　sensor　model.　In　this　paper,　a　new　kernel　latent　features　adaptive　extraction　and　selection　method　is　proposed.　Ensemble　empirical　mode　decomposition　(EEMD)　is　used　to　decompose　these　mechanical　signals　into　multiple　time　scales　sub　signals　with　different　physical　interpretations.　These　sub-signals　are　transformed　to　frequency　spectra,　and　then　kernel　partial　least　squares　(KPLS)　algorithm　is　used　to　extract　their　kernel　features.　Integrated　with　mutual　information　(MI)-based　feature　selection　method,　a　new　define　index　is　exploited　to　select　the　important　sub-signals　and　their　latent　features　adaptively.　The　shell　vibration　and　acoustic　signals　of　an　experimental　laboratory-scale　ball　mill　in　the　mineral　grinding　process　are　used　to　validate　the　proposed　approach.　(C)　2016　Elsevier　B.V.　All　rights　reserved.