A　ball　mill　is　a　heavy　mechanical　device　and　its　safe　operation　affects　the　entire　grinding　process.　Mill　load　is　a　key　index　in　the　optimum　operation　of　the　grinding　process,　but　it　cannot　be　measured　directly.　In　industrial　practice,　operational　experts　normally　estimate　its　value　based　on　their　experiences　and　the　mechanical　signals　produced　by　the　ball　mill.　In　this　paper,　we　proposed　a　heterogeneous　selective　ensemble　method　by　using　a　multiscale　mechanical　frequency　spectrum.　The　multicomponent　adaptive　decomposition　algorithm　is　first　used　to　decompose　the　original　shell　vibration　and　acoustic　signals　into　sub-signals　with　different　timescales.　Then,　selective　ensemble　(SEN)　kernel　projection　to　latent　structure　algorithm　is　used　to　model　the　spectral　data　of　these　sub-signals.　Furthermore,　the　latent　features　of　multiscale　spectra　are　extracted　to　construct　SEN　models　based　on　fuzzy　inference.　Finally,　the　two　types　of　heterogeneous　SEN　models　are　fused　by　using　information　entropy.　The　main　contribution　of　this　study　is　that　the　proposed　soft-sensing　model　has　a　dual-layer　ensemble　structure　that　can　fuse　multi-source　information　in　different　mechanical　sub-signals　with　physical　meaning.　Moreover,　the　proposed　model　can　simulate　the　fuzzy　cognitive　behavior　of　domain　experts　in　the　mineral　grinding　process.　The　effectiveness　of　the　method　is　verified　by　the　shell　vibration　and　acoustical　data　of　a　laboratory-scale　ball　mill.