The　mapping　relationship　between　the　mill　load　and　the　multi-component　mechanical　signals　generated　by　the　ball　mill　of　the　mineral　grinding　process　is　non-deterministic　and　complex.　With　the　inherent　filtering　function　of　the　human　ear,　the　operating　expert　can　effectively　estimate　the　mill　load　and　its　internal　parameters　for　their　familiar　mill　in　the　actual　industrial　process.　In　order　to　obtain　multiple　single-mode　sub-signals　with　physical　meaning　and　complementary　characteristic,　this　paper　proposes　a　single-mode　sub-signal　selection　method　based　on　variational　modal　decomposition　(VMD)　and　predictive　performance.　At　first,　based　on　prior　knowledge,　the　value　of　decomposition　layers　required　to　perform　VMD　is　determined.　Then,　VMD　is　used　to　decompose　the　original　mechanical　signal　into　multiple　time-domain　single-mode　sub-signals　with　different　bandwidths　and　time　scales,　and　further　are　transformed　to　the　frequency　domain　to　obtain　candidate　single-mode　sub-signal　frequency　spectrum.　Finally,　based　on　these　candidate　spectral　data,　a　serial　of　candidate　sub-models　for　mill　load　parameter　prediction　are　constructed,　and　a　series　of　selective　ensemble　models　are　built　for　obtaining　reduced　single-mode　sub-signal　frequency　spectrum,　The　final　single-mode　sub-signals　with　the　biggest　complementary　characteristics　are　selected　based　on　the　practical　requirement.　The　effectiveness　of　the　method　is　demonstrated　by　comparative　experiment　simulations　based　on　the　shell　vibration　signal　of　a　laboratory-scale　ball　mill.