Accurate　detect　mill　load　inside　the　heavy　rotating　ball　mill　is　one　of　the　key　factors　for　realizing　operation　optimization　and　control　of　the　mineral　grinding　process.　By　using　multi-source　mechanical　signals,　such　as　vibration　and　acoustical　signals　at　different　location　of　the　ball　mill　system,　to　construct　mill　load　parameter　forecasting　(MLPF)　model　has　been　a　hot　and　focus　recently.　However,　a　few　researches　about　the　relative　accuracy　and　reliability　contribution　ratio　of　these　different　multi-source　signals　are　addressed.　It　is　necessary　to　select　the　suitable　mechanical　signal　at　different　industrial　applications.　Aim　to　this　problem,　multi-source　mechanical　signal　analysis　based　on　linear　latent　structure　MLPF　model　is　studied　in　this　paper.　At　first,　Fast　Fourier　transformer　(FFT)　is　used　to　obtain　mechanical　frequency　spectrum　of　the　time　domain　vibration　and　acoustic　signals　with　characteristics　of　high　dimension　and　strong　co-linearity.　Then,　project　to　latent　structure　(PLS)　models　are　build　based　on　these　spectra　data.　Finally,　a　new　defined　metric　is　used　to　estimate　relative　accuracy　and　reliability　contribution　ratio　of　these　multi-source　signals　based　on　generalization　performance　and　structure　parameter　of　different　prediction　model.　Experiments　based　on　a　laboratory-scale　ball　mill　are　used　to　valid　the　proposed　method.　©　2018　IEEE.