Residue　is　one　of　the　main　factors　affecting　the　reliability　of　sealed　electronic　equipment.　The　traditional　particle　collision　noise　detection　(PIND)　method　is　only　suitable　for　the　detection　of　the　presence　or　absence　of　residues,　but　the　particle　size　identification　of　the　residue　is　important　for　tracking　the　source　of　production　and　improving　the　process.　In　this　paper,　the　endpoint　detection　algorithm　based　on　spectrum　variance　is　used　to　extract　the　residual　pulse　signal,　which　is　beneficial　to　the　concentration　of　characteristic　quantities.　Fisher＇s　discriminant　method　was　used　to　conduct　dimensionality　reduction　processing　and　clustering　analysis　of　characteristic　parameters,　and　BP　neural　network　was　used　to　realize　the　classification　of　residual　particle　size,　and　the　recognition　accuracy　could　reach　93.75%.