The　features　traditionally　extracted　from　hysteresis　loops　are　highly　sensitive　to　both　the　variation　of　case　depth　and　uncontrollable　factors　in　repeated　testing　cycles,　thus　increasing　the　difficulty　in　predicting　the　tensile　force　applied　on　surface-hardened　steel　rods　with　different　case　depths.　In　this　study,　in　order　to　eliminate　the　influence　of　such　high　sensitivity,　a　case　depth-insensitive　feature　(CDIF)　was　proposed　to　characterize　the　tensile　force,　and　a　particle　swarm　optimization　(PSO)-optimized　neural　network　was　used　to　establish　the　correlation　between　the　CDIF　and　tensile　force　in　order　to　predict　the　tensile　force　applied　on　steel　rods　with　different　case　depths.　Five　classical　features　(including　remanent　magnetic　induction　intensities,　coercive　force,　hysteresis　loss,　maximum　magnetic　induction,　and　distortion　factor)　and　the　CDIF　were　successfully　used　to　characterize　the　tensile　force.　Then,　the　linear　regression　model　and　PSO-optimized　neural　network　model　were　used　in　turn　to　establish　the　relationship　between　each　feature　and　tensile　force　to　predict　the　tensile　force　applied　on　steel　rods　with　different　case　depths.　The　CDIF　was　insensitive　to　the　variation　of　case　depth　and　linearly　correlated　with　the　tensile　force.　Even　though　the　CDIF　is　affected　by　the　unknown　and　uncontrollable　factors　in　repeated　testing　cycles,　the　PSO-optimized　neural　network　model　based　on　it　can　be　used　to　accurately　predict　the　tensile　force　applied　on　steel　rods　with　different　case　depths　with　a　prediction　error　of　0.67%.