In　lower-limb　rehabilitation　equipment,　the　prediction　of　the　knee　joint　moment　using　surface　electromyography　signals　is　an　important　method　of　motion　intention　recognition.　To　improve　the　viability　of　control　by　human-computer　interactions　and　to　reduce　the　complexity　of　the　knee　joint　moment　prediction　model,　this　paper　presents　a　prediction　model　for　knee　joint　moment　based　on　artificial　neural　networks,　in　which　the　knee　joint　angle,　the　knee　joint　angular　velocity,　and　a　pair　of　surface　electromyography　signals　from　the　antagonistic　and　agonistic　muscles　of　the　knee　joint　are　selected　as　inputs.　Two　public　databases　that　include　the　walking　data　of　hemiplegic　patients　and　healthy　people　are　used　to　test　the　effect　of　muscle　pair　selection　on　knee　joint　moment　prediction　under　non-isometric　contraction.　The　dependence　of　the　model　on　speed　and　the　individual　is　also　tested.　The　correlation　coefficient　and　the　mean　absolute　error　are　used　as　performance　indicators.　The　results　demonstrate　that　the　proposed　model　can　predict　the　knee　joint　moment　well.　Across　the　difference　of　speeds　and　subjects,　the　choice　of　muscle　pair　has　no　significant　effect　on　the　prediction　of　the　knee　joint　moment.　Compared　with　previous　research,　the　proposed　model　simplifies　the　measurement　parameters　and　the　signal　processing　process,　reducing　the　number　of　sensors　used　in　practical　applications,　which　increases　the　safety　and　the　fluency　of　the　lower-limb　movement.