The　accurate　detection　and　identification　of　collision　states　in　industrial　robot　environments　is　a　critically　important　and　challenging　task.　Deep　learning-based　methods　have　been　widely　applied　to　collision　detection;　however,　these　methods　primarily　rely　on　dynamic　models　and　dynamic　threshold　settings,　which　are　subject　to　modeling　errors　and　threshold　adjustment　latency.　To　address　this　issue,　we　propose　MomentumNet-CD,　a　novel　collision　detection　method　for　industrial　robots　that　leverages　backpropagation　(BP)　neural　networks.　MomentumNet-CD　extracts　collision　state　features　through　a　momentum　observer　and　constructs　an　observation　model　using　Mahalanobis　distance.　These　features　are　then　processed　by　an　optimized　three-layer　BP　neural　network　for　accurate　collision　identification.　The　network　is　trained　using　a　modified　Levenberg-Marquardt　algorithm　by　introducing　regularization　terms　and　continuous　probability　outputs.　Furthermore,　we　developed　a　comprehensive　acquisition　system　based　on　the　Q8-USB　data　acquisition　card　and　the　QUARC　2.7　real-time　control　environment.　The　system　integrates　key　hardware　components　including　a　MR-J2S-70A　servo　driver,　ATI　six-dimensional　force/torque　(F/T)　sensor,　and　ISO-U2-P1-F8　isolation　transmitter,　and　the　corresponding　software　module　is　developed　through　MATLAB/Simulink　R2022b,　which　achieves　the　high-frequency　real-time　acquisition　of　critical　robot　joint　states.　The　experimental　results　show　that　the　MomentumNet-CD　method　achieves　an　overall　accuracy　of　93.65%　under　five　different　speed　conditions,　and　the　detection　delay　is　only　12.16　ms.　Compared　with　the　existing　methods,　the　method　shows　obvious　advantages　in　terms　of　the　accuracy　and　response　speed　of　collision　detection.