In　this　paper,　a　novel　convolutional　neural　network　named　FilterNet　is　proposed　to　learn　feature　representation　directly　from　raw　magnetic　Barkhausen　noise　(MBN)　for　quantitatively　evaluating　plastic　deformation.　In　the　proposed　model,　three　modified　filter-type　convolutional　kernels　are　adopted　to　extract　appropriate　feature　expressions　from　raw　MBN.　In　the　experimental　part,　a　special　experimental　device,　which　could　change　the　angle　between　the　electromagnetic　sensor　and　specimens,　was　designed　to　simulate　the　case　that　the　performance　of　a　trained　model　declined　due　to　the　change　in　the　angle.　MBN　signals　at　11　angles　from　left　to　right　were　measured.　The　performances　of　multiple　linear　regression　(MLR),　BP　neural　network　(BP),　and　the　FilterNet　were　analyzed　with　mean　absolute　error　(MAE)　and　root-mean-square　error　(RMSE).　MAE　and　RMSE　of　the　FilterNet　were　smaller　than　the　other　two　models　at　different　angles　and　reveal　the　outperformance　of　the　FilterNet.