As structural health problems are becoming more and more important, a neural networks model is introduced to detect structural damage. The structural modal flexibility matrix can be accurately constructed by the natural frequency and modal information. All elements of changes in the modal flexibility matrix are looked on as inputs of the networks. Damage locations and extents are both considered with different outputs in the present study. A simply supported truss structure is studied with different damage cases. To localize damage, one case is chosen as location input/target pairs to train the present BP network model. But to identify damage extent, two cases are chosen as extent pairs to train. Although modals of BP neural networks with different outputs are presented for different damage detecting schemes, it is more difficult to ascertain damage extent than location. The results indicate that the present BP neural network modal can effectively detect damage of structures with changes in the flexibility matrix between the intact and the damaged cases.