In the present paper we propose a new method, the Penalized Adaptive Method (PAM), for a data driven detection of structure changes in sparse linear models. The method is able to allocate the longest homogeneous intervals over the data sample and simultaneously choose the most proper variables...