The deformable structure will be deformed by the force during the drilling process， and the state of the tool cannot be judged according to the applied feed distance. During the cutting process， the tool is cut into different positions of the deformable structure to generate vibrations with different amplitudes. The vibration signal collected by the acceleration sensor is processed and analyzed by fast Fourier transform（FFT）， and the drilling process can be divided into five states. The variation coefficients of all harmonic components are calculated， and then the amplitude of some harmonic components is selected as the feature for condition monitoring. Combination of linear transformation and logarithmic transformation is used to achieve the normalization of the input data. The normalized results are imported into the support vector machine and the back propagation（BP） neural network for drilling state classification. Experiments show that the classification accuracy of the support vector machine is above 85% in all drilling states， and the classification accuracy of some drilling states reaches 100%. The classification accuracy of the BP neural network classifier is slightly lower than that of the support vector machine. According to the identified state， the tool is stopped when it is close to reaching the other side of the deformable structure， and the remaining thickness of the workpiece is measured to verify the proposed method.