Wavelet neural network (WNN) is used widely in tool wear detection, but the curse of dimensionality and shortage in the responding speed and learning ability is brought about by the traditional models. An improved WNN algorithm which combined with modified particle swarm optimization (PSO) was presented to overcome the problems. Based on the cutting power signal, the method has been used to estimate the tool wear. The Daubechies-wavelet was used to decompose the signals into approximation and details. The energy and square-error of the signals in the detail levels was utilized as characters which indicated tool wear, the characters were input to the trained WNN to estimate the tool wear. Compared with BP neural network, conventional WNN and genetic algorithm-based WNN, a simpler structure and faster converge WNN was obtained by the new algorithm, and the accuracy for estimate tool wear has been tested by simulation and experiments.