On-line tool wear estimation in turning through sensor data fusion and neural networks

This work investigates a method which uses sensor data fusion and neural networks for on-line flank wear estimation. This method, unlike most of the existing ones, predicts the gradually increasing flank wear on a cutting tool continually.; The proposed method involves four important steps: sensor data acquisition, sensor data preprocessing, sensor data representation, and flank wear estimation from sensor data using a neural network. In this approach proper sensor data representation is crucial for obtaining accurate flank wear estimates.; During cutting, force, vibration, and acoustic emission signals are monitored. These sensor signals are digitized and preprocessed using a set of bandpass filters to improve the signal to noise ratio. Force and vibration signals are represented by univariate or multivariate autoregressive moving average models. Acoustic emission signals are represented by discrete wavelet transforms. The patterns created from the coefficients of either autoregressive moving average models or discrete wavelet transforms are input to a specially designed neural network architecture to compute flank wear estimates. This neural network is designed by combining Kohonen's feature maps, radial basis functions, and recurrent neural networks. This network satisfies two important design considerations: it reduces demand on supervised training time and data, and it remains fault tolerant to fluctuations in metal cutting.; The performance of the flank wear estimation method is studied by conducting a set of turning experiments on AISI 6150 steel. The results indicate that the proposed estimation method provides accurate flank wear estimates for the range of operating conditions that were used during neural network training data generation. These results also indicate that the development of crater wear on a cutting tool adversely affects the accuracy of flank wear estimates. The root mean square of estimation errors is in the range of 0.0010 to 0.0017 inches. This makes the proposed flank wear estimation method very attractive for real-world applications.; This work stands unique in representing force and vibration signals by autoregressive moving average models, representing acoustic emission signals by discrete wavelet transforms, developing a novel neural network architecture, and testing the proposed flank wear estimation method with extensive experimentation.