In this work, a comprehensive study on the different soft computing based models using different strategies for indirect monitoring of drill flank wear has been presented. Two new soft computing techniques, which were not applied earlier to condition monitoring, have been used for the first time for prediction of drill wear. One such technique is the normalized radial basis function network, and the other is a hybrid technique, called fuzzy radial basis function network. The standard back propagation multilayer neural network and the standard radial basis function network have also been considered for the purpose of comparison. Experimental set-ups have been developed to carry out several drilling experiments on the mild steel workpieces using high speed steel drill bits with different combination of drill diameter, spindle speed and feed-rate. During the experiments, signals have been acquired from most commonly available pool of sensors (e.g. dynamometer, current sensor, accelerometer, acoustic emission sensor, and rotational speed sensor). Time-domain statistical parameter estimation, fast Fourier transform and wavelet packet transform techniques have been applied on these acquired signals for extracting wear sensitive features. The selection of final features, which were to be used in different strategies, was done through experimental sensitivity analysis. Different strategies, i.e., different combinations of wear sensitive extracted features from sensor signals and the process parameters, have been formulated. Comparative evaluations of all the soft computing models with respect to drill wear prediction performance for various strategies have been reported. The results show that the dynamometer is the most effective sensor in drill wear prediction, followed by current sensor, accelerometer and acoustic emission sensor. No improvement of the prediction performance has been observed for sensor fusion over the performance of a single effective sensor. Integrating noisy signal features from different sensors actually deteriorates the prediction performances of all the artificial neural network models when a single sensor is already effective. The three best combinations of the strategies and the soft computing techniques that produced minimum predictive error have been selected and recommended for online implementation.