The development of an automated tool condition monitoring system has been the subject of research for quite some time. Today, the need for such a system is greater than it once was because of the shortage of skilled workers, higher machining speeds, increase in precision machining, and the need to lower downtime. Most of the proposed tool condition monitoring systems suffer from problems such as sensitivity to process noise, difficulty in on-line implementation, inability to accommodate variations in cutting conditions, unsuitability for shop floor implementation, and difficulty in the interpretation of results.

This thesis describes a novel approach for tool condition monitoring in turning. A neural network-based system with hierarchical architecture including static and dynamic neural networks is used for this purpose. The wear estimation is carried out by measuring the cutting force components. The changes in cutting force components are related to the progress of tool wear.

The overall system acts as a non-linear state observer with tool wear components as the states, the cutting conditions as input, and cutting force components as the outputs. The static neural networks represent the output equations while dynamic neural networks emulate the state equations. Each wear component is monitored by a subsystem comprising a static and a dynamic neural network. These subsystems share information about the tool wear component they are monitoring and their error in estimating the cutting force components is used for on-line training of the dynamic neural networks.

The on-line training and the adaptability properties of neural networks ensure that changes in machining parameters can be accommodated. Neural networks are known to be noise tolerant and can therefore work with noisy signals originating from the cutting process. To further reduce the uncertainty in the wear predictions, cutting force ratios are used as the output of the static neural networks. This helps to reduce the effects of variations in the turning process.

The proposed system has been tested using a variety of cutting conditions both in simulation and experimental setups. The results have been satisfactory and the system has been able to provide reasonable tool wear predictions albeit with some variation in accuracy among wear components.