The smoke emission from diesel engines is now regarded as a very significant problem because of greater public consciousness of environment damage, and in some countries legislation is now in force limiting the particulate generation. This thesis describes work on the feedback control of diesel smoke, in which a new smoke sensor was used to make real-time smoke detection feasible during engine transients. Experimental results demonstrated that this control system significantly reduced the smoke 'puff' normally generated by a sudden power demand. An advanced identification method is presented, which models the engine based on experiments. This method is dependent on an understanding of the sampling action of the fuel injector on a diesel engine and utilizes a synchronization technique for acquiring the information of the engine dynamics. Parameter estimation was used to derive the models of speed-to-fuel rack and smoke-to-fuel rack. The nonlinearity of the model of smoke-to-fuel rack was investigated in a preliminary way. Two kinds of controllers, non-adaptive controllers and self-tuning controllers, were developed and implemented on a real engine. The non-adaptive controllers, designed with a combination of PI control algorithm and optimisation of the engine performance, were of relatively low complexity and had the ability to work effectively at light-load engine conditions. The self-tuning controllers associated with a smoke predictor were applied to solve the problems of the system nonlinearity and enable the control system to work in a wider range of engine operations. The problem associated with self-tuning controllers were demonstrated but the possibility of using such controllers was proven.