Diagnosis of Plasma Equipment using Neural Network and Impedance Match Monitoring

A new methodology is presented to diagnose faults in equipment plasma. This is accomplished by using neural networks as a pattern recognizer of radio frequency (rf) impedance match data. Using a match monitor system, the match data were collected. The monitor system consisted mainly of a multifunction board and a signal flow diagram coded by Visual Designer. Plasma anomaly was effectively represented by electrical match positions. Twenty sets of fault-symptom patterns were experimentally simulated with variations in process factors, which include rf source power, pressure, Ar, and $O_$2 flow rates. As an input to neural networks, two means and standard deviations of positions were used as well as a reflected power. Diagnostic accuracy was measured as a function of training factors, which include the number of hidden neurons, the magnitude of initial weights, and two gradients of neuron activation functions. The accuracy was the most sensitive to the number of hidden neurons. Interaction effects on the accuracy were also examined by performing a 2$^$4 full factorial experiment. The experiments were performed on multipole inductively coupled plasma equipment.