Dynamic Bayesian Network-Based Approach by Integrating Sensor Deployment for Machining Process Monitoring

Many condition monitoring systems based on artificial intelligence process models for machining process monitoring have been developed intensively. However, given that machining processes are very complex (i.e., nonlinear and nonstationary), there is still no clear methodology to acquire machining monitoring systems allowing machining processes to be optimized, predicted, or controlled. In this paper, the coupled hidden Markov model, based on dynamic Bayesian networks, is proposed to monitor a machining process by using multi-directional data fusion and to analyze the effect of the sensor layout on the monitoring accuracy. The features extracted by a singular spectrum and wavelet analysis constitute the input information to the system. The technique is tested and validated successfully by using two scenarios: tool wear condition monitoring (initial wear, gradual wear, or accelerated wear) for the milling process and surface roughness accuracy grade prediction (accuracy grade 9, accuracy grade 8, or accuracy grade 7) for the turning process. In the first case, the maximum recognition rate obtained by the single-sensor placement for tool wear is 83%, whereas in the case of the three-sensor placement, the model recognition rate is 89%. In the second application for turning, the maximum recognition rate obtained by the single-sensor and the double-sensor placements for surface roughness accuracy prediction is 77% and 85%, respectively. In the case of the three-sensor placement, the model recognition rate is 89%. The proposed approach can also be integrated into the diagnosis architecture for condition monitoring in other complex machining systems.