Integrating neural networks with influence diagrams for multiple sensor diagnostic systems

Abstract

The objective of this research is to develop an adaptive architecture for the fusion of sensory information for diagnostic reasoning. The system architecture envisioned uses influence diagrams to provide a symbolic representation of the manufacturing or process system model. The independence encoded in influence diagrams is utilized in defining the structure of the neural networks to reduce the system complexity. Neural networks are then used as a learning mechanism for estimating and updating conditional probabilities from a set of training data. To this end, we have designed two types of probabilistic neural networks: probabilistic Hamming nets (PHNs) for discrete variables and self-organizing probabilistic neural networks (SOPNNs) for continuous variables. A PHN is basically a modified Hamming net which implements pattern matching in the first layer and estimates conditional probabilities in the second layer. This model uses a Bayesian learning rule to update the weights of networks. A SOPNN is a neural network with a fixed-size self-organizing net for clustering and a modified Specht's model for estimating the probability density functions (PDFs). The idea is to represent similar training patterns in the same cluster by its centroid and then use these centroids for PDF estimation. Several sets of simulated and real sensor data were employed to test the effectiveness of the proposed models. The results were analyzed and compared with those obtained by the conjunctoid, Specht's model and the polynomial Adaline (Padaline) neural models as well as Bayesian learning rules. The effectiveness of the integrated network for reducing the system complexity both in representation and in inference is also demonstrated. Based on the results of this study, our proposed architecture has the following advantages: (1) The structure of the neural networks is well defined and less complex. (2) Our neural models (PHNs and SOPNNs) are of fixed-size. (3) Both models are very general, not limited to pre-specified parametric or unimodal distributions. (4) The SOPNN is self-organized. (5) The learning is one-pass and incremental. (6) Training and performance are very fast. (7) Partial data are better utilized. (8) Subjective information can be easily incorporated. (9) The integrated network can be flexibly reconfigured to respond to on-line requests. Since our models do not employ the time-consuming backpropagation and simulated annealing algorithms, this integrated network appears more promising for real-time applications.