This paper describes models and methods for estimating the acoustic reflectivity of the welded interfaces between spot-welded sheets from normal-incidence pulse-echo ultrasound signals. The simple geometry of the problem allows an abstraction that does not resort to complex wave equations. Instead, a reflectivity model predicts the timing and amplitude of the echoes arriving at the probe. This reflectivity model is nested in a signal processing model; recovering reflectivity first requires deconvolution to recover discrete impulses from the probe signal, then processing these with the reflectivity model. Reflectivity maps of spot welds generated with this model show promise for predicting weld quality.

Information processing in sensor networks, with many small processors, demands a theory of computation that allows the minimization of processing effort, and the distribution of this effort throughout the network. Graphical model theory provides a probabilistic theory of computation that explicitly addresses complexity and decentralization for optimizing network computation. The junction tree algorithm, for decentralized inference on graphical probability models, can be instantiated in a variety of applications useful for wireless sensor networks, including: sensor validation and fusion; data compression and channel coding; expert systems, with decentralized data structures, and efficient local queries; pattern classification, and machine learning. Graphical models for these applications are sketched, and a model of dynamic sensor validation and fusion is presented in more depth, to illustrate the junction tree algorithm.

In a highly collaborative research and development project with mining and university partners, sensors and data-analysis tools are being developed for rock-mass characterization and real-time measurement of mineral content. Determining mineralogy prior to mucking in an open-pit mine is important for routing the material to the appropriate processing stream. A possible alternative to lab assay of dust and cuttings obtained from drill holes is continuous on-line sampling and real-time x-ray fluorescence (XRF) spectroscopy. Results presented demonstrate that statistical analyses combined with XRF data can be employed to identify minerals and, possibly, different rock types. The objective is to create a detailed three-dimensional mineralogical map in real time that would improve downstream process efficiency.