Diagnostic Information Fusion for Manufacturing Processes Kai Goebel & Vivek Badami GE Corporate R&D Niskayuna, NY 12309 goebelk@crd.ge.com; badami@crd.ge.com Amitha Perera Rensselear Polytechnic Institute Troy, NY 12380 perera@cs.rpi.edu Abstract - This paper addresses diagnostic information fusion for situations where several diagnostic tools are used to estimate a single system state. These estimates will always disagree to some extent and it is the task of the fusion module to provide an estimate which is more reliable than the best of the diagnostic tools. To that end, a fusion process was developed which performs a weighted average of individual tools using confidence values assigned dynamically to the individual diagnostic tools. These confidence values are derived from validation curves which are designed using individual a priori tool information and which are centered about the previous system estimate. In a further step, the fusion output is smoothed leading to additional performance improvement. In experiments, data were gathered from a high speed milling machine and fed through several developed diagnostic tools. Key words: fusion, information fusion, diagnosis, soft computing, fuzzy fusion. Introduction and Background The need of manufacturers to produce inexpensive quality products has resulted in increasing demand for unattended and/or automated manufacturing systems. One problem in automating machining is how to deal with common malfunctions and disturbances such as tool wear, chatter, and tool breakage. Tool wear is a highly non-linear process which is hard to monitor and estimate. To avoid costly damage due to tool wear or breakage, manufacturers use conservative operating procedures to prevent these malfunctions [1]. However, these result in less efficient and more costly production. A number of diagnostic techniques attempt to deal with theses problems, including neural networks [2], clustering algorithms Burke [3], Kohonen’s Feature Map [4], fuzzy logic [5], and influence diagrams [6]. To achieve further performance improvement, hybrid systems were proposed to overcome shortcomings of individual systems, such as fuzzy-neural systems [7]. Hybrid use of above mentioned techniques and other soft computing principles for diagnostics and prognostics are given in Bonissone and Goebel [8]. In a similar spirit, fusion techniques combine different methods to overcome shortcomings of individual tools. This paper proposes one fusion method based on fuzzy validation gates. Diagnostic Fusion via Validation Gates The method developed is a two-level system consisting of a number of diagnostic classification systems on the first level and a managerial fusion unit on the second. The data are fed into each of the first level units, and their output is combined in the second level to produce a single, better solution (Fig. 1). NN1 NN2 RM NNBR COMBINE INPUT CLASSIFICATION Fig. 1: The system architecture To address some of the problems outlined above, we propose the fusion of diagnostic estimates via fuzzy validation curves called Fuzzy Diagnostic Validation and Fusion (FUDVAF). This technique is related to the FUSVAF (Fuzzy Sensor Validation and Fusion) algorithm developed for sensor fusion [10, 11, 12]. The fusion algorithm uses confidence values obtained for each diagnostic output from validation curves and performs a weighted average fusion. With increasing distance from the predicted value, readings are discounted through a non-linear validation function. The predicted value in the FUDVAF algorithm is obtained through application of an exponential weighted moving average time series predictor The confidence value which is assigned to a particular diagnostic output depends on the specific Proceedings of the 2nd International Conference on Information Fusion, Fusion ’99, vol. 1, pp. 331-336, 1999