Automated rotor welding processes using neural networks

#The present invention relates to automated welding systems and, in particular, relates to arc welding systems for rebuilding turbine rotors using neural network computer technology.

#The possibilities of applying logic to industrial processes to enhance their efficiency has long been recognized, even prior to the advent of electronic logic circuits and computers. One form of control that has advanced along with computer technology is pattern recognition. For example, U.S. Pat. No. 4,589,140--Bishop discloses a system that uses a video camera to compare magnified images of circuit boards. Similarly, a video surface inspection system disclosed in U.S. Pat. No. 4,675,730--Adomaitis et al. compares surface patterns to reference images and determines if an inspected surface falls within pre-determined minimum or maximum "grey level" values for the image.

#Pattern recognition concepts are also useful when the "pattern" is not a video image of an object, but is instead a signal related to a characteristic. Thus, the sensed vibrational patterns of a machine tool can be analyzed using a pattern recognition system to determine tool breakage or other abnormal conditions as disclosed by U.S. Pat. No. 4,918,427--Thomas et al.; U.S. Pat. No. 4,853,860--Thomas; and U.S. Pat. No. 4,724,524--Thomas et al.

#FIG. 1 is a partial diagrammatic illustration of a prior art rotor welding process.

#FIG. 2 is a partially diagrammatic, partially schematic illustration of the welding process of the present invention.

#FIG. 3 is a schematic of the training/retraining aspect of the welding process of the present invention.

#The present invention applies neural network technology to enhance the above-described automated rotor welding process or other arc welding processes. The present invention enables automated monitoring of the weld arc, and uses the pattern recognition capabilities of the neural network computer's adaptive information processing architecture to automatically differentiate between normal and abnormal weld arcs. Referring now to FIG. 2, a partially diagrammatic, partially schematic representation of the present invention is shown. The weld arc 10 is preferably monitored through a UV/intensity reduction optical filter 20 by a video camera 30. A video monitor 40 may also be provided in certain embodiments, and is particularly useful during the early training and retraining stages described below. Monitored video images are digitized to be processed by an image preprocessor 45, which reduces the voluminous image data to the important features discussed below with reference to FIG. 4. In the image preprocessor 45, the digital intensities in the various image pixels are compared. The intensity comparison is based upon the pattern of lines of constant intensity as discussed below with reference to FIG. 4. This is preferably accomplished either in a fast digital computer or in a parallel processing array.

#Reduced image data include the location of maximum intensity and the changes in that location over time. This can be represented as an average location, present location, and difference between the average and the present locations. Since the location of the weld needs to be indexed periodically, the average location at a previous point in time is an important parameter. As seen in FIG. 4, much of the information about the arc 10 is contained in its size, direction, and symmetry. This information is developed by determining the locus of points where the intensity falls below a certain fraction of the maximum intensity. Thus, as seen in FIG. 4 lines of constant intensity d.sub.1A -d.sub.1B are found in a typical arc. The number of fractions depends upon the desired quality of differentiation between normal and abnormal arcs. Factors of ten will give a coarse differentiation. Factors below two require considerable extra computing and are seldom valuable. The most reasonable starting point is factors of the square root of ten (3.16). Computing the distance from the point of maximum intensity to the various lines of constant intensity provides a spectrum of distance versus the number of occurrences, which represents the size of the arc 10. The direction of the arc 10 can be determined by the average vector from one end to the other of a line of constant intensity. Since there will be a line on each side of the arc 10, the average of these two lines estimates the centerline. This averaging is then preferably performed over all lines of constant intensity, as shown by the dotted line in FIG. 4.

#The length of the arc 10 will be determined by the filler wire or electrode 14 and the pool of molten metal 11. These are both at a lower temperature than the arc 10 and represent an abrupt change in temperature. This can be recognized as two lines of constant intensity being very close together. Symmetry is determined by the difference in distance from the centerline to the pairs of lines of constant intensity, from side to side and end to end. Thus, differences between the distance d.sub.1a and d.sub.1b indicate the side-to-side asymmetry and differences between distances d.sub.1A -d.sub.1C indicate the end-to-end asymmetry. These differences are scaled to the arc size and organized as a spectrum of side-to-side difference versus the number of occurrences, and as a spectrum of end-to-end difference versus the number of occurrences. Changes in symmetry are the most relevant data, so the spectra are examined to determine the difference between the present spectrum and the average over the last minute or other period of time. As noted above, the size of the pool of metal 11 is also an important parameter and is determined as the radius of a line of constant intensity. This intensity is selected dynamically as the lowest of the bunched intensity lines used above to determine the end of the arc.