Fuzzy ART/RRR-RSS: a two-phase neural network algorithm for part-machine grouping in cellular manufacturing

In this paper an efficient methodology adopting Fuzzy ART neural network is presented to solve the comprehensive part-machine grouping (PMG) problem in cellular manufacturing (CM). Our Fuzzy ART/RRR-RSS (Fuzzy ART/ReaRRangement-ReaSSignment) algorithm can effectively handle the real-world manufacturing factors such as the operation sequences with multiple visits to the same machine, production volumes of parts, and multiple copies of machines. Our approach is based on the non-binary production data-based part-machine incidence matrix (PMIM) where the operation sequences with multiple visits to the same machine, production volumes of parts, and multiple identical machines are incorporated simultaneously. A new measure to evaluate the goodness of the non-binary block diagonal solution is proposed and compared with conventional performance measures. The comparison result shows that our performance measure has more powerful discriminating capability than conventional ones. The Fuzzy ART/RRR-RSS algorithm adopts two phase approach to find the proper block diagonal solution in which all the parts and machines are assigned to their most preferred part families and machine cells for minimisation of inter-cell part moves and maximisation of within-cell machine utilisation. Phase 1 (clustering phase) attempts to find part families and machines cells quickly with Fuzzy ART neural network algorithm which is implemented with an ancillary procedure to enhance the block diagonal solution by rearranging the order of input presentation. Phase 2 (reassignment phase) seeks to find the best proper block diagonal solution by reassigning exceptional parts and machines and duplicating multiple identical machines to cells with the purpose of minimising inter-cell part moves and maximising within-cell machine utilisation. To show the robustness and recoverability of the Fuzzy ART/RRR-RSS algorithm to large-size data sets, a modified procedure of replicated clustering which starts with the near-best solution and rigorous qualifications on the number of cells and duplicated machines has been developed. Experimental results from the modified replicated clustering show that the proposed Fuzzy ART/RRR-RSS algorithm has robustness and recoverability to large-size ill-structured data sets by producing highly independent block diagonal solution close to the near-best one.