Hybridize Machine Learning Methods and Optimization Techniques to Analyze and Repair Welding Defects via Digital Twin of Jidoka Simulator

Lean thinking is interested in identifying and resisting defects that affect business safety, like welding defects of the cooling pipe exposing the chilled foodstuffs parcels to spoilage, posing a danger to the land transportation investment. Four clusters are used to identify welding flaws’ severity to keep them from getting worse via mapping their finite element meshes (MFEM) and using the high-frequency current ring (HFCR) lean technique to identify their size, depth, growth direction ( $\boldsymbol {{R}_{s}}$ ), and shape deployment via prediction to assign the flaws as Susceptible (S), and preludes to transfer to a Quarantine cluster (Q) to take one decision whether rescued (R), Isolated (I) or Eliminated (E). We are creating a “digital Jidoka twin system” (SQ(R/I/E)) with a controller segment programmed with machine learning (ML) algorithms that use the MFEM’s huge and uneven data to sort defects and their causes. Hybridising the Random-Forest algorithm with Dingo optimisation and called Regulated Random Forest (RRF) to precisely identify defect clusters and then predict the welding defect growth rate ( $\boldsymbol {{R}_{s}}$ ) using the Cat-boost optimiser, which is enhanced by a beetle search mechanism called CatBAS. The RRF is superior to Apriori, ECLAT, and FP Growth by 23.98%, 7.44%, and 8.38%, respectively, while CatBAS is superior to XG-boost by 94.62% in response time with a 1.04175% error that activates the treating stage-V quickly. The SQ(R/I/E) increased parcel rescues by 38.2% and reduced financial losses. Protecting chilled foodstuffs transport from spoilage serves the SDG (2).