Università degli studi di Modena e Reggio Emilia

Multi-criteria inventory classification groups inventory items into classes, each of which is managed by a specific re-order policy according to its priority. However, the tasks of inventory classification and control are not carried out jointly if the classification criteria and the classification approach are not robustly established from an inventory-cost perspective. Exhaustive simulations at the single item level of the inventory system would directly solve this issue by searching for the best re-order policy per item, thus achieving the subsequent optimal classification without resorting to any multi-criteria classification method. However, this would be very time-consuming in real settings, where a large number of items need to be managed simultaneously. In this article, a reduction in simulation effort is achieved by extracting from the population of items a sample on which to perform an exhaustive search of best re-order policies per item; the lowest cost classification of in-sample items is, therefore, achieved. Then, in line with the increasing need for ICT tools in the production management of Industry 4.0 systems, supervised classifiers from the machine learning research field (i.e. support vector machines with a Gaussian kernel and deep neural networks) are trained on these in-sample items to learn to classify the out-of-sample items solely based on the values they show on the features (i.e. classification criteria). The inventory system adopted here is suitable for intermittent demands, but it may also suit non-intermittent demands, thus providing great flexibility. The experimental analysis of two large datasets showed an excellent accuracy, which suggests that machine learning classifiers could be implemented in advanced inventory classification systems.

Inventory control is one of the main activities in industrial plant management. Both process owners and line workers interact daily with stocks of components and finite products, and an effective management of these inventory levels is a key factor in an efficient manufacturing process. In this paper the algorithms k-means and Ward's method are used to cluster items into homogenous groups to be managed with uniform inventory control policies. This unsupervised step reduces the need for computationally expensive inventory system control simulations. The performance of this methodology was found to be significant but was strongly impacted by the intermediate feature transformation processes.

The failure prediction of components plays an increasingly important role in manufacturing. In this context, new models are proposed to better face this problem, and, among them, artificial neural networks are emerging as effective. A first approach to these networks can be complex, but in this paper, we will show that even simple networks can approximate the cumulative failure distribution well. The neural network approach results are often better than those based on the most useful probability distribution in reliability, the Weibull. In this paper, the performances of multilayer feedforward basic networks with different network configurations are tested, changing different parameters (e.g., the number of nodes, the learning rate, and the momentum). We used a set of different failure data of components taken from the real world, and we analyzed the accuracy of the approximation of the different neural networks compared with the least squares method based on the Weibull distribution. The results show that the networks can satisfactorily approximate the cumulative failure distribution, very often better than the least squares method, particularly in cases with a small number of available failure times.

The balancing of assembly lines is one of the most studied industrial problems, both in academic and practical fields. The workable application of the solutions passes through a reliable simplification of the real-world assembly line systems. Time and space assembly line balancing problems consider a realistic versions of the assembly lines, involving the optimization of the entire line cycle time, the number of stations to install, and the area of these stations. Components, necessary to complete the assembly tasks, have different picking times depending on the area where they are allocated. The implementation in the real world of a line balanced disregarding the distribution of the tasks which use unwieldy components can result unfeasible. The aim of this paper is to present a method which balances the line in terms of time and space, hence optimizes the allocation of the components using an evolutionary approach. In particular, a method which combines the bin packing problem with a genetic algorithm and a genetic programming is presented. The proposed method can be able to find different solutions to the line balancing problem and then evolve they in order to optimize the allocation of the components in certain areas in the workstation.