Introduction of an Artificial Intelligence Support System for Production Planning in Manufacturing

The development of intelligent manufacturing methods for production planning is driven by the economic objective to decrease the necessary time-to-volume of the production and thus, to achieve a faster market entry of new products. The phase of production planning can be accelerated by implementing AI support systems which enable an evaluation of early planning data. Most useful is the application of AI support systems during the rough planning of production systems within the Digital Factory paradigm since changes can be implemented cost-efficiently during this phase. Knowledge-based systems with embedded ontologies are hereby the foundation for the development of AI solutions in order to enrich data-driven approaches with context and semantic information.

As a result from the review of the body of literature, it is shown that the use cases of process and resource planning utilizing AI models are two under-researched fields in production planning. This research gap is the result of limited data availability, high costs of data pre-processing, and limited interpretability of results. Rule-based or mathematical optimization models are therefore commonly applied in practice but both approaches require high manual efforts and expert knowledge during development. ... mehrThe research question of this thesis therefore focuses on the development of an AI model in the field of process and resource planning which does not rely on manual expert knowledge and is also highly interpretable.

In order to provide a data-driven AI model for the identified use cases, modified Hopfield neural networks for anomaly detection and classification tasks are introduced. The model is based on the main principles of using a database with only correct instances of planning data to circumvent labeling, the usage of ontological information in order to derive context and similarities between planned and existing production systems, and a high interpretability of the results. The core idea of modified Hopfield neural networks is the modeling of planning data as a graph network. The layout of modified Hopfield neural networks uses the activation of the edges instead of vertices in contrast to conventional neural networks. In order to predict the class membership of objects, the model utilizes an energy definition adapted from the Lenz-Ising model. A high energy is the result of uncommon features of an object and thus, indicates anomalous properties. The weights of the edges are adjusted during the training procedure using a Widrow-Hoff learning rule. As a result, the network layout enables anomaly detection and classification tasks and additionally a recommendation procedure to correct planning data, a similarity measure between classes of the observed production systems, and a highly interpretable decision process.

The model is applied in three use cases from the fields of process planning in manufacturing, synthetic data set generation in production planning and simulation of machinery sequences, and, as an excursus, Leukemia diagnosis using gene expressions. Modified Hopfield neural networks show a high precision across all use cases. In particular, the model outperforms Autoencoders, Isolation Forest, and One-class Support Vector Machines in anomaly detection tasks. In addition, the interpretability and build-in recommendation feature show the practicability of the model which a conventional "black box" model cannot provide.

The application of support systems utilizing a modified Hopfield neural network in the early fault detection during the planning of production systems is therefore recommended. Future research is necessary to prove the modified Hopfield neural network model's applicability across multiple domains and to target a fully autonomous production planning based, among others, on the concepts developed in this thesis.