Manufacturing companies are among the biggest beneficiaries of digital transformation. Access to ever-increasing amounts of data offers numerous opportunities to increase efficiency, improve quality, and reduce costs. Artificial intelligence (AI) solutions are already being used successfully in various areas of the company to achieve these goals and continue to be intensively researched. An important area of application for AI solutions is production planning and control (PPC). The Hanoverian Supply Chain Model (German: Hannoveraner Lieferkettenmodell, HaLiMo) structures the tasks and processes of PPC according to their temporal and logical sequence (see Figure 1). Within HaLiMo, different use cases of AI solutions can be located.

Three such use cases are described below. This description covers delivery date forecasting in the context of order management, material demand forecasting in the context of secondary requirements planning, and control production using intelligent agents.

Since the "next day delivery" offers of Amazon and Co., short and reliable delivery times are a matter of course. Being able to offer and maintain short delivery times is also of great importance for manufacturing companies. Order management in manufacturing companies has two important functions. On the one hand, delivery dates are negotiated and a fixed date is promised as part of the order clarification process at the start of order processing. On the other hand, order coordination tracks the progress of customer orders and communicates a delay in the event of unrecoverable disruptions. To support these two tasks, AI can be used to predict delivery dates.

Traditional methods for determining delivery dates are usually based on simple statistical methods that take into account only a few influencing factors, as well as expert knowledge. With the help of AI methods, it is possible to use the master and transactional data available in companies through the use of modern ERP and MES systems (e.g., confirmation data, routings, warehouse movements) and use it to forecast delivery dates. Large amounts of data can be searched in a very short time, patterns and trends can be detected and delivery dates can be calculated automatically. The AI algorithms used are usually far superior to statistical methods. However, it should be noted that the GIGO (Garbage In - Garbage Out) principle applies here as well. The algorithms can only process what is given to them. If the master data is poorly maintained or the feedback data is incomplete and of low granularity, the output will be correspondingly poor. It should also be noted that AI does not always determine the correct delivery date, but rather the date that is most likely according to the underlying models. However, since an AI model cannot account for all environmental conditions such as absenteeism or unusual delivery bottlenecks, it is necessary to verify the delivery date before communicating it to the customer.

In secondary requirements planning, the dependent requirements are determined on the basis of the production program planning. Dependent requirements are raw materials as well as parts and assemblies for the production of finished products. After calculating the gross and net requirements, taking into account the planned and actual stocks, the procurement type is assigned. The result of this allocation is a proposal for external procurement and/or in-house production.

Demand can be determined using a variety of methods. These methods include deterministic demand estimation, heuristic methods, and stochastic methods. Deterministic forecasting is well suited for stable and predictable production environments where there is little uncertainty. Heuristic methods, based on experience and estimation, are useful when data is limited or quick decisions are required. Stochastic techniques use statistical models and probability distributions to forecast demand. Forecasts are based on historical data and mathematical models that estimate future demand based on probabilities and trends. Stochastic techniques are useful when demand is unpredictable or highly volatile.

A major challenge in the application of stochastic methods is the modelling of stochastic relationships. In reality, stochastic relationships can be very complex. Several variables and factors may be involved, interacting and influencing each other. Consequently, modelling such complex relationships requires a lot of effort. The application of AI methods can address this challenge. A key advantage is that AI methods can analyse large amounts of data and identify patterns, trends, and relationships that may not be obvious to human analysts. Through advanced machine learning algorithms, AI can extract relevant features (characteristics) from the data and incorporate them into modelling. This helps create more accurate models, improving the accuracy of demand forecasting and increasing the efficiency of PPC.

Production control is responsible for dispatching production orders and controlling them during the production process. The primary objective is to complete the processing of production orders according to the existing production plan, taking into account any changes due to adjustments. To achieve this goal, the orders must be released according to a uniform logic and placed in a meaningful sequence. In this context, the integration of environmental factors represents a major challenge for companies since they must be taken into account in addition to the traditional economic and production logistics factors in order to survive in the market in the long term.

AI provides a variety of methods that have the potential to address the complexity described above. Established production control methods are often based on expert knowledge, heuristics, or operations research models. An advantage of novel AI methods lies in the simultaneous consideration of a large number of features with extensive data sets. In addition to the GIGO principle explained above, it should be noted that feedback data and especially the number of data measurement points have a significant impact on the accuracy of the results of AI analyses. In the course of Industry 4.0 and the associated technological advances, such as radio frequency identification (RFID), an essential basis for cost-effective data collection and, thus for AI applications has been created.

The use of reinforcement learning is one of these innovative approaches to deal with the complexity of the production control process. It is a machine learning technique in which an intelligent agent learns to make sequential decisions through interactions. With the help of Q-learning or policy gradients, the agent can be trained for holistic self-production control, taking into account existing and new constraints (economic, logistical, environmental).

Initial analysis shows that such novel approaches can meet requirements better than established industry standards and thus improve the logistical performance of production control.

Tobias Hiller, Jonas Schneider, and Tabea Marie Demke of Institute of Production Systems and Logistics (IFA), Leibniz University Hannover

With the help of AI methods, it is possible to use the master and transactional data available in companies through the use of modern ERP and MES systems and use it to forecast delivery dates. 

Since an AI model cannot account for all environmental conditions such as absenteeism or unusual delivery bottlenecks, it is necessary to verify the delivery date before communicating it to the customer.

The use of reinforcement learning is one of the innovative approaches to deal with the complexity of the production control process. It is a ML technique in which an intelligent agent learns to make sequential decisions through interactions.