AI in Manufacturing: Separating Fact from Fiction – A Look at Modern Optimization Techniques

From Linear Programming, Heuristics, Simulations, to Machine Learning – This Article Explores the Evolution of Production Optimization and AI in Manufacturing.

In today’s production planning and control, efficiently utilizing resources and minimizing costs are crucial to a company’s success. AI in manufacturing, along with other algorithms, plays an increasingly significant role. Advanced Planning & Scheduling (APS) systems employ a variety of mathematical methods to provide robust solutions for complex scheduling tasks. These methods complement each other to form a comprehensive optimization strategy.

Understanding the Problem: Not All Mathematical Optimization is AI in Manufacturing

APS systems are powerful tools for optimizing manufacturing processes. They enable detailed planning of production orders by considering various factors such as delivery dates, resource availability, and production constraints. The necessary data is sourced from leading ERP and/or MES systems, which the APS software uses to create an optimized production plan.

Detailed planning is particularly challenging due to the need to balance conflicting objectives. For instance, achieving a short throughput time or high on-time delivery rate often conflicts with the goal of minimizing capital tied up. APS systems like GANTTPLAN model these conflicting objectives as cost functions and seek an optimal solution that minimizes total costs. Different priorities are set to meet the unique requirements of each company.

Many tools in manufacturing and production planning, now categorized under AI in manufacturing, are actually long-established mathematical optimization methods. Both dialogue-oriented, step-by-step plan improvements using heuristics and algorithms, and mathematical solver-based optimization methods have been in use for a long time. It is only with the structured data collection from APS and peripheral systems like MES that AI applications for pattern recognition in manufacturing become feasible. These applications include predicting deviating process durations or forecasting more realistic delivery dates.

Before the Advent of AI in Manufacturing: The Historical Development of Optimization Tools

Linear programming (LP) is one of the oldest and most fundamental methods of mathematical optimization. Developed in the 1940s, it quickly found application in production planning due to its ease of implementation and solution. LP is particularly well-suited for problems with linear relationships, such as determining the optimal production amount while considering limited resources like material and labor time. Shortly after LP was developed, Integer Programming (IP) was introduced to model discrete decisions. This method extends LP by limiting solutions to integer values, which is especially useful for decisions like the number of units to produce or the selection of specific machines.

In the 1950s and 1960s, these methods were further refined. Mixed-Integer Linear Programming (MIP) combines continuous and discrete variables, providing a more flexible and powerful solution for complex problems. This method was increasingly used in production planning to optimize both continuous manufacturing quantities and discrete decisions such as machine utilization. Non-Linear Programming (NLP) was also developed during this period to solve problems with non-linear relationships. This method extends the capabilities of LP and MIP by allowing more complex and realistic models of manufacturing processes. A real-world example is the optimization of production volume in the pharmaceutical industry, where production costs are a non-linear function of the production volume.

In the 1970s and 1980s, Heuristics and Metaheuristics became popular for providing quick solutions to very large and complex problems. The Nearest Neighbor Method, a simple and intuitive heuristic, is widely used in production planning. For example, when planning the production sequence, the next order with the shortest changeover time for the current production is always selected to reduce total changeover time and increase production line efficiency. While this procedure quickly provides an initial solution, it can lead to suboptimal results in complex scenarios. Metaheuristics such as Genetic Algorithms offer an advantage by striving for ever better solutions through an evolutionary search. They mimic natural evolutionary processes where the best solutions prevail. In production planning, genetic algorithms can be used to optimize the sequence of orders to minimize total processing time while maximizing machine utilization.

Application of Optimization Tools in Production Planning

In the 1990s, the focus shifted to accounting for uncertainties and variability. Stochastic Programming was developed to provide robust solutions under uncertainty by modeling different scenarios. This method is particularly important in tactical production planning, where demand and other factors may still be uncertain. Robust Optimization also became popular during this time, offering solutions that work well under a range of conditions. This method reduces the risk of poor decisions and is useful in machine-intensive manufacturing environments such as the automotive, electronics, and mechanical engineering industries. The goal is to remain robust in the face of disruptions or failures of production machines or delivery delays.

In the 2000s, Constraint Programming and Simulations became increasingly used in production planning. Constraint programming is particularly powerful when a large number of constraints must be considered, such as machine availability and maintenance times. Simulations, especially Discrete Event-oriented Simulation (DES), provide detailed insights into system behavior and enable the investigation of different scenarios.

Machine Learning Algorithms are one of the latest developments in production optimization. They are particularly useful for identifying patterns in large amounts of data and making predictions. AI in manufacturing leverages these algorithms to enhance production planning and optimization, making processes smarter and more efficient.

AI in Manufacturing: Challenges of Data-Driven Production Optimization

Detailed planning based on heuristic models reaches its limits because it does not sufficiently consider the dynamic nature of manufacturing processes. Static assumptions about process times and resource availability can lead to an increasing discrepancy between planning and reality. Unforeseen events such as machine failures or material bottlenecks often necessitate adaptive replanning. Furthermore, dynamic influencing factors, such as employee learning effects or changing room temperatures, are often not taken into account.

A data-driven approach is a good way to overcome these challenges. By using Machine Learning (ML), valuable insights can be gained from historical production data, enabling more precise forecasts for future processes. ML models are continuously updated with new data to improve their accuracy and relevance.

Specifically, this means that relevant production data is collected, processed, and modeled using suitable algorithms (such as time series analysis, regression, or classification) that can predict the future development of process times, resource requirements, and other relevant parameters. These models can then be integrated into detailed planning to enable dynamic adaptation of production plans to changing conditions.

The advantages of such a data-driven approach are clear: higher accuracy of forecasts, greater flexibility in planning, proactive identification of problems, and an overall better basis for decision-making in production planning.

However, there are also specific challenges to be overcome, such as ensuring high data quality, selecting suitable algorithms, and accounting for the necessary computing power.

Concrete Use Cases: More Realistic Planning and Better Delivery Date Forecasts with the Help of Machine Learning (ML)

Machine Learning (ML), a subfield of artificial intelligence, is a highly topical research area. It describes the ability to extract valuable information and patterns from historical raw data. Thanks to increasing computing power and the networking of numerous intelligent sensors in manufacturing, sufficient data is available for analysis. However, ML-optimized production planning is only within reach when structured data from all planning-relevant systems is collected and available.

The possible applications of ML in production are many and varied, ranging from more precise estimates of target and process times, better delivery date forecasts, and make-or-buy decisions to quality control applications. Another important application is root cause analysis (RCA). This method is used to identify and eliminate the main causes of problems in manufacturing processes, such as machine downtime, quality defects, or inefficient operations. A clear example of RCA can be seen in a case where the machine speed always dropped at certain times. This always occurred when deliveries entered the production hall through open roller shutters. The cold draught caused the machine speed to drop immediately. The objective of RCA is therefore not only to identify the symptoms but also to find and eliminate the underlying causes to prevent future problems.

More Precise Determination of Target and Process Times

The precise determination of target and process times is a central and often critical aspect of production planning. Traditionally, these times are estimated by experienced foremen, which on average leads to useful results. However, the range of variation of the potential parameters in the planning process is not always adequately reflected because the triggers and their complexity cannot always be identified.

Another problem arises when experienced employees leave and no equally qualified colleagues can be found due to a shortage of skilled workers. Furthermore, estimates made by people can often be biased and subjective.

The use of machine learning (ML) aims to estimate standard and process times more realistically, taking into account a wide range of characteristics and influencing factors. This leads to significantly improved planning quality. ML can correct errors in the planning model that could otherwise only be remedied by regular, time-consuming, and resource-intensive replanning.

Better Delivery Date Forecasts

Improving the quality of delivery date forecasts is closely linked to determining more precise target and process times, which is more important today than ever before. Fluctuations in demand, unexpected disruptions in manufacturing, or supply bottlenecks can cause significant problems in the supply chain. Machine learning (ML) offers an innovative solution to significantly increase the accuracy of delivery date forecasts.

In this process, both future capacities are taken into account and historical production data is used to train the ML algorithms. This data includes information about past orders, machine runtimes, supplier failures, and external factors such as seasonal fluctuations. Based on this data, the ML model learns to recognize complex patterns and relationships that are often difficult for human analysts to identify.

This enables more realistic predictions of throughput times, probabilities of delays, dynamic adjustment of production plans, and the detection of anomalies. With the help of ML algorithms, deviations from normal patterns in the production data can be identified, such as unusually long processing times or increased scrap rates. These anomalies can indicate potential problems that can be resolved using root cause analysis (RCA).

Make-or-Buy Decision Support

As outlined above, Machine Learning (ML) can significantly enhance make-or-buy decisions in the production environment by enabling data-driven analysis and predictions. The process begins with collecting historical data on various aspects of production, including costs, supplier performance, and quality metrics. This data is then prepared and formatted appropriately.

After that, ML algorithms are trained with this data to identify patterns and correlations relevant to decision-making. Once trained, the ML models analyze the data and provide predictions about the costs, risks, and benefits of in-house versus outsourced production. These predictions help drive informed decision-making. Companies can simulate different scenarios based on these predictions to make optimal decisions that minimize both costs and risks. This data-driven approach enables companies to make informed and efficient make-or-buy decisions, optimizing their production processes and increasing their competitiveness.

AI in Manufacturing: DUALIS Collaborates with Partners on Cutting-Edge Production Optimization Research

The REPLAKI (REalistic Planning with AI) project, which was launched in January 2023 and is funded by the Federal Ministry for Economic Affairs and Climate Protection, is investigating the challenges of ‘batch size 1’ in the context of volatile and often insufficiently digitised value chains in the automotive sector. In addition to DUALIS, ten other prominent partners from research and industry, including Galfa GmbH & Co. KG, have joined forces in this initiative.

The aim of the project is to analyse interdependencies in historical process data using artificial intelligence (AI) and machine learning (ML) and to use them to increase the predictive accuracy of production plans. The focus is on more realistic forecasting of delivery dates and process durations as well as data-supported planning of new parts.