BMW Group

Quality is not negotiable – especially with premium products such as automobiles. To achieve 100 per cent quality, products are meticulously inspected, errors noted and reworked. Nowadays, analysis is aimed at uncovering procedural and conceptual insufficiencies. Up to now, static reports from numerous key performance indicators (KPIs) have been at the forefront of daily quality work in plants.

The trend towards individualization and derivatization increases the complexity of production systems. Electrification and hybridization bring completely new production processes with them. The mixed operation mode in automobile factories – the integration of different derivatives of all drive forms on one line – leads to a permanent ramp-up situation.

Lean production has been established as a dogma for quality work in production. Lean processes focus on the consistent reduction of waste ('Muda') and aim for maximum robustness. 'Poka Yoke' plays a vital role as a design principle for fail-safe processes and systems. Continuous improvement pursues the permanent optimization of production processes regarding quality and efficiency.

So far, quality work has taken place based on key figures, which are used to evaluate the production processes and the market. These key performance indicators are designed in such a way that they reflect the performance of production over several production stages. This approach, however, is in clear contrast to the philosophy of decentralized defect prevention and elimination. Besides that, it is becoming increasingly difficult to use the aggregated KPIs for targeted defect reduction. At the BMW Group, all quality characteristics are recorded (semi- ) automatically or manually in a database, depending on process suitability. This data is often analyzed in spreadsheet programs and made available in the form of static diagrams or key figures (also digital) on the shop floor as part of visual management.

An innovative approach is to look at errors not only in aggregated form but also directly on the shop floor in order to gain immediate insights from data analysis on-site. The aim is always to prioritize according to Pareto logic – those topics are considered first that show maximum criticality according to the combined frequency of occurrence and cumulative effort. Therefore, it must first be determined how the term criticality can be operationalized for the respective application case. For quality management in automobile production, for example, the duration of a rework activity can represent an essential measure. Ideally, the total quality costs generated are recorded to derive the relevant fields of action as precisely as possible.

With the help of Visual Analytics applications, the production and quality specialists are enabled to carry out their data analysis by interacting with this data in self-service. Because the process specialist carries out the analysis directly at the process, two essential goals have been achieved: immediate location of currently occurring errors (depending on the timeliness of the data acquisition or provision) and direct plausibility checking of data records in the database. This includes, on the one hand, the recording of pseudo errors which are generated by risk-averse checks 'as a precaution'. On the other hand, production errors may occur that are not correctly reflected in the quality database. Both phenomena induce additional inspection costs and torpedo the effective application of data analytics in production.

The aim is to consciously shift all potential complexity of data visualization into the software system so that it empowers and supports the user. The user is deliberately a production specialist without IT expertise. The focus is on software applications for visualization that are understandable without training and provide a high degree of robustness. In concrete terms, this means that an appropriate IT system must be able to handle mass data and ensure minimum response times.

An agile approach is indispensable for the development of the corresponding visualization or dashboarding. Only if the user on the shop floor assumes the role of a product owner a Visual Analytics solution guarantees effective self-service analytics. A central need for action – poor data quality – arises particularly in the early phase of such a development project. This can only be improved step by step and continuously.

In addition to the identification of data errors within the visualization, Artificial Intelligence applications play a central role. Deep learning-based object recognition is used to drastically reduce pseudo errors compared to conventional image recognition. Also, machine learning methods are used to provide evidence of anomalies in data. Based on decision trees, these forecasts serve, among other things, to suggest suitable defect classifications to the production specialist during the recording. This not only significantly improves the quality of data input, but also greatly shortens the input time for the production specialist.

To put in a nutshell,

• Visual Analytics can be seen as the enabler for the transformation towards data-driven process optimization.

• Lean first – digitization brings added value to production, especially when lean production is already a reality. In this case, for example, dashboarding can accelerate optimization.

Self-service as a competitive advantage: We understand innovative digitization solutions as a further offer in the tool stack for employees. Digitization and analytics will find their way into production when systems are user-friendly and efficient to operate.

• Artificial Intelligence can make a significant contribution to the construction of visual analytics – especially in the detection of anomalies and potential data defects in data sets.