As digitalization progresses, the processing of growing volumes of data presents manufacturing companies with major challenges. Machine learning, a sub-domain of artificial intelligence, can be used to generate valuable knowledge from data. When applied in industry, this process – termed “industrial data science” – may translate into a major competitive advantage in the future.
Under the heading of “Industry 4.0”, digitalization of industrial production has progressed by leaps and bounds in recent years. In conjunction with the growing networking of production processes, this gives rise to increasingly large data volumes. These contain valuable knowledge that can be leveraged for example for process optimization. Conventional methods are however no longer sufficient to extract this knowledge.
In order for the potential for innovation nevertheless to be exploited, innovative technologies of artificial intelligence (AI) are brought to bear. AI exploits mathematical and computational methods in order to develop solutions to specific application problems, for example for the optimization of production, quality management or automation. The interdisciplinary technology of machine learning is of particular importance in these areas of application. Serving as an interface between information technology, mathematics and statistics, it uses algorithms in order to “learn” from data and to generate generically valid information from them.
Industrial Data Science
The application of machine learning in industrial production is termed industrial data science (IDS). IDS combines computational, mathematical and statistical methods with domain-specific expertise in production (cf. Bauer, N.; Stankiewicz, L.; Jastrow, M.; Horn, D.; Teubner, J.; Kersting, K.; Deuse, J.; Weihs, C.: Industrial Data Science. Developing a Qualification Concept for Machine Learning in Industrial Production: European Conference on Data Analysis (ECDA) 2018). This interdisciplinary approach paves the way for innovative solutions to existing problems and the support of decision-making processes. The objectives of IDS applications may be descriptive (“What has happened?”), explanatory (“Why has it happened?”), predictive (“What will happen?”) or prescriptive (“What is to be done?”). The complexity of the application increases with each of these steps; at the same time however, the information content also rises.
Factors for success
Three factors for success have emerged for the performance of practical IDS projects:
Interdisciplinary composition of project teams: Data scientists possess sound expertise in data processing and management and in machine learning. Domain experts have a far-reaching technical understanding and are intimately familiar with their products and processes. Both groups are absolutely crucial to the success of projects; they do not, however, always speak the same language. The citizen data scientist, an expert with interdisciplinary training, can serve as a mediator to overcome these barriers to communication.
Structured approaches for projects: the CRISP-DM model (Cross Industry Standard Process Model for Data Mining) has become the established standard for this purpose in numerous industrial projects (cf. Chapman, P.; Clinton, J.; Kerber, R.; Khabaz, T.; Reinartz, T.; Shearer, C.; Wirth, R.: CRISP-DM 1.0. Step-by-step data mining guide, SPSS Inc. (2000)). CRISP-DM divides projects into six phases, which are completed in succession and if necessary iteratively: (1) Business Understanding (understanding the business and process), (2) Data Understanding (understanding and interpreting data), (3) Data Preparation (evaluating the data quality and preparing the data), (4) Modelling (selecting algorithms and generating models), (5) Evaluation (of models) and (6) Deployment (documenting the results and implementing solutions).
Data maturity: This should always be evaluated with reference to the project objective. As the level of automation and autonomy of the desired solution rise, so also do the requirements for data capture and quality. Pilot studies and analyses of the potential can generally be achieved with a lower data maturity, i.e. with greater manual effort. The data maturity should not therefore be seen as an obstacle; rather, its improvement should be regarded as potential for future projects.
In numerous product, process and system-related applications, artificial intelligence offers huge potential to optimize important production parameters. Opportunities for its application are however not limited to production, but exist across all sectors. Artificial intelligence can also optimise business processes and thus become an advantage for a company. The aim must therefore be above all to identify and exploit the possibilities for intelligent use of AI.
University Professor Dr.-Ing. Jochen Deuse
M. Sc. Jacqueline Schmitt