Data acquisition and data analytics for prognostics and health management in machine tools

Abstract

The transformation of the manufacturing industry under Industry 4.0 shifts decision-making and optimization towards data-driven analytics. In the manufacturing industry, data-driven condition monitoring and predictive solutions for equipment, tools, and processes are becoming increasingly important, as they have proven useful in other industries such as automotive, aerospace, or marine. However, their application and penetration in machine tools and most manufacturing environments are still relatively low, due to various factors: besides the required know-how, which many machine manufacturers do not possess, diverse machine and control types, software revisions, and the lack of standardized protocols and semantics lead to a general lack of interoperability of data acquisition and analysis approaches. To meet these requirements, a versatile approach to increasing availability, performance, and quality of machine tools has been investigated, developed, implemented, and validated, taking into account technology- and industry-specific limitations and constraints. The solution consists of a part related to data acquisition structures and a part related to the analysis of the large variety of machine tool and component types. The data acquisition addresses both the use of standardized interfaces (OPC UA) to provide high-frequency measurement data, as well as the retrofitting of existing equipment or components without a control unit with data acquisition units. These acquisition units are designed to meet the data quality requirements for condition monitoring and diagnosis purposes. The data analysis part describes the processing of measurement series, depending on the operating characteristics of the components (constant, controlled-constant, and variable). For this purpose, two demand-driven analysis approaches are described in detail, which allow condition diagnoses using artificial intelligence depending on the operating characteristics of a component or a process: For processes or components with (at least partially) stationary operating conditions, an approach based on supervised machine learning has been developed. This essentially consists of an artificial neural network, which is trained on the healthy or normal state under variation of environmental influences. Subsequently, the output of the neural network is compared with the actually measured values. Deviations between the model output and the actual values provide an indication of a change in the system, which is not due to environmental influences but to wear, for example. For components or processes with variable operating conditions, an approach based on unsupervised machine learning has been developed. For comparable segments, such as quasi- stationary operating conditions or test cycles, one or more sensor signals are recorded. These time series recordings are mapped to a discrete space for comparability using so-called features. Selected features span a multi-dimensional space in which a measurement can be defined as a point defined by its feature values. The position in space thus reflects the state of the process or component in this abstract feature space. Measurements at the same states accumulate into dense areas, while changes in the state due to wear or defects move significantly away from them. By analyzing the resulting clusters, not only deviations from an original state can be detected, but also clusters of known faults or wear states can be learned, which allow the identification of a previously known type of fault or wear in addition to an obvious deviation. Both approaches are based on data-sparse measurement series collection and low-effort model creation. In addition, both approaches cover the possible operating conditions of processes and components (stationary, partially stationary, and variable), making it a holistic principle for monitoring and analyzing components.