KI für Zeitreihen: Sequenzielle Daten

From the series: "The Evolution of Artificial Intelligence for Sequential Data"

What are sequential data?

Sequential data are information organized in a specific order, where this order is essential for understanding and analyzing the data. This applies, among other things, to natural language in the form of written or spoken text, but particularly also to time series, which are generated, for example, by sensors.

What are time series and where do they occur?

Time series play a particularly important role in technical contexts. They consist of measured values that are captured, for example, by sensors. Each measured value is assigned a unique timestamp. This not only gives the data an order, but also implicitly provides information about the time intervals between measured values. An example of a simple time series, where a measured quantity is recorded over the duration of a manufacturing process, might look like this:

The measured values could reflect, for example, the process force of a machine tool, the pressure of a press stamp, or the electrical power consumption of a system. From the figure, we can identify several characteristic observations:

• Sampling interval: The data were recorded at regular intervals. Depending on the technology used for measurement and storage of the measured values, time series can also have irregular intervals.
• Value range: Positive and negative values occur in single-digit magnitudes. In practice, the physical unit is also relevant.
• Periodicity: The values are not constant but show a sequence of peaks and troughs that differ in intensity. The time duration of the "peaks" and "troughs" is also not uniform.
• Trend: Based on the short time period presented, no long-term trend can be observed. In practice, trends play a major role in forecasting from long-term data and in comparing samples far apart.
• Noise: Additionally, the data are "noisy" due to random processes or dependent on other influencing factors, so no clear pattern emerges.

Why is the analysis of time series important?

The identified characteristics are typical examples of challenges that arise in the analysis of time series, particularly when AI is used for this purpose. It is not a trivial task, but rather a field where new methods are still being intensively researched. In doing so, the following questions and objectives are typically addressed:

• Pattern recognition: Robust identification of recurring sequences, for example for long-term monitoring or anomaly detection
• Assessment / Classification: Quality statements and error assignment for recommendation systems or for root cause analysis
• Forecasting: Extrapolation of existing data into the future, for example for early detection of critical states or other special events

Suitable AI models for time series can help, for example, reduce scrap, plan necessary maintenance in advance, or optimize energy supply coverage from local generation facilities. By contributing to fast and precise decision-making, they have the potential to increase operational efficiency in many areas.

In the next posts in this series, we will take a closer look at various analytical and AI models for sequential data.