A time series is a sequence of data points, or observations, that are indexed and arranged in chronological order. Typically, these data are collected at successive, equally spaced intervals, such as hourly, daily, monthly, or yearly measurements. The primary purpose of constructing and analyzing a time series is to identify non-random patterns in a target variable. This process enables data-driven forecasting and decision-making, allowing organizations to anticipate events and optimize resource allocation.
In the manufacturing sector, for example, within the cosmetics industry during an ingredient blending stage, time series analysis can be used to monitor input and output weights, process duration, and mixing homogeneity. Time series are often decomposed into fundamental components to uncover underlying patterns, facilitate interpretation, and enhance analytical accuracy. The traditional approach to time series analysis involves decomposition into the following components:
- Trend: represents the long-term movement within the time series, indicating an overall upward, downward, or stable direction;
- Seasonality: exhibits a fixed and known period; some time series may contain multiple seasonal components;
- Cycle: fluctuations of increases and decreases that occur without a fixed or known frequency, which distinguishes cycles from seasonality due to their irregularity;
- Residual (Irregularity/Noise): represents the “noise” or unpredictable component of the time series that cannot be explained by trend, seasonality, or cycles;
- Stationarity: a series is considered stationary when its statistical properties, such as mean and variance, do not significantly change over time.
Additionally, time series data are extensively used to monitor infrastructure performance, hardware, or application events. Because these data are machine-generated at short intervals, they allow for rapid identification of and response to unexpected changes as soon as they occur.
Tools, Languages, and Visualizations for Time Series
For the creation and analysis of time series, the most widely used programming languages with robust ecosystems are Python and R. Both offer an extensive range of libraries and tools dedicated to data manipulation, visualization, and modeling. Various libraries and frameworks have been developed to facilitate time series analysis, each with its own particularities and focus areas. Two of the most commonly used are Facebook Prophet and Statsmodels:
- Facebook Prophet: user-friendly and robust in business scenarios involving seasonality and irregular data;
- Statsmodels: a well-established Python library for statistical and econometric analysis, with strong support for time series modeling.
In addition to tools and languages, visualization enables the identification of patterns, trends, and anomalies that may not be immediately obvious in raw data. For visualization to be effective, it is essential to define the context, understand the audience, and select an appropriate visual format. The most common types of charts used to represent time series data are:
Chart type | Main purpose |
Line chart | Display changes in one or more quantities over time. |
Area chart | Visualize the aggregate of time series and the contribution of each series to the total. |
Heatmap | Identify seasonal and daily patterns over extended periods. |
Decomposition Chart | Disaggregate the time series into trend, seasonality, and residual components. |
Following the example of a cosmetics production line, such as perfume manufacturing, it is possible to generate visualizations that provide relevant insights by considering variables from the production process, including:
- Essential oil extraction line: steam distillers, cold-press extractors, and solvent extraction systems;
- Mixing and maceration line: mixing tanks, agitators, and temperature control systems;
- Filtration and purification line: high-precision filters;
- Filling and packaging line: automated filling machines, sealing machines, labelers, and quality control systems;
- Quality control line: gas chromatographs and spectrometers.
Practical Example: Time Series in The Fragrance Industry
In practice, programming tools such as Python and R enable the direct implementation of routines for data acquisition, processing, and modeling in industrial settings. In the context of the fragrance industry, sensors installed on distillation units, blending tanks, and thermal control systems continuously generate data such as temperature, pressure, flow rate, and chemical composition. These data can be stored in databases and structured as time series.
Taking a steam distillation unit as an example of equipment used in perfume manufacturing, several variables can impact extraction efficiency. Among these, the most prominent are:
- Chemical composition of essential oils;
- Steam pressure and temperature;
- Extraction yield;
- Solubility of the compounds;
- Boiling points of the components.
The information collected during the process can be organized using different types of charts, depending on the analysis objective. For example, if the goal is to monitor how the chemical composition of essential oils changes over time — with measurements taken every 5 seconds — a line chart is suitable for visualizing these changes in a continuous and time-resolved manner:
Alternatively, to examine the relationship between chemical composition and boiling points of the components, a scatter plot is more appropriate. This type of representation facilitates the identification of correlations between two variables, which can be useful for detecting patterns or clusters of compounds with similar characteristics:
This analysis is relevant because the chemical composition reveals which aroma compounds have been effectively extracted from the raw material. In turn, boiling points help define the optimal temperatures for extraction, enabling maximum yield without causing degradation of volatile and heat-sensitive compounds.
Time Series vs. Event-Based Analyses
Building a time series involves technical steps that, while fundamental, can be quite complex. The first challenge lies in data acquisition and synchronization, especially in industrial settings, such as perfume manufacturing, where different sensors record information at varying frequencies and in diverse formats. To ensure a unified timeline, it is necessary to standardize the data through resampling or interpolation.
Next, preprocessing demands careful attention to data cleansing, outlier removal, and fault handling, ensuring that noise does not compromise the analysis. Furthermore, many statistical models, such as ARIMA, require stationarity, which calls for transformations like differencing or normalization of the series.
Modeling is another critical stage. The appropriate choice of tools for building the time series (programming language, libraries, etc.) depends on the data type, the series’ behavior, and the forecasting goals.
Finally, integration with the production process represents an additional challenge. It is necessary to automate pipelines, continuously update models, and connect them to control systems. When properly implemented, these time series become decisive tools for production optimization, predictive maintenance, and quality improvement in the cosmetics industry.
Given the complexity involved in constructing time series—which encompasses continuous and synchronized data acquisition, through modeling and integration with production processes—it becomes clear how valuable these series are for predictive analytics. They also play a fundamental role in deepening the understanding of system behavior over time.
In contrast, there are more targeted approaches, such as event-based analysis, which focuses only on specific moments of interest. Time series provide a continuous overview that enables the identification of patterns, trends, and causal relationships. Event-based approaches, on the other hand, are triggered only when an unexpected occurrence takes place, such as a sudden change in boiling point during the heating of a mixture.
This type of data collection can be useful when the focus is exclusively on anomaly detection since it centers on specific events. However, by disregarding the intermediate data that precedes these changes, there is a risk of misinterpreting the situation. This happens because the full context leading to the observed peak or anomaly is not fully understood.
Learn more about ST-One.