Time-Series Image Encoding for Interpretable Anomaly Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for anomaly detection in multi-variate time series sensor data face challenges due to complex temporal and spatial dependencies, limited datasets, and loss of physics-based relationships when applying data augmentation, leading to difficulties in interpreting and identifying anomalies in industrial datasets.
Innovation Solution
A method that converts time series data into a graphical representation using value-dependent colors, brightness, or patterns, allowing for easy analysis by existing AI/ML techniques, which are domain-agnostic and require minimal computational resources, and enables the generation of additional datasets through difference images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If time series data is transformed into image using conventional methods (e.g., Gramian Angular Field), then the data can be analyzed by deep learning techniques, but the interpretability of results is reduced and domain-specific knowledge is required
Solution Approach 1:
The patent segments the time series data into discrete time stamps and represents each time stamp as a separate image element with specific visual characteristics (color, brightness, pattern) that directly correspond to the data values. This segmentation maintains the temporal structure while enabling automated analysis without losing interpretability, as each image element can be traced back to its corresponding time stamp and value range.
Solution Approach 2:
The patent uses color, brightness, and pattern variations in image elements to represent different data values and their relationships. These visual encodings preserve the original data characteristics while making them accessible to automated image analysis algorithms, thereby maintaining interpretability while enabling automation.
2Quantity of substance
If data augmentation techniques are applied to limited datasets, then the dataset size is increased, but physics-based relationships are lost
Solution Approach 1:
The patent transforms univariate time series data into a two-dimensional visual representation where the x-axis represents time and the y-axis represents data values encoded through color, brightness, and pattern. This dimensional transformation allows the data to be analyzed as images while preserving the underlying physics-based relationships, as the visual encoding maintains the temporal and value relationships present in the original data.
3Measurement precision
If complex deep learning models are used for anomaly detection, then detection accuracy is improved, but computational requirements increase
Solution Approach 1:
The patent introduces an intermediary visual representation layer that transforms time series data into images with specific structural characteristics. This intermediary representation makes the data more amenable to analysis by simpler image processing algorithms and machine learning models, thereby reducing computational requirements while maintaining detection accuracy. The visual encoding acts as a mediator that bridges the gap between raw sensor data and analytical algorithms.
Data Source
AI summary
A computer-implemented method of converting time series data of at least one operational data source of a technical system into an image, includes providing a time series data of at least one operational data source of the technical system as a series of values of successive time stamps, wherein the values of the series of values vary over time. The method further includes the steps of assigning for each of the considered time stamps of the respective time series data either a value-depending color, a value-depending brightness, a value-depending pattern or a combination thereof to an image element of the image, and sequencing the image elements along a timeline, preferably without spaces between the image elements, to form for each time series data a set of linearly arranged image elements of the image.