Multivariate Time Series to Image Transformation for Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Analyzing multivariate time series data from multiple sources is challenging due to its complexity and lack of correlation, making it difficult to extract useful information, especially as the number of data sources increases in manufacturing and other environments.
Innovation Solution
Transforming multivariate time series data into image data using techniques like Fourier-based transformations and lasagna plots, which can then be processed by image processing models such as convolutional neural networks to generate predictions and identify anomalies.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If multivariate time series data is analyzed directly, then the analysis can be performed on the original data format, but the complexity of analysis increases and useful information extraction becomes difficult
Solution Approach 1:
The patent introduces an intermediary transformation process that converts multivariate time series data into image data representation. This intermediary step allows complex temporal patterns to be visualized and processed by image-based machine learning models, thereby improving prediction accuracy while managing analysis complexity through a specialized transformation layer rather than direct complex analysis of raw time series data
Solution Approach 2:
The patent transforms data from one dimension (time series) to another dimension (spatial/image representation). By converting temporal data into spatial image formats, the system leverages the strengths of image processing algorithms to extract patterns that are difficult to detect in traditional time series analysis, thus improving measurement precision without directly increasing analysis complexity
2Loss of information
If more data sources are added to improve analysis comprehensiveness, then more information is available, but the task of analyzing multivariate time series data becomes even more challenging
Solution Approach 1:
The patent merges multiple data sources by transforming them into a unified image representation. Instead of analyzing each data source separately or dealing with their individual complexities, the transformation process combines them into a single image format that can be processed by image-based models, thereby maintaining information from multiple sources while reducing analysis difficulty
Solution Approach 2:
The image transformation serves as an intermediary that handles the complexity of integrating multiple data sources. This intermediary representation allows comprehensive information from multiple sources to be preserved while providing a unified structure that is easier to analyze than the original multivariate time series data
3Productivity
If traditional analysis methods are used on multivariate time series data, then the analysis process is straightforward, but the speed and accuracy of predictions are limited
Solution Approach 1:
The patent substitutes traditional mechanical/time-series analysis methods with image processing-based analysis. By replacing conventional analysis approaches with image-based machine learning models, the system achieves both faster processing speeds and higher prediction accuracy, as image processing algorithms can parallelize computations more effectively and detect patterns more accurately
Data Source
AI summary
An image-based prediction service of a provider network may receive multivariate time series data (e.g., from different sensors for a machine) from a remote network of the client. The multivariate time series data includes different time series of data that are obtained from different data sources (e.g., sensors) over a time period. The image-based prediction service may transform the multivariate time series data into image data that represents an image. The image-based prediction service may then process the image data using one or more image processing models to generate a prediction (e.g., machine failure within a week). The image-based prediction service may send the prediction to the remote network.