Time Series Recurrent Pattern Detection via Frequency Domain Decomposition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning models trained on time series data often struggle to detect and incorporate recurrent patterns, such as seasonality, due to noise and changing data characteristics, leading to prediction errors. Additionally, existing methods may fail to detect recurrent patterns at specific time periods like the beginning or end of a month, and can be inaccurate when dealing with calendar or fiscal year-based data.
Innovation Solution
The system addresses these issues by decomposing time series data into the frequency domain to generate frequency-based signals, and then uses signal processing to identify recurrent patterns. It defines custom intervals within the time series data to improve sensitivity in detecting patterns around specific time periods. Furthermore, the system converts date-based values to numeric values to define units of periodicity, and fine-tunes the order of frequencies using random sampling and statistical averaging to enhance pattern detection and forecasting accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If standard signal processing is applied to decomposed time series data, then frequency components can be identified, but recurrent patterns at specific time periods (beginning or end of month) are undetected due to interval misalignment
Solution Approach 1:
The patent segments the time series data into multiple overlapping intervals with different time granularities (e.g., weekly, bi-weekly, monthly intervals). This segmentation allows the system to capture recurrent patterns at different time scales and prevents information loss at interval boundaries by ensuring that patterns occurring at specific time periods are included in at least one interval's analysis.
Solution Approach 2:
The patent transforms the time series data from the time domain to the frequency domain using Fourier transform, adding a frequency dimension to the analysis. This dimensional transformation enables the identification of recurrent patterns by analyzing frequency components, thereby improving detection accuracy for patterns that may be obscured in the time domain due to noise or interval misalignment.
2Reliability
If machine learning models are trained on raw time series data, then they can learn general relationships, but they fail to detect recurrent patterns due to noise and changing data characteristics
Solution Approach 1:
The patent performs preliminary signal processing and frequency domain transformation on the time series data before feeding it to the machine learning model. This preliminary action decomposes the data into frequency components and identifies recurrent patterns in advance, making the patterns more detectable and reducing the difficulty for the model to learn from noisy raw data.
Solution Approach 2:
The patent introduces frequency components as an intermediary representation between the raw time series data and the machine learning model. This intermediary transformation converts difficult-to-detect recurrent patterns in the time domain into more detectable frequency domain features, thereby improving the model's ability to identify patterns and make accurate predictions.
3Measurement precision
If date-based values are used directly in frequency analysis, then temporal information is preserved, but periodicity detection is inaccurate for calendar or fiscal year-based data
Solution Approach 1:
The patent transforms date-based values into numeric representations that explicitly encode temporal periodicity information (e.g., day of year, week number, month). This parameter transformation maintains the essential temporal structure while making the periodicity detection more accurate by providing consistent numeric inputs that reflect the regular cycles in the data.
Data Source
AI summary
The disclosure relates to systems and methods of detecting recurrent patterns such as seasonality in time series data to improve the forecasting performance of machine learning models. The time series data may be encoded into interval encodings that capture recurrent patterns that may occur at certain time periods that would otherwise not be detected. To perform frequency-based analysis, the intervals may be converted to numeric values, then decomposed from a time domain to a frequency domain to form frequency components such as sine and cosine waves. Relevant frequency components that are influential are identified through Monte Carlo Markov Chain analysis and/or Gaussian Mixture Model analysis. The identified frequency components may represent recurrent patterns detected in the time series data and are used to configure an order value that defines the number of frequency components used for future decomposition and analysis for forecasting.