Industrial Time Series AIGC Model With Time-Frequency Synthesis
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing industrial time series data analysis models fail to leverage frequency domain information, resulting in poor data quality and inaccurate feature extraction, making it difficult to generate high-fidelity predicted time series data.
Innovation Solution
An industrial time series AIGC fundamental model that processes raw time series data by capturing information in both the time and frequency domains through an encoding module, decoding module, and synthesizing module, utilizing cross-channel and cross-temporal attention mechanisms to generate high-fidelity target time series data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If existing diffusion models are used to generate time series data, then the generation process can be simplified, but the accuracy of extracted features and fidelity of predicted time series data deteriorates due to ignoring frequency domain information
Solution Approach 1:
The patent transforms the time series data from the time domain to the frequency domain using Fast Fourier Transform (FFT), adding a new dimension of analysis. This allows the model to capture frequency components (periodicity, rhythm) in addition to time-domain patterns, thereby improving feature extraction accuracy and generating more faithful time series data without significantly increasing process complexity.
Solution Approach 2:
The patent segments the time series data into multiple frequency components through spectral analysis. By decomposing the time series into distinct frequency bands (e.g., using wavelet transforms or FFT), the model can process and generate each frequency component separately, improving overall accuracy while maintaining manageable complexity through modular processing.
2Device complexity
If existing diffusion models focus only on time domain, then the model structure remains simple, but the fidelity of generated time series data deteriorates due to ignoring rich frequency domain information
Solution Approach 1:
The patent merges the time domain and frequency domain representations by combining time-series data with its frequency spectrum as dual inputs to the diffusion model. This fusion allows the model to leverage both temporal patterns and frequency characteristics, significantly improving data fidelity while the combined representation structure remains manageable through efficient integration of both domains.
Solution Approach 2:
The patent introduces frequency domain information as an additional dimension to the traditional time domain model. By augmenting the model inputs with spectral features (via FFT or wavelet transforms), the model gains access to rich frequency patterns that enhance generated data fidelity without requiring a complete redesign of the fundamental model architecture.
3Productivity
If time series data is processed without frequency domain analysis, then the processing process remains simple, but the quality of extracted features deteriorates due to ignoring frequency components
Solution Approach 1:
The patent segments the processing into distinct stages: time-domain preprocessing, frequency-domain transformation (FFT/wavelet), feature extraction from both domains, and synthesis. This segmentation allows efficient processing of each stage independently, maintaining high productivity while achieving precise feature extraction through multi-domain analysis.
Solution Approach 2:
The patent adds frequency domain analysis as an additional processing dimension that operates parallel to time-domain processing. By using efficient transforms like FFT with optimized complexity, the model extracts precise frequency features without significantly slowing down the overall processing speed, achieving both high productivity and high measurement precision.
Data Source
AI summary
An industrial time series AIGC fundamental model, an electronic device, and a computer program product. The model includes: an encoding module, configured to capture first relevant information of raw time series data in a frequency characteristic domain along a channel dimension, and downsample the first relevant information to obtain encoded time series data, where the raw time series data includes multiple types of raw vibration data, and each channel collects one type of raw vibration data; a decoding module, configured to capture second relevant information of the encoded time series data along a time dimension, and upsample the second relevant information to obtain decoded time series data; a synthesizing module, configured to synthesize the decoded time series data according to a trend and a time period respectively, and then perform concatenation processing to obtain the target time series data.