Transformer-Based Anomaly Detection in Time Series Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional deep learning models for detecting abnormalities in time series data, such as CNN-based and RNN-based models, face challenges in processing time series data effectively due to increased neural network size and specialization in lattice data, rather than time series data, and require multiple decoders which complicates the process.
Innovation Solution
A method using a transformer-based artificial neural network that masks and restores time series data to minimize neural network volume and improve detection performance, involving pretreatment, embedding, and training with a generator and discriminator to learn normal distributions and detect abnormalities without needing multiple decoders.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If CNN-based deep learning model is used for detecting abnormality in time series data, then the model can process data effectively, but the model becomes specialized for lattice data rather than time series data and increases neural network size
Solution Approach 1:
The patent extracts and removes unnecessary components from the neural network architecture. Specifically, it eliminates the need for multiple decoders by using a single decoder that processes all time series data uniformly, thereby reducing neural network size while maintaining abnormality detection performance
Solution Approach 2:
The patent makes the neural network model universal by designing a single decoder that can handle all types of time series data regardless of resolution or format. This universal approach replaces the specialized multi-resolution decoders, reducing complexity while improving adaptability to different time series data
2Adaptability or versatility
If RNN-based deep learning model with multi-resolution decoders is used to capture time features, then the model can process time series data with various resolutions, but the size of the neural network increases
Solution Approach 1:
The patent merges multiple specialized decoders into a single unified decoder. Instead of having separate decoders for different resolutions, the single decoder processes all time series data uniformly, combining the functionality of multiple components into one to reduce neural network size while maintaining versatility
Solution Approach 2:
The patent changes the architectural parameters by removing the multi-resolution decoder structure and adopting a single decoder with uniform processing capability. This parameter change simplifies the network while preserving the ability to handle various time series resolutions through a unified approach
3Reliability
If multiple decoders are used in the deep learning model, then the model can capture time features with various resolutions, but the device complexity and processing overhead increase
Solution Approach 1:
The patent extracts and removes the unnecessary multi-resolution decoder structure, keeping only a single decoder that captures time features effectively. This extraction eliminates redundant components while preserving the essential time feature capture capability
Solution Approach 2:
Instead of using multiple decoders to handle different resolutions (the conventional approach), the patent inverts the approach by using a single decoder that uniformly processes all data. This inversion simplifies the architecture while maintaining or improving detection performance
Data Source
AI summary
A method of detecting an abnormality in time series data according to an embodiment of the present disclosure is performed in a computing device including one or more processors and a memory storing one or more programs executed by the one or more processors. First masking is performed to cover a portion of input time series data with a mask. First-restored time series data in which the time series data is restored is generated by inputting the first-masked time series data to a generator. A difference between the first-restored time series data and original time series data is calculated. Second masking is performed to cover a portion of the time series data with a mask on basis of the calculated difference. Second-restored time series data in which the time series data is restored is generated by inputting the second-masked time series data to the generator.