CPC-Based Time Series Anomaly Segmentation With Local Latent Scoring
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current anomaly detection methods in machine learning systems are ineffective in identifying anomalous regions within time series data, particularly in domains beyond images, and lack the ability to adapt self-supervised learning for detecting anomalies in a subsequence level.
Innovation Solution
The proposed solution involves a Local Neural Transformations (LNT) approach that learns to embed and augment time series data locally, combining Contrastive Predictive Coding (CPC) with neural transformation learning and Hidden Markov Models (HMMs to derive anomaly scores for each time step, enabling the detection of anomalous regions through a dynamic deterministic contrastive loss and smoothing of scores to identify continuous anomalous regions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If current anomaly detection methods are used, then general anomaly detection is possible, but detection precision for anomalous regions in time series data is insufficient
Solution Approach 1:
The patent applies local quality by transitioning from global anomaly detection to local region-level detection. The system divides time series data into subsequences and generates anomaly scores for each local region, enabling precise identification of anomalous segments rather than treating the entire series uniformly. This local-focused approach improves detection precision while maintaining adaptability to time series characteristics.
Solution Approach 2:
The patent segments the time series data into multiple subsequences and processes each segment independently to generate localized anomaly scores. This segmentation allows the system to capture temporal patterns and anomalies at appropriate granularities, improving both precision in identifying anomalous regions and adaptability to different time series structures.
2Ease of operation
If self-supervised learning is applied, then learning without labels is enabled, but detection at subsequence level is not achieved
Solution Approach 1:
The patent introduces dynamic temporal context modeling that adapts to varying time series patterns. The self-supervised learning framework dynamically adjusts to local temporal dependencies and generates context-aware anomaly scores for each subsequence, achieving both ease of operation through label-free training and precision in subsequence-level detection.
Solution Approach 2:
The patent adds a temporal dimension to the self-supervised learning process by incorporating time-series-specific transformations and contextual information. This dimensional enhancement allows the model to capture temporal patterns while maintaining self-supervised learning benefits, achieving accurate subsequence-level anomaly detection without requiring labeled data.
3Productivity
If deep learning baselines are used, then general pattern recognition is achieved, but performance on time series anomaly detection is insufficient
Solution Approach 1:
The patent transforms the anomaly detection problem by changing the output parameter from binary classification to continuous anomaly scoring. This parameter transformation enables the model to produce graded anomaly assessments for each subsequence, improving detection performance while maintaining broad domain applicability through the flexible scoring mechanism.
Solution Approach 2:
The patent introduces an intermediary anomaly scoring mechanism that bridges deep learning capabilities and time series-specific requirements. The scoring system acts as a mediator between general pattern recognition and domain-specific detection needs, enhancing performance while preserving versatility across different applications.
Data Source
AI summary
An anomalous region detection system includes a controller configured to, receive data being grouped in patches, encode, via parameters of an encoder, the data to obtain a series of local latent representations for each patch, calculate, for each patch, a Contrastive Predictive Coding (CPC) loss from the local latent representations to obtain updated parameters, update the parameters of the encoder with the updated parameters, score each of the series of the local latent representations, via the Contrastive Predictive Coding (CPC) loss, to obtain a score associated with each patch, smooth the score to obtain a loss region, mask the data associated with the loss region to obtain verified data, and output the verified data.