Local Neural Transformations for Time-Series Anomalous Region Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing anomaly detection methods struggle to effectively identify anomalous regions within time series data, particularly in unsupervised learning scenarios where labeled data is scarce, and lack the ability to detect anomalies on a subsequence level rather than as a whole sequence.
Innovation Solution
The method employs Local Neural Transformations (LNT) that combine Contrastive Predictive Coding (CPC) with neural transformation learning, using a dynamic deterministic contrastive loss (DDCL) to score each time step and mask anomalous regions, leveraging self-supervised learning to enhance anomaly detection.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional anomaly detection methods are used on time series data, then the system can detect anomalies in the overall sequence, but it cannot effectively identify anomalous regions at the subsequence level
Solution Approach 1:
The patent divides the time series data into multiple local patches or subsequences, allowing the model to process and detect anomalies at a finer granular level. This segmentation enables region-level anomaly identification rather than treating the entire sequence as a single unit, directly addressing the limitation of traditional methods.
Solution Approach 2:
The patent transforms the time series data into a latent representation space through neural transformations, adding a dimensional transformation layer. This allows the model to capture complex temporal patterns and anomalies that are not apparent in the original time domain, improving detection precision without linearly increasing complexity.
2Adaptability or versatility
If unsupervised learning is used for anomaly detection, then the system can operate without labeled data, but it struggles to effectively identify anomalous regions
Solution Approach 1:
The patent employs self-supervised learning where the model learns to reconstruct normal time series patterns from unlabeled data. By training on the assumption that most data is normal, the model learns meaningful temporal structures and can then identify deviations as anomalies, achieving both unsupervised operation and region-level detection accuracy.
Solution Approach 2:
The patent uses dynamic thresholding and adaptive parameter adjustment in the loss function to improve anomaly detection in unsupervised settings. By dynamically adjusting the contrastive loss parameters based on the learned latent representations, the system can effectively distinguish anomalies without labeled data while maintaining high precision.
3Adaptability or versatility
If local neural transformations are applied to each patch, then the system can achieve diverse transformed vector representations, but the computational complexity increases
Solution Approach 1:
The patent applies neural transformations to segmented local patches rather than the entire time series at once. This segmentation reduces the computational burden per transformation while still achieving diverse representations through the aggregation of transformed patches, balancing versatility and energy consumption.
Solution Approach 2:
The patent applies transformations selectively to local regions that are most likely to contain anomalies or contain the most informative patterns. By focusing computational resources on critical regions rather than uniformly transforming all data, the system achieves diverse representations with reduced energy consumption.
Data Source
AI summary
A anomalous region detection method includes receiving time series data being grouped in patches, encoding, via parameters of an encoder, the data to obtain local latent representations, determining a representation loss from the local latent representations, transforming the local latent representations associated with each patch, via at least two local neural transformations, to a series of diverse transformed vector representations, determining a dynamic deterministic contrastive loss (DDCL) from the series of diverse transformed vector representations, combining the representation loss and the DDCL to obtain updated parameters, updating the parameters of the encoder with the updated parameters, scoring each of the series of the diverse transformed vector representations, via the DDCL, to obtain a diverse semantic requirement score associated with each patch, smoothing the diverse semantic requirement score to obtain a loss region, masking data associated with the loss region to obtain verified data, and outputting the verified data.