BiD-STM Anomaly Detection Using Bidirectional Stochastic Transition Matrices
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for detecting anomalous behavior in time-series signals, such as employee attrition and cyber-attacks, lack efficiency in distinguishing between nominal and anomalous dynamics, leading to suboptimal early detection and intervention.
Innovation Solution
The BiD-STM technique involves converting time-series signals into sequences of symbols, computing stochastic transition matrices for both forward and backward sequences, and concatenating eigenvalues to generate a final feature vector, which is then used for efficient classification of anomalous behavior.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional methods are used for detecting anomalous behavior in time-series signals, then the detection process is simpler, but the accuracy in distinguishing between nominal and anomalous dynamics deteriorates
Solution Approach 1:
The patent segments the time-series signal processing into distinct components: forward stochastic transition matrix computation, backward stochastic transition matrix computation, and eigenvalue analysis. This segmentation allows each component to be optimized independently, improving overall detection accuracy while managing complexity through modular processing steps.
Solution Approach 2:
The patent applies inversion by computing stochastic transition matrices in both forward and backward directions. By analyzing the time-series signal from both temporal perspectives and combining the results through eigenvalue concatenation, the method enhances its ability to distinguish anomalous dynamics that might be invisible from a single direction.
2Measurement precision
If both transient and steady state components are leveraged for detection, then the accuracy of anomalous behavior detection is improved, but the computational complexity increases
Solution Approach 1:
The patent extracts the essential dynamic characteristics of time-series signals by computing stochastic transition matrices that capture both transient and steady-state behaviors. By focusing on the eigenvalues of these matrices, the method extracts the most informative features while discarding redundant computational details, thus improving accuracy without proportionally increasing computational power requirements.
Solution Approach 2:
The patent transforms the raw time-series data into a different parameter space through stochastic transition matrices and eigenvalue decomposition. This parameter transformation allows the system to capture complex temporal dynamics (both transient and steady-state) in a compressed feature representation, achieving high detection accuracy with manageable computational resources.
3Measurement precision
If the bidirectional stochastic transition matrix approach is used, then the separation between nominal and anomalous behaviors is improved, but the processing time increases
Solution Approach 1:
The patent performs preliminary action by pre-computing the stochastic transition matrices and their eigenvalues from the time-series signal. This preprocessing step organizes the data into a structured format that facilitates efficient anomaly detection, reducing the computational burden during actual detection operations and thereby mitigating processing time increases.
Data Source
AI summary
In some aspects, the techniques described herein relate to a method including: receiving a time series signal; converting the time series signal into a sequence of symbols; computing a first stochastic transition matrix from the sequence of symbols; reversing the sequence of symbols, wherein the reversing generates a backward sequence of symbols; computing a second stochastic transition matrix from the backward sequence of symbols; computing a first vector of eigenvalues based on the first stochastic transition matrix; computing a second vector of eigenvalues based on the first stochastic transition matrix; and computing a final vector of features based on the first vector of eigenvalues and the second vector of eigenvalues.