Fleet Asset Monitoring Using Residual Time-Series Anomaly Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing condition monitoring systems face challenges in distinguishing anomalous changes in machine behavior from expected changes, particularly in high-dimensional data spaces, leading to increased complexity, false positives, and difficulty in interpreting and visualizing results, especially when dealing with continuous and discrete variables.
Innovation Solution
A method involving Gaussian Processes to model dependencies between time series data, generating residual values, and identifying anomalies by offsetting these residuals to account for central tendencies and functional dependencies, thereby isolating anomalous data points.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If multivariate analysis tools are used to analyze data from multiple variables simultaneously, then the ability to understand relationships between time series data is improved, but the model becomes increasingly unstable and yields a greater number of false positives as the number of time series grows
Solution Approach 1:
The patent segments the multivariate analysis into multiple univariate analyses. Instead of analyzing all variables simultaneously in a high-dimensional space, the system performs separate anomaly detection for each variable (time series) independently. This segmentation reduces model complexity and instability while maintaining the ability to detect anomalies across multiple variables by comparing each variable's anomaly score against a threshold derived from the overall distribution of anomaly scores.
2Measurement precision
If multivariate analysis tools are used to analyze data from multiple variables simultaneously, then the ability to understand relationships between time series data is improved, but the higher dimensional input space makes the result difficult to interpret and visualise
Solution Approach 1:
The patent segments the complex multivariate analysis into simple univariate analyses that are easy to interpret. Each variable is analyzed independently, producing simple anomaly scores that can be directly compared. This segmentation transforms the high-dimensional interpretation problem into multiple simple one-dimensional interpretations, making the results easily visualizable and understandable for end users.
Solution Approach 2:
The patent employs visual indicators (such as color-coded anomaly scores or highlighted data points) to represent the anomaly detection results. By transforming the abstract numerical anomaly scores into visual signals, the system makes it easy for users to identify and interpret anomalous patterns across multiple time series without needing to understand complex multivariate relationships.
3Adaptability or versatility
If different multivariate algorithms are used to detect different kinds of multivariate behaviour, then the detection capability for various phenomena is improved, but the complexity of the system further increases
Solution Approach 1:
The patent employs a universal univariate anomaly detection approach that can detect multiple types of anomalies (point anomalies, contextual anomalies, and collective anomalies) across all variables using the same algorithm. This single unified method replaces the need for multiple specialized multivariate algorithms, reducing system complexity while maintaining versatility in detecting different kinds of anomalous behavior through the common anomaly scoring mechanism.
4Measurement precision
If data is partitioned into subsets based on the value of the discrete parameter, then separate analyses can be run on each portion of the data, but information is partitioned and it becomes difficult to take a holistic view on what is happening
Solution Approach 1:
The patent merges the results from individual univariate analyses into a unified anomaly detection framework. By combining the anomaly scores from multiple variables and comparing them against a globally derived threshold, the system maintains a holistic view of the system state while still benefiting from the detailed analysis of individual variables. This merging approach prevents information loss and enables comprehensive anomaly detection across the entire multivariate system.
Data Source
Figure 1~2
Figure 3~4
Figure 5~6
AI summary
A method and data processing system for identifying data anomalies in data obtained by a condition monitoring system from a fleet of industrial assets are described. A time series of data values is obtained for each asset in the fleet and an average time series for the fleet is determined, based on the obtained time series. For a selected asset in the fleet a model of the dependency between the obtained time series and the average time series is generated, and a further time series is generated for the selected asset. Anomalous data values are identified from the further time series.