Multivariate Autoregression for Sparse-Data Sensor Anomaly Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In predictive maintenance for IoT, standard classification algorithms are impractical due to the lack of or minimal training data, and they fail to detect anomalies that have not been seen in historical data, especially in cases of sparse or no historical data availability.
Innovation Solution
The use of multiple auto-regression models, specifically nonlinear auto-regression models implemented using neural networks like LSTM or GRU, to predict sensor values and detect anomalies by comparing predicted and observed data, with an anomaly score indicating the presence of anomalies based on threshold scores.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If standard classification algorithms are used for anomaly detection, then the system can detect known anomalies from historical data, but the system fails when minimal or no training data is available and cannot detect previously unseen anomalies
Solution Approach 1:
The patent pre-processes sensor data by computing statistical features (mean, standard deviation, skewness, kurtosis) and creating lagged variables before modeling. This preliminary transformation of the data enables the autoregressive models to effectively capture temporal patterns even with minimal training data, resolving the contradiction between reliability and data availability
Solution Approach 2:
The patent transforms the original sensor data into different parameter representations including lagged variables (t-1, t-2, t-3), rolling statistics, and standardized features. These parameter transformations enable the model to detect anomalies through multiple statistical perspectives, improving reliability without requiring additional training data
2Adaptability or versatility
If multiple autoregression models with different data subsets are used, then the system can detect previously unseen anomalies and improve robustness, but the device complexity increases
Solution Approach 1:
The patent divides the available sensor data into multiple distinct subsets (first subset, second subset, third subset) and trains separate autoregression models on each. This segmentation allows the system to capture different statistical patterns from different data portions, improving adaptability to various anomaly types while maintaining manageable model complexity through modular architecture
Solution Approach 2:
The patent creates multiple autoregression models that can collectively handle diverse anomaly detection scenarios. Each model processes different data subsets but follows the same autoregressive framework, providing universal applicability across various anomaly types and conditions while maintaining consistent methodology throughout the system
Data Source
AI summary
Methods, systems, and computer-readable storage media for receiving a time-series of data values associated with a plurality of sensors, each sensor generating at least a portion of the time-series of a respective data value, providing a plurality of auto-regression models, each auto-regression model being provided based on a respective first sub-set of the time-series of data values used as input, and a respective second sub-set of the time-series of data values used as training data during a training process, receiving respective data values associated with a time from and generated by each of the plurality of sensors, determining respective predicted values for each of the auto-regression models, and selectively indicating that an anomaly is present in the system based on respective predicted values for each of the auto-regression models, and the respective data values associated with a time.