State-Transition Graphs for IoT Anomaly Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Detecting anomalies across multiple IoT assets is challenging due to the difficulty in extracting modes from IoT data in an unsupervised manner, differentiating between modes and anomalous states, and handling incomplete data caused by sensor failures or data acquisition issues.
Innovation Solution
The method involves constructing state-transition graphs for each device to identify modes and anomalous states, and using inter-asset comparisons and dynamic knowledge graphs to detect anomalous devices, even with incomplete data, by applying a domain-agnostic definition of modes and leveraging asset relationships.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional anomaly detection methods are used for multiple IoT assets, then individual asset monitoring is possible, but detection accuracy and efficiency deteriorate due to the large number of assets and difficulty in extracting modes from IoT data
Solution Approach 1:
The patent combines multiple individual asset analyses into a unified group-level analysis by constructing a state transition graph that integrates data from multiple assets. This merging approach enables simultaneous detection of anomalies across all assets while identifying common patterns and modes, thereby improving detection accuracy without proportionally increasing system complexity.
Solution Approach 2:
The state transition graph serves multiple functions: it extracts modes from individual assets, detects anomalies within each asset, identifies group-level patterns, and handles incomplete data. This multi-functional approach allows a single framework to address various detection needs across diverse IoT assets, improving precision while managing complexity through universal methodology.
2Loss of information
If comprehensive data collection is performed for all assets, then complete analysis is possible, but data completeness deteriorates due to sensor failures and data acquisition issues
Solution Approach 1:
The patent performs preliminary actions by pre-defining a domain-agnostic set of modes that represent typical operational states across assets. These pre-established modes serve as reference frameworks that enable meaningful anomaly detection even when data is incomplete, as the system can compare observed states against the predefined mode structure rather than requiring complete data to establish what is normal.
Solution Approach 2:
The state transition graph acts as an intermediary that bridges incomplete asset-specific data and the domain-agnostic mode definitions. By mapping observed states to the predefined mode framework, the graph enables reliable anomaly detection despite data gaps, effectively mediating between incomplete observations and comprehensive analysis requirements.
3Measurement precision
If domain-specific mode definitions are used for each asset type, then precise mode identification is possible, but adaptability deteriorates due to the need for domain-specific knowledge and customization
Solution Approach 1:
The patent employs a domain-agnostic set of modes that can be universally applied across different asset types and domains. This universal mode framework maintains precision by capturing essential operational patterns while enabling broad adaptability to various IoT contexts without requiring domain-specific customization, thereby resolving the trade-off between precision and versatility.
Data Source
AI summary
Methods, systems, and computer program products for anomaly and mode inference from time series data are provided herein. A computer-implemented method includes receiving time-series sensor data for each one of a group of devices; extracting a set of states for each device in the group from the time-series sensor data; constructing a state-transition graph for each of the devices, wherein each of the state-transition graphs comprises nodes corresponding to each state in the set and edges corresponding to a probability of transition between the extracted states over time; identifying, for each set, a given state as one of: a mode, a normal state and an anomalous state based on the state-transition graph; and detecting one or more anomalous devices in the group by computing similarities between different devices in the group, based at least in part on the determined state-transition graphs.