Sensor Causal DAGs for Root Cause Analysis in Noisy IoT Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Automated root cause analysis of anomaly events in high-dimensional sensor data from IoT devices in industrial settings is hindered by the complexity and irregularities of sensor data, such as heterogeneous formats, errors, and high correlation, which existing methods fail to address effectively.
Innovation Solution
A system that preprocesses sensor data by aligning it to a unified global reference time, applying data summarization and interpolation techniques to generate meta-nodes, and builds an optimal Directed Acyclic Graph (DAG) to model causal dependencies, allowing for the determination of anomaly causes through back-tracking in the graph structure.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual study and identification of causal relationships among sensors is performed, then accuracy of root cause analysis is improved, but time consumption and cost increase prohibitively with growing data dimensionality
Solution Approach 1:
The patent replaces manual mechanical analysis with automated computational methods. Specifically, it uses machine learning algorithms to automatically discover causal relationships among sensors, substituting human expert analysis with algorithmic processing that scales efficiently to high-dimensional data without proportionally increasing time or cost.
Solution Approach 2:
The patent introduces an intermediate computational layer between raw sensor data and root cause identification. This layer includes data preprocessing modules, feature extraction algorithms, and causal inference engines that transform raw high-dimensional sensor data into meaningful causal relationships, enabling automated analysis while maintaining accuracy.
2Productivity
If automated root cause analysis is implemented, then productivity and scalability are improved, but handling of data irregularities and heterogeneity becomes more challenging
Solution Approach 1:
The patent segments the complex data processing task into distinct modular stages: data preprocessing (handling missing values, outliers, and synchronization), feature extraction (identifying relevant sensor patterns), causal relationship discovery (building causal graphs), and root cause identification (tracing anomalies to sources). Each module handles specific aspects of data irregularities independently, making the overall system more manageable and scalable.
Solution Approach 2:
The patent transforms raw sensor data into standardized parameters and features through preprocessing operations. By converting heterogeneous sensor readings into unified temporal and statistical parameters, the system simplifies subsequent causal analysis while maintaining the ability to handle diverse data types and irregularities.
3Reliability
If high-dimensional sensor data is analyzed in detail, then completeness of anomaly detection is improved, but computational cost and complexity increase exponentially
Solution Approach 1:
The patent extracts only the most relevant features and causal relationships from high-dimensional sensor data rather than analyzing all dimensions equally. By identifying and focusing on key sensors and causal pathways that contribute most to anomaly detection, the system maintains high reliability while significantly reducing computational requirements through selective feature extraction and dimensionality reduction.
Data Source
Figure 1A~1B
Figure 2
Figure 3A~3B
AI summary
One embodiment of the present invention can provide a system for identifying a root cause of an anomaly event in operation of one or more machines is provided. During operation, the system can obtain sensor data from a set of sensors associated with the one or more machines, convert the sensor data into a set of sensor states, build an optimal DAG based on the set of sensor states to model causal dependency; determining, by using the DAG, a probability of an anomaly state of a target sensor given a state of a direct neighbor sensor, and determining a root cause of the anomaly event associated with the target sensor by back-tracking the anomaly state in the DAG.