Multi-Domain Graph Forecasting for Missing Predictive Maintenance Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
AI models for predictive maintenance are inaccurate when faced with missing or inaccurate data, as they rely on high-quality training data, and existing methods struggle to effectively handle irregular multivariate time series with missing values.
Innovation Solution
A multi-domain graph-guided network approach that learns graph structures from incomplete input data, combines time and frequency domain forecasts using a variable-wise mixture mechanism, and aligns forecasts to capture domain-invariant similarities, enabling accurate predictions despite missing data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If AI models are trained using traditional methods with complete data assumptions, then training accuracy is improved, but performance degrades when facing missing or inaccurate data
Solution Approach 1:
The patent applies preliminary action by pre-training the graph-guided network with synthetic complete data to learn optimal graph structures and temporal patterns. This pre-training enables the model to handle missing data effectively during actual operation, as the model has already learned the underlying relationships and can infer missing values based on the learned graph structure and temporal dependencies.
Solution Approach 2:
The patent introduces an intermediary mechanism through the graph-guided network architecture that mediates between complete training data and incomplete operational data. The graph structure serves as an intermediary representation that captures relationships between variables, allowing the model to transfer knowledge from complete training data to handle missing data scenarios by inferring relationships through the graph structure.
2Device complexity
If traditional AI models are used for predictive maintenance, then model simplicity is maintained, but they fail to handle irregular multivariate time series with missing values
Solution Approach 1:
The patent applies universality by designing a graph-guided network that can handle multiple types of input data patterns (complete data, missing data, irregular time series) within a single unified architecture. The model universally processes both regular and irregular multivariate time series data, making it adaptable to various real-world scenarios without requiring separate models for different data conditions.
Solution Approach 2:
The patent transforms the problem by adding a graph structure dimension to the traditional time series analysis. Instead of only analyzing temporal patterns in a single dimension, the model incorporates relational patterns through graph structures, creating a multi-dimensional representation that enables handling of missing values and irregular data by leveraging both temporal and relational dimensions simultaneously.
3Reliability
If graph structures are learned from incomplete input data, then ability to handle missing data is improved, but computational complexity increases
Solution Approach 1:
The patent applies preliminary action by pre-computing and storing graph structures during training on synthetic complete data. This pre-computed graph structure information is then reused during inference on incomplete data, avoiding the need to compute graph structures from scratch for each new data point, thus reducing online computational complexity while maintaining the ability to handle missing data effectively.
Data Source
AI summary
Systems and methods for handling missing data with multi-domain graph-guided networks. Graph structures can be learned with masked dimension extension based on incomplete input data obtained from monitored entities to generate inferred graphs. Time and frequency domain forecasts generated based on the inferred graphs with a variable-wise mixture mechanism can be combined to generate combined forecasts. The combined forecasts can be aligned to time and frequency domains to obtain final forecasts that capture domain-invariant similarities between variables. A corrective action generated with multi-domain graph-guided networks for the monitored entities based on the final forecasts can be performed.