Predictive Maintenance Alerts With GNN Explainability
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional predictive maintenance systems, particularly those using graph neural networks (GNNs), face inefficiencies due to their opaque nature, making it difficult for users to comprehend the reasoning behind maintenance alerts and predictions, which can lead to unnecessary downtimes and resource wastage.
Innovation Solution
A computer-implemented method that combines graph neural networks with sub-symbolic explainers and large language models to generate predictive maintenance alerts. This method processes the graph neural network's predictions to identify influential edges and features within a sub-graph, and then transforms this information into a natural language maintenance justification, providing transparency and clarity to users.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If graph neural networks are used for predictive maintenance, then prediction accuracy is improved, but system transparency deteriorates
Solution Approach 1:
The patent introduces an explainer module as an intermediary between the graph neural network and the user. This module processes the complex internal representations and generates human-interpretable explanations, thereby mediating between the high-accuracy black-box model and the need for transparency without compromising either aspect
2Reliability
If complex machine learning models are deployed, then predictive capability is improved, but ease of operation deteriorates
Solution Approach 1:
The explainer module serves as an intermediary that translates complex model predictions into understandable explanations for operators, maintaining high predictive capability while improving ease of operation through interpretable output
3Measurement precision
If detailed analysis is performed to improve prediction accuracy, then processing time increases
Solution Approach 1:
The system pre-processes and structures maintenance data into graph representations beforehand, and the explainer module is pre-configured with explanation templates and rules. This preliminary preparation enables fast generation of explanations during actual prediction without sacrificing accuracy
Solution Approach 2:
The system segments the explanation generation process into distinct components: identifying influential nodes and edges, retrieving their meanings from the knowledge base, and formatting explanations. This segmentation allows parallel processing and reduces overall computation time
Data Source
Figure 1~2
Figure 3
Figure 4
AI summary
A graph (G) has nodes (V) which represent entities of the industrial system (SYS) and edges (E) representing relations between the entities of the industrial system (SYS). A graph neural network (GNN) processing the graph (G) calculates a class prediction (P) for at least one entity of said industrial system (SYS). A sub-symbolic explainer (GNN-EXP) processes the class prediction (P) to identify edges (E) between nodes (V) and associated features (X) of nodes (V) belonging to a sub-graph (SG) within the graph (G) having influenced the class prediction (P). A large language model (LLM), equipped with a plugin (PL) for accessing the graph (G) and receiving a prompt (PR) including the sub-graph (SG), transforms the sub-graph (SG) into a maintenance justification (MJ) in natural language. A user interface (UI) outputs a predictive maintenance alert (PMA) along with the maintenance justification (MJ). As a result, the domain experts or operators of the plant can understand why a failure will occur and act accordingly. The natural language explanations reduce the cognitive workload, improve the transparency and trustworthiness of the maintenance process greatly, and can also add additional context not given by the graph neural network.