Battery Diagnosis Model Monitoring With Local and Global Feature Importance
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning-based state diagnosis models for secondary batteries provide only global explanations for time series data, which can decrease reliability when battery data characteristics change, necessitating a system that can also provide local explanations and evaluate causal relationships among parameters.
Innovation Solution
A learning model monitoring apparatus and method that evaluates local and global importance of features using Shapley values and causal discovery algorithms to analyze battery state diagnosis models, enabling reliable diagnosis even with changing battery characteristics.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of information
If a method evaluating the global importance of features for a model trained on battery time series data is employed, then the model can provide global explanations for time series data, but the reliability decreases when the characteristics of battery data change
Solution Approach 1:
The patent segments the evaluation of feature importance into two distinct levels: local importance evaluation for individual time series data points and global importance evaluation aggregated from local evaluations. This segmentation allows the system to provide both specific local explanations and overall global explanations, resolving the contradiction by maintaining reliability through local context-aware evaluation while preserving global explanation capability.
Solution Approach 2:
The patent implements local quality by evaluating feature importance locally for each individual time series data point rather than uniformly across all data. This local evaluation approach adapts to changing data characteristics at each point, maintaining reliability even when battery data characteristics change, while still enabling global explanation through aggregation of local evaluations.
2Device complexity
If only global explanations are provided for time series data, then the model structure remains simple, but the system cannot adapt to changing battery data characteristics
Solution Approach 1:
The explanation system is segmented into local importance evaluation modules that operate independently on individual data points and a global aggregation mechanism. This segmentation enables the system to adapt to changing battery characteristics through local evaluations while maintaining manageable complexity through automated aggregation processes.
Solution Approach 2:
The system implements self-service through automated local importance evaluation that adapts to each data point's characteristics without requiring manual intervention. The automated aggregation of local evaluations into global explanations further reduces complexity, allowing the system to maintain both adaptability and simplicity.
3Measurement precision
If local importance evaluation is performed for each cycle using Shapley values, then the local explanation accuracy improves, but the computational complexity increases
Solution Approach 1:
The patent applies preliminary action by pre-calculating and storing Shapley values for feature importance evaluation. This pre-computation approach enables accurate local explanations when needed while reducing real-time computational complexity, as the intensive calculations are performed in advance rather than during deployment.
Solution Approach 2:
The system uses copying by pre-computing and storing Shapley value results that can be reused for multiple evaluation purposes. This copying approach maintains high measurement precision for local explanations while reducing computational complexity during actual model monitoring, as the expensive calculations are performed once and reused.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A learning model monitoring apparatus according to an embodiment disclosed in this document may include an acquisition unit configured to acquire time series data of a battery, an extraction unit configured to organize the time series data and extract at least one feature value corresponding to at least one feature for each of a plurality of sub-time series data corresponding to different cycles, a learning unit configured to training a state diagnosis model capable of diagnosing the state of the battery based on the extracted at least one feature value, a first importance evaluation unit configured to evaluate local importance of the at least one feature for the trained state diagnosis model in each cycle, and a second importance evaluation unit configured to evaluate global importance of the at least one feature for the trained state diagnosis model based on the local importance.