Battery Anomaly Diagnosis Using Latent Reconstruction Error
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional battery diagnosis methods struggle with detecting anomalies due to long-term cycle tests and misdiagnoses from time series data with noise, increasing the risk of device damage from battery failures.
Innovation Solution
A battery diagnosis device using an artificial intelligence model that learns battery state patterns through an encoder, decoder, discriminators, and calculators to diagnose anomalies based on time series data and reconstruction errors, calculating an anomaly score for precise battery health assessment.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional long-term cycle tests are used to diagnose battery anomalies, then comprehensive battery data is collected, but the diagnosis process requires excessive time and manpower
Solution Approach 1:
The system performs preliminary learning of normal battery patterns during the charging process before anomaly detection is needed. The encoder learns the normal charging pattern of the battery in real-time during charging, so when anomaly detection is required, the comparison can be made immediately without waiting for long-term cycle tests to complete.
2Measurement precision
If time series data with sensor noises is used for battery diagnosis, then detailed battery state information is obtained, but misdiagnoses occur due to noise interference
Solution Approach 1:
The encoder extracts only the essential features from the noisy time series data by converting it into a latent variable representation. This extraction process separates the meaningful battery state information from the sensor noises, retaining only the relevant patterns for accurate anomaly detection.
Solution Approach 2:
The system creates a reconstructed version of the battery charging pattern from the latent variable and compares it with the original time series data. This copying approach allows the system to identify anomalies by detecting deviations from the learned normal pattern, effectively filtering out noise through the reconstruction comparison.
3Device complexity
If traditional battery diagnosis methods are used, then simple comparison of pre-test and post-test capacity is performed, but previously undetected abnormal patterns are missed
Solution Approach 1:
The encoder acts as an intermediary that transforms the raw time series data into a latent variable space where normal patterns can be learned and represented compactly. This intermediary representation enables the system to detect complex abnormal patterns that would be invisible in the original data space, while the decoder reconstructs the data to allow comparison and anomaly identification.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A battery diagnosis device according to one embodiment disclosed herein includes an encoder configured to convert time series data on a state of a battery into a latent variable, a decoder configured to convert the latent variable into first reconstruction data in a time domain, a first discriminator configured to discriminate the time series data with the first reconstruction data and output a result of the discrimination, a first calculator configured to compare the time series data with the first reconstruction data and calculate a reconstruction error, a second calculator configured to calculate an anomaly score based on a result of the discrimination and the reconstruction error, and a diagnosis instrument configured to diagnose the anomaly of the battery based on the anomaly score and a critical value.