[ Paper Review ] Capacity Estimation Through Knowledge Transfer

Capacity Estimation of Lithium-Ion Batteries for Various Aging States Through Knowledge Transfer

• Journal : IEEE Transactions on Transportation Electrification ‘21 (IF 7.2)
• 17 citations

 

[ summary ]

• electrochemical knowledge를 학습하는 inverted bottleneck network(IBN)을 기반으로 리튬 이온 배터리의 aging states에 대한 새로운 capacity estimation 방식 제안
• 물리적으로 가능한 파라미터 조합에 대해 sythetic data 생성 (배터리의 특성 정보를 knowledge transfer) → AEHS 알고리즘 사용하여 파라미터 최적화
• IBN 학습
  → input : voltage/current/temperature profile
  → output : estimated capacity, attention map
• Estimation score로 estimation의 신뢰도 판단
  → input : volotage profile, attention map
  → output : estimation score

---

 

1. Problem Statement

[ Capacity Estimation Method ]

• Nonphysiochemical battery model
  • Equivalent circuit model (ECM), the electrochemical impedance spectroscopy (EIS)-based model, other empirical algorithms

→ Limitations in expressing electrochemistry-based aging states of the batteries

• Physicochemical battery model
  • Recently developed
  • Require considerable computational resources

 

 

[ Proposed Model ]

• Fully convolutional network (FCN), which is referred to as an inverted bottleneck network (IBN)
  • Increase memory efficiency and reduce computational complexity
  • Attention map is used for the interpretation, which visually expresses the points where the neural network focuses on the input data

 

 

2. Contribution

• Accurate capacity estimation of various aging states of batteries with knowledge transfer to memory-efficient neural network
• Estimation score based on input profile-attention map analysis for providing the reliability of the estimation results

 

 

3. Methodology

• Electrochemical knowledge transfer through generating synthetic data
• Get estimated capacity & attention map though IBN model
• Validate reliability of estimation through estimation score using attention map

 

 

1️⃣ Generate Synthetic data to transfer the knowledge of electrochemical lithium-ion battery model

• Target battery : NMC battery || parameter estimation algorithm : AEHS algorithm
• Generate synthetic data evenly over the valid model parameter space
  • Parameter space is determined to cover the dynamics of the target battery
• Aging states of the battery can be represented by the combination of the model parameters
  • ⇒ provide a diverse representation of the effects of the aging states

 

[ Parameter sampling equation ]

: Specify the sampling range close to the dynamics of the real battery

 

Interpretation of Eq.5 

• When the sign of k is positive, it is assumed that the value of that parameter decreases with aging 
• Variation rate F is independently assigned to each parameter as a fixed value 
• Larger D_i indicates accelerated parameter variation → high strength of the i-th aging mechanism 
• Sampling in proportion to $\Delta{Cap}$ → to get more diversely and distinctly synthetic data 

 

[ Implement ]

• Current Input driving profile
  • Urban dynamometer driving schedule (UDDS)
  • Worldwide harmonized light vehicle test procedure (WLTP)
  • Inspection and maintenance driving cycle (IM240)
• Diverse & dense synthetic data → obtain input, output profile pairs at different initial SOCs and capacities
• Data generation result
  • UDDS : 257×10^4, WLTP : 314×10^4, IM240 : 36×10^4

 

 

2️⃣ Get estimated capacity, attention map using IBN model

: Propose neural network-based estimation scheme with sufficiently small memory size

 

 

[ IBN model Structure ]

• IBN consist of
  • (1) convolution layers
  • (2) inverted bottleneck layers : pointwise - depthwise - pointwise convolutional layers
    (→ memory efficient) (Fig. 3)
  • (3) global averaging pooling (GAP) layer
• Batch normalization layer follows each pointwise and depthwise convolutional layer → for efficient training and regularization
• Stochastic gradient descent with a warm restart (SGDR) method → for an effective learning rate setup
• # epoch : 300 / minibatch size : 256
• Bottleneck Layer (from MobileNetV2)

 

 

[ Workflow of IBN ]

• Input profile : 1-D images composed of three channels
  • normalized voltage value, normalized current value, normalized temperature values
• Output : Estimated capacity, corresponding Attention map
  • Attention map highlights the points that the neural network notes as important

 

 

[ Capacity Equation ]

: Obtained from GAP layer

 

 

 

[ Attention map Equation ]

: Represent the points that the neural network focuses

: Obtained from the last convolutional layer and weights

 

 

 

 

3️⃣ Measure reliability using an estimation score

When the input profile has more excitation components, more information is available to aid the estimation

→ reliability of capacity estimation result can be drived indirectly

 

 

[ Estimation score equation ]

: Represent the correlation between feature of the input profile and the interest of the neural network

 

 

 

4. Experiment Result Anaylsis

 

[ Experiment Setting ]

 

 

 

 

Experimental Data의 80%는 neural network 훈련에 사용 / 20% 만 test 데이터로 사용

 

 

4.1 Capacity Estimation for Synthetic Data

 

[ Capacity Estimation with IBN ]

• Object : proposed neural network can accurately estimate the capacity by learning only synthetic data
• 80% for training / 20% for testing
• Result : proposed neural network provides good estimation accuracy for the test dataset
  • Mean absolute error (MAE) of 0.052 Ah
  • Root mean square error (RMSE) of 0.073 Ah

 

[ Visualize the decision basis of neural network ]

 

 

• Graph : voltage, current, temperature profile
• Band : corresponding attention Map

→ Trained neural network tends to focus on the rapidly changing areas in the input profiles

 

 

 

 

[ Estimation Score With Attention Map ]

• Mean value of estimation error is almost zero
• 99% confidence interval gradually decreases with an increase in the estimation score

⇒ capacity estimation is more reliable when the corresponding estimation score is high

 

 

 

 

 

4.2 Capacity Estimation for Experimental Data

 

[ Capacity Estimation with IBN ]

• #12 37Ah-standard NMC batteries with a capacity from 40.46 to 24.51 Ah

 

• Blue : only learned the experimental training data without the knowledge transfer
• Red : learned the experimental training data + synthetic data (knowledge transfer)
  • Improved results achieved
    • Mean value of the estimation approaches to the reference capacity
    • 95% confidence interval of the result also narrows

 

[ Estimation Score With Attention Map ]

• Error probability distribution had a smaller variance for increasing the estimation score (same with synthetic data)

 

 

[ Performance of IBN and Knowledge Transfer ]

• Model with superscript “1” : use only experimental data
• Model with superscript “2” : use both synthetic + experimental data

• Proposed neural network had the smallest MAE, RMSE
• Electrochemical information was injected more into the neural network by additional training with the synthetic data → ‘knowledge transfer with synthetic data’ is effective

 

[ Computational Burden ]

• IBN was implemented more efficiently than the electrochemical model in the same CPU environment

 

5. Conclusion

• Propose a new neural network-based capacity estimation scheme based on knowledge transfer
• IBN accurately estimate the capacity of different aging states within a short computation time
• Attention map from the IBN indicates the features of the given voltage profile that is informative for capacity estimation.
• Estimation score was calculated by considering the attention map and input voltage profile to determine the reliability of the capacity estimation result

 

---

Background Knowledge

 

[ Electrochemical Model ]

: Represent the internal states and electrochemical reactions of the battery

 

• Pseudo-2-D (P2D) model
  • x-axis : flow direction of lithium-ion | pseudo-r-axis : radial direction of solid particle

 

 

[ Geometric & Electrochemical parameters ]