Detecting faulty lithium-ion cells in large-scale parallel battery packs using current distributions

Figures

• 

  Fig. 1. Schematic of the cell-level battery model used to build the packs. a Equivalent circuit model and b module of n cells connected in parallel.

• 

  Fig. 2. Discharging behaviour of three cells connected in parallel. Simulated currents (a) and states of charge (b) for three parallel-connected cells during a 1C discharge.

• 

  Fig. 3. Simulated currents, state-of-charges, and voltages for 74 parallel-connected cells during a 1C discharge. A Fresh Tesla Model S module: currents (a), states of charge (b), and voltages (c). Each line corresponds to a cell. B Aged Tesla Model S module: currents (d), states of charge (e), and voltages (f). Each line corresponds to a cell. C Pack with one faulty cell: currents (g) and states of charge (h), with the solid cyan line indicating the defective cell among 74 cells.

• 

  Fig. 4. Discharge simulation results of two parallel-connected DFN models using the parameter values in Table 1 and Table 2. Case 1 corresponds to the nominal setup, with Rctc = 5.7 × 10−4 Ω for cell 1 and Rctc = 7.7 × 10−4 Ω for cell 2. Case 2 additionally has an anode reaction rate coefficient k three times greater in cell 2 than cell 1. Case 3 has an anode active particle radius three times greater in cell two than cell one and the fourth row is plotted every 25 seconds, with dark lines corresponding to the start of the simulation and light lines to the end. a–c are the voltage responses for the three cases, d–f are the branch current densities, g–i are the anode reaction rates in time, and j–l are the distribution of the reaction rates in space across the anode.

• 

  Fig. 5. The impact of cell faults on pack-level current distributions. Comparison between the current distributions of a faulty pack (a) and the same pack without the faulty cell current (b) during a 1C discharge. Each line corresponds to a cell, with the solid cyan line in (a) corresponding to the defective cell.

• 

  Fig. 6. Change in pack-level current distributions as the faulty cell’s resistance increases. Simulations of Tesla Model S battery pack module when all cells are healthy (a) and when one of the cells is faulty (b). The value of the faulty cell resistance (rfaultycell) varies between cases b.1, b.2 and b.3. It is defined as a multiple of pack mean resistance μr, and significantly impacts pack current distribution and hence the performance of the fault classification algorithm.

• 

  Fig. 7. Histograms of feature for healthy and faulty packs of the simulated dataset. a gives the distribution of feature f1 defined in (11), and similarly b in (12), c in (13), d in (14), e in (15), and f in (16).

• 

  Fig. 8. Accuracy score of fault-detection classifier (SVM) as a function of Ns (the number of current sensors within the 74p1s parallel pack). For each number of sensor plotted, 50 classifiers were trained, with the crosses representing the median accuracy and the bars representing the standard deviation of the accuracy across the tests.

• 

  Fig. 9. Application of SVD classifier to detect faults in experimental pack data with four RJ-LFP72174204-280 cells connected in parallel. As detailed in Methods Section, four fault cases were considered, with Cases 1–3 being faulty and Case 4 fault free. a–d show experimental current data, e–h show scaled currents, i–l show fault predictions with Ns = 1 and m–p show fault prediction with Ns = 3.

• 

  Fig. 10. The open circuit voltage curve used for the Panasonic NCR 18650B cells. Comparison between the digitised OCV curve from ref. 58 (blue) and its polynomial interpolation (red).