Thermal Fault Detection of Lithium-Ion Battery Packs Through an Integrated Physics and Deep Neural Network Based Model

A variety of faults and failures may occur in a battery pack over time, many of which are hard to determine based on conventional measurements. For example, a blockage in the cooling system would result in higher temperatures, but a fault would typically only be triggered if temperatures exceed a protection limit. This work proposes a model based method for detecting and identifying thermal faults in a battery pack well before any protection limits are reached. The method can detect a failed temperature sensor, a failed or incorrectly operating coolant pump, or a blockage of coolant for a single battery module. Faults are determined through an algorithm which compares measured and modeled temperatures and classifies the fault type accordingly. The core of the algorithm is a high accuracy temperature estimation model, which is realized through a physics based thermal model with a feedforward neural network. To experimentally demonstrate the concept, the algorithm is applied to a 72 cell air cooled battery pack with eight temperature sensors. Faults were applied to the pack during electric vehicle drive cycles and the algorithm was able to detect and identify each fault type within about ten minutes. In practice, the algorithm could trigger a warning light in the vehicle and report an error code describing the failure, much like a cylinder misfire error for an internal combustion engine vehicle. This would allow the driver to have the vehicle serviced before the fault resulted in a condition which makes the car undrivable. This work goes beyond prior studies in this area by proposing a new battery pack thermal modeling method and by focusing on difficult to identify faults, such as that of a single battery module.

The proposed statistical battery pack fault detection and identification method is based on an integrated thermal lumped parameter and feedforward neural network (LP+FNN) temperature estimation model. The LP+FNN model is benchmarked against more conventional physics-based lumped parameter (LP) and feedforward neural network (FNN) temperature estimation models. The development of the proposed LP+FNN models is discussed, then the proposed model is compared to the LP and FNN models using healthy drive cycles at various thermal and driving conditions. Next a statistical-based fault detection and identification method is built using the proposed LP+FNN thermal model. Finally, the proposed fault detection and identification method is validated at different thermal faults.

The concept of the integrated LP+FNN model is introduced and is used to model the temperature of an air-cooled 72-series battery pack with 5Ah cell capacity . The LP model part mimics the heat transfer equations in the pack. On the other hand, the integrated LP+FNN model is used to model the temperature for each cell. A correlation analysis is used to select the best inputs for the proposed LP+FNN model, including filtered current with 1 mHz first order butter worth filter corner frequency, cell SOC, air inlet temperature , and the LP model cells’ temperature rise . The FNN model shares a similar structure and inputs to the LP+FNN model, except not using the LP cells’ estimated temperatures . The proposed thermal LP+FNN model is trained and benchmarked versus LP and FNN models using fault-free standard test cases, including driving and charge profiles ranging from 4 to 10 C-rate . The LP+FNN shows significantly better temperature estimation accuracy than the LP and FNN, especially for aggressive thermal conditions such as US06&6C testing profile at 15 _C . Overall, the integrated LP+FNN model shows promising accuracy with less than 1 ºC average RMSE, typically 33%, 76% less than LP and FNN, respectively, and 2 ºC MAXE for all studied test cases and cell.

Finally, a fault detection and identification method is proposed by monitoring the LP+FNN estimated and measured temperature of eight cells in the 72-series air-cooled battery pack . The eight temperature sensors are selected on the two modules dividing the pack into six submodules (SMs). The method works by monitoring the residual between the measured and estimated temperature and recording fault flags in each sensor based on a cumulative log-probability function. Then based on the number and nature of the recorded flags, the fault type is determined.

One test case includes a long driving scenario containing a series of US06&6C, UDDS&4C and LA92,8C. This test includes fan off, normal and higher than setpoint fan speed faults. After 24 minutes from the beginning of the test, the method shows the first high flag, and within 10 minutes of a waiting window, the proposed method is able to detect eight high flags in the system. Since five or more high flags are recorded in the system, a fan failure or low airflow fault is declared. The fault is then cleared after one hour from the beginning of the test, and the residuals go back to the normal thresholds. Then the fan speed is increased after approximately one hour from the fan-off fault clearance. Again, the proposed method takes nine minutes to record the first low flag, and after ten minutes of the waiting time window, eight low flags are recorded, declaring airflow lower than the setpoint fault. A repeated HWFET and 10C charge at 15 _C ambient with the airflow blocked on submodule#3 containing cells#24 to 36 is also used to test the proposed method. The proposed method is also able to detect high flags in both cells after 22 minutes from the beginning of the test.