Parallel lithium-ion battery connection fault diagnosis based on complete ensemble empirical mode decomposition and temporal convolutional network-attention

The rapid and accurate fault diagnosis has always been challenging in parallel battery systems. In particular, the non-measurable cell current and small voltage fluctuations make it extremely difficult to diagnose some faults that occur inside it. To solve this problem, a fault diagnosis method is proposed based on complete ensemble empirical mode decomposition (CEEMD) and temporal convolutional network enhanced by an attention mechanism (TCN-Attention). Firstly, the CEEMD algorithm is used to decompose the fault voltage into the intrinsic mode functions (IMFs). The IMFs with obvious entropy changes are chosen to reconstruct the voltage, and thus some weak fault information is enhanced. Subsequently, three types of connection faults are generated from a parallel equivalent circuit model, and the fault diagnosis model is trained by the TCN-Attention algorithm. The experimental results demonstrate that the fault diagnosis accuracy of the model reaches 98.37%, better than the models trained by other algorithms. Furthermore, the method's feasibility is demonstrated in different aging cycles, maintaining an identification accuracy of 96.02% even after 500 aging cycles. • The 2RC parallel model is built by loop current method to simulate three connection faults in different aging cycles. • The voltage signals are decomposed and reconstructed by CEEMD based on IMF entropy change for better fault diagnosis. • The TCN-Attention model is used to identify three connection faults with 98.37% accuracy, 96.02% across aging cycles.