Abstract：The insulated gate bipolar transistor（IGBT），as the core device of power electronics system，is

widely used in industrial control，transportation and new energy power generation due to its high efficiency and

high switching frequency. However，the internal bonding wire is vulnerable to thermal stress and current shock

during long-term operation，which has become one of the main reasons for IGBT module failure. A hybrid model

combining convolutional neural network（CNN），attention mechanisms，and long short-term memory（LSTM）

was proposed to accurately evaluate the health of bonding wire. Short-circuit current data were collected by cutting

the bonding wire experiment，and the health state was divided into three categories：healthy，damaged and faulty

based on the short-circuit current deviation. CNN was used to extract the local characteristics of the short-circuit

current，and the attention mechanism focused on the abnormal change of the key time step. LSTM captured the

time-sequence dependence of the short-circuit current，so as to realize the accurate classification of the failure state

of the bonding wire. The results show that the model has high classification accuracy on verification set and can

distinguish the different health states of bonding wire effectively. The research results provide scientific basis for

health monitoring and failure diagnosis of IGBT module，and have important engineering application value.

      Key words：insulated gate bipolar transistor（IGBT）device；bonding wire；convolutional neural network

（CNN）；long short-term memory（LSTM）network；health status assessment