To improve the accuracy of circuit chip fault diagnosis, the efficiency of hyperparameter setting and the efficiency of feature extraction, an improved arithmetic optimization algorithm (IAOA) based on temporal pattern attention mechanism (TPA) is proposed to optimize the bi-directional long and short-term memory network (BiLSTM) for circuit fault diagnosis. Firstly, IAOA is employed to search for the optimal hyperparameter combinations of BiLSTM to improve the diagnostic accuracy of the model. Then TPA is used to extract important features and assign weights to enhance the model feature extraction capability. Finally, the infrared temperature data collected by the infrared camera is inputted into the optimal diagnostic model to achieve circuit board chip fault diagnosis. The experiments are verified by using 0~30 V adjustable regulated power supply circuit board. The results show that the model for circuit chip fault diagnosis is as high as 98.27%, which can achieve high accuracy fault diagnosis for circuit board chips.