Internal short circuit failure may cause thermal runaway, which poses a huge threat to the safe operation of lithium-ion batteries. Therefore, it is crucial to conduct research on internal short circuit diagnosis. This paper introduces an internal short-circuit diagnosis method comprising a voltage prediction approach based on temporal convolutional network and a residual evaluation method based on sliding window cumulative summation. Fault diagnosis can be accomplished by assessing the residual discrepancy between the predicted and measured voltages. This method only requires normal battery data to train the model without high-quality fault data, which is practical and feasible. Experimental verification confirms that the method can swiftly and accurately detect faults.