Abstract: To diagnose the issues related to ammonia engine ignition failures and misfires caused by difficulties in ignition and unstable combustion， a diagnostic method based on instantaneous speed signals integrating improved ensemble empirical mode decomposition（EEMD） with image recognition technology was proposed. The instantaneous speed signal was decomposed into several intrinsic mode functions（IMFs） using improved EEMD. By combining fast Fourier transform（FFT） and Pearson correlation coefficient（PCC）， the IMF11 component sensitive to faults was selected to identify ammonia ignition failure （ammonia fuel fails to be ignited by the pilot diesel） and misfire （both diesel and ammonia fail to combust） faults. Furthermore， the IMF8 time-domain component was converted into polar coordinate images， and the lightweight convolutional neural network EfficientNet-B1 was employed for training， ultimately achieving precise localization of the cylinders with ammonia ignition failure and misfire. The results showed that under normal engine operation， the energy of IMF11 was weak， whereas ammonia ignition failure or misfire faults significantly amplified this low-frequency component. The frequency-domain amplitude of IMF11 could serve as a key feature for identifying ammonia ignition failure and misfire faults. The main frequency of the IMF8 component was closest to the rotational frequency of the ammonia engine. Converting the time-domain signal of the IMF8 component into two-dimensional polar coordinate images clearly revealed the firing phase information. The study demonstrates that the lightweight EfficientNet-B1 model enables precise localization of the faulty cylinders affected by ammonia ignition failure or misfire in medium-speed ammonia engines.