Although machine learning-based intelligent detection methods have made many achievements for diesel engine misfire diagnosis, they suffer from a certain degree of performance degradation due to the strong environmental noise and the change of working conditions in the actual application scenarios. To tackle this issue, this study presents a deep twin convolutional neural networks with multi-domain inputs (DTCNNMI) for diesel engine misfire diagnosis under strong environmental noise and different working conditions. Vibration signals of engine cylinder heads at different speeds are collected experimentally and fed into input layers of the proposed model. It constructs three input layers to combine automatically extracted time-domain, time-frequency-domain and hand-craft time-domain statistical features, improving the model performance. Twin convolutional neural networks with large first-layer kernels are presented for extracting multi-domain information of vibration signals and resist influence of environmental noise and the change of operating conditions on the final diagnosis results. The effectiveness of the proposed approach is validated on the datasets collected experimentally and by making comparison with the existing representative algorithms. The results demonstrate that the accuracy and anti-noise performance of the proposed DTCNNMI outperforms the existing algorithms. On the constructed four datasets under different working conditions, the proposed method achieved at least 97.019% accuracy even when signal-to-noise ratio was −4 dB, which is much higher than those of the other methods. Moreover, the accuracy of the proposed DTCNNMI on datasets A, B, C and D is at least 61.177%, 33.334%, 17.646% and 29.961% higher than other methods, respectively.