To improve the intelligent operation and maintenance of ships, the research on the perception of engine room auxiliary equipment (A/E) will realize AUTO-0 and replace the engineers. Due to compact layout of the cabin and the different shapes of equipment, the Single Shot MultiBox Detector (SSD) that uses the traditional convolution method for feature extraction cannot adapt to the complex environment. Therefore, an improved A/E detection algorithm based on SSD is proposed. Firstly, we collected the images of actual ships, annotated them manually, and used the K-Means clustering algorithm to analyze them. Secondly, we replaced the classification loss of SSD with the focal loss, balancing the positive and negative samples in the dataset. Finally, we added repulsion loss for the overlapping targets to improve the detection effect of the model against dense occlusion & overlap. The experiment is based on the dataset containing 18185 training object boxes, and 2044 test object boxes, the mean average precision (mAP) in this paper reaches 78.95%, which is 5.63% higher than the original SSD, and the detection boxes are more suitable.