Increasing recognition of the impact of shipping on air pollution has led the International Maritime Organization (IMO) to establish Sulfur Emission Control Areas (SECA) to reduce emissions. Within SECA, ships must switch to low-sulfur fuel or use a scrubber technique to clean their exhaust gases. Conventional monitoring methods are limited by detection range, real-time data availability, and challenges in source attribution. This study describes a monitoring system that combines single-particle mass spectrometry (SPMS) with deep learning to overcome these shortcomings. SPMS can reveal the chemical composition of individual airborne aerosol particles, with the capability to detect emissions over several kilometers, enabling real-time pollution source identification. To automatically process the complex mass spectral data, a convolutional neural network (CNN) was designed, achieving 92 % accuracy in classifying 13 distinct classes of abundant aerosol particles. The results demonstrate that the proposed detection system enables to automatically classify aerosol particles from multiple sources. Of particular concern in this study is the in-situ analysis of particles from ship exhaust plumes, to rapidly identify ships running on polluting heavy fuel oil. Focusing on unique particles containing vanadium (51V+/67[VO]+), nickel (58/60Ni+), and iron (54/56Fe+) ions, designated as V-rich class, the real-time classification makes it possible to reliably detect particles from heavy fuel oil (HFO) combustion. In addition, to locate the emission sources, the CNN's predictions are linked to local wind measurements and ship trajectories provided by the Automatic Identification System (AIS). During a one-week monitoring period, 21 ships passing the measurement site 80 times in distances of up to ∼1.3 km were detected using HFO.