Extending electronic shot counters with machine learning-based ammunition discrimination for targeted maintenance

Purpose - Weapon fleets degrade differently depending on usage. Electronic shot counters provide armorers with accurate tracking and allow more efficient preventive maintenance practices. We propose a machine learning technique for general-purpose ammunition type discrimination, thereby enhancing targeted maintenance. Design/methodology/approach - We study an experimental shot counter deployment to understand its impact on maintenance practices. We then extend the existing EDGAR machine learning technique to solve the discrimination problem in a generic way on a weakly-labelled dataset, requiring only the total counts for each shot type. By repurposing intermediate neural network activations, we simplify training and minimize computational overhead. We evaluate our approach against a widely used live/blank discrimination algorithm. Findings - We show a 94% improvement in instance-level error rate and perfect burst-firing performance. This generalizes across weapon platforms without hyperparameter adjustment. The feature incurs as little as an 8% overhead (2.8 ms on a 64 MHz ARM Cortex-M4F). Originality/value - Compared to existing techniques, it promises applicability to a broader range of weapon configuration discrimination tasks, including platforms previously deemed too complex or constrained. Additionally, we examine how performance scales with dataset size to offer practical data collection guidelines, a major challenge in this field. This technique supports a new generation of shot counters and targeted maintenance, thereby reducing costs, preventing incidents and increasing operational availability.