Identification of Crystallization State of Ammonium Nitrate Aqueous Solution Based on EfficientNet Optimization Algorithm

The quality of aqueous ammonium nitrate solutions is a crucial factor affecting the production quality of industrial explosives. Accurate identification of crystallization states and precise recording of crystallization temperatures in these solutions are of paramount importance during the manufacturing process. Convolutional Neural Networks (CNNs), leveraging the principle of "translation invariance" in images, can effectively learn salient features, making them suitable for addressing the problem of automatic detection of crystallization states in aqueous ammonium nitrate solutions. This study proposes EfficientNet-blast, an optimized algorithm based on EfficientNet, specifically designed for this purpose.EfficientNet-blast utilizes deep neural network architecture, EfficientNet, with integrated Squeeze-and-Excitation (SE) modules as its foundation. By increasing the number of feature channels in each layer while strategically removing multiple MBConv layers in depth, the model reduces the number of parameters, lowers FLOPS (Floating Point Operations), and accelerates detection. After training on a curated dataset of crystallization states, EfficientNet-blast achieves a remarkable recognition accuracy of 99.83% thanks to its improved convolutional and pooling characteristics.A comparative analysis of EfficientNet-blast with existing image recognition algorithms such as ResNet-50, MobileNetV2, ShuffleNet, DenseNet, EfficientNet, and Swin Transformer, using the same crystallization state dataset, reveals that EfficientNet-blast comprehensively enhances accuracy while maintaining a relatively low parameter count and without significantly increasing FLOPs. It sustains a high detection speed, with FPS (Frames Per Second) consistently at a level suitable for real-time monitoring. The experimental results show that the algorithm EfficientNet-blast can quickly and accurately identify the crystallization state, and the measurement deviation of the crystallization point temperature is less than or equal to ±0.3℃. The measured crystallization point temperature dispersion degree is small, the measurement fluctuation value is small, and the stability is high. This study offers a practical solution for enhancing quality control in industrial explosives production and holds promising applications in related fields.