基于卷积神经网络研究的应用

通过分析火控计算机的工作原理，选取了 12 个主要参数作为故障预测的输入， 针对输入数据的特性，选用了卷积神经网络（CNN）建立故障预测模型，针对传统卷积神 经网络模型存在的问题，加入批标准化层提高网络训练效果。通过实例验证，在数据充足 的条件下，针对训练弹的改进后卷积神经网络的火控计算机故障预测模型，预测准确率为 93.1% 化层 is added. Through practical verification, under the condition of sufficient data, the improved CNN-based fire control computer fault prediction model for training rounds achieves a prediction accuracy of 93.1%. (2) The convolution kernel in the CNN is transformed into atrous convolution. By applying this modification, the fault prediction model's generalization ability and accuracy are enhanced for the fire control computer. For the case of training rounds, the prediction accuracy of the model using atrous convolution reaches 96.9%, demonstrating its effectiveness in capturing patterns and improving performance. (3) To address the issue of limited data for armor-piercing rounds, transfer learning is integrated with the CNN model. Using the fire control computer fault prediction data from training rounds as the source domain and armor-piercing rounds as the target domain, the CNN model undergoes transfer learning adjustments. The prediction accuracy of an unmodified CNN model for armor-piercing rounds, under data scarcity, is 82.1%. However, when incorporating transfer learning, the accuracy improves significantly to 96.1%, highlighting the value of transfer learning in adapting the model to new scenarios. (4) A user-friendly fire control computer fault prediction software is developed, featuring a human-computer interaction interface that streamlines the process of adding, deleting, and conducting fault prediction data operations. This software enhances the efficiency and usability of the fault prediction system. In conclusion, this research delves into the application of CNNs in the field of fault prediction for fire control computers. By analyzing the system's working principles, selecting appropriate input parameters, and employing advanced techniques like batch normalization, atrous convolution, and transfer learning, the model's performance is significantly improved. The study also contributes a software solution that simplifies data management and prediction tasks. These advancements contribute to the reliability and readiness of military equipment, ultimately supporting the nation's defense capabilities.