Makine Öğrenmesi Yöntemleri ile İmha Değerlendirme
Özet
Kill assessment is the ability to decide whether the relevant threat has been hit or not as a result of an initiated engagement, and to evaluate whether the initiated engagement has been successful or not. The destroyed-not destroyed information specified for the threat is critical for the proper management of user attention, costs and engagement. This detection is so far from the ranges and magnitudes that can be observed by the human eye today. In the kill assessment, it is aimed to make decisions independently of the human figure in the decision-making process by using machine learning and deep learning techniques. Thus, the assessment reduces errors and waste of resources.
In this study, which was carried out on the assessment of kill using machine learning and deep learning algorithms, a simulator was developed that generates radar signals as synthetic data and simulates physical events such as kill - miss, due to the lack of real-world radar data. The use of raw radar data as an input to classification algorithms has not been found appropriate due to disadvantages such as being susceptible to noise and overfitting, insufficient equipment in the training process and long training time. For this reason, these features are provided as input to the algorithms by performing feature extraction processes. These feature types are divided into three as the perfect feature set, unperfect feature set and the ambiguity graphs feature set.
The kill assessment process is expressed by dividing these three classes, which indicate different situations: Pre Hit, Miss and Kill. Kill is an event that takes place in the order of instantaneous or milliseconds by its nature. For this reason, when the balance in the data set of the classes is supervised, the data density of the moment before the hit or the moment of the miss is much higher when compared to the data density of the moment of kill. For this reason, various methods that provide class balance were investigated and the training process was carried out with feature sets that provided class balance. SMOTE, ROS, RUS, Near-Miss, SMOTE-ENN, SMOTE-Tomek methods were used to provide class balance.
Kill assessment algorithms were tested by using unbalanced and balanced data sets as inputs, respectively. By comparing which method is more accurate and best in terms of performance and accuracy in making a correct kill decision by using adaboost, gradientboosting, decision tree, random forest, support vector machines, multi-layer perceptron from machine learning algorithms and neural networks and LSTM from deep learning algorithms detected.
When algorithm performances are analyzed, it was concluded that the Random Forest algorithm performed the most successful classification with a rate of 99 % in all three ROS applied data sets. It has been seen that ensuring class balance is a factor that increases performance in all classification algorithms. As a result of this thesis, it has been concluded that a successful kill assessment can be made by using the features extracted from the radar signals with artificial intelligence algorithms.
