Abstarct: With the growth of the software industry, the focus has shifted towards enhancing the quality and reliability of software products. Software testing is commonly used for this purpose; however, testing all possible inputs of a software due to time and cost constraints is impractical. To address this issue, the utilization of software error prediction systems has emerged. The goal of software error prediction is to identify software modules prone to errors baxsed on certain underlying features of the software project before the actual testing process begins, thus aiding in achieving the desired software quality with optimized cost and effort. This is crucial as software errors can have serious consequences on system performance and reliability. One challenge in the field of software error prediction is that previous models often worked on old, small, or limited datasets, resulting in limited effectiveness for modern software. To tackle this issue, this research utilizes the BugHunter dataset, comprising 15 open-source Java software projects, extensively used by software engineers. Various machine learning techniques such as decision trees, naive Bayes, logistic regression, random forest, etc., have been employed to detect software errors in this dataset. However, the next challenge faced is the low accuracy of existing learning models on this dataset. This research aims to enhance the performance of error prediction models by leveraging neural network techniques and convolutional network architectures to improve accuracy. Using convolutional network architectures, important features and error patterns are extracted from the training data. The model is capable of automatically detecting software errors and predicting their likelihood of occurrence. To prevent model bias, an up-sampling technique is employed. The best F1-score obtained from the implementation of the proposed model on the BugHunter dataset is 89%. Evaluation results demonstrate a 2% improvement over the best previous prediction model.