Nowadays, the software is used almost in every field. The software industry is evolving rapidly, and developing quality software is challenging for developers. Defects in the software are the primary reason for degrading the quality. Software defects prediction is a difficult task. Though researchers have developed several methods to overcome this issue that support the problem, there is room for improvement in software defects prediction. The curse of dimensionality occurs when research is undertaken without feature reduction. Therefore, after consulting several other works of a similar caliber, a hybrid machine learning approach that combines Principal Component Analysis (PCA) for feature optimization with Multi-layer perceptrons for classification has been designed to overcome the hurdle. The proposed model reduces the time complexity of the traditional methods by reducing the number of attributes used for classification. We have used the PROMISE dataset KC1 & CM1 repository from the NASA directory (KC1 2109 observations & CM1 344 observations) to perform our research work. We have divided the datasets into training (1476 from KC1 & 240 from CM1 observations) and testing (633 from KC1 & 104 from CM1 observations). We have employed precision, recall, F-measure, and accuracy in evaluating the model. Using the proposed model, we reduced the time complexity and increased the accuracy by 6% using the KC1 dataset and the previously proposed models. In addition, the proposed model receives a greater score on additional criteria. The primary limitation with Multi-layer perceptrons is that it requires a linearly separable dataset, which can make classifications relatively inflexible. Soon, several new strategies may be introduced to overcome this limitation and maintain a soft-based categorization margin.