Software Fault Prediction (SFP) is a crucial strategy in software quality assurance, aiming to reduce costs and optimise resource allocation by identifying potentially faulty software modules based on key project characteristics. Despite significant advancements, SFP methods often reach a ‘performance ceiling’ due to the constraints imposed by the limited size of datasets available from public repositories and the difficulties in selecting optimal software metrics for specific application domains. Furthermore, while traditional machine learning approaches have predominantly been used to predict fault proneness, the possibilities offered by more sophisticated techniques like Deep Neural Networks (DNNs) remain underutilised. This study introduces a novel application of DNNs, trained using Error-type Metrics, to predict faults in open-source software projects. These metrics are application-independent and have been shown to enhance predictive accuracy by enriching the training data with a broader range of information, thus helping to break through the existing performance limitations. Our empirical findings indicate that models trained with Error-type Metrics significantly outperform those using traditional CK metrics, achieving improvements of up to 40% in AUC and ROC scores. Our method outstrips even the latest DNN models incorporating advanced self-attention mechanisms, achieving performance gains of up to 17.86%.