Fault detection in software engineering is gaining attention due to the severe impacts of the faults in the software and malfunctioning of software used in diverse domains. Among the various faults, the class overlapping and the data distribution difference are one of the major issues that are required to be properly addressed. To deal with these issues, we propose a concept on cross-version software fault detection with data selection (CV-FDDS). This scenario is much reasonable and significant to detect the above-mentioned issues by predicting the faults in the new versions based on the labelled data of previous versions. The proposed framework works on identifying the faults in the newer versions of the software projects on the basis ofolder version labels through the use of learning strategies. Initially, the dataset is pre-processed using an appropriate data filter. The older and newer files are treated in two different ways to accomplish this task. To deal with the existing or the older files, an adaptive spiking atom search recurrent neural network classifier (ASARNN) has been proposed through which the training and testing data are automatically selected. This is done by assigning higher weight values to the much relevant and noise-free older versions. Based on the training and testing, the newer versions comprising faults are detected using the proposed kernel optimized extreme learning machine (KOELM). The implementation of the developed approach is implemented on PYTHON. The implemented approach performance is compared against the conventional methods with respect to the major performance indicators. The implemented nodel achievesa geometric-mean (G-mean) of about 0.7815, f-measure of about 0.775 and balance is about 0.8055 respectively.