Enterprises are facing unprecedented challenges in computerized data processing, especially in the efficient processing and accurate prediction of multidimensional time series data. To solve this problem, based on Faster RCNN algorithm, this paper studies feature extraction network in depth, and compares three data processing methods: ZF-Net, ResNet and ResNet (DWSR) model based on depth separable convolution. First, the DWSR model optimizes the traditional ResNet architecture with deep separable convolution, giving it greater ability to capture multi-dimensional features. Compared with ZF-Net and ResNet, DWSR is not only more complex in structure, but also enhances the flexibility and adaptability of the model by introducing multi-convolution kernel topology. This design enables DWSR to extract information more efficiently in the face of diverse inputs. DWSR also adopts the packet convolution technology, which not only achieves high accuracy, but also significantly improves the calculation efficiency of the model. Compared to traditional ResNet, DWSR maintains excellent performance while reducing parameters. Especially in the aspect of multi-dimensional feature extraction, DWSR shows better performance than ZF-Net and ResNet. In addition, in order to facilitate information transfer between different groups, an improved channel mixing mechanism is proposed, which further enhances the feature fusion effect. The model also uses Mish activation functions to prevent the collapse of low-dimensional data, thus ensuring that features are efficiently expressed. The experimental results show that the Top 1 error rate and Top-5 error rate of DWSR model are lower than that of traditional ResNet on various practical and public benchmark data sets, and the complexity of the model is reduced. Compared with existing models such as ZF-Net and ResNet, DWSR not only performs more prominently in multidimensional feature extraction, but also shows obvious advantages in overall efficiency. This proves that DWSR can provide more accurate and efficient solutions in processing multi-dimensional feature extraction tasks of enterprise computerized data.