Table-to-text generation aims to produce coherent and natural language descriptions from structured tabular data while maintaining semantic integrity. This process often requires converting the table into a linear sequence to facilitate model processing and learning of its patterns. However, existing studies overlook a critical characteristic of table linearization, namely that the same table can be linearized into different sequences. Recognizing this characteristic, we incorporate a multi-task learning mechanism into table-to-text generation. Leveraging the inherent heterogeneity of linearized table sequences, we propose a Structure-focused Enhancement Transformer (SfE-Transformer) architecture based on multi-task learning. The framework incorporates an auxiliary encoder and a structural decoder. The auxiliary encoder processes heterogeneous sequences from the same table in parallel with the primary encoder, thereby enriching the decoding process with enhanced structural and content-aware representations. The structural decoder connects to the primary encoder to perform table sequence transformation tasks, effectively augmenting the primary encoder's structural perception capabilities. Experiments on the ToTTo and HiTab benchmarks demonstrate that our framework outperforms most baseline methods, thereby validating its effectiveness.