Firmware has become a critical attack surface in industrial control systems (ICS), where vulnerabilities can enable privilege escalation, remote control, or persistent backdoors in embedded devices. Accurate semantic modeling of firmware functions is therefore essential for effective vulnerability detection and similarity analysis across architectures and optimization levels. This paper proposes BiAttnNet, a bidirectional attentive semantic feature extraction framework for firmware function representation. The model first employs the Skip-Gram algorithm to embed instruction streams while preserving operand-related semantics through instruction normalization. It then integrates a BiLSTM network to capture bidirectional dependencies within long instruction sequences, followed by a self-attention mechanism that adaptively focuses on critical instructions, enhancing the discriminative power of the semantic representation. Compared with existing methods, BiAttnNet demonstrates superior robustness and generalization in cross platform firmware analysis. More importantly, the proposed framework effectively bridges low level instruction semantics with global function understanding, laying a foundation for future semantic structural fusion in large scale firmware vulnerability detection.