In the context of new quality productive forces, enterprises must leverage technological innovation and intelligent management to enhance financial risk resilience. This article proposes a financial distress prediction model based on deep learning, combined with a CNN, BiLSTM, and attention mechanism, using SMOTE for sample imbalance and Hyperband for hyperparameter optimization. Among four CNN-BiLSTM-AT model structures and seven mainstream models (CNN, BiLSTM, CNN-BiLSTM, CNN-AT, BiLSTM-AT, CNN-GRU, and Transformer), the 1CNN-1BiLSTM-AT model achieved the highest validation accuracy and relatively faster training speed. We conducted 100 repeated experiments using data from two companies, with validation on 2025 data, confirming the model’s stability and effectiveness in real-world scenarios. This article lays a solid empirical foundation for further optimization of financial distress warning models.

As a key pillar of China’s economy, the manufacturing industry faces sustainable financial risk management challenges as it undergoes digital and green low-carbon transformation. However, existing financial risk prediction models often suffer from limited accuracy, insufficient robustness, and a suboptimal activation function design. In this study, we investigated advanced deep learning architectures to address these limitations, and we introduced a novel composite triple activation function (CTAF) framework to enhance predictive performance and model robustness. We began by evaluating several deep learning models, such as CNNs, BiLSTM, CNN-AM, and BiLSTM-AM, demonstrating that CNN-BiLSTM-AM achieved the highest performance. On the basis of this model structure, we proposed a CTAF, a composite activation mechanism that combines two distinct functions applied to the raw input x, effectively mitigating gradient instability and enhancing nonlinear expressiveness. Through ablation experiments with different composite activation functions, we verified that the CTAF consistently outperformed alternatives. Meanwhile, the mainstream activation functions and CTAF were applied to different layers for comparison, further verifying the CTAF’s advantages in various structures. The optimal configuration was achieved when tanh was used in the CNN and Dense layers and the CTAF (tanh_relu) was applied in a Lambda layer after a BiLSTM layer, resulting in the highest accuracy of 99.5%. Furthermore, paired t-tests and evaluations on cross-industry datasets confirmed the optimal model’s stability and generalizability.