With the rapid proliferation of the Industrial Internet of Things (IIoT) as the foundation of Industry 4.0, ensuring secure and efficient data handling has become crucial. Federated Learning (FL) has emerged as a promising solution by (1) enabling distributed model training across multiple heterogeneous IIoT devices and (2) preserving data privacy by keeping sensitive information local while only sharing model updates with the central FL server. However, FL systems remain vulnerable to security threats such as targeted data poisoning attacks, particularly label-flipping attacks, which significantly degrade global model performance as the proportion of malicious clients increases. In this work, we investigate the impact of label-flipping attacks on IIoT FL systems and address the limitations of previous studies, which lack hybrid defense mechanisms tailored to IIoT use cases. We propose a novel hybrid defense approach that integrates the Elliptic Envelope Algorithm (EEA) with quorum voting to effectively detect and filter out poisoned model updates from malicious clients during FL aggregation. To evaluate the effectiveness of our method, we conduct extensive experiments on a benchmarked IIoT dataset under simulated label-flipping attacks. Our proposed defense achieves a reduction in root mean square error (RMSE) from 178 (under attack) to 59 after applying the hybrid defense, achieving an approximate 67% improvement in accurate prediction of the Remaining Useful Life (RUL) of industrial engines compared to baseline aggregation methods without defense. These results demonstrate that the proposed hybrid defense substantially enhances the robustness and reliability of FL in IIoT environments, maintaining high model accuracy even under adversarial conditions.