Artificial Intelligence (AI) is evolving rapidly, and Knowledge Graphs (KGs) are widely used in tasks such as smart question answering, recommendation, and search engine optimization due to their semantic search, reasoning, and knowledge aggregation capabilities. However, in industrial applications with complex and heterogeneous data, the quality of KGs suffers, leading to data inconsistency and decision-making errors, which poses a challenge to intelligent manufacturing. We propose an adversarial learning-based error detection method for industrial knowledge graphs (GAN-MKED). This method extracts features from multimodal industrial data including documents, images, and sensor data using a fine-tuned BERT model for document encoding, SENet for image feature extraction, and a combined BiLSTM and FFT approach for sensor signals. A multimodal attention mechanism fuses these features, while a relation-aware graph attention network (R-GAT) aggregates neighborhood information. Spectral Normalization GAN (SN-GAN) introduces random noise to generate pseudo-neighborhood features, improving robustness and stability. The difference between real and generated features is computed and used within an adaptive thresholding mechanism to detect errors effectively. Experimental results on a self-constructed industrial KGs dataset demonstrate that GAN-MKED achieves high accuracy and robustness in error detection for industrial knowledge graphs.