In recent years, the level of industrial automation has made tremendous progress with the support of information science and technology. Multiple sensor devices can collect a large amount of industrial safety data, which often contains a wealth of information about working conditions and operating modes. By mining the hidden information in this time-series data for anomaly detection and prediction, we can not only grasp the safety level of industrial equipment in real time and ensure its normal operation but also provide effective guidance for the industrial production process and improve industrial productivity. This paper proposes a real-time monitoring and prediction method for industrial safety data based on deep learning techniques and in combination with actual industrial safety data. In this paper, a novel anomaly detection model for industrial safety data is presented, which incorporates bidirectional long short-term memory (BiLSTM) and deep neural networks (DNN). Unlike existing models, this approach takes into account the temporal correlation among industrial safety data features and addresses the challenge of dealing with multiple features. BiLSTM is utilized to capture the inter-feature correlations, while DNN is employed to extract more complex features. Furthermore, to improve the consideration of feature importance, an attention mechanism is integrated into the network. Extensive experiments are conducted to demonstrate the reliability and effectiveness of the proposed method.