Fault diagnosis based on deep learning is an important research direction in the intelligent management of industrial safety. Failures frequently occur in the actual industry production, leading to reduced production efficiency, and in severe cases, production stoppages or even casualties. Due to the complex and variable production environment, fault features are difficult to extract and recognize, andreal-time monitoring and rapid diagnosis are required at industrial sites. Traditional fault diagnosis methods typically rely on expert experience for feature extraction and pattern recognition, which makes them difficult to adapt to the complex and dynamic industrial environments. To address these issues, a deep learning based multimodal industrial fault diagnosis(DMIFD) method is proposed. Extreme gradient boosting (XGBoost) is employed to select process parameters related to industrial faults, which are then used as the multimodal input data for the model. The deep extreme learning machine (DELM) is used to extract nonlinear and high-dimensional features from the production process parameters to identify equipment in abnormal states. The key parameters of DELM are optimised using the frost and ice optimisation algorithm (FIOA) to achieve optimal model performance. The RIME-DELM module outputs equipment samples in normal states, while the samples of abnormal equipment are further input into a deep belief network (DBN) and a least minimal squares support vector machine (LSSVM) to perform specific fault type classification. The proposed method is applied to the aluminium electrolysis production process to validate its effectiveness. Field conducted in an aluminum electrolysis plant show that the model achieves an abnormal state detection accuracy of 97.96%, an F1-score of 0.975 3, a fault type diagnostic accuracy of 96.75%, and a Macro-F1 score of 0.944 7. Compared with common deep learning models and through ablation experiments, the DMIFD model demonstrates higher diagnostic accuracy and provides more accurate and reliable support for fault diagnosis in practical industrial production.