Traditional fault diagnosis methods usually involve complex data processing as well as pattern definition, long diagnosis and analysis times, and poor interpretability, which are not applicable to real-time production control. In this paper, we propose an equipment operation fault diagnosis method based on a Monte Carlo tree search and large language modeling. By modeling the equipment operation fault problem as a search space, representing the possible operation states of the equipment and the different causes of faults as states, and defining the operation behaviors and decision-making behaviors as the transfer of states, and then using the Monte Carlo tree search to simulate the operation logic tree of the equipment in a specific state, the operation of the equipment, the faults detection, and the execution of the operations can be simulated. By using the prior knowledge as well as the reasoning capabilities in the large language model, the operation of the device in the current state is simulated, and the results of the simulated operation including possible faults and other critical information are generated. Finally, the results of the simulation are evaluated by assigning an appropriate score to each state to indicate the value in solving the problem. By combining the production capabilities of the language modeling and outputting the natural language diagnostic explanations and suggestions, the interpretability as well as the real-time nature of the diagnostic method are effectively improved, and the inspection efficiency as well as the detection accuracy are enhanced. The proposed method was employed in the fault diagnosis for large-scale coal mining equipment. Compared with the traditional Monte Carlo tree search method, the fine-tuned Llama2 language model inference method and the Llama2 direct inference method, the accuracy of our new method is improved by 5.9%, 12.2% and 23.3%, and the task success rate is improved by 9.3%, 21% and 28%, respectively.

The vibration signal of mechanical equipment is non-linear and non-stationary, and it is difficult to fully reflect the operation state of equipment through traditional fixed threshold alarm method. While the early warning method based on multi-feature parameter fusion relies on manual experience to extract features, which is difficult to ensure the accuracy of the extracted features and cannot achieved good early warning effect. To solve this problem, a feature self-learning method based on adversarial auto-encoders is proposed in this paper, which encodes high-dimensional monitoring data in normal state into low-dimensional vectors with certain statistical laws and uses it as a benchmark to detect abnormalities in the operating state of the equipment in time by measuring the difference between the encoded features of real-time monitoring data and the benchmark. The actual application cases of reciprocating compressors show that the proposed method can detect the weak signs of equipment fault at the early stage, and realize early warning. At the same time, by comparing with the Auto-Encoders network-based warning method and the Dirichlet process mixture model-based warning method, It is verified that the method in this paper has more advantages in terms of warning accuracy and warning time.