<p>In addressing the issue posed by few-shot working condition classification, this paper proposes a meta-learning framework built upon a multi-head attention autoencoder. The framework integrates a prototypical network and the model-agnostic meta-learning algorithm to efficiently adapt to classification tasks. The autoencoder effectively captures intra-class and inter-class dependencies among samples through the multi-head attention mechanism. In the task preparation phase, the data is partitioned into a support set and a query set. At the same time, the model-agnostic meta-learning algorithm enhances the model’s adaptability to new working conditions via rapidly fine-tuning the parameters with the support set. Classification is performed using a prototypical network, calculating the cosine similarity between query samples and prototypes of each class for working condition category prediction and evaluation. To emphasize the practical significance, the proposed framework is specifically validated on the Tennessee Eastman process and gold hydrometallurgy datasets, demonstrating significant improvements in classification accuracy and adaptability to new working conditions. Accurate classification of working conditions is critical for fault diagnosis, operational optimization, and predictive maintenance in complex industrial systems. In practical applications, the solution can identify abnormal working conditions in time, minimize downtime, reduce operating costs, and ensure safety. By achieving superior classification performance compared to traditional methods, our approach enhances the reliability and efficiency of process monitoring and decision-making.</p>