<p>Accurate assessment of blast furnace conditions is a crucial component in the blast furnace control decision-making process. However, most adversarial models in the field currently update the parameters of the label predictor by minimising the objective function while maximising the objective function to update the domain discriminator's parameters. This strategy results in an excessive maximisation of the domain discriminator's loss. To address this, a friendly adversarial training-based tri-training furnace condition diagnosis model was proposed. This model employed a convolutional neural network-long short-term memory-attention mechanism network as a single-view feature extractor and used decision tree methods as three classifiers to compute the cosine similarity between features and representative vectors of each class. During the knowledge transfer process, the classifiers in this model have a specific goal; they not only seek to maximise the entropy of the target domain samples but also aim to minimise the entropy of the target domain samples when they are misclassified, thus resolving the trade-off in traditional models where robustness is improved at the expense of accuracy. Experimental results indicate that the diagnostic accuracy of this model reaches 96%, with an approximately 8% improvement over existing methods due to the inner optimisation approach. This model provides an effective and feasible solution for the efficient monitoring and diagnosis of blast furnace processes.</p>