The Flexible Job Shop Scheduling Problem (FJSP) is central to smart manufacturing, yet standard algorithms often prioritize productivity (makespan) at the expense of cost and reliability. This paper introduces t-MOHHO, a collaborative optimization framework designed to equilibrate machine load, processing costs, and delivery timeliness alongside throughput. By incorporating an adaptive Student’s t-distribution mutation operator and a non-linear energy escape mechanism, t-MOHHO effectively navigates high-dimensional search spaces. Extensive validation on 10 MK benchmark instances reveals that t-MOHHO demonstrates significant advantages over classic HHO, MOPSO, and MOEA/D across most metrics. Notably, in comparison to the state-of-the-art NSGA-III, t-MOHHO executes a clear trade-off: it trades marginal makespan efficiency for substantial reductions in cost and tardiness. Specifically, on the large-scale MK10 instance, t-MOHHO reduces total tardiness by 56.2% and lowers processing costs by 3.4% compared to NSGA-III. These results demonstrate that t-MOHHO can strategically sacrifice maximum speed to secure superior punctuality and cost-efficiency, making it a robust decision-support tool for Just-in-Time (JIT) production environments.

Recent research has demonstrated that various network architectures combined with Transformers exhibit outstanding performance, particularly in models integrating Transformers with CNNs, which have made significant strides in the field of fault diagnosis. However, existing models that combine Transformers and CNNs fail to effectively utilise multi-scale convolutions for feature extraction and suffer from channel weight information loss, which weakens the network's feature extraction capability. Therefore, this paper proposes an MK-ACFormer framework for mechanical fault diagnosis to address these issues. Firstly, a multi-scale convolutional attention module is designed to extract features with multiple local receptive fields from vibration signals and capture inter-channel correlations, thereby optimally distributing channel weights. Secondly, the ECA module is employed to fuse features extracted from different scale convolutions with those derived from the Transformer, adaptively adjusting feature channels to assign varying weights and reduce information redundancy. Experimental results indicate that the proposed method is more robust compared to recent Transformer-CNN integrated approaches and CNN-based improvements for fault diagnosis, further validating the model's effectiveness through performance analysis. Finally, transfer experiments on small sample data demonstrate the model's capability in employing transfer learning methods.