The key bottleneck hindering performance improvements in Sound Event Detection (SED) lies in its heavy reliance on strongly labelled data with precise temporal annotations. However, audio data collected in real-world scenarios typically presents a multi-source, heterogeneous supervision format comprising strongly labelled, weakly labelled, unlabelled, and soft-labelled data. Significant variations in annotation granularity and supervision strength make it difficult for models to establish a unified representation and fully exploit fine-grained temporal information. To address these issues, this paper proposes a semi-supervised sound event detection method that integrates pre-trained acoustic representations with pseudo-strong labelling learning, from a unified optimisation perspective. It models multi-source heterogeneous supervision as a joint optimisation process involving explicit supervision, consistency constraints, and pseudo-labelling supervision. In terms of model architecture, Transformer-based pre-trained acoustic representations are integrated with a CNN-RNN detection network to achieve multi-scale feature fusion, thereby enhancing the model’s global semantic modelling capabilities and local spatio-temporal resolution. Regarding training strategies, a progressive two-stage fine-tuning mechanism is designed; through an ‘feature adaptation–joint optimisation’ pathway, stable alignment is achieved between the pre-trained representations and the frame-level detection task. In terms of data utilisation, we propose a pseudo-strong label inference mechanism driven by multi-model consistency. Through prediction fusion and temporal structure reconstruction, this mechanism converts weakly labelled and unlabelled data into high-quality frame-level supervision signals, which are subsequently utilised for training lightweight models, thereby achieving data-driven knowledge transfer. Experimental results on the DCASE 2024 Task 4 datasets (DESED, MAESTRO Real and AudioSet Strong) demonstrate that the proposed method effectively enhances model performance. Specifically, the CRNN-ATST model achieves PSDS1 and mpAUC scores of 0.531 and 0.740 respectively, representing improvements of 14.8% and 9.3% over the baseline. Following the introduction of pseudo-strong labels, the lightweight LightSED model exhibits a mere 3.46% decline in overall performance despite an 88.6% reduction in parameters, whilst metrics such as PSDS2 remain on par with those of large models. The results demonstrate that our method balances detection performance with model complexity whilst enhancing the utilisation efficiency of heterogeneous supervised data, thereby providing an effective solution for sound event detection under conditions of scarce strong labels and resource constraints.