Visual–frequency bimodal fusion networks for oscillation detection and quantification in industrial control loops

Oscillations arising from performance degradation in industrial control loops threaten operational safety and efficiency. Existing automated detection methods often rely on fixed assumptions and single-modality signals, limiting their robustness and generalization. In this work, we propose a visual–frequency bimodal fusion network (VFBFN) for oscillation detection and quantification. VFBFN employs a dual-branch encoder consisting of a two-dimensional (2D) convolutional neural network (CNN) for image-based signal embeddings and a one-dimensional (1D) CNN for frequency spectra to extract complementary features. These features are integrated through a bidirectional guidance-enhanced fusion module to enhance discriminability. The detection model employs a softmax-based classification head to classify signals into non-oscillation, regular oscillation, and time-varying oscillation. Experiments on synthetic data, two public benchmarks, and a self-constructed dataset comprising 119 real industrial loops demonstrate an average accuracy of 96.77% across the four datasets and consistent superiority over state-of-the-art baselines, together with strong robustness to varying signal lengths, noise levels, intermittent behaviors, and slow drifts. Based on the same VFBFN architectural template, three quantification models are independently instantiated and trained to estimate oscillation characteristics, including cycle count, sparseness, and regularity, enabling fine-grained oscillation assessment.