DAC-TS: a decoupling adaptive cooperative approach for fault diagnosis of time series data from cross-modal sensors

Abstract

Sensor data in industrial applications typically contains severe noise interference, temporal patterns across multiple scales, and physical properties that vary across different domains. To address these challenges, this study proposes a Decoupled Adaptive Cooperative method for industrial time series classification, termed DAC-TS. The core components of DAC-TS are the Multi-Scale Residual Bidirectional Temporal Convolutional Network (MSR-BiTCN) and a Variational Mode Decomposition (VMD) module optimized using an improved Sparrow Search Algorithm with discrete parameter constraints, termed Di-SSA-VMD. MSR-BiTCN learns discriminative temporal representations through parallel multi-scale convolutions and bidirectional temporal modeling. Specifically, the multi-scale convolutional branches capture broadband temporal patterns, while the bidirectional mechanism integrates historical and future contextual dependencies. Furthermore, an Efficient Channel Attention (ECA) module is introduced. This module adaptively recalibrates information features along the temporal and channel dimensions to achieve robust classification. To improve the input representation under extreme noise conditions in rotating machinery, a Di-SSA-VMD module is used as an adaptive signal enhancement component. It performs reconstruction using selection weights in a physically meaningful discrete parameter space. Experiments on three cross-modal industrial sensor datasets demonstrate that DAC-TS exhibits strong robustness under severe noise interference. The results also show that it has cross-domain generalization ability. These findings demonstrate that DAC-TS provides an effective decoupling scheme for reliable classification of industrial time series data.