A dual-decoupling network for anomaly detection in automotive transmissions under time-varying operating conditions

As a core component of the automotive powertrain system, the operational state of the transmission is directly related to the overall safety performance of the vehicle. However, during actual operation, transmissions are often subjected to time-varying operating conditions (TVOCs). The frequent changes in these conditions cause significant distribution shifts in the collected vibration data, leading traditional anomaly detection (AD) methods to be prone to “false positives” or “false negatives”, which compromises their engineering applicability. To address this issue, this paper proposes a novel Dual-Decoupling Network (DDN) to effectively mitigate the interference of TVOCs on AD performance, thereby enhancing detection accuracy and robustness. First, based on the causal generation mechanism of transmission vibration signals, this paper deconstructs the vibration signal into four main components: operating condition (OC) information, basic information, abnormal information, and noise information, and constructs a corresponding phenomenological model. On this basis, a DDN architecture, consisting of explicit and implicit decoupling stages, is designed. In the explicit decoupling stage, a mapping relationship between operating condition parameters and the normal vibration response is established to decouple the basic information. In the implicit decoupling stage, a novel adversarial loss is designed to further achieve deep decoupling of operating condition features and abnormal features, thereby extracting robust abnormal features that are invariant to changes in OCs. From the perspective of optimization theory, it is derived that the abnormal features extracted when the implicit decoupling model reaches its optimum are independent of the OC information. Finally, the proposed method is systematically validated on multi-level experimental platforms at the part, component, and vehicle levels, and compared with several existing advanced methods. The experimental results show that the proposed method can accurately identify and provide timely warnings for abnormal states of automotive transmissions under TVOCs, demonstrating significant potential and value for engineering applications and promotion.