Predictive Maintenance for Gantry Crane Shafts: Enhancing Sensor Features With Dynamic Dual-Domain Competitive Attention GRU Autoencoder

Anomaly prediction is an important aspect of predictive maintenance for port machinery and equipment. Traditional anomaly prediction techniques process the raw time- or frequency-domain data, which makes time-series data difficult to reconstruct. In particular, time-series data with weak features, such as sensor data, are difficult to model. Furthermore, the absence of sufficient fault-labeled data for industrial processes means that classification models are difficult to train. Therefore, this study introduces an advanced temporal anomaly prediction model, which uses a gated recurrent unit (GRU) autoencoder to process time- and frequency-domain data in parallel. The feature match and temporal consistency of the data are modeled using a bi-domain competitive attention module and bi-directional loss function, respectively, which reduces the mean squared error by 15.43%. Weak fault features in the sensor data are enhanced using dynamics simulations, which address the problem of missing fault labels by using a threshold diagnostic model instead of supervised learning. Thus, the recall rate is increased by 9.04%. An attention storage pool approach is also used to mitigate the prediction decay caused by external variables. Overall, the proposed method can improve the efficiency and safety of industrial machinery by addressing the challenges involved in modeling time-series sensor data with weak features.