Multi-sensor fault diagnosis of automated guided vehicles in container terminals using MCNN-SEGate-Transformer

Automated guided vehicles (AGVs) provide horizontal transport between quay-side and yard-side systems in automated container terminals. Reliable fault diagnosis is important for maintaining transport continuity and supporting targeted maintenance. However, operational multi-sensor monitoring data are high-dimensional, contain short-term temporal variations, and include variables with unequal diagnostic relevance. Sensor responses may also overlap across fault types. This study develops MCNN-SEGate-Transformer, a multivariate time-series model for AGV fault diagnosis, in which MCNN denotes a multi-level convolutional neural network. Multi-level one-dimensional convolutions extract local details and higher-level fault representations. The squeeze-and-excitation-based gate (SEGate) module recalibrates fault-sensitive feature channels and retains shallow information through gated residual fusion. A Transformer encoder models global temporal dependencies across the input window, and the features are fused for diagnosis. The model is evaluated using operational monitoring data from an automated container terminal, covering normal operation and ten fault types. MCNN-SEGate-Transformer achieves 97.19% accuracy and a 97.03% F1-score. Ablation results show that both SEGate and the Transformer encoder improve performance over the basic MCNN. Shapley additive explanations characterize the global and fault-specific contributions of monitoring variables to model outputs. The model also has the fewest parameters and requires the fewest floating-point operations among the deep learning models evaluated.