CMformer: A condition monitoring method for industrial equipment utilizing both channel and temporal dependencies

Industrial equipment operating states are typically unobservable and must be inferred from noisy, strongly coupled multi-sensor data, making operational state estimation for predictive maintenance a high-dimensional latent-state inference problem. Existing time-series segmentation methods for industrial equipment often assign static weights to sensor channels or process them independently, which limits their ability to capture dynamic cross-sensor dependencies across operating stages. They also struggle to efficiently model long-range temporal patterns over complete industrial cycles, leading to inaccurate state labels and blurred change-point localization. This study aims to improve industrial time-series segmentation by jointly modeling dynamic cross-sensor interactions and long-range temporal dependencies within a Transformer framework. We propose CMformer, a Transformer-based model for industrial time-series segmentation (TSS) and operational state detection that incorporates a Segment-Wise Channel-Enhanced Embedding (SWCE) module, which treats sensor channels as tokens within each temporal segment and performs segment-wise channel self-attention and temporal compression to obtain compact, information-dense tokens. In addition, a Point-to-Segment Attention mechanism links each time point to multi-scale temporal segments, balancing local detail with segment-level context while keeping attention complexity tractable. Experiments on real-world industrial datasets demonstrate that CMformer consistently outperforms strong baseline methods across multiple segmentation metrics, yielding more accurate operational state labels and more precise prediction of state transitions. The enhanced change-point detection capability is valuable for early fault indication and condition-based maintenance scheduling, helping to reduce unexpected equipment downtime in complex industrial environments.