Abstract: Gait phase estimation serves as the perceptual interface between human motor intent and exoskeleton torque delivery, yet wearable sensing signals vary substantially across users, and generic models degrade markedly when deployed on unseen subjects. A cross-subject adaptive gait phase estimation method fusing insole pressure, inertial measurement units and surface electromyography is proposed. A functional coordinate system is first constructed from the quasi-static gravity reference and the plantar pressure baseline so that installation-induced attitude discrepancies are removed at the input stage. A multi-resolution temporal convolutional encoder combined with graph attention fusion is then designed, supervised jointly by gait phase, phase rate, gait events and locomotion mode, where a sine-cosine dual-channel parameterization suppresses discontinuity at phase wrapping. Finally, a low-rank online adaptation module updates a small subset of parameters on the frozen shared backbone using cycle consistency and event sparsity priors as self-supervised objectives, with update magnitude constraints and a safety fallback mechanism. Experiments on a public multimodal gait dataset and a self-developed lower-limb exoskeleton platform show that, under the leave-one-subject-out protocol, the continuous phase root mean square error reaches 3.21%±0.87% of the gait cycle and the heel-strike timing error reaches 18.6±6.2 ms, reduced by 22.1% and 30.6% respectively compared with the non-adaptive generic model; the single-cycle inference latency is 6.8 ms, leaving sufficient margin within a 100 Hz control cycle; in the assistance experiment, the integrated electromyography of the rectus femoris and gastrocnemius decreases by 16.3% and 13.8% respectively. The results indicate that minute-level personalization can be achieved without target-user labels, providing a usable perceptual interface for human-exoskeleton collaboration.