Limited downlink bandwidth constrains the continuous transmission of satellite telemetry channels. Transmitting a selected subset of channels and reconstructing the remaining channels reduces the amount of data requiring downlink transmission. However, independent recovery networks duplicate temporal processing, while static models cannot adapt to persistent or time-varying reconstruction errors after deployment. To address these problems, this paper proposes a telemetry data compression method based on multi-task long short-term memory reconstruction with adaptive online calibration. First, a shared long short-term memory network jointly reconstructs multiple untransmitted channels through task-specific output heads. Then, target references are transmitted in rotation, and a lightweight residual module uses each reference to update the output correction of the corresponding channel while the reconstruction network remains frozen. Finally, the method is evaluated on real satellite telemetry data by comparing it with independent long short-term memory networks and deterministic baselines and by analyzing its sensitivity to the reference interval. At a normalized transmission volume of 60%, affine calibration reduces the average normalized reconstruction error of the shared model by 0.107. The calibrated shared model attains reconstruction accuracy comparable to that of calibrated independent long short-term memory networks, while reducing the parameter count by 66.2% relative to three independent networks with the same per-network hidden size.

satellite telemetry,data compression,multi-task learning,online calibration