US 12,452,692 B2
Elastic cooperative inference architecture and method for UAV cluster
Chao Dong, Nanjing (CN); Yuben Qu, Nanjing (CN); Hao Sun, Nanjing (CN); Feiyu Wu, Nanjing (CN); Weiqing Ren, Nanjing (CN); Qihui Wu, Nanjing (CN); and Lei Zhang, Nanjing (CN)
Assigned to NANJING UNIVERSITY OF AERONAUTICS AND ASTRONAUTICS, Nanjing (CN)
Filed by NANJING UNIVERSITY OF AERONAUTICS AND ASTRONAUTICS, Nanjing (CN)
Filed on Nov. 2, 2022, as Appl. No. 17/979,008.
Claims priority of application No. 202211083483.7 (CN), filed on Sep. 6, 2022.
Prior Publication US 2024/0080679 A1, Mar. 7, 2024
Int. Cl. H04W 16/26 (2009.01)
CPC H04W 16/26 (2013.01) 3 Claims
OG exemplary drawing
 
1. An elastic cooperative inference architecture for an unmanned aerial vehicle (UAV) cluster, wherein the UAV cluster comprises a plurality of heterogeneous UAVs with computing capability; wherein, the elastic cooperative inference architecture comprises:
in an early stage of inference, a node is selected as a master node;
the master node obtains a performance information and connection of each UAV at each time point;
a plurality of exit points are set in a plurality of parts of a model; the master node adjusts a model size according to QoS requirements of an inference task, determines an exit point, and ends a model inference in advance to reduce unnecessary calculations;
the master node divides the model or data used for inference according to the performance information and connection of each UAV, and transmits task allocation information and node connection information to each node;
wherein a round of cooperative inference under a time point t comprises:
S1: a UAV i of the master node first measures the performance information Di+1,t, Di+2,t, . . . Di+k,t of available K UAVs, wherein t is a number of time points, and t is an integer greater than or equal to 1;
S2: the master node adjusts the model size according to the QoS requirements of the inference task, determines the exit point, assigns the inference task Ti+1,t, Ti+2,t, . . . , Ti+k,t and transmits the task allocation information and the node connection information to each node:
wherein Ti+n,t represents an inference task assigned to a node n at the time point t, and Tt represents a total inference task of the time point t; and
S3: each node obtains inferential data according to an assigned task and sends a result to subsequent nodes.