Page 200 - IJOCTA-15-4

P. 200

Z. Chen et al. / IJOCTA, Vol.15, No.4, pp.738-749 (2025)
Table 2. Experimental results of different algorithms for air cargo loading on wide-body aircraft.

B777 B787
Algorithm
MAC(%)↓ INDEX(%)↓ TIME(s) ↓ MAC(%) ↓ INDEX(%) ↓ TIME(s) ↓
Deepseek 24.30±0.82 10.00±6.51 2.14±0.20 24.32±0.72 10.00±5.43 2.17±0.18
Llama 24.20±0.58 9.17±4.73 1.04±0.01 24.31±0.61 9.77±4.58 1.04±0.00
FastLoader Phi 24.22±0.64 9.33±5.05 1.36±0.01 24.12±0.41 8.20±2.82 1.39±0.05
Qwen-1.5 24.38±0.85 10.68±6.84 1.87±0.10 24.10±0.48 7.97±3.44 1.68±0.35
Qwen-3 24.14±0.62 8.65±5.08 1.76±0.49 24.19±0.72 8.79±5.86 1.70±0.46
COM 25.39±0.98 16.21±5.80 0.02±0.01 25.24±0.83 15.76±5.03 0.04±0.01
Combinatorial
IOM 25.30±0.96 15.84±5.64 0.03±0.01 25.16±0.80 15.61±5.03 0.03±0.01
optimization MLIP 25.25±0.97 15.75±5.78 0.05±0.01 25.17±0.84 15.58±5.20 0.12±0.01
HGA 23.25±0.47 9.14±3.25 110.21±25.5 23.64±0.24 9.19±4.82 183.91±20.16
Heuristic GA-normal 23.24±0.47 9.26±3.11 118.30±17.98 23.68±0.24 9.56±4.36 185.54±20.95
search DMOPSO 23.36±0.38 9.40±4.10 135.21±26.36 23.56±0.31 9.25±3.78 269.48±28.15
PSO-normal 23.39±0.69 9.30±3.47 160.19±24.63 23.38±0.69 9.21±4.78 275.49±20.24
ULD records the weight, cargo type and cargo the information of two main cargo types: Unit
hold correspondence between cargo and the cargo Load Device (ULD) and Bulk. The raw mani-
hold. Bulk records the weight, cargo volume, fest is transformed into a high-resolution descrip-
and the bulk information, such as mail, luggage, tion of every freight item so that optimization
etc. The LLM engine consists of aircraft type and algorithms can consume the data with no fur-
cargo constraints. The cargo constraints include ther cleaning. (1) Each record is first classified
all important constraint information for the cargo as either a ULD or bulk cargo. It is then en-
loading scenario. riched with all operational attributes: mass, ULD
Second, the search space reduction is based contour code (or bulk volume), special-handling
on the existing heuristic search methods. We de- tags such as “dangerous goods” or “perishables,”
sign two modules, including a classic cargo load- and any stated stowage preferences or prohibi-
ing solver and a search space reducer. The cargo tions. (2) The constructor also references airline
loading solver (§ 3.3) generates its search space and aircraft documentation to enumerate which
and takes cargo information and cargo loading cargo hold positions are structurally or geometri-
features as input. Then, the solution is com- cally compatible with that piece. The output is a
pleted through classic heuristic search method. pair of dictionaries keyed by unique cargo IDs. (3)
The search space reducer (§ 3.4) calculates the Items that violate a structural rule at this stage
score of each solution in the search space. At the (e.g., a heavy ULD that exceeds every individual
same time, driven by LLM, it uses the important hold limit) are flagged immediately.
constraint information contained in the LLM en-
gine to simplify the search space.

Engine building. The second step assem-
3.2. Cargo loading constructor
bles an LLM engine that fuses aircraft cargo hold
Cargo loading constructor converts the cargo in- information and complex constraints into a sin-
formation and complex constraints into simpli- gle data structure. (1) Numeric constants are
fied modeling data and inputs them into the next loaded from the trim-sheet database and stored
stage. There are two main steps in the cargo in a structured map. The numeric constants in-
loading constructor, including cargo information clude the longitudinal coordinates of every hold,
building and engine building. individual weight capacities, and the center-of-
Cargo information building. In this step, gravity (CG). (2) Complex constraints are then
we design a dictionary data structure to record appended as human-readable clauses: examples
742

195 196 197 198 199 200 201 202 203 204 205