Maximizing Throughput of a Parcel Sorting System by Using Simulation

Overview

A tilt tray sorter is a high-speed sortation device that sorts small and medium-sized parcels. Tilt tray sorters are commonly used in automated parcel sorting, distribution, order fulfillment centers, and baggage handling.

The automated parcel sorting consists of several processes. The parcels get inducted onto trays and move along the conveyor. When the parcels reach their destination, trays tilt, and the parcels get ejected.

Problem

Automation Intelligence is a consulting company delivering value-engineered solutions in the areas of automation, machine learning, and artificial intelligence. Noorjax Consulting is a simulation consulting company that generates feasible solutions for diverse industries. Automation Intelligence, together with Noorjax Consulting, worked on improving the parcel sorting systems.

In practice, the tilt tray sorters typically had multiple induction areas. The parcel sorting systems were mostly built using the first in, first out (FIFO) logic. Without optimization, suboptimal, greedy FIFO assignment used more trays than necessary. Additionally, the trays could have different lengths.

Moreover, discharges could not happen back-to-back. After one parcel was discharged, some time was required to reset that lane. This could be from one to four trays, depending on the speed of the conveyor and the size of the object that was removed.

When the FIFO assignment was used, if the second yellow parcel was placed after the first one going to the same destination, the second parcel was going to recirculate. Recirculating parcels prevented new parcels from entering the system. This created a bottleneck that should not exist.

By making smarter decisions and using another induction lane to put packages on these trays, the recirculation could be avoided.

The way parcels were assigned to trays using FIFO logic decreased total throughput by 10%. That was major for parcel sorting facilities and distribution centers of big retailers when they were overrun during the holidays.

The current logic with the different sizes of the lanes did not consider the size of the parcels related to the size of the trays.

As illustrated in the picture below, if the red parcel was first, the greedy algorithms would place it on two smaller trays, followed by the yellow and blue parcels. This was the naive utilization method.

As for the optimized utilization method, where the size of the parcels was taken into consideration, the red parcel would be placed on the larger tray. This second method allowed for four parcels instead of three.

Both of these methods needed multiple trays to pass before another parcel could be discharged from the same lane.

Solution

Automation Intelligence and Noorjax Consulting used mathematical programming to resolve these problems. The mathematical model had 3 components: decision variables, objective, and constraints. The decision was whether or not a particular parcel should go to a particular tray. For instance, if there were 20 parcels and 50 trays, 1,000 decisions would have to be made about which parcels should be on which trays.

The objective was to maximize throughput. This means getting as many parcels as possible into the system. When the mathematical model is being created, the developers should indicate how these system variables are related and what the constraints are.

The mathematical optimization was only one of the components of the solution architecture. The specialists still needed the solution to test their assumptions. For this purpose, they chose AnyLogic simulation software.

In the picture below, an example of the tilt tray sorter made by Automation Intelligence and Noorjax Consulting is shown. This sorter has 4 induction lanes on the left and right sides, as well as 10 discharge lanes on the top and bottom.

AnyLogic was used in this project for several reasons:

1. Data object input allowed the model developers to make a system flexible and parametrize everything in the simulation. This enabled engineers to modify the physical and operational parameters of the system.

   In terms of flexibility, this system could be applied to any existing circular conveyor. The data object contained all system characteristics, including overall length, width, radius curve, the number and size of different types of lanes, etc. When parcels came into the system from the input lane, it was possible to identify the probability for each size of a particular parcel. Custom distribution in AnyLogic could be used to accommodate that. Thus, the engineers developed a flexible project thanks to AnyLogic simulation software.

   AnyLogic also gave an opportunity for customization. For instance, there was an option in the model to put one object on two trays of the conveyor at the same time, thanks to the AnyLogic Material Handling Library. Also, it was possible to create different sizes of trays that could accommodate objects in the way, that was needed based on optimization, for different facilities' configuration. This flexible system could be used in any facility.

2. The developers could use a set of blocks to create agent-based models and then transform them into discrete-event simulation models.

3. AnyLogic sent the information about the parcel arrivals, their sizes, and trays to a solver using Gurobi and Python. The solver was agnostic. It received a REST API call and returned back the assignments of parcels and trays using REST API.

   

   Using REST API for tray assignments

Results

The picture below illustrates the naive and optimized tray utilization. The tray utilization is the number of trays that are full. For the optimized algorithm, there was consistently lower tray utilization. The optimized algorithm is minimizing recirculation, and as a result, parcels were getting to their destinations on the first go around and then getting discharged.

As for the FIFO method, there were going to be too many recirculating parcels that were taking up more trays. It was not going to leave enough room for additional parcels to enter the system. As the developers started to push that up, the naive method froze, whereas the optimized method had additional capacity to take on more parcels.

The 10% increase in throughput was reached by building out additional capacity to absorb more parcels through minimizing recirculation.

The case study was presented by Ari Siesser, of Automation Intelligence, and Felipe Haro, of Noorjax Consulting, at the AnyLogic Conference 2022.

The slides are available as a PDF.