Roadmap to Membership of RICS - Pathways and Routes
Evaluation of harvesting and packaging operations in a Greek tomato-production greenhouse, with the use of model-based method
1. Final MSc Thesis Presentation
Evaluation of harvesting and packaging operations in a
Greek tomato-production greenhouse, with the use of
model-based method
21/1/2016, Ioannis Moutsinas
4. Harvesting and packaging
Harvesting: sorting of trusses according to number of
red tomatoes
• Simple sorting: in two classes (Lucia and
standard); in carbon boxes (6.5–7kg); separate
packaging in processing room; several products
• Intensive sorting (packaging): two classes –up to
six classifications + box weight calibration in
processing room; in carbon boxes (5.5 - 7kg)
Criteria:
● Harvest type: if expected yield > 50 tons
combined harvest & packaging in path and vice
versa
● Classification: size, shape, colour
4
7. Problem description
High labour costs + unstable labour efficiency + limited
trained personnel use personnel optimally
First step: simulate labour processes re-design
This study: Packaging in Path (PinP) (sub)model
Objective: evaluate performance and efficiency of
alternative (work) methods for the in-path packaging
operation using a model-based method
7
8. Research questions
1. Is packaging in path (PinP) sub-model able to represent the
harvesting and packaging operations in a Greek tomato
greenhouse?
2. What is the performance of the harvesting and packaging
operations in path level, within the Agritex Company for the
years 2012, 2013?
3. What is the best work method to harvest and package Idooll
tomato cultivar in a (Greek) greenhouse?
8
9. Materials & Methods
Problem approach
In-path packaging methods of typical cultivar (Idooll)
IDEF3 analysis
Behaviour and labour registration data analyses
Define and simulate list of work scenarios
9
10. Idooll cultivar
Tomato weight: 140-150g
Diameter ≈ 60 mm
Tomatoes per truss: 3-5
Rows (2013): 150-314 (both)
10
Quality
classes
Name Capacity (kg)
1 Lucia Big Red 5.5
2 Lucia Small Red 5.5
3 Lucia Big Semi-Red 5.5
4 Lucia Small Semi-Red 5.5
5 Blue Pair 6.5 – 7
6 Blue Single 6.5 – 7
11. Data processing
1. Videos & scripts of harvesting and packaging in
greenhouse (grower)
2. Labour registration data (Nomad)
1. IDEF3 process analysis (AllFusion process modeller) &
Behaviour analysis (Noldus Observer XT)
2. Descriptive statistical analysis 2012 (Excel)
11
12. IDEF3 process analysis
First pre-modelling step
Method to record network of relevant actions in a
process within a context of operation(s)
Raw data: video and scripts by grower
Goal: Record combined harvest and packaging in-path
operation + Weight calibration process in packaging
room model structure
12
13. Data analysis
General (all cultivars) and specific (only Idooll) harvest
analyses
● Trace and assess yield effects on harvesting
Simulation dates
● Relations between harvested yield and harvest rate
or harvest duration per path input and
verification data
Data filtering
Behaviour analysis
● Coding scheme of movements + videos observation
Probability distributions of basic processes (cut,
prune, allocate)
13
14. PinP submodel
14
Discrete event system: System’s state transition ruled by
asynchronous discrete incidents (events); environment:
entities, attributes, events, resources, queues
represents the Harvest & Packing process (Idooll) at path
level + the weight calibration in the processing room
Entities (resources): harvester, trolley, truss, box
Matlab, Simulink environment, SimEvents toolbox
15. Model structure
15
Inputs
Greenhouse dimensions
Production system dimensions
Plant density
Daily yield
Daily path schedule
Probability density functions
Run settings
Initial state
Velocity vectors of operators
PinP model
Harvest & packaging
processes
Reports
Outputs
Job cycle times
Product throughput
Labour times
Transportation times
Performance parameters
Subsystems:
1) Harvest and package in path (harvester, trolley)
2) Transportation (tractor driver)
3) Processing room (worker)
16. Simulation scenarios
Aim: test model’s functionality and flexibility + find most
effective work method
Simulation of single harvest session: single (average)
harvester, trolley, tractor driver, worker, group of paths
Simulation date: May 20th, 2013
Comparison on yield and time related parameters
16
Scenario Description Packaging
Box
Classifications
Weight
calibration in PR
S0 Reference Manual 6 Yes
S1 Simplified sorting method Manual 2 Yes
S2 Automated packing in path Automated 6 Yes
S3 Complete H&P in path Manual 6 No
17. Model calibration & validation
Reference scenario
Dates: May 20th, March 20th, April 14th, April 24th
Complete harvest single path by avg. harvester-trolley
Measured vs simulation data
● Measured data: daily average times for a path;
optional data filtering applied (outliers)
17
Performance indicator Units
Yield per path kg
Yield per path (boxes) -
Time to harvest a path minutes
Harvest rate kg min-1
Box filling rate min box-1
Average box weight kg
Truss quality ratio -
21. 21
Time parameter Value (sec)
S0 S1 (%) S2 (%) S3 (%)
Total harvesting duration 3188 3163 (-0.8) 2406 (-32.5) 3963 (+24.3)
Mean time to harvest truss 2.34 2.29 (-2.1) 2.33 (-0.4) 2.55 (+8.9)
Time interval for truss harvest 8.86 (0.984) 8.93 (+0.8) 6.68 (-32.6) 12.4 (+39.9)
Net packing duration 647.6 761.5 (+17.5) 0 1672 (+158)
Mean time to pack truss 1.9 (0.02) 2.15 (+13.1) 0 5.21 (+174)
Time to fill a trolley 3504 3341 (-4.8) 2548 (-37.5) 4314 (+23.1)
Time to sort trusses NA 1334 NA NA
Time to calibrate boxes’ weight 464 466 (+0.4) 429 (-8.2) NA
22. Discussion
Despite limited parameter comparison, model showed
good adaptation to reality
Despite limited applicability, alternative scenarios
simulation results showed good model adaptation
Overall quality of labour data can be characterized as
average but not unreliable
● Relation equations not reliable for use
Inter-planting did not affect simulation results
S1: increased time in processing room (+22 min)
S2: decreased time in path (-32%)
S3: increased time in path (+24%)
22
23. Conclusions
23
1. Model: Good adaptation to reality
● 85-115%; 85-123%
● Flexible to yield changes time changes (avg.
worker)
2. Harvest demand dependent on crop productivity (yield)
two peaks: February & July (two cultivation periods)
● Harvest rate and time proportional to yield
● Harvest time per truss reversed proportional to
yield
● Coefficient of determination (R2) not high enough
3. Best scenarios: S2, S0, S3, S1
● Significant variations in time parameters:
(+29.5%; -25%; +8.7%)
● Yield variations not significant
24. Recommendations
24
Model performance:
● Expand model for other cultivars
● Incorporate with GWorkS model
● Improve behavior analysis with more video footage
● Simulate multiple trolley visits in a path
● Introduce alternative scenarios for tomato processing
Grower:
Apply reference scenario whenever possible
Evaluate the possibility to introduce automation in
path
Educate and motivate workers to use Nomad
Expand Nomad’s database
26. Agritex cultivation practises
Crop replacement process
● Regular: complete crop removal and installation of
new separately for the two greenhouse sections (2
cultivation periods); applied in 2012 and before
● Inter-planting: gradual new crop integration in
February; tested in paths 310-614; applied in 2013
26
34. Calibration results
34
Time parameter Value (sec)
Initial Difference (%) Calibrated
Total harvesting duration 3132 +1.8 3188
Net harvesting duration 906.5 -6.8 844.7
Mean harvest time per truss (std. error) 2.7 (0.054) -15.3 2.34 (0.0062)
Time interval for truss harvest (std. error) 9.40 (0.555) -6.1 8.86 (0.984)
Min – max of time interval 2.17 - 130.4 - 2.39 – 340.7
Net packing duration 867 -33.8 647.6
Mean time to pack truss (std. error) 2.6 (0.0136) -12.1 2.32 (0.0196)
Time to bring new boxes 174 +86.8 325
Time to fill a trolley 3277 +6.9 3504
Time to calibrate boxes’ weight 434 +6.9 464
35. Calibrated parameters
Mean & std. of probability distribution of the number of
tomatoes per truss decreased reduce total path yield
Std. of probability distribution of tomato colour reduced
increase Lucia trusses
Lucia box capacity reduced increase number of boxes
35
Editor's Notes
Introduction: current situation, research questions, problem, objective
Materials and methods: what was done and how in order to approach the problem and reach its solution or objective
aerial view of the greenhouse; structure of grh; Compass for orientation;
a) the mini-plum tomato (Ardiles) = 366 – 444 (even) b) the cocktail tomato (Brioso and Ornella) = 1-70+72-148 (even) and 71-149 (odd); and c) the large-tomato truss (Idooll)
Planting density: 2.8 - 3.6 plants m-2
Harvest is executed with trolleys filled with boxes of two different major classes
Harvest type depends mainly in expected harvested yield and therefore the time period (winter usually harvesting is done in two major classes and they are sorted in processing room; in summer all in path).
Classifications according to size, colour, shape. Colour: according to orders (4+1, 3+2 etc.); Trolley capacity in Lucia and blue boxes and other boxes. The different classifications appear in the next slide.
Intermediate actions: refill with empty boxes (remaining tomato quality, harvesting history); buffer trolley in middle of distance; change path
In the context of this study the PinP submodel was created, as a response to the main objective of this research which is to... [objective]
2) 2012 is a typical year while in 2013 inter-planting method was applied for the crop replacement procedure -> interesting to compare and see also its effects on the harvest + packaging performance. However , this is not related to model -> it will not be described thoroughly during this presentation.
How to approach the problem to reach its solution.
Focus on the combined harvesting and packaging in-path operation of Idooll cultivar
Conduct an IDEF3 analysis to record the logistics of all processes to use it as model base
Film videos on site to use for the behavior analysis in order to determine the durations and std. dev. (pdf) of stochastic processes and use them as model inputs
Filter and Process labour data to use it as model input or extract measured data which will be used for the model validation.
Blue = standard = box colour
How to handle data to serve the model creation purpose?
Harvest video recording: total duration of less than 10 minutes;
General info: scripts and other general video recordings which were transformed into notes
Labour registration data: no later data due to company’s experimentations on the planting process which distorted the quality and accuracy of measured data
Descriptive stat. analysis: 1) harvest effects -> how crop productivity affects processes like harvesting; 2) trace yield increments -> use as indications to select date for further model validation; 3) relation equations -> model input/substitute or verification data -> unsuccessful; 4) comparison of 2012 and 2013 -> detect effects of inter-planting method.
The first step towards the model creation is to record and archive all the involved processes.
Its results are presented here since they are a necessary prerequisite of the modelling process
-Specific harvest: investigates the relations between harvested paths and harvested trusses; the harvesting duration (min) and the total harvested yield (kg); the harvest performance (kg h-1) and the total time to harvest a path (min) -> find best simulation dates + crosscheck model’s validity based on yearly fluctuation of yields.
-Behaviour analysis: coding scheme -> enable movements interpretation; type, mean and std. of prob. Distr.; tom van zundert dataset -> check definitions of common actions -> use those only
-Data filtering: basic and advanced on data irregularities (prolonged breaks, multiple registrations, extravagant harvested kg)
Entities are passive objects that represent people, parts, tasks, etc. They travel through the blocks of the flowchart where they wait in queues, get delayed, processed, seize and release resources, split, combine, etc. The term resource designates a system element that provides service. Resources are usually capacity-limited, so entities compete for their use and sometimes must wait to use them, experiencing delay as a result.
Entities: Real-world elements functioning as resources (Harvester, trolley, truss, box; worker, tractor driver)
Harvest Performance parameters: harvested plants, yields (kg), boxes, trusses etc.
Throughput: trusses/kg/boxes per min
Inputs: loaded through Matlab script or inside model blocks (with italics these inputs are introduced inside the model).
Refill method: 4 boxes
Simulation goal: to sim an entire path using only one trolley
Date selected based on harvest demand (total hours of harvesting)
PinP model modifications? Model assumptions? Model inputs, outputs?
Even though the simulation is executed for multiple paths the calibration was executed for the average path -> one complete was selected to avoid making reductions of simulated values
Measured data: daily average to reduce time variation among paths -> measured data more flexible than model; average worker
-the simulation and the sub-model capabilities may seem limited, however results close to reality both for the calibration and the validation trials conducted
-model’s adaption level to alternative work methods was good since the simulations were run with success, the results agreed with the initial hypotheses and their comparison with the reference scenario’s results showed clear effects of the changed parameters on the model outputs
Despite poor behavior analysis
Result of the use of average worker
Criteria: harvest time and yield
S2 more theoretical, still lot of obstacles, trade-off between cost to apply it and save from labour work,
Total time difference per scenario
R^2 low -> equations cannot be used as model input or output verification data
Units of behavior -> action
Junction control name: exclusive OR -> exactly one preceding/following action must be completed/executed
The core of this study
2) Harvested paths = trolley visits to paths
3) Not clear relation -> due to the multiple trolley visits in one path which lowers the average time spent in the path (also related to the trolley capacity)
2012 graphs; positive and negative relations;
* Compare simulation with fit curves => not possible due to low R2 (R2 too low to be used as model inputs or validation outputs)
* Indicate that measuring alone is not enough