Course: Special Ship Automation Systems 2025
Homework Assignment (HW)
System Modeling Using Experimental
Measurement Data
Instructor: Assoc. Prof. Georgios Papalambrou
April 2025
1 The Problem
The experimental hybrid diesel-electric propulsion installation of the HIPPO-2
facility at the Laboratory of Marine Engineering (LME) consists of a Diesel en-
gine with a maximum power of 261 kW and an electric motor with a power of 90
kW, which are connected to the same shaft with an electric brake used to apply
load to the system. The HIPPO-2 installation and the parameters measured
during its operation are shown in Fig. 1. During experiments, numerous mea-
surements are taken, which, although very useful for monitoring, fault diagnosis,
etc., are quite difficult to obtain during the operation of an actual propulsion
system. To achieve these goals, virtual sensors are often proposed as a solution,
working in parallel with the other sensors of the installation. Additionally, we
often aim to create a digital twin of the experimental setup or actual propulsion
system to simulate the machine under various operating conditions and perform
diagnostic checks. These requirements—virtual sensors and digital twins—can
be addressed using neural network models.
2 Tasks to Solve
In this assignment, you are asked to develop such neural networks by following
these steps:
1. Analyze the data using various graphical representations.
2. Filter and transform the data.
3. Split the data into training, validation, and test sets.
4. Develop 4 simple neural networks to predict the following parameters:
(a) NOx,
(b) Fuel Consumption,
(c) Intake Pressure,
1
Figure 1: HIPPO-2
(d) λ value
using a combination of the remaining parameters (excluding the four being
predicted).
5. Evaluate the results on the test set and provide plots of the training
progress (Training loss vs. iteration, Validation loss vs. iteration).
6. Create new parameters by combining the original variables (feature engi-
neering).
7. Develop deeper or more complex networks (with justification) to
improve model fitting and generalization.
8. Evaluate the new results on the test set and provide training progress
plots.
2
3 Notes:
1. The data for the assignment can be found on the course website on Helios.
2. For step 4, use 1-2 hidden layers with at most 100 neurons. Create
one network per parameter.
3. Justify the criteria used to select the new parameters.
4. The network architecture can be of your choice. You may use ideas from
the literature (cite sources if applicable).
5. For optimizing the networks in step 7, use Optuna [1] or Weights & Bi-
ases (WandB) [2] for hyperparameter tuning and for experiment track-
ing. Alternatively, create a small hyperparameter grid manually and track
the training process in appropriate folder format.
6. All plots must include axis labels, curve labels, grids, and must not
be screenshots.
7. Code must be in Python.
8. Code must include comments explaining the process.
9. Presentations must be between 30-35 slides.
10. The assignment will be presented in two phases:
• Progress presentation: 16-23/5/2025.
• Final presentation deadline: 30/6/2025.
References
[1] T. Akiba et al., “Optuna: A Next-generation Hyperparameter Optimization
Framework,” in Proc. 25th ACM SIGKDD Int. Conf. Knowl. Discov. Data
Min., 2019, pp. 2623–2631. DOI: 10.1145/3292500.3330701.
[2] L. Biewald, “Experiment Tracking with Weights and Biases,” 2020. [Soft-
ware]. Available: [Link].