Personal website

Pig farm monitoring

In this project, a proof of concept is delivered to demonstrate the feasibility of monitoring pig noises in a barn. There is a device in the barn that records sound 24/7. These sounds are labeled according to individual sounds such as sneezing, fighting and coughing.

The first goal is to be able to monitor these individual sounds automatically. Later when this is fully implemented, the goal is to be able to distinguish positive and negative pig noises from each other.

The name of the company for which this is being implemented nor the product that is used cannot be mentioned yet.

The GitHub repository is private because this project was made for a company that has all its legal rights.

Deeplearning on avocado images

During my third-year internship I had to carry out an individual assignment. I ended up at ExperienceData. They have a machine that can objectively assess avocados on quality and ripeness. This is done using the avos-mini. The avos-mini is a machine in which an avocado can be placed. The machine analyzes the avocado through different techniques. One of these techniques is images to which smart software is applied. During this measurement, the fruit remains completely intact.

I have researched different deep learning model architectures and compared the performances with each other.

In addition, I developed a tailor-made data augmentation library for the images for the deep learning model.

During this internship I learned a lot about how deep learning models are used in practice, learned more about Convolutional Neural Networks, computer vision in general and all the possible details of data augmentation on images.

I had a great time at Experience Data and would like to thank everyone for all the wise lessons and the opportunity they have given me.

The GitHub repository is private because this project was made for a company that has all its legal rights.

Broken fence detection

During this project, everything was built from the ground up. We needed to create a deep learning model that can determine if a farmer's fence is broken or not.

The first step was to collect data on fences. We got access to a drone so it was easy to collect videos of fences. After that, we started to create a convolutional neural network (deep learning model). Our first drone footage consisted of chain-link fences. We discovered some problems, the cnn wasn't able to learn the chain-link patterns. Afterwards we found out that this was because there is too little contrast between the fence and the background. We tried lots of filters to increase the contrast, but without any positive results.

Now we were back at step one! We needed to collect data from new fences. We chose a wooden fence with more contrast. This decision was the best we ever made. Our deep learning model started to recognize the wooden fence very precisely. Now we could proceed to optimize the model.

For the final convolutional neural network we used a U-net architecture. We finetuned the learning rate and the optimizers and added some data augmentation to make the model more robust. The model was able to recognize the fence almost perfectly.

The picture above shows how the workflow of the model development works. We started with frames. The frames were labeled in the online tool Labelbox. Now we had a dataset with masks and true images. We used data augmentation on the whole train dataset. And last but not least we trained the convolutional neural network with the U-net architecture on the train data. This gives us a deep learning model capable of recognizing fences from images.

We then wrote software in which we gave a video as input and we got back a video where the fence is predicted and marked in red. (see the video below).

Now that we could recognize the fences we could start to recognize defects in fences. We marked a break in the fence pattern as a defect. The deep learning models gave a predicted mask as output. We could use that mask to determine breaks in the pattern. We made a second dataset to create a second model to determine the breaks in the fence pattern. The image below shows an example of the created dataset.

To determine the breaks in the fence patterns we used a simple convolutional neural network as a classifier. The classifier was able to predict 90%+ of the test data correctly! This was an amazing success!




The GitHub repository is private because this project was made for a company that has all its legal rights.

Intersection

This project was done for the Dutch company Vialis. They wanted to know how well we could simulate their intersections.

Vialis provided us with all the necessary data of the intersection. For example, the coordinates of all the objects on the intersection like traffic lights and sensors. We even got data on when sensors were activated. For example, when a car was waiting for the traffic lights to go green.

We started by plotting all the useful objects like traffic lights and sensors on a map. We visualized this in the picture above.

After that, we made the decision to make the simulation in Python with the library Mesa. We started with a live visualization of all the sensors going on and off. We first needed to plot the objects on a canvas. We scaled all the coordinates between 0 and 100 to make it fit. This resulted in a real-life ratio between objects.

We succeeded in visualizing the sensors in real-time going on and off. Now the hard part has started!

We managed to reverse engineer the sensor logics, but this was harder than expected. We succeeded in making an agent-based simulation of the intersection but it was not 1:1 realistic. There were too many factors to make it as realistic as needed, especially in our time frame.

We started to do some research with our simulation to search for improvements in the traffic light logic. This research wasn't a great success. The simulation missed some features because we were out of time. As a result, we were unable to draw firm conclusions from our research.

In the end, we managed to develop a very good simulation of the intersection but we were out of time to add some other important factors. The best thing about our written software is the reusability. If we were asked to simulate another intersection we could do it instantly, just by importing the data!

Body conscious

For this project I made a graphical user interface with JavaFX. The software asks for all your body characteristics, such as height and weight. With this, the software makes calculations that predict how many calories you need to eat to reach your body goal.

Neural networks

This was my favorite course during my education. We learned everything about neural networks in detail. It all started with the perceptron. Then we learned how backpropagation works. When we managed to understand backpropagation in detail we learned about all the activation functions like the sigmoid, relu, and tanh.

When the basics were understood we were asked to develop our own neural network library in Python from scratch. (see the GitHub icon for the code!)

In the end, I learned about convolutional neural networks (deep learning) for the first time.

This post is in my portfolio because I had an amazing time learning everything about neural networks and this was a start for the real AI world to me!

ProRail

This project was made for the data science team of ProRail. It was the first real-life project where I was able to use machine learning.

When there is a railroad failure, ProRail wants to make an estimation of the time an outage will last, this helps the planners and travelers. When a malfunction is detected, data is collected and put into the system. They used people to guess how long the outage will last, but they want to automate it with computers.

ProRail gave a huge dataset with lots of unusable data which first needed to be cleaned. After the cleaning process (which was the hardest part) I started trying many machine learning algorithms. I soon found out that regression was the best option, but there are so many types of regression. Like linear regression, random forest regression, and decision tree regression. They all gave some good results but I chose linear regression because other regressors weren't much better and linear was the simplest.

The final results were not how I expected them to be. But my teachers told me that it was due to the data and that a better result was not possible. I even managed to get a better result than the teacher could ;).



This github repository is private because ProRail owns all the legal rights.