• 

• 

• 

• 

• 

• 

• 

With the sensor specific encasing complete, all that was left was to put those two sub-assemblies together and have an enclosure to hold the raspberry PI, breadboard and all the electrical wires.

A few aspects of this design:

Airflow was required to keep the electrical components at a stable level so slots were added to create a path for air to exit and enter the enclosure.

The fan will be suspended over the Raspberry PI.

The LED screen will be face up on the cover along with the push button.

The breadboard will be attached to the cover of the enclosure. This is a stipulation of the short wire lengths. Similarly, there is a spot for the raspberry PI to be added to the cover of the enclosure, as a fail safe, in case placing the raspberry PI on the bottom of the enclosure does not work due to wire length or other unforeseen circumstances.

Overall, the design is definitely more aesthetically pleasing. The numerous lessons over the initial designs and the prototypes have all come together to be implemented in this final product. Now all that is left is to get all the components together, finish 3D printing the remaining pieces and set-up the final product.

Onward to symposium!

Welcome to the sixth and final instalment of the overview series! We will be discussing the next steps.

The textile classification sensor as it is right now is great for many reasons as discussed in the fifth instalment. In order to improve the following steps will be taken.

Work will be done to:

• Sort between even more types of textiles, such as leather.
• Sort textile blends (i.e. 50% Cotton / 50% Polyester)
• Improve accuracy of the system (larger image database for machine learning algorithm)
• Install into factory mechanical system
• Use consistent camera setup between training data and inference
• Get a better camera and IR sensor
  

Thanks for tuning into the overview series! We hope it summarized the work done over the last two terms. 

Group 19 is signing off for now!

Welcome back to the overview series! In the fifth instalment the results from the benefits of the current design of the sensor are discussed.

During the brainstorming and prototyping of this project, multiple designs were considered. In the end a classification sensor was decided on as the one the team will pursue. The following are the main reasons for this. 

A sensor can easily be integrated into any process! This makes it very desirable for a wide variety of companies. It can be integrated into a factory line in any manner that works best for that case. It can easily be used as a handheld solution for smaller scale applications. Furthermore, this solution is non-invasive which means that it leaves the textile in question in the exact same state as where we began. 

We hope that this design really enables the industry to become better and more green!!

Welcome back to the overview series! In the fourth instalment the final results from the computer vision model is discussed. 

Our final iteration of the computer vision model is a ResNet50 multi-class classification model that classifies between cotton, denim, and polyester. During training, regularization and data augmentations (blur, affine, etc) were used to artificially increase the data size and avoid overfitting. The computer vision model achieved an accuracy of 80.83% on the test set which is comprised of 10% of the fabrics dataset for cotton, denim, and polyester. Due to the Identifybre camera apparatus being different from the one used to collect the training data, in practice the model accuracy is somewhat lower. Being able to invest in a better camera and also collect a much larger dataset to train on would drastically improve the accuracy! For the prototype these are definitely promising results. 

The first figure is that of a confusion matrix of the test set and the second figure showcases the difference in the camera quality.

• 

After the initial testing prototype. The team was ready to put the camera sensing and IR sensing together. Using the 3D printed parts from the last design, the team put the parts together to make a make-shift demo prototype. This prototype was functional and able to detect clothing.

This prototype will be used to demo our product to our adviors on March 6th.

Again, this prototype was by no means the end result. The prototype was needed again to give the team room for improvement for the final day. The prototype gave the team a better idea of how close all the components needed to be. The team spent some time, thinking about this prototype and what would be needed in the final design/product.

A few lessons learned from this prototype:
1. A fan is needed for the Raspberry PI
2. Both the IR sensor encasing as well as the camera encasing need to be the same height.
3. The final design should be portable. The idea is that we can demo it well.
4. Manufacturing needs to be thought about. The wires etc, need to be hidden. The astethetic appeal of the prototype was severally lacking
5. 3D-printing is a great method for manufacuting the enclosure but the tolerancing needs to be adapted to the lack of precision that comes with 3D printing. The high variance needs to be designed around in order to make all the components fit snug.

Those are the major lessons the team took from the prototype. This prototype was submitted and used for the final day demo. Since the encasing for the camera and the IR sensor was already complete and seemed to work well, the only design needed to be finalized was for the overall enclosure.

Welcome back to the overview series! Here is the third instalment that goes through the solution developed.

The textile classification sensor can be broken down into three main components: Mechanical, electrical, and software. 

The mechanical design of the sensor is a portable, custom-fit, aesthetically pleasing design that was 3D-printed. After initial designs, a few iterations of prototyping, and consideration of feedback from these trials, the final CAD design was established. These can be seen in the images below. 

The electrical design of the classification sensor consists both of a camera and an IR sensor. The near-infrared spectroscopy sensor is connected to a raspberry pi through I2C. It measures how various materials absorb and reflect different wavelengths of light. The camera is used to take close up images of the article of clothing. This is to analyze the texture of the textile. The raspberry pi is used to interface with the peripherals and run the production software. 

The software design is described in the following flow chart. The image taken by the camera is passed into a convolutional neural network trained on close-up images of cotton, polyester, and denim clothing. The data from the near-infrared spectroscopy sensor and color information from the camera are passed into a heuristic algorithm to aid in the prediction of wool. The results from the machine learning model and the heuristic algorithm are fused to make a final prediction about the clothing’s fibre composition. 

All three components work together to create a textile classification sensor!

Welcome back to the overview series! Here is the second instalment that goes through the main problems we found.

Fast fashion has propelled a culture of continuous buying and throwing away of clothing. Rather than the traditional 2 cycles per year, fast fashion puts out an average 50 cycles per year of new clothing lines, resulting in millions of tons of textile waste each year. 85% of used textiles go to landfills rather than being recycled. There are little ventures recycling clothing as it currently is a manual process done by hand by workers. Furthermore, different types of textiles require different recycling processes. 

Moreover, the fashion industry emits 10% of the world’s carbon emissions. This is more than international flights and maritime shipping combined!

Another important note is that this industry is the world’s 2^nd largest consumer of water, and produces 20% of the world’s industrial water pollution. The amount of water required to make 1 cotton shirt = water 1 person drinks for 3.5 years! This number increases with the type of clothing as well. The amount of water required to make 1 pair of jeans is the same amount of water 1 person would drink for 10 years! Synthetic textiles also pollute when washed as they release micro-plastics into the ocean (500 000 tons/year). 

Overall, this industry is in need of some innovative solutions to tackle the issues present!

This is the first post in the overview series which will be a series that summarizes the journey the group went to in order to create the final textile classification sensor!

The journey to choosing this project started off with the desire to create something that would help to make the world a better place. As engineers we have a duty to create designs that are ethical and increase the well being of others.

Our diverse group of engineers all have a passion for giving back to the community and for fashion. We felt that tackling the textile waste problem that is occurring would be a perfect fit. 

As more and more research was done, it became apparent that there was little to no advancements with automated recycling of textiles. All the textiles that have been created to this date, still exist on the planet and that amount still keeps increasing. We wanted to find a solution to dealing with all this waste.

This led to the creation of our objective for the project. The objective is: Design a textile sensor that will enable a variety of companies to recycle fabrics in order to reduce wasted resources, recycle scrap fabric, save company costs, and lower the fashion industries environmental footprint. 

Attempts were made to improve the accuracy of the detection of wool using computer vision but the average recall remained around ~40% on the validation dataset. We suspect this is largely due to the lack of original samples for wool (only 90 in our dataset) in addition to the similarity of the texture of wool and cotton. Despite the use of data augmentation and regularization, the accuracy of wool detection remained largely the same. Thus, wool is being dropped from the computer vision aspect of the project for the time being and denim is being added as a replacement. Denim was selected as the third material due to it having the largest number of samples in the dataset (behind polyester and cotton).

Now that an initial prototype was complete, the team decided that 3D printing would be the best way to custom-make encasings for each of the sensors. The IR sensor would have it’s own encasing specific to it’s features. The camera similarly would have it’s own encasing.

The camera encasing was left at 2 pieces. One for the camera to be placed in, then another enclosure on top to seal the camera in place. The IR sensor would require 3, one component for the IR sensor, another for the tunnel holding the LED’s and a last component on top to enclose everything again.

Each of the models was custom designed to be incorporated in the 2nd prototype. This will be the encasing’s for each sensor that will be used in the demo to our advisors.

The above images are the 3D printed bottom component of the camera sensor. For the most part, the print went great. The design itself held well. The prototype was functional and the team was impressed with the result.