Inspiration
One of the biggest obstacles in recycling, a $500 billion dollar industry, is its high costs. One third of these costs come from the human labor needed to sort and separate different recyclable materials. In attempt to reduce the cost of recycling, we created a way to automate the recycling sorting process upstream at the civilian level.
What it does
Upon disposal of PET plastic, paper, or aluminum cans, AutoSort sorts and separates the materials. As the trash is dropped through the trash chute, it crosses the pathway between a red LED and photoresistor, generating a unique analog signature that is dependent on the type of recyclable item. The analog signals are read as ADC values on an ATMega328p, which classifies the trash as paper, plastic, or aluminum. Based on the classification, the microcontroller writes a pair of servo motors to move to a certain position. The servo motors control two doors in the bin that direct the trash to secondary bins assigned to contain either plastic, paper, or metal.
How we built it
We built AutoSort by focusing on three components: sensing, classification, and actuation.
Sensing: We needed the sensor to provide analog outputs dependent on the material of the trash and not the size, shape or weight. We also wanted our final product to have as little moving parts as possible. In order to meet these criteria, we considered radar and optical sensors. Although we determined that the radar sensor was great at distinguishing metals from other materials, we found that it could not output distinct enough analog signals between paper and plastic. On the other hand, using a red LED and a photoresistor gave us distinguishable analog output among all three classes of materials. Thus, we ultimately chose to use the red LED and photoresistor and determined the average ADC readings for each material by repeatedly dropping sample materials through the trash chute in front of the sensor.
Classification: The ATMega328p polls ADC readings and stores these readings in a 40-integer array once the sensor picks up under 5 consecutive ADC measurements above the ambient baseline reading. Once the array is full, the average is taken and compared to threshold ADC values determined from prior experimentation. The microcontroller makes a classification when the average ADC value crosses one of the threshold values.
Actuation: Upon classifying the material, the microcontroller outputs pulse-width modulation signals to control servo motors to move two doors, directing the material to the correct categorized bin.
We chose AutoSort's method of actuation because it was a simple, space-efficient, and time-efficient way of actuation without requiring moving parts in front of our sensor. We also chose this design because it allows gravity to do all the work to move the trash. in addition, the design works well with readily-available servo motors provided by the Bioengineering Lab.
AutoSort's physical body was designed in SolidWorks and built by laser-cutting sheets of acrylic and press-fitting the pieces together.
Challenges we ran into
The biggest challenge we ran into was getting consistent sensor output distinct to each material. We tried several analog signal processing methods to make the output of the radar sensor more robust, including signal amplification and high-pass filtering. Despite these efforts, we could not get robust enough output from the sensor. To top it off, the radar sensor ended up breaking at the last minute. In fact, using the LED-photoresistor setup was a last-minute adjustment in response to this. And ironically, this worked greatly in our favor, even simplifying our code.
Another challenge that we faced was building the physical body of AutoSort from scratch. Neither Frank nor I are mechanical engineers and have limited experience with machining and designing physical systems. In addition, designing and building the body was more complicated than expected, especially figuring out how to attach the doors to the motors and the motors to the walls. However, overcoming this challenge was simply a function of time to learn new skills and iterate through different ideas.
Accomplishments that we're proud of
We are especially proud that our final product can perfectly sort plastic bottles, aluminum cans, and paper cups. AutoSort used zero dollars of our budget and we made due with the resources already provided for us at Penn Engineering. Furthermore, the project is a direct application of concepts we learned throughout the class such as ADC on the ATMega328p, output compares, interrupts, and timers. In that sense, it was a great capstone project for ESE 350.
What we learned
From building the physical parts of AutoSort, we gained and improved practical experience and skills in mechanical design and rapid prototyping. We also learned important lessons in completing a project against a deadline. We not only needed to plan for unexpected road blocks and bugs but we needed to learn to adjust and design flexibly with that in mind. This proved especially useful in overcoming our issues with radar sensing.
What's next for AutoSort
Future improvements for AutoSort would increase the types of materials that it can sort. The physical size can be scaled up and more experiments to determine average ADC readings can be performed to accommodate more classes of materials. In addition, more sensors such as the radar sensor can be added in series so that sensors can specialize in identifying one type of material. The algorithm can also be improved by looking for more features in addition to the average of the 40 ADC readings. With improved feature selection, machine learning methods can be applied to classify the material.
Log in or sign up for Devpost to join the conversation.