Welcome!

Microservices

Currently I work at a Bank, developing mainly backend solutions for Brazil's instant payment system, also kwown as Pix. Basically, transactions between different banks pass through the Central Bank, which regulates and defines the APIs that banks should use to make Pix transactions.

First, the debtor participant sends a message to the Central Bank stating the desire to initiate a transaction. Then, the Central Bank sends a message to the creditor participant, informing about the payment request. If the receiving bank accepts the payment, he answer with another message to Central Bank, which then notifies both the participants about the success of the transaction.

The project utilizes microservices architecture, dividing the components as Orchestrators (Integration microservices) and Atomic microservices. Orchestrators handle the integration and routing business logic, making the necessary calls to the atomic components, and listening/posting to MQ queues. Atomics handle the database connection, saving and retrieving data, among other more specific tasks.

Every component is implemented following the Hexagonal (Ports and Adapters) pattern, which focus on segregating as much as possible the Application layer from the Domain (business logic) layer. The main goal is to prevent coupling and dependency between services, providing an easy way to connected to their software environment by means of ports and adapters. For instance, if it were necessary to change the DBMS, this could be done by making changes only to the adapter layer, without changing the other layers.

Automation with NodeRed

During my internship at a soda factory, I started to learn about process Automation and control. The factory had a lot of equipments which weren't connected to any external network, only the OT network, so the data acquired was not storaged. From that, the Automation team began to study ways to access and store data aiming to improve the process control management. We utilized NodeRed, which is a programming tool for wiring together hardware devices, APIs and online services. NodeRed is built on Node.js, and offers a browser-based flow editor, allowing to write flows mixing JavaScript code and custom nodes. There are many libraries of nodes available, providing an easy way to connect with multiple interfaces. The image below shows how one of the created flows looked like.

The first step access the data is the Ethernet/IP node, responsible to connect with the PLC and access the internal variables to read their values. There is also specific nodes to interface with Allen Bradley/Rockwell or Siemmens PLCs. After that, we could save the values to a database. We utilized the InfluxDB node, due to it's simplicity and focus on storing time-series data. After storing, dashboards were built with Graphana to monitor in real time the behaviour of variables such as pressure, temperature and concentration.

In addition to continous data, we read boolean variables from the PLCs, indicating if some valve was on/off, if some process was ocurring or not, and also discrete numbers which represented which tank was selected, what recipe and what volume would be produced, since the production was controlled using a Human Machine Interface (HMI). For that type of variable we opted for a relational database, MySql. NodeRed also provides a built-in tool to create dashboards, called NodeRed UI. We created some UIs to show which tanks were filled with what product, the volume produced, and the date of production. This helped to manage production without relying on book annotations.

It was also possible to better manage Cleaning in Place (CIP) routine, which is the process of cleaning the tanks from the inside, circulating chemical products such as caustic soda, each one during a certain amount of time and temperature. The CIP process has a determined frequency to be done. Since the process of doing the CIP was controlled by the PLC, we could access the variable that stored which tank was selected to be cleaned, and store the number and current date. The Dashboard then calculates how many days are left untill the next mandatory CIP, and that number is presented with different background colors to indicate how much time is left (green parts of the image above). During the CIP process there are many programmed stages to occur, each stage represented by a variable, so we also calculated the time spent at each stage to assess the quality of the process and find improvement opportunities.

Finally, we connected a TV to the OT network to display the dashboards in the control rooms, and we could also use Tablets to access the network via WiFi. This project was part of a new phase of automation at the factory, and once again I could see the difference it made in people's routine, making data collection and analysis faster, more accurate and more reliable.

Low Code with Power Apps and Power Bi

Also during my internship, I develop an App using the Low-Code platform Power Apps, maintained by Microsoft. The goal was to test a prothotype and see if people would like to use it. Since I worked at a soda factory, the App was designed to manage the tasting sections made for quality assurance of the product. The process consists of tasting some samples of the week's production, which were previously collected and arranged in a laboratory for trained and qualified employees to taste. Each employee gives a rating for each sample and makes observations about the strengths and weaknesses of the samples. All notes were made on paper, and later transcribed to Excel.

The App focused on making this process more agile. First, the employee in charge of the laboratory created a "tasting section", which represent the collection of samples to be analized at that session. Each section had it's own number of samples, and each sample had it's own information, such as bottle size, expiry date, temperature, brand, and batch number. After creating the section, the tasters were able to access it in their cellphones and avaliate each sample, giving a score, and giving notes about the intensity of flavors. Besides, the App had a little section to see the avalitions of the latest section, and another section to see the expected characteristics of each brand.

After all the tasters had tasted and avaliated their samples, the data was sent to a Sharepoint List, which was used to storage all data. We also created an integration between Sharepoint and a Power Bi Dashboard. The metrics calculated by the dashbaord were used in meetings and discusions about improvement points for the product, mainly focusing in understanding what causes the lowest scores. We could filter by date, flavors, worst ratings, and others.

The App was develop in about 2 weeks, and the first adjustments were made very quickly. I partnered up with the laboratory responsible, who knew the business and taught me, and I proposed to him how the App could be develop. It was a great experience, because I could see the difference that it made in the laboratory routine, and I enjoyed transcribing the business logic into a new solution. The project was awarded as one of the best internship projects among Latin America factories, and later the development of an official, more scalable and robust solution was started.

Python and Image Processing

As my final paper, I conducted a research investigating the possibility of using a Raspberry Pi to perform an automatic inspection of the quality of expiration date on bottles. This idea came during my internship, where I saw the importance of this inspection on the production line. The equipment which does the inspection verifies every bottle, assuring that the expiry date is present, although, it doesn't verify if it's readable (not blurried) or incorrect.

My project utilized a domestic treadmill to simulate an industrial conveyor belt. The sample bottles were placed manually at the start of the treadmill. At the end was placed a Raspberry Pi, combined with an infrared sensor to detect the bottles, and a RaspCamera V2 to capture the images.

After the capture, a Python script processed the digital images using libraries such as OpenCV to manipulate the image and enhance quality for the proposed goal, performing operations such as Sharpen, Background Removal, Denoise, Adaptive Threshold, Dilate, Deskew, and others, ending with an OCR operation to extract the text of the expiry date region of the bottle. The set of captured images was used to test image processing methods. Different techniques and sequences were tested until the final combination was determined.

The extracted text was then avaliated as correct/incorrect using Regex Expressions, according to the possible patterns of each bottle tested. For instance, if a bottle has fixed letters/digits in certain positions, these fixed characters must be found. If it has 4 digits in a certain position, then 4 random digits must be found, but not letters.

Finally, I started to integrate the "Capturing image" algorithm with the "Processing image" one. Since the methods utilized in the images required more computational capacity then Raspberry Pi could provide, I utilized Web Sockets to comunicate between the Raspberry (client) and my PC (server), sending the captured images to the PC while the Raspberry waited the response. After receiving the conclusion of the analysis for that sample (OK/NOK), the whole process starts again and continues on loop. Here are some of the techniques utilized to process the images.

Sharpen

Threshold Binarization

Background Removal

Denoise

Dilation

Deskew

Arc Distortion

After graduation, my mentor during the research helped me to find an adequate journal to publish our work, and we submitted an article to the international publisher Digital Manufacturing Technology, which was accepted. The article is available here. The final paper is available at the digital repository of my University, here.