-
-
Our LSTM model's prediction of pressure values 5 hours in advance in the event of a machine failure.
-
PAM Chatbot User Interface. Fetching and summarizing mock data from the SQL server.
Inspiration
Our journey with PAM began through in-depth conversations with automation engineers, who highlighted the critical need for precise machinery diagnostics and predictive maintenance solutions in industrial operations.
What it does
PAM, standing for “Predictive Analytics & Maintenance”, is a chatbot created to help assist companies by detecting faulty machinery and devices. By utilizing LLMs for a streamlined machine diagnosis interface, developers and engineers can be efficient and proactive in performing maintenance. PAM cleverly connects the LLM to a predictive machine learning model to learn patterns in machine time-series data, allowing users to predict failure-indicating metrics up to 5 hours ahead of time. Additionally, our connected database can store financial information (cost of repair and cost of machine downtime) and return this information when a potential failure is detected. By identifying potential dangers in advance, companies can save huge amounts of money, and focus time and resources towards their goals.
How we built it
Our application is built by combining a variety of components. Our front end was designed in a Python library called Reflex, which is essentially React.js in Python. To create our chatbot application, we leveraged OpenAI's ChatGPT API and utilized prompt engineering to fine-tune the model. To perform predictive analytics, we utilized Tensorflow/Keras to train a long short-term memory model (LSTM), a form of recurrent neural network, to predict future pressure values of a motor. Using moving averages, we identify severe deviations in future pressure values from previous stability to alert potential failures in our machines. To analyze our data, we constructed our backend using FastAPI and MySQL for the database, which allowed us to store and send machine data to our AI Model. Then, our model will predict any potential issues with the machine, and send this data to the LLM.
Challenges we ran into
The biggest issue we ran into was finding the data for our model. It was difficult to find sufficient data that we could use for our predictive analysis. Due to the time constraint, it would have been impossible for our team to create our dataset, so we spent lots of time locating a synthetic dataset to use.
Accomplishments that we're proud of
We are incredibly proud of the predictive LSTM model we built in Tensorflow. This model was able to predict pressure data up to 5 hours in the future, which was a huge accomplishment due to the limited available data. We are also proud of the connectivity of the system we built, as we were able to successfully integrate multiple pieces, such as the frontend, LLM app, backend server and database, and our LSTM Model.
What we learned
Through the development of PAM, we gained insights into the intricacies of predictive machine analysis and the power of AI-driven solutions in industrial settings. Our project illuminated the importance of comprehensive data acquisition strategies and the critical role of backend infrastructure in facilitating seamless operations. Additionally, we developed strong collaboration and project management skills. At all times, our project had a lot of moving pieces that required constant communication to ensure a final product would result.
What's next for PAM - Predictive Analytics & Maintenance Chatbot
In the future, we hope to expand this model to accept a wide variety of parameters and perform better diagnostics and predictions based on machine documentation. With this, we can have more specialized analysis that can benefit manufacturing and assembly line developers, and help companies reduce the loss associated with faulty machinery.
Log in or sign up for Devpost to join the conversation.