Project Name

Compressor Short Cycling Detection Using Machine Learning

Industry

Manufacturing

Technology

Machine Learning, Python, AWS

Overview

Our client operates within the HVAC (Heating Ventilation and Air Conditioning) Service Provider, where they specialize in high-quality heating, ventilation, and air conditioning systems. They have a critical business imperative to optimize compressor performance and reliability by harnessing AI/ML techniques, with a specific focus on the detection of short cycles in their compressor units.

Challenges

The client encounters a challenge stemming from disruptions or anomalies within the dataset, impacting their ability to work with reliable data.
The client encounters a substantial challenge stemming from improperly labeled data, which hinders their ability to address the issue of detecting short cycles in compressors effectively.
Identifying and detecting short cycles in the compressor poses a significant challenge for the client.

Our Solution

We provide a comprehensive solution to our client that helps to solve all the complex challenges arising that are:

We employed data extraction and data cleansing techniques to assist the client in overcoming their challenges related to data quality, ensuring more accurate short-cycle detection.
Leveraging the XGBoost Library, we developed and trained a robust machine-learning model to process the enhanced data, aiding in the identification of short cycles in compressors.
We facilitated the deployment of the solution by integrating it with GitHub and utilizing AWS CodePipeline, streamlining the process and ensuring efficient code deployment to AWS auto-scaling.
To optimize resource utilization, we implemented a scheduler that dynamically scales up or down the number of running services in Amazon Elastic Container Service (ECS) based on predefined time triggers.
We created a service within AWS ECS to run multiple services as containers within a Fargate cluster, enhancing the efficiency and manageability of the deployment.
The AWS Application Autoscaling service was configured to operate for a predefined duration, automatically adjusting the number of running services based on demand. Once the designated time elapsed, the service gracefully stopped running to optimize resource usage.

Data Flow Diagram

Conclusion

By leveraging AI and ML technologies for Compressor Diagnosis, we developed a model with an F1 score of 0.86 and implemented it in production. This transition marked a significant shift from a “fail and fix” approach to a more intelligent solution, allowing our client to proactively monitor and improve model performance, ultimately leading to more informed and effective decision-making.