How to Create a Machine-learning Model in Your Enterprise in Six Simple Steps
|John Wallace | Director of Innovation, Retail Solutions
Emerson Commercial & Residential Solutions
This blog summarizes an article from our most recent E360 Outlook, entitled “Applying Machine Learning for Facility Management.” Click here to read it in its entirety.
Machine learning is a subfield of computer science that refers to a computer’s ability to learn without being programmed. Although machines should be able to learn and adapt through experience, human interaction is still needed to produce desired results. Today, many facility management applications — for refrigeration and HVAC systems, for example — have taken a supervised learning approach that utilizes historical data to train an algorithm and predict an outcome from a series of inputs.
To create your own supervised-learning model, businesses can take these relatively simple six steps:
- Define the problem. It’s critical to have a keen idea of the problem you are trying to predict or solve, and establish well-defined goals of the application.
- Develop a data collection strategy. Data collection is achieved via inputs from a variety of information, including: temperatures, pressures, on-off activities (from motors, etc.) as well as the actions that result from these inputs. Your goal will be to predict the action that will occur for a given set of inputs. Data will be used to both train the learning model and validate the model’s performance.
- Create machine-learning models. Based on the training data collected and available inputs, you can create a machine-learning model that uses specific algorithms (math) to predict an action. Since different types of models may perform better or worse for a particular data set, you might need to create multiple models (different math) and then pick the one that performs best based on your data.
- Establish a standard. How closely does your model predict the action or result that came out of your training data? A perfect model would anticipate the result every time. While that usually doesn’t happen, the goal is to get as close as possible to achieving the desired results, and then use that model as a standard moving forward.
- Test the validation data. Based on the validation data from step two, evaluate the performance of your model. If the validation data doesn’t match up, you may need to step back and select a different training model, and then validate the data again. This is an intricate process. When and if the results do not match expectations, you may have to start from the beginning. Make sure you are collecting the right types of data before running the process again.
- Utilize the machine-learning model. Upon completion of your efforts, you should have a model that can be used to predict an action or result based on the available inputs. At some point, input parameters may change or another system modification may be required; in this event, you will need to go back periodically and update the model based on new data.