Democratizing machine learning with a no-code platform

In today's world, artificial intelligence (AI) is transforming many aspects of our lives. Developing AI typically requires high-level programming skills. However, no-code AI offers a user-friendly way to create AI models without coding. Nowadays, data is everywhere. We can generate it in databases using devices such as smartphones, smart watches, cameras, etc. Additionally, computers, including no-code AI, are increasingly helping data scientists find the best architecture for machine learning (ML) models. In the future, we may even have autonomous AI. SMS group recognizes the importance of machine learning in improving a blast furnace’s operational efficiency and sustainability and uses it to develop its digital solutions.

No-code AI is an innovative technology that enables the creation of AI models without any coding knowledge. Users design models visually with a graphical interface, democratizing AI for a wider audience. This entails the partial automation of tailored data pre-processing and model training. The objectives are to foster quick iterations and accessibility for non-data scientists. No-code AI is ideal for initial ML model development and typically addresses 80% to 90% of customer needs. Data scientists can customize the first iteration, incorporating domain knowledge to meet 100% of requirements.

Looking at the development of the AI model pipeline, traditional machine learning involves data scientists, domain experts, and data, and so has an interruptive effect. In contrast, traditional no-code AI only involves data and the robot, simplifying the process. While traditional machine learning requires data preparation, features engineering, and model selection, no-code AI only requires model selection based on the relevant use case. As a result, no-code AI enables more efficient processes, speeding up development and deployment as well as creating and scaling machine learning-driven products. In addition, it reduces costs and improves profits for businesses of all sizes.

AIXpert is a no-code platform created by Paul Wurth, a company of SMS group, for training machine learning models. The back end is optimized to handle time series data, which reduces processing times. The target user is a domain expert in an industrial process. AIXpert is part of the DataXpert toolset, which includes RulesXpert (rule engine for knowledge creation) and BIXpert (tool for developing visual applications).

In traditional no-code AI, the process involves the use of a robot and data without the need for a domain expert. In contrast, AIXpert aims to combine the expertise of domain experts with data science. Its functionalities enable domain experts to validate datasets, incorporate domain knowledge, and proceed to model training. While automation focuses primarily on the model training aspect, the development focus in AIXpert is on giving domain experts the ability to leverage their expertise in data cleansing, a requirement for training good machine learning models. Moreover, the tool focuses on data analytics and visualization to enable comprehension of the dataset, thus ensuring its suitability for training machine learning models.

The core idea behind AIXpert is to create a tool that supports domain experts in constructing predictive models effortlessly and deploying them seamlessly for production, streamlining the data preparation process using their domain expertise. The knowledge of domain experts is vital for laying a solid foundation for AI products as they possess an in-depth understanding of the data and the ability to process it in a manner that enhances model training.

The functions of AIXpert include displaying and analyzing signals, editing parameters, specifying signal groups, cleaning data, and visualizing data distribution and quality. It is possible to image the correlation between input and output using a matrix that helps domain experts determine if uploaded data is suitable for training the model. CSV files can be uploaded, or the time series data source loaded from a database using connectors implemented in AIXpert.

There are two modes for training ML models: The first one involves manual training with data, requiring specific domain expert training. The second mode automates the model training and is typically the initial iteration to train a model quickly, put it in production, and show the customer that the product generates information and can be billed. We can train four types of models:

• Regression model: This model can predict a numerical value. For example, a trained regression model can predict the hot metal temperature with a time horizon of three hours, thus maintaining stability in the blast furnace.

• Classifier model: When a classifier makes a prediction, it assigns a non-numerical value called a class. A classifier can predict sudden changes in blast furnace permeability and represent them visually with circles. It displays the class prediction over time and provides an overview of the prediction error for each class.

• Auto encoder: An auto encoder can be used to train the model to detect anomalies by learning the relationships between input signals.

• Pattern detection: The model allows users to detect a pattern within a time series by comparing it to an available reference pattern selected by the domain expert.

With the "Switch to autoML" feature, the robot assists domain experts in training the best ML model by autonomously identifying it. Experts can specify the number of models to be trained, generate predictions before deployment, and, by shifting each input value and analyzing its impact on the output of the model, understand if the model has learned the known relationship between input and output. The model can be exported in RulesXpert.

In the near future, AIXpert will include more tools to help domain experts understand and effectively validate datasets for machine learning training. These tools will include advanced visualization and algorithms for statistical analysis. AIXpert will feature advanced pattern detection to assist domain experts in identifying patterns in time series to gain new insights. These labels can be used to train a machine learning model that can forecast the occurrence of anomalies in a process. Advanced pattern detection for data labeling is a crucial aspect of predictive and prescriptive maintenance.

At SMS, we believe that dedicating AIXpert to domain experts is the key to creating optimal machine learning models for industrial applications. This will put us in a competitive position to develop machine learning models for industrial use cases with the help of our domain experts.