Predictive Quality and Anomaly Detection in the Steel Industry.

In the steel industry, product quality is one of the best indicators of the stability of the production process. Producing with a  consistent, predictable and adequate quality is a major ambition in industrial processes. Being able to predict what the quality of the product will be from the data collected during its processing makes it possible to detect, identify and possibly correct sources of instability in production. Hence the need to field systems of  predictive quality, predictive quality, to prevent and resolve issues that may compromise material characteristics.

In the case of steel manufacturing, in fact, resistances to mechanical stresses are among the most significant variables to be taken into account. In large-scale manufacturing plants, testing extensively and accurately is both economically and time-consuming. In the context of steel manufacturing, predictive quality projects make it possible to improve aspects of product testing.

How Machine Learning models can help improve product quality

This requires managing the large amount of data generated by sensors on production facilities. The main challenges of this phase are as follows:

• accommodate and make data available in near-real-time: an IoT-oriented architecture is essential to channel sensor data to ad hoc databases that can offer high performance to the artificial intelligence models that will have to interpret them.
• cleanse the data from the structural uncertainty present in the process: real processes are subject to constant evolution over time to which maintenance, renewal or plant replacement events along with unpredictable behavior due to external factors contribute.
• select the most significant variables for quality prediction: in the immense stream of data coming in from sensors, only a portion carries significant information to the prediction of product quality so feature-selection is essential to reduce the amount of data without loss of useful information.

By applying these best practices, the massive and disorderly flow is transformed into a consistent dataset, essential and relevant to be processed into accurate process modeling, or a machine learning algorithm trained on the data history. The model learns to output one (or more) product quality indices from the processing data collected along the production chain making testing more extensive and automated.

Regesta Lab’s approach to predictive quality

During the development process, we kept an eye on the “explainability” of the models, understood as the ability to trace the causes of too low or too high a prediction provided by the models themselves. We know it is as important to train the models properly as it is to have the ability to learn from the models ourselves once they are developed.

Results were produced through the use of supervised models trained on a history of products and related quality tests used as labels. The supervised approach involves training a model by providing as input to it a portion of the historical data along with the desired outputs for that data. In this way, the model learns from the input-output examples provided by adapting its internal parameters. One of the obstacles to be overcome in taking this route is to associate test results (labels) with the data that participated in the production of the tested product.

One step further: on-line anomaly detection

Timeliness and accuracy are essential requirements for intervention, by repair teams, during unforeseen situations such as failures, malfunctions or unwanted plant shutdowns. In this regard, we have supplemented the work done for steel quality prediction with an automatic anomaly recognition mechanism. The purpose is to provide a near-real-time alerting system that can provide effective reporting regarding the stability and status of the production process.

Such an anomaly recognition mechanism can interact with predictive models for product quality by allowing the correlation of detected anomalies with product quality.

This approach was conducted in an unsupervised manner, relying on the self-similarity of process steps during mass production of steel workings. Unlike a supervised model, an unsupervised model cannot rely on the desired input-output examples to adjust its parameters. This situation is often encountered in the recognition of anomalies in real processes since there is no real formal definition of anomaly on which to build the above examples.

Therefore, the experience and sensitivity of domain experts who are able to validate and compare the results of unsupervised techniques is essential. This approach  human-in-the-loop made it possible to greatly improve performance based solely on the statistical properties of the variables.

Predictive quality techniques and solutions are essential in all fields where processing does not include the possibility of subsequent corrections. A significant example is the steel industry. I  models made by Regesta Lab make it possible to prevent and recognize anomalies in the material produced.