Mixer

With long productive processes and huge batches, the challenge for Mixer was reducing out-of-standard products.

Therefore the organisation needed a software application allowing the monitoring, control and prediction of the quality of finished products, together with additional efficiency and productivity parameters of the plant. The overall project aimed to: 

• Intercept production drifts in progress, to guarantee prompt intervention on the process;
• Reduce waste and optimise yield (with obvious positive effects on overall productivity). 

Following an initial exploratory phase of the data transmitted by the machinery and historicised on existing systems, an important cleaning and preprocessing phase was carried out. The identified anomalies required important corrections, to ensure, among others, the congruence of the information recorded at each production step (an essential aspect to move to the modelling phase with PCA).

The cleansed and rationalised data were collected in a relational database, which was developed in an attempt to model the structure, entities and relationships necessary to later save the production data and models.

This phase of the project thus provided an organic and organised view of the production data.

In the second phase, the incorporation of MSPC (Multivariate Statistical Process Control) techniques on top of the MES (Manufacturing Execution System) enabled the extraction of all the plant’s efficiency and productivity parameters, thus significantly improving the control of the entire production process, both descriptively and predictively.

Control is performed by means of principal component analysis (PCA) of the data, presented as:

• multivariate control charts: these allow the identification of potential failures in the dumps, with a focus on the tags that contributed to the errors themselves, thus isolating the most responsible ones;
• analysis of scores and loadings: these allow monitoring the trend of the principal components or their scatter plot by discerning the behaviour of the individual dumps and of the tags that determined it.

Before the analysis of the principal components, a technical diagnostic report of the Production Order is generated. This step, aiming at verifying the quality and consistency of the data, is a necessary condition for the correct execution of subsequent model creation and validity tasks. 

To enable real-time intervention in production, rather than just post-process control, a module was implemented that projects production data onto reference models every five minutes. This approach allows to prevent errors that could lead to deviations from the required quality levels.

In fact, the software application developed makes it possible not only to monitor the entire production process but also to create

• a model at the end of each production cycle (production order closing); 
• a model for each product dump (dumping closure).

The near-real time projection of production data onto the reference models, calculated using machine learning techniques, allows any production drifts to be intercepted and corrected in progress by the operator.

The reference model for each Production Order is selected based on indices of both product affinity and the quality of the model itself. This enables the system to make increasingly effective predictions.