MUNICH, GERMANY, March 18, 2021 /
aivis currently comprises three product groups. Insights is used for the automated creation of reports to explain why something is happening. Predict enables, among other things, the automated creation of virtual sensors to determine what is happening. Prevent provides continuous protection against process interruptions, for example.
All of them have in common that their usage does not require any data science expertise. PerfectPattern is thus democratizing the utilization of artificial intelligence in the industrial sector with aivis. On its website, the company provides information on how to get started.
Under its previous name Pythia, parts of the aivis technology have already convinced several well-known companies including VOITH and MPDV.
In the manufacturing industry, artificial intelligence is primarily used to optimize production processes. This mostly involves analyzing data in order to identify patterns and thus predict process flows. This enables early intervention when efficiency losses or even production interruptions are impending.
Until now, this has always required data science knowledge. This meant that process engineers and domain experts had to rely on appropriate support for complex tasks such as data cleansing, data analysis and model creation.
aivis applications perform these tasks on the basis of unprocessed and unsynchronized raw data in an unsupervised, fully automated manner and at high speed. Fabian Rüchardt, CEO of PerfectPattern, comments: “This enables data owners for the first time to use artificial intelligence methods on their own with little effort. At the same time, the result is expressed in an understandable way – this is AI in plain language. Thereby, we are initiating a new era in the use of artificial intelligence.”
Many machine learning platforms basically work like search engines for mathematical methods and solution approaches. They then often suggest several methods from which users have to choose. Christian Paleani, Chief Scientist at PerfectPattern, says: “aivis does not ask users to pick the best mathematical method to solve the current problem – aivis is already the best method. It is based on a very comprehensive and powerful mathematical methodology that is perfectly adapted to the industry’s data challenges and issues.”
The underlying theory of aivis is a further development of kernel machines with methods from stochastic differential geometry and quantum field theory. These allow, for example, hyperparameters to be calculated geometrically directly. Paleani summarizes: “Our methods not only eliminate a lot of effort in model creation, the result is also transparent and interpretable.”
In the areas of data analytics and predictive analytics, aivis offers various advantages over conventional approaches such as neural networks. For example, the technology uses a very efficient and targeted learning strategy that allows to deliver results very quickly while avoiding the well-known problems of overfitting and hyperparameter tuning. Although it can process huge amounts of data very quickly, the amount of data required to find results is actually comparatively small.
Also, data analysis and model building used to be time-consuming, expensive, and required extensive data science expertise. aivis works quickly and unsupervised with any amount of unfiltered raw data.
This means that production experts are now able to work independently, efficiently and with much less training data than before. The system generates the necessary models unattended and without the need for manual intervention.
Even in complex scenarios with learning data from thousands of data sources, a result can usually be calculated within a few minutes – whereas with conventional methods this takes hours.
In industrial AI applications, the number of relevant process parameters can become very large. Paleani explains: “Previous approaches hit their limits with several hundred parameters – aivis can easily handle thousands of parameters. In addition, aivis also provides indicators for causes of trends – an important prerequisite for identifying why something is happening in a production process.”
Currently, aivis comprises the following three product segments:
aivis Insights: The applications in this product group analyze data with regard to the question of why. They search the data for hidden patterns and uncover relationships, dependencies, reactions, causes and influencing factors. On this basis, it is possible to explain why something happens. The result is a clearly understandable report that addresses the relevant question and often brings an “aha” moment.
aivis Predict: In these applications, everything focuses on the question of what. It is answered with easily created software modules whose core is a trained mathematical model.
Probably the best-known application of this is the creation of virtual sensors. Based on easily determinable input variables, they continuously determine a target variable that indicates what is currently happening. Virtual sensors are mostly used where the utilization of a physical sensor is difficult or impossible or laboratory measurements are to be replaced by live measurements.
Another application are so-called estimators (regressors), which determine a target variable similar to a look-up value for easily determinable input variables. The underlying mathematical model takes into account even the most complicated dependencies. These estimators are used, for example, to determine set-up costs, mechanical strength or viscosity.
Classifiers as the third application of aivis Predict indicate, for example, whether the relevant production parameters are currently in a desired target corridor or are at risk of breaking out.
aivis Prevent: These applications are used to prevent adverse progression in production processes such as material- or machine-related process malfunctions, interruptions and failures. It is based on live monitoring of process data to detect problematic situations at an early stage and recommend countermeasures. In this way, costly production malfunctions or interruptions can be minimized or prevented entirely.
In prevention based on seen faults, first the historical data is analyzed to identify the different situations with their respective reasons that led to the specified fault. Targeted countermeasures can be defined for each of these situations. In regular operation, the problematic situations are identified on the basis of the live data stream and the appropriate countermeasures are recommended.
Prevention based on unseen errors utilizes the historical data to learn the normal behavior of the system. Based on the live data stream, it then continuously analyzes whether the system is still in a normal state or deviating from it.
Getting started with aivis is easy – nothing more than random data sets in table or time series form and a question or data analysis request associated with the data are necessary. PerfectPattern provides information on its website about the nature, scope and possible use of the data.
In the near future, a demo app will be available to give an easy and quick introduction to the technology and a hands-on overview of the capabilities.
aivis is not the result of a single, isolated approach, but the combination of many subsets of mathematics and theoretical physics.
During development, many fundamental ideas and concepts were tested for their effectiveness. Some were incorporated, many were discarded, or some were extended by newly invented approaches.
Paleani says: “Our goal was to create a system that would enable people working in the manufacturing industry to quickly gain insights from their data, even without data science knowledge, to draw data-driven, transparent conclusions and build software solutions.”
