Predictive maintenance describes the prediction of future wear or failure of industrial equipment and machine components based on sensor data and mathematical algorithms. Critical operating parameters and patterns in the data that indicate failure are automatically detected. In this way, pinpoint preventive measures can be initiated, and periodic maintenance and expensive downtime can be reduced to a minimum.
The Hackathon: In January 2017, ANDRITZ and PIONEERS DISCOVER invited seven select software startups from Europe, including 7LYTIX, to a Big Data Analytics Hackathon in Graz / Austria. In these three days, five core issues in predictive maintenance had to be solved in the best feasible way. Andritz's industrial plants are equipped with many sensors that permanently measure plant performance.
Accurate lifetime predictions: By identifying and using the best machine learning algorithm on a test dataset, 7LYTIX was able to identify patterns in the sensor data and develop a predictive model that accurately predicts the timing of upcoming machine failures over the following 20 days. The prediction accuracy achieved was outstanding with an F1 value of 0.96 (0 = bad, 1 = perfect prediction).
Decisive operating parameters: In addition, 7LYTIX was able to determine the operating parameters that have the greatest impact on the failures, maintenance and operating costs of Andritz's industrial plants. As part of this, it was immediately possible to determine the production phases in which the plants work most cost-effective and which variables play a decisive role.
Benefits: The benefits for manufacturing companies are obvious: cutting down on expensive downtime. Reduce maintenance. Reduce production costs. However, suppliers of industrial equipment and machines such as ANDRITZ also can make precise product improvements based on usage data.
Is a machine able to differentiate between vegetarian and non-vegetarian dishes based on recipe images and descriptions?
Methods for recognizing and classifying natural objects, such as non-vegetarian recipe ingredients, form their own scientific discipline. However, companies also attach significant importance to automated object recognition using deep learning algorithms. So also, a supplier of innovative household and kitchen appliances, who wanted to test with 7LYTIX, which possibilities exist for the automated classification of recipes. For this, 7LYTIX demonstrated the classification of vegetarian and non-vegetarian dishes based on recipe images and recipe descriptions.
Our approach: As experts for the diverse use of Artificial Intelligence, 7LYTIX identified the best model for correct recipe classification among several artificial neural networks and unsupervised deep learning algorithms. The challenge in this project was the very small data base. Only 90 recipes, including recipe picture and description were available for the development, the validation and the testing of the classification model. To nevertheless achieve a high degree of model accuracy, the data set has been extended by additional image adaptations by rotating and mirroring existing recipe images.
Results: The best results were provided by the so-called unsupervised deep learning algorithm Doc2Vec, which achieved an accuracy of 80% - a good result regarding the small data base (see figure).
Benefits: Diversity of recipes is a crucial factor when it comes to deciding which smart food processor to buy and is therefore an important additional sales appeal. The more recipes are tuned to the device, the higher is the convenience that it brings and the more attractive it becomes for the potential buyer. Using the AI solutions of 7LYTIX, the well-known supplier of innovative household and kitchen appliances is now able to generate a large variety of different recipes for its smart kitchen appliances, which are completely automatically enriched with additional vital information.
Do you want to be able to do, what Amazon does?
Amazon, Zalando, youtube - these successful companies have one thing in common: they have an intelligent recommendation system. A software system that predicts how strong a customer's interest in a product is. With this knowledge, they can automatically generate personalized product recommendations for each customer which match their personal interests and needs. As a source of information for the prediction, the system uses product data and customer attributes, as well as the search and purchase history of the customer and similar customers.
Task: Austria`s leading online job platform is constantly aiming at making it easier for users to search for jobs or employees. For this purpose, a specific recommendation system should be developed and implemented which would display personal job recommendations to every job seeker.
Our approach: Through our sound expertise in the fields of recommender systems we developed a job recommendation system by using artificial intelligence. Therefor several methods such as Collaborative Filtering, Graph-Based Algorithms and the Natural Language Analysis (NLP) have been combined to achieve the highest possible degree of personalization.
Result: The recommendation system of 7LYTIX automatically determines the preferences of a job seeker such as position, area, industry, place of work or company size. Therefor no explicit user input or filter settings are required. Already after his first search query the user receives personalized job recommendations which get more and more accurate during the ongoing search.
Benefit: By offering its users a significantly extended service that saves effort, time and money, the leading online career portal karriere.at is further increasing its competitive edge. For applicants, the process of online job search is now much more convenient. But also advertising companies face the advantage that they are more likely to reach the right candidates faster with their job ads.
Advantages: Simplified job search and increased success of job ads.
Can traders achieve significantly higher returns using AI?
Today, Artificial Intelligence is everywhere -and it is has arrived in the trading world. Rather than being the product of humans, profitable trading strategies are formed by computers. The machine – based on a framework defined by its creators - observes and learns the way the market works independently and forms what it considers to be the best strategy to attain the highest profit. A professional UK-based commodity market maker has challenged us to outperform the earnings of his trading strategy with our deep, neural-network-based AI trading, taking controlled risk management into account.
Task: The daily strategy in trading options of commodities should be improved upon. For this purpose, a web application was developed, which issues daily trading recommendations and evaluates the fund's performance.
Available Data: 7LYTIX had approx. 2.5 years of price data for two related goods and options (scrap and rebar) available. The option period extends from 1 to 12 months.
Our approach: Neural networks were trained on yield maximization using deep learning techniques by establishing correlations between two related parameters (steel options on scrap and rebar), while taking controlled risk control into account. Therefore, the first 2 years of the dataset were used to evaluate the performance the last 5 months.
The result: Yield was increased by 13% over the test period of 5 months, assuming the same capital employed. The volatility of the strategy is a dynamic parameter that can be adjusted to the risk aversion of the client. This flexibility allows significantly higher yields - with only a slight change in the capital use.
Outlook: 7LYTIX is interested in partners from the finance community to develop together the broader use of Artificial Intelligence for fund management and trading. We see great opportunities beyond the use of time series in raw material options trading through the automated inclusion of globally available qualitative and quantitative information into the trading strategy.