Vast numbers of sensors record data in machines and plants. When analysed correctly, these data can improve manufacturing processes and guarantee high-quality products. The Industrial Analytics business unit develops the models required to do so in close cooperation with customers, and customers can participate in the data and the model development themselves.
Various data sets, called features, can be drawn from machine and plant data. These data can be evaluated automatically using artificial intelligence (AI). This, for example, includes the temperature, pressure, power consumption and vibrations. Experience from prior projects shows that the machines and plants typically are already recording all important data. In most cases, no additional sensors are required. The actual challenge is recovering hidden information from the data and to recognise relevant correlations. This is where Industrial Analytics from Weidmüller comes into play.
Detecting and classifying anomalies
There are many causes that may disrupt the smooth operation of a plant. These include, for example, air bubbles in the cooling circuit resulting in lower cooling capacity or gear backlash causing imprecise movements. The Weidmüller data scientists develop models using artificial intelligence that recognise such deviations from normal behaviour, that is anomalies, in real-time data. The scientists use historical data as a reference which provide a typical pattern for the operation of a machine over a set period of time.
During the anomaly classification, recognised deviations are then placed into categories from Important to Unimportant and important anomalies are assigned to a cause for the error. Machine operators can use this information to react to problems faster and even recognise malfunctions which may have otherwise gone undetected. A faster diagnosis ultimately reduces downtimes, which results in lowering costs and an optimised production output.
Feature engineering recognises complex patterns
Weidmüller received the German Innovation Award 2018 in the “Excellence in Business to Business” category for the integrated approach of Industrial Analytics. Dr Markus Köster, head of Research and Development in the Industrial Analytics business unit (l.), and Tobias Gaukstern, head of the Industrial Analytics business unit (r.), accepted the award in Berlin.
Feature engineering is an important technology for developing reliable AI models. In this approach, measured values are considered in complex statistical correlations. For this purpose, for example, correlation coefficients are formed which represent interrelated changes of two or more features over the course of time. The data scientists use historic machine data to develop new features. The goal is to recognise deviating patterns even better and more reliably than would be the case simply using the raw data. One example: high-frequency signals, such as from vibration measurements or frequency converters can be split up into different frequency ranges with their corresponding components of the output signal based on mathematical methods. The model learns the signal components characteristic for the normal behaviour of a machine. These components are a better indicator for possible malfunctions than the original signal.
It’s up to all of us
Since the data sets must be interpreted and evaluated based on the concrete machine or process behaviour, feature engineering requires comprehensive application knowledge. The data scientists’ expertise, the mechanical engineer’s or machine operator’s application know-how as well as the knowledge already acquired are all equally important for finding answers that will result in a practical solution. Only an application expert can assess whether or not an anomaly actually represents a machine error. The expert helps the data specialists to construct the algorithms that correctly describe the normal operational status as well as possible deviations and anomalies.
Models based on AI are currently already in use for numerous applications such as packing machines, in filling technology and materials handling as well as for robotics. At Weidmüller, these models result in software tailored towards the individual user. The software constantly monitors and predicts the behaviour of the machine and applies the data as well as the results of the analysis to a visualisation. UI experts design the user interface individually so that every customer gets a solution matching their field of application.
The visualisation makes it easier to keep on top of the current status of the machine. For this purpose, individual time ranges can be viewed and tagged with information which should be included in future data assessment. In this example, the yellow highlighted areas show potential anomalies that the algorithm identified to the user. Users can also look at these areas to indicate whether or not this in fact is an anomaly. In this manner, the model continues to learn and can classify future statuses more precisely.
However, a new AI-based model is not initially able to depict all potential future mistakes and statuses of a plant, especially when they are not or only very rarely contained in the historic data. The Industrial Analytics modules are therefore designed in such a way that users can update, expand and refine their model themselves over time. The Weidmüller data scientists will of course provide support to customers, if required.
Using features for success
Feature engineering is the key to the success of an Analytics solution. Weidmüller combines the requisite application knowledge and technical expertise on the physical correlations with data science know-how. Thanks to the option to develop AI-based models independently, mechanical engineers and machine operators can significantly increase their model performance without revealing their domain knowledge.
Keith Atkinson
Automation Sales Engineer