Detecting Anomalies with Artificial Intelligence

Welcome to AI.DA! Digital Operator detecting anomalies

DIG:ITA created a MOSAICO digital workforce equipped with a set of digital Competences, in the sense of the set of knowledge, skills, attitudes, abilities and strategies, that are required when we are using Information and Communications Technology and digital media

The future of work is changing the face of industry as you know it: it’s shaped by a profound transformation, driven by the meshing of the digital and the physical world, the emergence of new design and production techniques, and a seismic shift in the role that human beings play in the production process.

AI.DA Digital Operator

DIG:ITA created a MOSAICO digital workforce equipped with a set of digital Competences, in the sense of the set of knowledge, skills, attitudes, abilities and strategies, that are required when we are using Information and Communications Technology and digital media

AI.DA Digital Operator provides anomaly detection monitoring outliners in alarms and events related industrial production process.

“Through DIG:ITA you can hire #iDO & AI.DA Industrial Digital Operators to take on high-volume and repeatable tasks, so human employees can concentrate on higher-value jobs. Once you bring #iDO & AI.DA into your organisation, you’re on a path to build a hybrid workforce, where digital and human workers collaborate for greater business value.”

Anomalies Detection background

Anomaly detection (aka outlier analysis) is a step in data mining that identifies data points, events, and/or observations that deviate from a dataset’s normal behaviour. Anomalous data can indicate critical incidents, such as a technical glitch, or potential opportunities, for instance a change in consumer behaviour.

AI.DA. Project- Anomalies Detection in industrial production process

The AI.DA approach is a novelty in the industry for its approach and its algorithms.

First of all, anomaly detection is frequently applied to detect outliers on a single process variable, knowing if a parameter value is in the expected range or not. This is obviously useful to understand if a sensor is working correctly, but tell us little of the real production process. What we really want to know if our process is working in a normal behaviour and, if not, where is the problem.

AI.DA achieves this purpose by analysing all the events in the production plant through its algorithms which continuously learn and evaluate how the process is working. Furthermore, AI.DA is able to identify and instantaneously signal the presence of anomalies: this is a real value added since it allows to resolve immediately the process problems and to avoid further more critical and more expensive failures.

Duration and Frequency anomalies

In order to understand if an event is anomalous or not, two different parameters are estimated and analysed by the algorithm: the duration and the frequency of the occurrence. Since these two metrics describe the current behaviour of that event, they are compared with its history in order to evaluate if they are anomalous or not. If an event occurs too often, or too little, if it has been activated for a time too short or too long it’s all addressed by these features of AI.DA. These metrics are estimated and updated, learning from changes in the plant behaviour during time.

Group anomalies

The real core of AI.DA is in its Neural Networks and training methods to detect group anomalies. Specifically, group anomalies are behaviours that are not expected in more than one event across the production plant. For example imagine if suddenly one motor stops, this will led to a multitude of alarms and events. So far is all perfectly normal, but what if one event does not show up? What if others, unexpectedly, instead do? With group anomalies AI.DA monitors all the events in the production process and estimates correlations and causation to detect if some signal differs from the expected behaviour. We also talked about multiple Neural Networks because this features leverages the state of art of Reinforcement Learning using two models to describe and learn from the production process.

Leveraging the multiple information coming from duration, frequency and group anomalies, AI.DA represent the state of art of intelligent anomaly detection systems, completely autonomous in monitoring a production process and communicating what could be anomalous.

AI.DA notifications

In addition, when anomalies are detected in the industrial production process, AI.DA is also able to instantaneously signal them through a specific notification which directly arrives to the plant manager and the operators. For example, below is reported the notification of an anomaly concerning the Limestone Loading Valve. Specifically, AI.DA reports the anomaly also by showing the trend of that alarm during the last month. In addition, the notification includes also some insights which AI.DA gives to the plant operators in order to deal with the anomaly detected and to avoid further future occurrences.

AI.DA data exploration over VOYAGER

DIG:ITA data scientist could also benefit from the predictions of AI.DA thanks to the VOYAGER Data Exploration tool, digging deeper to understand the causes of the anomalies and to intervene. Having access to all the raw data and being deeply integrated, it’s easy for the data scientists to evaluate a situation and immediately communicate an anomalous event that could lead to unwanted consequences.