FacilityConneX Data Scientist, Zlatko Vasilkoski Explains How Neural Networking Methodologies can be Applied to Early Warning Detections Today!
Big data has been around us for some time already. We currently live in the era of internet of thing (IoT), and most of the big data is, and will be related to IoT. Related to this, obvious question arises. What do we do with all this IoT big data and how do we make sense out of it? A data-driven insight can predicatively ensure a cutting-edge advantage over the competitors. But how do we achieve this in an industrial context and what tools do we use?
Here is our case study. We recently visited a customer, a producer of thin films, to address the issues they have with their unplanned production downtime due to variety of defects in their manufacturing process. This is a very typical problem that can be pictured by any producer of plastic, glass, fabric or pretty much any product where the manufacture lacks visibility of the sensor data stream from their entire production process, thus not being able to analyze all the downtime scenarios they experience.
Our approach was to organize and integrate the streaming sensor data from their entire production line into a single resource that can further be tracked and analyzed. The insight into their historical production line data revealed variety of patterns leading to downtime events. By analyzing these patterns several defect scenarios become apparent. That pointed the way to the use of predictive analytics that, by monitoring the real-time production line patterns will give the customer a confidence bounds for a downtime event.
This is just one of many illustrations of how predictive analytics can bridge the step from reactive to efficient proactive monitoring system control. But the ultimate goal is to build data-driven adaptive IoT apps that will be applicable over variety of industrial data sets.
Our Approach to Analyzing Data
The typical data analysis contains the following steps.
- Accessing different data types (time series, images, text, analog etc.) and staging the data – extracting data to historian database.
- Organizing and preprocessing data for data manipulation, such as time series-functionality, high performance merging and joining of data sets, domain-specific time offsets and join time series without losing data and intelligent data alignment and integrated handling of missing data.
- Data mining and analysis, which typically involves statistical methods and variety of machine learning techniques. Most of these techniques provide preventive component, while few of the machine learning techniques can also identify a predictive component.
Machine Learning in a Nutshell
Machine learning is a field of computer science that study pattern recognition. Machine learning algorithms and techniques give computers the ability to
- learn without explicitly being programmed for a task,
- get better with experience and
- make predictions based on what is been learned.
These are features that we typically associate with how living organisms learn, thus the name. As mentioned there are variety of machine learning methods. The choice of which particular method to use depends on the type of the problem (equipment, industrial process) we need to be analyzed.
Relating to our case study, the producer of thin films, the identification of the defect probability distribution for a particular defect type required statistical methods. But one of the key defect identification components was a visual identification of the defect shape and an association with the defect type. This required a need for a machine learning technique that can recognize a particular shape from the historical imaging database of the customer.
Certain machine learning techniques are appropriate for some types of problems while some are not. It would be great if there was a machine learning technique that can be applied to most of the problems encountered in the industrial internet. In fact, a technique that comes very close to this requirement is the neural networks.
Neural Networks in a Nutshell
The neural network (artificial neural networks) models are a family of machine learning models that capture the feature of learning. They are inspired and based on the operation of biological neural networks.
In a nutshell, the input to the neural network consists of a set of input training data and a set of corresponding desired responses. An error is composed from the difference between the desired response and the system output. This error information is fed back to the system to adjusts the system parameters in a systematic fashion.
The neural network solutions are extremely efficient in terms of development time and resources, and in many difficult problems artificial neural networks provide performance that is difficult to match with other technologies. Relating to our case study example, sieving through the customer’s large historical defect data, the visual classification of the defect shapes can be easily accomplished with our own FCX NuNet classifier. Our classifier was not designed specifically for classifying images but it did not take too much development time to be used for this purpose.
Predicting Time Series for the IOT
Most of the time the data we deal in IoT consists of time series and we are concerned with the estimates of lag dependence in these time series. This kind of analysis is typically done to create predictive analytics that projects the conclusions to the near future.
Neural networks that have been used already for several decades are suitable tool for the IoT needs. Especially this is true of the more recently developed class of neural networks that go under the name of deep learning, one of which is the Recurrent Neural Networks class or RNN for short. While in a traditional neural network all of the inputs (and outputs) are independent of each other, the RNN make use of sequential information in the inputs. As a result, the RNN can retain “memory” of previous events and utilize this memory to learn more complex temporal patterns. This makes RNN a great tool for analysis and predictive decisions in time series data.
A special variant of RNN is LSTM (long short term memory) networks. They try to mimic human long/short memory. In a sense if someone tells us a phone number we can remember it and dial it, but after some time we will forget it. The LSTM RNN’s achieve this by having repeating module in a standard RNN.
Why RNN’s are useful for IoT? Since they can unleash the cognitive deep learning capabilities when connected to a multimodal sensors. RNN and LSTM can be used for many IoT tasks such as for example coding of electrical activity to categorize activity peaks.
Neural networks are typically used in situations where the amount of data is extremely large and any existing pattern is hard to notice. An example of such data might be the energy consumption of a university campus with all of its facilities, and the complex dependencies on factors such as daily and seasonal external temperature and occupancy, among many others. A single neural network model can be trained on past data to find an optimal baseline pattern and notify on out of optimal anomalies in the system based on the streaming sensory data. In addition, the neural network model can provide forecasting which improves with the usage and opens opportunities of creating a predictive analytics. The list of industrial problems that a neural network can address is quite large and goes way beyond the facilities. Some of our customers include alternative energy producers and managing water resources for example. But it is important to notice that in each case an opportunity of savings can be identified and realized.
In conclusion, here at FacilityConneX, by utilizing the power of deep learning capabilities on the cloud platform, we’re working closely with our customers to demonstrate how these new capabilities can solve industrial problems and improve their performance. In particular, the use of neural network concepts in the IoT field gives a distinctive advantage to our predictive analytics across the industrial sector.
Zlatko Vasilkoski is a chief data scientist at FacilityConneX. Prior to that he has been developing the curriculum and working as a lecturer at the Boston University Metropolitan College Computer Science department since 2012. In addition, he has extensive college teaching experience at Tufts, MIT, Suffolk, BU and Bentley. He obtained his PhD. in physics from Tufts University working with David Weaver and Martin Karplus on computational implementation of the diffusion collision multi scale model of protein folding, to which the 2013 Nobel Prize for chemistry was awarded. He did his postdoctoral research work at MIT and Northeastern, in the area of computational physics and neuroscience. His work at the Department of Cognitive and Neural Systems, at Boston University was in the area of neural network’s learning laws, and his work at the Harvard Medical School was in the area of medical data analytics.