The Rise of Deep Intelligent Operation

As a practitioner in the AI industry, I have not spoken with a single client recently who does not have 'Intelligent Operation' in their vision and roadmap. Welcome to 2022! .. and this applies across industry domains.

• In the context of supply chain, it may mean optimised routes, optimised inventory, auto replenishment of stocks, just-in-time supply signals for dynamic fulfillment, air quality tracking for sustainability, and so on.
• In the health sector, it may mean 'Connected Patients' - where medical sensors can monitor and track patient conditions and trigger alert.
• In retail, banking, or hospitality, it may mean 'Connected Customers' - where consumers, customers and citizens expect robust and relevant experiences when they interact with organizations.

There is one critical success factor that underpins this; successful deployment and implementation of industrial IoT (IIoT) systems that connects the physical world to the digital world on a real time basis.

Intelligent Connected Operation with IIOT

Intelligent Connected Operation is a future looking vision that aspires to apply Industrial IoT (IIoT) sensors at scale to generate and analyse real time streaming data from various parts of the operation and enables automated (or assisted) decision making, or even in its advanced state, it may enable "things", that is machines, to make cognitive decisions and initiate action.

As an example, a control system in a plant detects an anomaly pattern coming up within next 10 mins, and therefore initiates an action - may be a SOS message to the operator, or even closing a valve to control the flow in a part of the plant. Or, a medical sensor indicates low blood sugar being dangerously low in a patient and calls an ambulance or presses an alarm bell for the carer. However, the IIoT sensors by themselves wont be able to achieve any of these unless and until the data in motion that they generate is captured and embedded in the backbone of the operation on near real time basis. And that, is not trivial.

The "Data Deluge" problem with IIoT

A large IIoT setup generates enormous amount of data. The implicit nature of IoT data is heterogeneous time series generated by a multitude of sensors, and therefore the collection, transformation, correlation analysis, and cross stream pattern discovery from this data requires advanced methods and algorithms.

As an example, a complex IIoT setup representing connected operation in a plant may have thousands of sensors; these sensors generate time series data of different characteristics, frequency, and behaviours. A key challenge is to collect, transform, store, and model that data in real time to enable the "Things" to make accurate decisions.?The “Things”, in particular Intelligent Systems, learn from historical data to anticipate future trends.?Only single numeric values from a sensor at a point in time won’t suffice; the inherent time series and the temporal patterns therein, as well as the interactions and correlations between multiple sensors are equally important to train these intelligent systems so that they can then identify patterns on real time basis and act accordingly.?Existing time series analysis methods struggle to handle this due to following reasons:

1. Firstly, classic time series analysis methods such as AR, ARIMA, SARIMA etc generally make early assumptions about the time series characteristics (a detailed take on those methods a can be found here), and also operate on posteriori basis on a large volume of a snapshot historical data, which will not work in these online real time prediction scenario where data is continuously streaming in and decisions need to be made on-the-fly.?
2. Secondly, machine learning techniques have been applied to solve this but that has issues too. E.g., time series data have temporal patterns and every point in time has some correlation with its past values, and therefore simply breaking a time series window into input variables will not generate desirable outcome. Standard neural networks have been applied to this, but they also ignore the same temporal dependency producing suboptimal results. In case of large multi-sensor scenarios, these methods do not work well in a real time setup.
3. Thirdly, the multivariate time series and their interactions need to be modelled using machine learning techniques. In case of complex system with a large number of sensors, multivariate statistical analysis such as principle/independent component analysis (PCA/ICA) are generally applied to discover the correlation between the large number of sensor data streams. However, PCA and ICA can only derive linear correlations while many physical systems have nonlinear relations among their components. Nonlinear multi-variate algorithms have scalability problem. Therefore, suitable methodologies and algorithms are needed to address this.

This is where Deep Learning comes in.

So, what is Deep Learning?

Simply put, deep learning is a subset of machine learning in which large artificial neural networks with deep layers of neurons imitate the inner workings of the human brain to extract patterns and insights out of data.

Over the the past years, deep learning has been largely applied in the domain of computer vision and speech recognition; driverless cars, voice control in consumer devices like phones, tablets, TVs, and hands-free speakers - are some example of this phenomenon. Recently, deep learning is being applied in time series research domain for IoT, and is a promising avenue wherein a multi-sensor IIoT system generating multiple time series can pass through a deep learning model that reduces the dimensionality of the whole datasets and derives an abstracted representation of the data that can then be taken up for any further ML analysis such as classification, regression, anomaly detection etc. Over the past years, significant progress has been made in deep learning application in IoT time series. With the advent of deep networks such as Convolutional LSTM (Long Short Term Memory) that can memorise temporal patterns of time series, many of the challenges mentioned above can now be addressed.

There are still areas that need resolution; e.g., deciding an optimal sliding window size for a multivariate time series is a challenge that needs attention. I intend to write more on these topics in future. But from practitioner's perspective, applying deep learning in IIoT is going to be a game changer in achieving Connected Intelligent Operations goals.

Recently, I applied a deep learning model (Convolutional LSTM autoencoder) to smart meter reading data to derive anomaly detection by looking at the reconstruction loss and it was good fun.

Have you tried deep learning in real life scenario? Please share your experience.