Historically, the primary role of data in the industrial setting has been fault detection and diagnosis (FDD). Today, companies are increasingly looking to their data sets as assets to influence their revenue and profits. With the rise of big data storage and processing, combined with the new capabilities of machine learning for prediction and recommendation, these data sets may be converted from inactive, latent assets to critical-path components of an overall production ecosystem.
Presently the term data exhaust is used in many contexts and is perhaps a bit of a cliché; it’s worth taking a moment to define this term. Data exhaust is a by-product of industry: measurements taken and recorded without the requirement that they be used. Initially such measurements were taken for the purposes of FDD. But today, as the communication and computing capabilities of industrial systems advance, these data assets may now be leveraged to harness untapped potential.
When we add new, augmentable data assets—collected through new sensors and Internet of things (IoT) devices—the combined data set holds an expansive new role in industry. In this report, we explore a few real-world examples of how this is done today and the opportunities for the future.
In their 2015 report, “What Is the Internet of Things?”,1 Mike Loukides and Jon Bruner describe the notion of frictionless manufacturing—how the process of bringing new products to market has radically changed from a paper- and negotiation-heavy process across multiple vendors to a seamless design-prototype-produce model enabled via software and the Internet.
There is an additional potential unlocked in this new model: autonomous control. In a functioning system, data, combined with machine learning, can lead to a conversion of once-human systems into real-time automation, which can identify and act on opportunities for efficiency and improvement.
The potential of this autonomous control through leveraged and augmented data is real. In the paper “Winning the Industrial Internet of Things,” Accenture estimates the global incremental effect of this shift at $7.1 trillion by 2030 in the United States alone.
The consensus in the applied data science community is that two benefits will surface through the move to data-driven command and control.
First, it is generally understood that in exploring data questions, invariably new, more interesting questions arise. In practice, this means as we seek to automate those systems around us, the roles humans play in the system become more interesting, cognitive, and creative. This move plays to the strengths of human nature in generating causal ties and hypotheses beyond the reach of machines. In short, jobs should become more interesting and less repetitive as a result of the change.
Second, a balance can be achieved through human + machine systems, where the role of humans is consistently upgraded, though never marginalized. The World Economic Forum cites the “collaboration between humans and machines” as a major driver for future economic growth globally.2
Consider the difference between IoT and historical data logging, and what we can gain from the IoT to generate value from data.
No doubt the current euphoria over “connected devices” and “smart appliances” generates collective eye-rolls among computational specialists in highly mechanized industries, where the idea of data-for-value-creation is as old as industries themselves. Since the advent of mechanization, dials and read-outs have performed the important role of informing the operator of the conditions of the machine. They also suggested, if implicitly, recommendations and predictions for performance improvement. With the advent of the Internet, industries were quick to capitalize on the potential to collect, transmit, and store data from product networks into centralized databases for the purposes of analysis and future product design.
A well-known system of this kind of feedback loop is the error-reporting infrastructure in Microsoft Windows, dating back to the first versions of the product. By collecting and analyzing logs from machines globally, Microsoft and its partners could quickly identify and act on issues in the field. Further, through the centralization of such data, the work of finding and fixing software and hardware bugs could be prioritized based on the “top” issues, meaning the work of engineers could be optimized in near time to have the greatest upside impact.
This example—and parallels exist in every industry from automotive to industrial HVAC—achieves only a part of the promise of IoT. The “log” process lacked the ability to proactively identify efficiency and opportunity through prediction and recommendation. For one, it relies on logs from networked machines, based on the logging capabilities of the machine itself. What if the machine is so inflicted that it cannot generate data about itself? Also, any action taken will only address “illness” within the system when a bug or failure occurs. This process is arguably the most critical and therefore presents the greatest opportunity in terms of short-term ROI. However, by definition it will be reactive to systems as built, as opposed to pro-actively generating new value.
Such systems only report history, and do nothing to predict future events or make recommendations based on the events. Logs provide directional trends, but log data will only reflect what was deemed worthy of collection at time of design, which may not provide the holistic understanding necessary for machine learning.
The IoT now enables us to augment FDD sources with sensors. Some might view IoT sensors as new packaging for an old concept. However, the capabilities of Internet connected devices (and equipment to augment these processes) can move us from reactivity to proactivity, from remediation to automation.
The key to such a shift is integration. Successful applications of IoT integrate with existing systems so as to revolutionize processes, while they “evolutionize” equipment upon which they reside. When we combine newly generated sensor data with historical logs, a new level of visibility is achieved. By using sensors to generate a response variable (the thing we wish to impact), we can design machine-learning systems that leverage sparse, low-quality data to achieve a high-quality outcome: performance improvement and value capture.
Let’s look at three specific examples. In each example, we’ll explore the current use of existing data, the new data enabled by IoT and the insights provided, and finally, the business advantage achieved.
In all three of the cases, The Data Guild served as a product co-development consultancy. We worked with companies assess existing data assets, build a strategy to leverage these assets, and deploy an ROI-generating solution to market. All of the examples leverage machine learning to produce a response based on near-time information.
Optimum Energy Corporation in Seattle, Washington, is revolutionizing energy efficiency in buildings by converting HVAC systems, traditionally based on set point and human control, to real-time machine learning and optimization systems. They do so not by a “rip and replace” model, but rather by attaching themselves to the as-built environment. Specifically, they place a Java Application Control Engine (JACE) alongside the Building Automation System (BAS) to take-over day-to-day control of the HVAC system (of which the BAS is one element) through set point optimization.
The outcome of moving from a set point (static) model to a learned model is revolutionary: massive savings in energy costs, while the approach simultaneously augments the existing system.
This type of an approach is imperative in most cases within the existing industrial world where the capital expenditure required for a wholesale replacement cannot achieve value in a reasonable payback period. However, we need not re-engineer an entire system in order achieve the benefit. New value can be created from old data. The core of the Optimum Energy platform is based on the Hartman loop, an existing, well-regarded approach to dynamic set point optimization in the world of HVAC.
However, there are limitations to relying on existing data sets available through the BAS to derive optimized set points and equipment sequencing. We observed that efficiencies could be furthered by augmenting the data set with a new, primary exogenous data source from a custom-built IoT device that went beyond the JACE system. In fact, such a sensor could be completely detached from the system it would ultimately control.
For instance, we found that in some cases, centrifugal chilled water systems run more efficiently when close to their surge point (the point at which the machine nears a backflow state, thus entering harmonic vibration). However, since surges were at best inconsistently measured by existing BAS systems, and often not at all, it was not a reliable data point upon which to build machine learning. To ameliorate this shortcoming, we developed a vibration sensor using GSM (cell network) backhaul, as opposed to the Ethernet system upon which the BAS resided.
Why was this an important design decision? Often in industrial settings, network security is paramount over all other design decisions. To gain entry into such a network requires considerable time and cost, and approval from many levels of authority. By contrast, a sensor that records a variable of interest using a backhaul network of its own simply requires the approval of its physical presence, not its network presence.
The surge point variable in itself was relatively meaningless. It wasn’t much good to know that a chiller was surging after it had surged, after any chance to remediate the situation had passed. However, when combined with other data points across the system, particularly as a response variable (the variable we wish to predict), the augmented variable opened up new value for existing data points to be used in a more profitable fashion.
In this example, operators were often fearful of operating equipment close to surge points, giving themselves a healthy buffer in daily operations sequencing. However, by using this data point as a response variable, then using equivalently time-stamped BAS data points as predictor variables, they might be able to model each chiller, update the models daily, and in turn achieve greater overall efficiency—and massive savings.
Through the extension of these systems with new sensing hardware, backhaul data collection, and adjacent control system, Optimum Energy was able to capture significant value with minimal disruptive cost.
This example comes from the heavy industrial environment, but it’s useful as an allegory for other industries, and even human systems. Where in your organization are you avoiding risk at the cost of efficiency or improved performance? What new data points can be created to provide a missing link between your data exhaust and business gain?
In a recent project, The Data Guild assisted a major Bay Area tech company to identify and optimize system performance bottlenecks within a multi-tiered global Software as a Service (SaaS) product—a real-time, on-demand video service.
Ostensibly, the challenge of working in software/server systems alone seemed less daunting than affecting massive industrial systems in the example above. However, a paradox exists in software systems: infinite flexibility and scalability introduces the possibility of infinite complexity.
In industrial mechanics, the opportunity to alter a system is limited by the capabilities of the system. For example, what controls are enabled? What physical changes can humans achieve in such a system? In software, these limitations vanish, enabling even novice coders to create complexity that makes fault detection and diagnosis challenging. A machine breakdown is easily noticed, but can the same be said for a data center?
In this project, the service logs between each tier of the service were analyzed to determine bottlenecks: which steps in the system were contributing most to limitations in performance? However, the location of observed bottleneck could not reliably be considered its source. To improve the situation, a systematic approach was developed to understand the relationship between different components of hardware and software. Through this covariance model, we could then make recommendations for system improvements.
In this case, the industrial exhaust was the server logs. These logs represent billions of transactional records of the input/output of the servers, detailing the work (CPU) and communication (bandwidth) in the context of a much larger system. In isolation, these logs held limited value other than local tuning, but through the integration of these sources, we were able to identify patterns, hierarchies, and relationships. Though a derivative of existing log data, the covariance data became the new source from which we could derive dependencies and ultimately, recommendations.
It’s worth pointing out that the designers of the logging system likely had little in mind beyond fault detection and diagnosis or software performance tuning. However, in a broader, complex networked environment, each transaction, when taken in symphony with logs, began to illuminate the nature of the larger system. This in turn gave us a path to follow in order to boost software performance in some applications and improve hardware performance in places requiring greater scale.
Another lesson from this project is the importance of data visualization. Humans are trained by evolution to recognize patterns in nature and respond to these designs for the purpose of optimization. In our history, those that understood the patterns of the seasons could optimize crop yield, for example; and those who observed movements of wildlife were more successful hunters and thus more effective survivors. In the current age, pattern recognition improves outcomes in nearly every problem we choose to tackle. Unfortunately, unlike weather or herd movements, signals in data can be difficult to pin down.
In data science, we split activities broadly into supervised and unsupervised learning. In the former, we “know what we’re looking for” and hope to achieve models to approximate that outcome. In the latter, we’re not sure what we will learn as we set off on the journey, but hope to identify new patterns that can help us understand our context in the form of meaningful classes or clusters. The utility of data visualization in this context cannot be understated. Spreading our data across a graph visualization in order to show relationships by providing multi-dimensional rendering or labeling with intelligent color-coding to isolate patterns can unlock value that otherwise may be non-intuitive.
In this project, we went step further: we developed a data sonification system to enable those in the network operations center (NOC) to hear their data center in the same way a factory floor manager might be able to hear her machines and understand when they are on or off, functioning smoothly, or having issues. Audio examples and more detail on this project can be heard in this podcast [insert link to Bruner/Turner podcast].
We did not use IoT to generate new input, but rather used a form of IoT to append the system to achieve new output. Instead of measuring an ambient characteristic (as was done in the prior example), here we created an ambient characteristic through sound to consume information. IoT need not be a read-only variable.
The profit creation through this type of approach is two-fold. While identifying opportunities for throughput optimization in a data center environment can immediately reduce capital expenditure for additional capacity (hardware), the approach is also extensible. It can save organizations from being overly dependent on one platform versus another through a deeper infrastructure investment.
Much has been written on the fallacy of the first mover advantage—in other words, the stricture that a business must be first to market to be the best in market. However, other priorities such as efficiency, cost management, finance structure, channel, and branding and marketing often become the critical factors of market dominance.
In data, there is a similar truth—it is not he who has the best data that wins, but rather he who is most clever in its application.
Proteus Digital Health, a San Mateo company, received FDA approval for an ingestible sensor that captured various critical attributes as it passed through a patient’s body. The sensor in turn passed information via Bluetooth low-energy (BLE) to a body patch, smartphone, or other device, which then uploaded to a centralized data collection service. The device put clinical-grade data in consumers’ hands for the first time, unlocking a wealth of new data usage scenarios.
While the notion of self-quantification is not new, this example highlights the importance of data centralization, in which the measurements from a larger group can improve the outcome for each individual. Understanding individualized measurements could assist various parties (doctors, nurses, families, or trainers), but the centralization of data enabled recommendations based on best practices identified within the broader population from whom the data was collected.
In receiving FDA approval for their device, Proteus had access to a wealth of new data points from a rich source: the human body. By applying the data to areas beyond their traditional markets (for example, the tracking of drug effects, both in trials and once drugs were released commercially), they were able to enter new markets and solve problems not originally conceived by the inventors of the data sources.
In this project, we were able to rethink the notion of addressable market for medical technology. Traditionally, medical devices are focused on the sick patients within the context of a medical facility. With lower costs and global connectivity, we can alter the market to focus on community health versus sickness. In this context, all patients are participants within a health community graph, as opposed to leaf nodes within the healthcare system.
Up to this point, we’ve looked at IoT from the evaluative state: what are the opportunities for implementation of IoT within a data-rich industry? As with any endeavor, though, the devil lies in the details: tactical considerations, when considered up front, can greatly improve the ROI of implementation.
Here are key factors to consider when doing any data analysis.
Time stamp matching is critical when blending systems, since it is the basis of comparability. Time stamps in data can be nasty quality issues, as they fall into the unknown space of the Rumsfeld Quadrant.3 In addition to basic questions such as “UMT or Local,” and “Daylight or Standard,” there are local issues that may affect reliability.
Even systems synchronized by “time servers” can have irregularities due to discrepancies in sampling, software issues in logging, or variance in granularity. Considering these issues early in design is critical to reliability in the performance of a system.
In addition to latency in local logging, the difference in time between data collection and its availability may render a data point moot for a given application. For example, if we cannot process the space between the front bumper of one car and the back of another in real-time, how well will our autonomous car perform?
Many opportunities for value creation in IoT are based on the invalid assumption that similarly designed equipment will perform with similar efficiency. When we learn about these differences and their trends, we can create value by understanding the uniqueness of industrial components.
Similarly, we should not assume that data points with same name and similar source mean the same thing. The key here is comparability: unless we truly understand the difference between the data we are comparing, we cannot stand behind the conclusions of such a comparison.
Those who write software appreciate the value of reproducibility: generally speaking, without changes in input, the output of code is the same, indefinitely. The same cannot be said of hardware. Over time, all hardware drifts and requires calibration; eventually, all systems die. This creates a challenge for those of us that seek to optimize such systems, as we must not only understand the meaning behind measurements today but also how they are expected to change tomorrow, and when the machine, and therefore our optimization, will no longer function.
Software instrumentation is remarkably reliable when compared to industrial hardware telemetry. However, in order to ameliorate fragility and harden learning systems in IoT, we cannot assume that data will be available from any source at any given point. Data can disappear for a myriad reasons: communication platforms/outages, expired security and registration, human error, or even solar flare-ups.4
In the design process, consider redundancy/duplicity and cross-validation to help address sparsity. In addition, the notion of “failing gracefully” should be built into IoT system design. In other words, ask the question: if I lose half my data, can I produce a result at least half as well?
Barring issues of quality, the question remains of how much data is needed to generate the desired result. The question is critical, since it can be a direct contributor to cost on many fronts. IoT data volume is directly correlated to storage, computation, and bandwidth costs.
No list would be complete without the consideration of data quality. In developing systems that mix old (log) and new (sensor) data, we’ve encountered data quality issues in all cases.
The questions outlined in Table 1-1 are necessary in the IoT design phase in order to ensure a right-sized solution.
| Question | Decision | Examples |
|---|---|---|
| “What data do I need in real-time versus near-time?” | Communication type/speed | Fiber optic, Ethernet, or cellular |
| “What data must be recalled instantly, versus with some latency?” | Storage type | Memory, disk, or tape backup |
| “What level of granularity is needed?” | Aggregation tiers | Raw, summarized by entity, or summarized by date |
| “How fast do I need a result?” | CPU type/location | Device, on-premise, or remote/cloud processing |
| What seasonality might be at play?” | Collection design | Duration of collection/sampling rates |
Security is a unique challenge in that, unlike other areas, is a problem not of building a capability/strength, but rather closing up a vulnerability/risk.
IoT creates a potential new vector of attack for systems when not properly integrated on a policy and technical level. It requires a right-sized approach. Design questions must be addressed at the communication layer. The type, sensitivity, and potential of risk implied by the data transmitted must be considered.
In designing IoT systems, risk should not be considered in isolation of the application, but rather in the potential when sources are coalesced, triangulated with other sources, or evaluated over time. For example, could proprietary production data be gleaned by evaluating power consumption with a particular product manufacturing line over time? A thorough threat modeling phase and proper remediation is critical to limiting the downside risk of deploying new sensing platforms.
Perhaps most importantly, the issue of cost cannot be understated. The implicit promise of the IoT is that it will radically reduce costs within existing systems and services. However, what is to be said of the integration costs of such learning systems themselves? To address concerns around capital expenditure for otherwise profitable projects, running pilot projects is often a way to prove, at the ground level, the payoff of implementation.
Indeed, modern IoT platforms are well suited to a gradual adoption model, where operators can implement pilots based on low-cost hardware and communication protocols that require very little up-front investment. Combined with the up- and down-scalability of cloud computing, achieving measurable, prototypical ROI in pilot engagements can help justify and target future deployment and spend.
IoT has the enormous promise of radically altering historically analog business through the integration of existing data, new sensor data, and real- and near-time machine learning. In some sense, IoT encapsulates the best of software and hardware. By combining the speed and scalability of software with the tangible outcomes of hardware, IoT and sensors will change the world and the role of humans in the world in ways not yet imagined.
However, we’ve come a long way since the advent of those systems that enable IoT scenarios. Here are three commonalities we’ve observed, key differentiators that influence success in implementation:
While vendors are quick to recommend hardware-heavy deployments, this can complicate organizations hoping to incrementally grow into the IoT sphere. More likely than not, there are data assets within your organization today that are well suited to building machine-learning systems. Though traditionally not considered IoT, data may be hiding as historical logs, unstructured data store, or even “serial out” data dumps from your hardware systems.
If you’re looking for a place to start, consider the data you’re currently using for FDD and work your way forward. Such an approach ensures that each additional layer of cost in IoT adoption achieves incremental improvement in ROI. Through iterative design cycles, organizations can achieve continuous value creation and agility. By enabling flexibility and extensibility in design phases, organizations future-proof their investments while capturing value in the short term.
Great ideas can come from unexpected sources. Similarly, critical data points may come from outside your organization. Consider the industrial HVAC example above. Imagine the difference in outcomes between the IoT/sensor system that considers the local weather forecast in load prediction for heating and cooling, versus the system that does not.
The advent of the Internet in the 1990s inspired the imagination of every industry. Information that had been locked on computing systems could be shared globally and standardized for interoperability, unleashing a wave of innovation.
The Internet of Things and the sensor technologies that support it hold an even broader promise in enabling the as-built world to communicate with itself in order to achieve previously impossible levels of best practices, efficiency, and opportunity identification. This revolution will be achieved through the evolution of as-built systems with existing histories of data. Data can be exploited in ways never imagined by its inventors to expand from local fault detection to global prediction, recommendation, and optimization through machine learning.
For more information on IoT and sensors, including audio samples to illustrate the examples discussed above, listen to a recent podcast, “The Sound of Sensors,” with O’Reilly’s Jon Bruner and The Data Guild’s Cameron Turner here.
1 http://www.oreilly.com/iot/free/what-is-the-internet-of-things.html
2 http://www3.weforum.org/docs/WEFUSA_IndustrialInternet_Report2015.pdf
3 http://micrex.com/rumsfeld-matrix-part-1/
4 http://www.dailymail.co.uk/sciencetech/article-2656010/Could-Friday-13th-phones-GPS-devices-knocked-sun-Impact-massive-solar-flare-expected-hit-Earth-tomorrow-disrupt-communications-satellites.html
