Production Monitoring and Data Mining –No Strip Mining Allowed!

Data Mining

Any manufacturing or process plant has many sources of production data, and that data, if kept, very soon becomes voluminous. If a facility has 1000 sensors on the plant floor, and data is taken at once a second interval from all those sensors, that's 1000 data points per second, 60,000 data points per minute, and over 86 million data points per day. Most plants have many more than 1000 sensors that can provide data to a data historian.

The concept of data mining is similar to finding a needle in a haystack. In fact, noted manufacturing process consultant Eliyahu Goldratt titled his 1991 book about data mining "The Haystack Syndrome," because he had discovered that traditional ways of mining data were about as effective as trying to find that needle.

In essence, data mining is the process of sifting historical data to find data that supports a premise, or produces a pattern. The implication is that only favorable data is discovered, making the process somewhat dubious. There are a variety of ways data can be mined, including manually, or with increasingly sophisticated analysis software that includes artificial intelligence and neural network modeling tools.

Conceptually, a data miner has to gather tons of dirt, rocks, and ore from his data mine, in order to be able to sluice and sift through the "non-data" to find the useful information, those "actionable nuggets" that allow the miner to understand his process, and optimize it based on the information he has mined. Note the difference between "data" and "information."

Unfortunately, the tools for data mining are not robust enough, generally, to permit easy extraction of data from most of the huge haystacks of data plants produce. In the vast majority of plants, this data is unused, unusable, and finally discarded. It has been pointed out time and time again that data is useless unless it can be turned into information—actionable information AND if that actionable information is actually used to effect improvements. This is one of the reasons that it is extremely difficult to show a significant payback in cost accounting terms for data optimization and MES projects: there is no payback until the money is spent, and the plant management actually begins to use the tools. If the data doesn't get turned into information, and the management doesn't use the tools to sift the information, there is no payback, regardless of the possibilities.

Machine Monitoring and Asset Management

Asset management is one of the most important ways data mining is used in manufacturing and process plants. Quality studies have found that critical asset failures are extremely costly. In a well known failure in the early 1990s, the failure of a level measuring sensor caused a complete plant shutdown for over eight hours that cost over $175,000 per hour. Predictive maintenance on the level gauge was not performed.

There are many tools for asset management, including sensors, data loggers, and similar devices. The majority of these devices are simple data-taking tools, and the intelligence of the asset management system is resident in the control system, or in a data historian that piggybacks on the control system. These large, server-based systems can be management-intensive, and require sophisticated operators, and are add-ons to the control system. Connecting to the control system is still sometimes problematic even in the era of OPC and the new OPC Unified Architecture.

Sometimes the control system vendor adds proprietary tools that make using a standard version of OPC impossible. Other times, it is just difficult to interface the data historian with the asset management system. Sometimes, too, the avalanche of data that comes from the control system is just so much data overload, and cannot be processed by the data mining tools available in the control system, or the data historian. When a moderately sized system can produce millions of data points a week, it makes sense to look at some other ways to collect data in a format that is more useable.

Another significant issue for data mining is bandwidth. It is not enough to depend on the loop sensors, the "one sensor per loop" that conventional single loop control requires. In fact, for accurate asset management, the control sensor may not be the parameter of interest. To keep a recirculation loop running, the maintenance department is far more interested in the health of the motor and the pump and the control valve than they are in the flowmeter, which, of course, is the control parameter in the loop.

Many of these parameters have no existing sensors. These parameters, such as vibration, motor temperature, pump noise, and so forth, need to be wired in to the control system. Installing and wiring in these additional sensors is expensive. So, many companies are considering the use of wireless to get these values back to the control system and thence to the data historian and thence to the data mining software and the asset management software systems.

Wireless Data Collection and Analysis is Here to Stay

Many companies are considering wirelessly enabling existing smart transmitters to extract asset management data, as well as adding sensors to motors and other devices via other communications protocols such as Ethernet. Companies are even considering installing purely wireless sensors for control loops. There are issues of security, network robustness, and signal latency that must be addressed before this effort will be completely practical, but existing intelligent field devices such as Advantech's field PCs can easily be made to operate wirelessly over Ethernet using existing wireless standards like IEEE 802.11a, b, g, and soon n.

Network reliability is becoming a significant problem. System integrators report that they are often called to their customers' work sites with complaints that their system isn't working, and when the integrator investigates, the problem is in the enterprise wide network, and its connection to the control system. Since asset managers often reside on the enterprise side of the network, this can mean real problems. The use of COTS field PCs communicating over Ethernet, whether wired or wireless, can simplify the maintenance of plant floor networks considerably.

Mining Gold Ingots Instead of Dirt

The analogy that works here is that between open pit or “strip” mining and down the hole mining at the rock face. Traditional data mining has been like an strip mining operation where huge volumes of dirt are moved, sifted and discarded to find the gold. In an underground mining application, however, very little is touched except the ore itself.

So, too, with data mining for machinery health and performance. Instead of bringing huge numbers of wired and wireless sensors back to the control system and thence to the asset management system, some companies are considering a different approach. This new approach is based on the concept that data bandwidth is finite in a plant setting.

Therefore, they are moving the computational capability from the asset management system into the field. They are using the power of networked and distributed "field intelligent" devices to transform that raw ore of data into the gold ingots of information, but at the sensor or machine level rather than in the control system. This is made possible by the use of advanced industrial computers that are designed to be robust enough to be used on the plant floor and intelligent enough to be able to run rules generated by human process domain expertise.

These devices are generally PC-based industrial computers that are cost-effective and use the embedded computing versions of Windows XP and Windows CE, and are .NET enabled for the easiest transfer of data to existing plant control systems and data historians.

Here is a figure that describes both the concept of intelligent field-based monitoring of machinery health, and an implementation of the concept based on Advantech systems and products as an example.

As can be seen from the figure, sensor signals are brought directly to the industrial PC via a variety of inputs. They are translated via machine interface I/O and immediately processed in the embedded computing environment. Using the embedded application software which contains the rules-based engine for machinery monitoring, and the web server operating in a .NET framework providing the connectivity for the actionable information. So in this concept, data is converted to actionable information at the device and field level, and then sent to the asset management system. This system can be proactive, updating operators and supervisors in a timely manner with key events and trends, while not clogging up the data highways with millions of data points per day.

This type of system can be retrofitted for reasonable cost to any existing machine or motor train, and this permits plant-wide production monitoring and machinery health monitoring. Solutions are also available to work with existing production automation structures.

Distributed Intelligence Using Ethernet is Communications Method Agnostic

The concept of distributed intelligence, of which this is an excellent example, is based on the existence on the plant floor of robust communications and efficient networking solutions. By adopting a defacto plant standard, .NET, and by using plant standard Ethernet for communications, this distributed intelligence system is communications methodology agnostic. That is, it doesn't care if the signal is carried on wires, by fiber optics, or by radio. Ethernet is a tried and proven communications topology that is ubiquitous in both the plant and enterprise environments. And since there are standards for Ethernet communications, proprietary solutions are not required.

That means that the technology, concepts, tools, and methodology are all in place to do intelligent machine monitoring and production monitoring at the plant floor level, either wired or wirelessly, and do it with standard COTS systems and standardized communications topology.