Improving and Refining the Big Data Search

Coming out of stealth mode earlier this year, Maana has now announced all the enhancements it’s been making the past several months to its analytics platform; and is moving from proof of concept to operationalization.

Trying to make sense of the data that abounds throughout industry, a wide variety of companies and their solutions are coming to the fore. One of those is Maana, a small startup that is gaining notice for its analytics platform that helps to operationalize Big Data insights.

First speaking with Maana at this fall’s GE Minds + Machines in San Francisco, the discussion focused in part on its relationship with GE and other early investors, and how its capabilities had been successfully tested through some of the largest players in several different industries. Now the company is taking its analytics platform to the next level, announcing further enhancements to its software and moving beyond proof of concept to operationalization.

Started in 2013, Maana has spent its first couple years building its platform around real-world experiences—building the scaffolding for the solutions its customers will need, says Donald Thompson, founder and CTO of Maana. “We have been extremely fortunate,” he adds, “to evolve our platform around a wide variety of use cases.”

Maana execs liken their platform to Google, and the way consumers have become very comfortable with starting their searches there—whether planning a trip, making a purchase, or what have you. In the industrial world, however, the search itself has to be rethought, Thompson says. Maana makes that search concept relevant in an industrial setting, dealing with data that is not textual in nature, and comes primarily from structured and semi-structured sources, he adds.

Maana uses its Emergent Semantic Graph and unified index strategy to search across all data silos and data types. Its user-guided, machine-assisted approach makes it easy and intuitive for users to find and draw correlations between data from disparate sources in the context of the asset or process to be optimized. The platform uses Apache Spark, supporting advanced machine learning algorithms and distributed compute.

The past six months of development have culminated in a more powerful end-to-end platform for discovery and operationalization of Big Data insights, Thompson says. The latest enhancements include flexible knowledge modeling, advanced analytics, operationalization of data mining models, and an intuitive user interface to navigate descriptive statistical analysis.

“We’re focused on an end-to-end solution for addressing the needs of these massive corporations who have collected data across hundreds of different systems and are really confronted now with how to make sense of it,” Thompson says.

A recent example of the capabilities Maana has been able to deliver comes in the form of accounts receivable for a Fortune 50 company. The company had all kinds of data from enterprise resource planning (ERP) systems, customer invoicing management, servicing, call centers, etc., but the systems were not designed to interoperate, Thompson explains. By crawling and indexing the data from the various components together, they were able to begin statistical understanding of issues.

Developing a model based on a 30-day trial, the company was able to show a 70 percent improvement over the same period the previous year, Thompson says, and has now made analysis part of the workflow for collectors. “We’re the intelligent component of that workflow,” he says. “Our aim is not to replace accounts receivable or the ERP system. We plug into the existing line of business solutions to provide that layer of intelligence that they’re currently lacking.”

The statistical information can be very useful, Thompson says, because you can use it to get at typically low hanging fruit. “For instance, close to 40 percent of invoices were in dispute just because of erroneous information to begin with,” he says, explaining that other statistics require correlation, quantifying value according to the effectiveness of a piece of equipment, for example, depending on time of year, product line, servicing, etc.

Sometimes, the fruit is not so low-hanging, requiring more analysis. In one particular recent case geared toward predictive maintenance, Maana gathered environmental information and performance information for some very expensive and complex equipment, each device with upwards of 300 sensors and each sensor submitting value every 5 ms. “The only significant statistical correlation we could find,” Thompson says,” was that the longer the equipment ran, the more likely it was to fail. Obviously not very useful. We needed deeper analytics.”

What Maana was ultimately able to do, though, was to learn a model of how the different sensors behaved leading up to certain kinds of failures. “We developed a very powerful time series search capability,” Thompson says, “over enormous quantities of sensor data.”

They looked at the pattern the sensors displayed leading up to a broken shaft event, for example. They then could look at all the other instances of that pattern to see if the same event followed. “Where they agree, hey, there might be something to this,” Thompson notes. “Then we tag all the instances of those, then we’re looking at a semantic level.”

Once that happens, Maana can turn that into a predictive model. “That’s something we’re doing to put it into the field so it’s taking the raw data as the sensors are generating them,” Thompson says. “We can get an early warning as failure is about to occur.”

This is particularly useful in the oil and gas sector, which has very expensive pieces of equipment operating in the wells, along with very challenging conditions that vary quite a bit. “It’s very expensive when failures occur, given their location and the logistics of repairing them,” Thompson adds.

The advances that Maana has made over the past few months are several, revolving around knowledge modeling, advanced analytics, improved user interface, and operationalization. The flexible knowledge modeling includes the newly patented Emergent Semantic Graph, which creates a continuous knowledge structure of assets for iterative insight. Other improvements within this group include assisted modeling, ontology support, knowledge representation and powerful query language.

Advanced analytics enhancements include continuous insights, temporal co-occurrence, temporal clustering, auto-parsing, topic modeling, and similarity search. Temporal co-occurrence finds relationships between events that happen together in time. As an example, Thompson points to an example in healthcare, where a patient’s medical records might be loaded with information about tests administered, drugs prescribed, etc., but with no connections between these events in the documents.

“It makes finding correlations very difficult,” he says. “But as a physician, you want to very quickly see through their medical history that relates to the problem you’re seeing them for.”

Maana is working with one of its customers on ways to mine that data and present information in ways that are more useful for clinicians. The platform is able to make more relevant correlations by mining a broader range of data, Thompson says. Some correlations made on a smaller scale could be spurious, he says, “but when you look at millions of records, the ones that are truly associative are very obvious. That builds a model. Given any one of those things, you can very easily create a lens on that record and quickly filter out the things that aren’t relevant.”

The similarity search that’s part of the enhancements is not as simple as it might sound. If you’re showing a coffee cup, for example, and told to look around the room to find similar objects, what do you find? Other coffee cups? Anything that can hold fluid? Anything the same color? So the question becomes: Similar in what way?

“In oil and gas, what does it mean to find a similar well? It depends on what you’re trying to do,” Thompson says. You can look at the cost of drilling vs. production rate; or hydraulic fracturing vs. deep sea exploration; or similar health or safety issues. “It can be a somewhat arbitrary thing.”

If an exploration company is getting ready to drill a new well, it can use similarity search to help with risk planning. “You first identify a set of offsets you’re going to use to plan for the risk,” Thompson explains. “Human experts start with one well, grab another well that they know to be similar, and train the system to go find other wells. We’re doing the heavy lifting to find similarities, and the human expert can judge and help.”

The user interface has also been made more intuitive to interactively navigate descriptive statistical analysis. Maana has introduced notebooks, advanced filters, and connected search to support data modeling and analysis from the interface. “Since the beginning of year, we’ve done a full redesign of the user experience,” Thompson says. “The system is far more intuitive, flows better, and is oriented around various intents and modes that the user goes through.”

Whereas Maana’s work so far has largely been proving out the platform, the company is entering the next phase, with customers “entrusting day-to-day operation to Maana’s engine,” Thompson says. “All of the pilots concluding this year have been approved for operationalization.”


More in IIoT