Print headline: Smarter Diagnostics
The loudest industry buzz has been about using big data and artificial intelligence (AI) for predictive maintenance, or turning unscheduled events into scheduled ones by forecasting likely failures. But surprise events still occur, and AI can also help troubleshoot them faster and more effectively.
Any tool that enables predictive maintenance also helps troubleshooting, as it often points to causes of likely failures. But to provide maximum diagnostic benefits, some AI techniques can also be used in different ways.
For example, natural language processing can translate mechanics’ plain-spoken inquiries into text that helps find answers. Or it can translate past maintenance reports into very helpful data, or digest repair manuals to quickly locate relevant advice. And machine learning can sift through past troubleshooting incidents to spot most-likely causes and the best steps to take.
Any device that enables line mechanics to collaborate with experts or to exploit documented expertise more easily, is a huge boon on the troubleshooting line. And sophisticated techniques for highlighting chronic problems can help focus the best resources, human and machine, on their resolution.
Major airframe OEMs have been active in providing troubleshooting help. Increasingly, startup and high-tech companies are becoming involved as well.
Boeing is now testing augmented reality on smart glasses to show mechanics hands-free, interactive 3D wiring diagrams, rather than forcing them to view two-dimensional, 20-ft.-long drawings and retain that information while doing repairs. Ken Sain, Boeing vice president of digital aviation and analytics, says tests show a 90% improvement in first-time fixes and a 30% reduction in repair time.
Of course, Boeing’s existing line-maintenance applications have long given mechanics immediate access to manuals, part numbers and other critical troubleshooting information. Toolbox, used by 200 customers, offers intelligent documents and visual navigation to aid diagnostics, record structural repairs and manage part and task cards. Toolbox combines manuals and airline-created content so users can search for information about parts and fault histories.
Boeing’s Maintenance Turn Time brings assistance to mobile devices on the flight line, enabling line techs to collaborate with engineers—for instance, by uploading photos of damaged parts.
But some innovators want to go even further. Five-year-old SparkCognition has been working with Boeing and the U.S. Air Force on both predictive maintenance and troubleshooting. Marketing manager Carlos Pazos says it uses machine learning to automatically build prescriptive models much faster than data scientists could. The company uses natural language processing to absorb the content of both repair manuals and the previous experience of mechanics.
These tools have proved 70-80% accurate in predicting failures on their first try, Pazos says. And they continually improve with experience. Applications can be implemented as turnkey solutions since the company employs both data scientists and subject matter experts in aerospace.
Casebank Technologies has been helping airlines with diagnostics since 1999 and now supports 10,000 aircraft operated by 300 companies, including some very large airlines, according to Phil D’eon, senior vice president of strategy. Casebank has two basic applications, SpotLight and ChronicX.
SpotLight stores data on symptoms, causes and solutions of component failures. Then, through diagnostic reasoning, it recommends optimal troubleshooting steps. The data come from both OEM manuals and individual customer’s experiences in fixing past defects. “It mines the maintenance records on what was done to fix a part and its success, and then that information is enriched by our own subject matter experts,” D’eon explains.
The application recommends diagnostic steps but not repair instructions. These are still given in the OEM manuals to which SpotLight links. “We find causes,” D’eon stresses. He says trials prove SpotLight can reduce troubleshooting time for a novice mechanic to half the time that would be required even by a veteran mechanic who did not have SpotLight’s guidance. And it saves 75% of the time that a novice would need without any help.
Casebank’s second application, ChronicX, detects and manages recurring defects, ranks chronic problems and highlights new trends in defects. It uses natural language processing to interpret unstructured data from pilot and maintenance records and then spots the clusters of recurring defects. Uploading fresh maintenance records every few hours, ChronicX alerts users to the latest emerging defect trends. “It detects recurring defects,” D’eon says. “Then you can put your A Teams on these defects.”
D’eon says combining SpotLight and ChronicX yields a unique solution that improves first-time fix rates and reduces no-fault-founds. Often, the Casebank tools are offered to airlines by OEMs such as Bombardier and Pratt & Whitney. So far, the applications have supported line maintenance only and have been used chiefly by airlines. D’eon predicts Casebank will begin helping with heavy maintenance in the future and that might extend its reach to shops.
Helmuth Naumer, an associate partner in IBM’s travel and transportation unit, says what he calls cognitive computing can enable mechanics to see, based on past experience, which troubleshooting steps are most likely to fix a problem. He estimates that can cut turnaround times dramatically, up to 90% in some cases. And it increases the first-time-fix rate, a big money-saver by preventing cancellations and delays.
Korean Air began using IBM’s Watson-powered tools for cognitive computing four years ago. The carrier started with a single fleet but soon extended the solution across all its aircraft. IBM is now working with other operators.
Naumer says IBM solutions work best for airlines that perform their own repairs because they have extensive repair data.
Another service, Acsis’s new CrossSense application, alerts engineers to recurrent problems that have not yet—but could soon—result in very expensive aircraft on ground incidents, cancellations and delays. Managers and engineers meet each morning to discuss yesterday’s big technical problems, explains Svetoslav Petrov, vice president for development. But Petrov says senior staff do not discuss, and may not even know about, the chronic faults that have not yet caused major difficulties but may soon do so.
That is the gap Acsis was developed to fill. The tool constantly draws the latest data from the airline’s maintenance-execution system. Then sophisticated algorithms spot potential problems. Alerts are sent to managers identifying specific problems and the aircraft affected by each problem.
When a manager clicks on a problem, the entire fleet’s experience with that problem appears. When the manager clicks again, Acsis displays all pilot and maintenance reports related to the problem, past troubleshooting steps and which parts were replaced or repaired each time. The tool even assigns a probability to a delay occurring on future flights due to each problem.
Acsis does not recommend specific troubleshooting steps. But it warns the senior staff, who are best-qualified to analyze chronic problems and respond, about their biggest risks.
Intel has just launched Saffron for maintenance support. Manager Gayle Sheppard says it works on associative memory, not machine learning, which needs a lot of data to build a helpful model. “Associative memory takes each event and doesn’t need data to train the algorithms. It can use a one-shot example to understand a similar situation.”
Saffron has two capabilities. Its Similarity Advisor looks for close matches to a maintenance issue and identifies paths to fixes based on even sparse experience of similar events. An early use of Similarity Advisor helped an aerospace manufacturer cut part searches from up to 4 hr. to about 5 min., by sorting through complex and changing part data to find the right part for each job.
Saffron’s Classification Advisor automates and improves accuracy of assigning maintenance issues to correct ATA codes. Manual assignments get only 70% of the first two digits and 23% of the next two digits correct. Saffron hits 80-90% accuracy on the first six digits, helpful in planning maintenance tactics.
Intel is working on new capabilities for Saffron, which is now being used by one aircraft OEM. “Stay tuned,” Sheppard advises. Implementation, including data ingestion, generating insights and refining them, takes six to 12 weeks.
Sydney-based 1Ansah has been helping a rotorcraft MRO in Australia troubleshoot problems for several years. It is now extending its solution to fixed-wing aircraft, according to CEO Anant Sahay.
1Ansah combines natural language processing and machine learning to enable mechanics to easily query technical documents and benefit from past experience with similar problems.
It first ingests maintenance manuals, component manuals, service bulletins, airworthiness directives and fault-isolation manuals, analyzes their content and creates an index specific to MRO. That enables the application to quickly locate all materials that might be relevant to a troubleshooting challenge. The software can now ingest even illustrated part catalogues and combine such data with 3D models to create an augmented-reality recognition engine. Sahay says his approach provides help well beyond what is available from major OEMs’ troubleshooting software.
He says machine learning uses statistical techniques to learn from experience. “It learns from the organization’s data, including technical publications, digital and handwritten logs, work orders and defect history, users’ behavior and aircraft behavior,” Sahay explains. He believes 1Ansah offers many benefits for airlines, including reducing repair times and improving the effectiveness of repairs.