Sectors of industry which operate constantly, such as manufacturing and shipping, aircraft and rail operations, often use a form of predictive maintenance known as condition-based maintenance (CBM). The machine is inspected while it is running, with the emphasis often being on rotating components such as bearings: these may be subject to vibration analysis, oil analysis or thermography.
Meanwhile, vehicle operators are making more and more use of predictive techniques, made possible by on-board sensors and the connected nature of modern trucks. For instance, tyre pressure monitoring has proven to be an effective form of predictive maintenance; a loss of pressure correlates pretty well to the likelihood of a blowout (as well as being bad for your fuel consumption, of course).
But other systems in a truck are more complex, and generate a mass of data — and given the diverse nature of truck operations and the vehicles themselves, this can be extremely difficult to interpret. So manufacturers are using artificial intelligence (AI) or machine learning (ML) to generate useful conclusions and recommendations.
HACKATHON
In 2016 Scania contributed two machine learning ‘datasets’ — databases populated with real-world information gathered from truck telematics — to an industrial challenge competition based at Stockholm University, but open to teams around the world. The data came from thousands of Scania trucks in operation, and the challenge was to diagnose faults in the air pressure system (APS).
The teams used a ‘training’ dataset to guide an ML algorithm, with the aim of determining which trucks had problems with the APS. This dataset contained 60,000 examples (each with up to 170 items of data) of which 59,000 were negative (no problem) and 1,000 were positive (there was a problem with the APS).
The criteria used to judge the algorithm were based on the cost of mistaken diagnoses — both false positives and false negatives. A false positive leads to the cost and inconvenience of inspecting the truck, only to find that there is no fault. A false negative, on the other hand, could lead to a failure, a VOR and substantial cost and downtime. By assigning a cost to each eventuality and ‘training’ the system to minimise the overall cost, you can develop a useful algorithm.
What’s surprising is that there is no need to know precisely what each item of data means (for example, a local oil temperature); the failure prediction simply relies on finding patterns in how the different data fields are correlated. So it’s possible to create a successful ML strategy without any human knowing exactly how it works.
In the 2016 competition, algorithms were tested using a separate ‘test set’ of 16,000 real-world examples, and the winner delivered the lowest overall cost — although subsequent experimenters have come up with even better algorithms.
FINDING THE SOURCE
But all that data must come from somewhere, and there is an increasing range of sensors available for different criteria. Temperature measurement is one obvious area, and is almost trivially easy now, with tiny sensors that combine a thermocouple with the amplification and interfacing hardware needed to talk to the truck’s telematics.
Vibration analysis usually involves mounting an accelerometer on the bearing to measure its displacement; these days the accelerometer is often triaxial (it can measure vibration in three orthogonal directions) and built into a single MEMS (micro electro-mechanical system) chip. Whereas traditional systems needed interpretation by an experienced professional, today that single chip can include hardware to record the vibration readings, perform sophisticated signal processing such as FFT (Fast Fourier Transform, which isolates the relevant frequencies from the ‘noise’ of surrounding components) and deliver a diagnosis.
Oil analysis is still usually a matter of taking a sample of lubricant and inspecting it with a spectrograph for contaminants, wear debris (often particles of the bearing material itself) and signs of oil breakdown. However, oil condition sensors are increasingly used on machinery; these generally use the principle that the dielectric constant of the oil changes as it oxidises. And more advanced types are coming into use, such as optical sensors which use Fourier transform infrared spectroscopy to compare the oil’s characteristics to that of known samples. Magnetic sensors can detect not just the presence of ferrous particles in the oil, but, with an inductive coil, discriminate between fine and coarse particles.
Some sensors, such as DSi’s Super Compact Air-X sensor, pictured above left, even use x-ray emissions to determine the precise density of the oil and detect whether it becomes aerated in operation (as air in fluid reduces lubricity and leads to oxidation) but at the moment this is a tool to use at the development stage.
Oil condition sensors can suffer from cross-sensitivity: measurements that indicate one condition (for example, water contamination) could also indicate a different problem (such as wear debris). So firms are having to develop ever more sophisticated machine learning algorithms to distinguish between these issues.
Shell offers a remote oil condition monitoring service called Shell Remote Sense, which is currently being marketed for off-road applications such as dump trucks. However, Shell says it is suitable for “any brand of engine, and any brand of oil … regardless of whether an existing telematics system is used” and says it offers ‘predictive analysis’. A sensor such as Bently Nevada’s VitalyX tracks criteria including base number, acid number and oxidation, and can detect fuel dilution and water contamination. A web-based dashboard indicates oil health and predicted life, and issues alerts when limits are reached.
MONETISING CONDITION
Of course, it’s not just the vehicle itself that is a factor: in some markets, Scania offers a reduction in monthly service contract costs for trucks which are driven exceptionally well, according to the driver score generated by the truck’s telematics.
Volvo Trucks offers what it calls ‘Connected Service Planning’, using data from truck diagnostic systems to schedule maintenance and liaise with dealers. Over 600,000 Volvo trucks already have connected systems, and the firm has an ‘Uptime Center’ located in Ghent, Belgium, where ‘real-time monitoring’ of major components is combined with AI techniques and a large technical staff (which operates in 16 languages). The Real Time Monitoring service is offered as part of Volvo’s Gold level of service contract.
Markus Efraimsson, Volvo Trucks’ vice-president for customer support, says: “By using machine learning, we can identify hidden patterns and predict malfunctions more accurately. That means not only can we almost eliminate the need for extra workshop visits, we can also look at exactly what is causing components to fail.”
Google Maps and similar connected services mean that cars and trucks already collect information on the state of other traffic, but with the multitude of sensors on board a vehicle — particularly if it has autonomous driving features such as GPS and road-scanning cameras or LIDAR built in — the vehicle could perform condition monitoring on the road itself. This sort of thing is already a feature of some railway vehicles, which are equipped with triaxial accelerometers for that purpose. Once vehicle-to-infrastructure (V2I) communication becomes more common, on-vehicle monitoring could help maintain the road network.
BOX: The P-F Curve
The P-F Curve is a useful concept for visualising the way in which a fault develops and could ultimately lead to a complete failure — and how corrective maintenance could avert that failure.
The chart, right, has two axes: time, and the condition of the equipment. We haven’t put a scale on these axes, but they demonstrate the idea. The curve shows the condition of the equipment starting at a high level, then declining, first slowly and progressively faster, until a rapid, catastrophic failure renders the equipment unusable.
Point A is the point at which actual failure begins — for instance, a bearing goes past the running-in stage, and begins to go out of tolerance. This is inevitable, but not in itself a problem.
More important are points P and F. The first marks potential failure — the point at which we begin to experience the equipment starting to deteriorate. F marks functional failure — when the equipment ceases to operate usefully. After this point, the machinery might exhibit excessive noise, odour and overheating, before it fails catastrophically.
The interval between the two points — the P-F interval — is where proactive condition monitoring and condition-based maintenance (CBM) can make a real difference. Checking vibration levels, lubricant quality and other criteria can extend the usable life of a machine dramatically.-Toby Clark
BOX: Data tech round-up
STRATIO, the real-time predictive fleet maintenance platform, is supplying bus operator Go-Ahead Ireland. The Stratio platform integrates predictive maintenance and advanced remote diagnostics with GAI’s current fleet management system.
After a successful proof of concept (POC) trial in 2021, GAI decided to extend Stratio’s service to its entire fleet. The results of the trial indicate that Stratio’s technology will aim to reduce GAI’s vehicle breakdowns by over 57% in the next three years. Stratio enables GAI to turn its vehicle data into intelligence to inform and optimise maintenance operations. Its AI platform remotely collects and analyses crucial indicators of vehicle condition, turning them into real-time actionable insights that allow GAI to predict breakdowns, it says. With Stratio, GAI will be able to reduce the frequency of preventive checks and plan the wear rate of components dynamically, rather than relying on more approximative mileage-based estimates, ultimately saving time and resources.
Vehicle safety technology provider Nauto and Navistar, US truck subsidiary of the Traton Group (corporate owner of Scania and MAN brands in Europe) have agreed a distribution deal to make Nauto’s predictive-AI vehicle safety solution available for new fleet vehicle purchases or as an upgrade for existing fleet vehicles throughout US dealers. Nauto’s AI technology tracks and analyses risk in real time. When it detects risks, it can provide preventative warnings which may give drivers extra time to respond. “Nauto’s partnership with Navistar is an important milestone in terms of commercial fleet industry adoption,” said Stefan Heck, CEO, Nauto.
ZF-owned telematics firm TRANSICS and commercial vehicle digital platform Rio (main picture) are deepening their technical collaboration. With newly developed interfaces and improved data exchange between the partners’ cloud-based solutions, customers can now decide on transport data sources and how information is presented at the front end.
The first company to benefit from the closer cooperation is the French logistics service provider, Jacky Perrenot (fleet vehicle pictured below). Its head of purchasing Christophe Dauzat says: “RIO and Transics have recognised that by connecting the two platforms, there is an opportunity to reduce the cost and downtime associated with retrofitting hardware, thereby affording us greater flexibility. New trucks from MAN are ready to head out on to the road immediately after the respective services are booked, rather than having to go back into the workshop for installation work.”
Jan Kaumanns, CEO of RIO, says: “In the future, a greater degree of networking and partnering will be required if transportation logistics operations are to be optimally efficient.”