1 May 2002
Smart sensor project takes flight
Boeing 'pressure belt' to measure airplane wing stress.
By Wayne Catlin, Lee Eccles, and Larry Malchodi
In 1997, several groups within Boeing, along with Endevco and Georgia Tech, started building a smart sensor–based device to measure pressure distribution on the top and bottom surface of an airplane wing.
Called a “pressure belt,” the device is a long, flexible circuit card with smart sensors located at intervals along the belt. Measuring airplane wing pressure distribution places some severe requirements on designing pressure sensing circuits. For example, the temperature range over which they must operate varies from –60°C at altitude to more than +50°C when sitting on the ground in the sun.
In addition, the device had to maintain a maximum uncertainty of 0.1% of full scale, referring to range of measurement. Measurements range from about 0.5 pounds per square inch (psi) to 15 psi, making its full-scale range 14.5 psi. The maximum allowable error, therefore, is 0.015 psi. Also, to remain within the boundary layer of the air flowing over the wing, it needed to be less than 0.1 inch thick.
|ASIC||application-specific integrated circuit|
|FIR||finite impluse response|
Each group involved in the development had a different objective. One group within Boeing wanted to demonstrate an array of microelectromechanical systems (MEMS) sensors, for which they received a Defense Advanced Research Projects Agency contract.
Georgia Tech helped with sealing the integrated circuits against the environment. The objective was to design a reliable product without hermetically sealing it in a metal can, a traditional way of protecting high-reliability parts. Instead, it encapsulated the active components in plastic.
Endevco and Boeing Commercial Airplanes testing organizations wanted to achieve two other objectives. First, they wanted to develop a product that could measure the pressure distribution across an airplane wing. (Endevco wanted to build and market it, while the Boeing people wanted a product they could use.) Second, they wanted to investigate techniques needed for a successful smart sensor network technology. The second objective is the primary subject of this article.
HEIGHT LIMIT DRIVES DESIGN
The project faced several requirements, but the one that drove most major design decisions was the height restriction. Because of it, the pressure sensor needed to be extremely small, leading us to choose a MEMS sensor as the sensing element. Also, because of the height limit, we implemented two application-specific integrated circuits (ASICs) for the active parts of the sensor module (for ASIC details, see sidebar).
We prototyped all the circuits before committing the design to an ASIC. We chose analog circuit components that matched the characteristics of the analog ASIC vendor’s circuit components. The ASIC vendor also drove the choice of 12-bit analog-to-digital converters (ADCs) because it was the best the vendor had to offer at the time.
The digital ASIC was prototyped using field-programmable gate arrays. A packaged version of the processor was used in the prototype. Endevco built the smart network sensor prototypes. Boeing Commercial Airplanes built the bus controller prototype.
The filter circuits and correction engine in the digital ASIC were key to accepting some choices made in the signal conditioning circuit. There were two problems. The cutoff frequency of the antialiasing filter was both too high and not programmable. The addition of the finite impulse response, or FIR, filter, which was designed using a mathematical model, solved both problems nicely.
There were two side benefits from the use of a digital filter. Normally, when a digital circuit has been placed in close proximity to a low-level signal, noise occurs in the analog signals. This was true in this case as well. However, it turned out to be a benefit instead of a problem. The digital filter removes noise from the output so well, it is one of the quietest signal conditioning circuits we have ever seen.
In addition, the noise on the input to the filter causes the filter’s output resolution to be greater than the input bit stream’s resolution. The ADCs are 12-bit devices, yet the digital filter’s output resolution is providing data with at least 14 bits of useable resolution. We worried about trying to get 0.1% accuracy with 12-bit ADCs, so when we discovered this phenomenon, it was a welcome relief, not to mention a surprise.
Another decision made in the design was to use an untrimmed sensor and an untrimmed ASIC. The sensor Endevco supplied normally comes with some extra resistors to help balance the output and correct some of the zero drift. We took these components out of the sensor, as we believed they would not be required.
An ASIC not trimmed can also provide some wildly varying component values. Absolute values of components in the analog ASIC can vary as much as 20% from device to device. Gains and offsets provided by these devices vary by a large amount, but that is acceptable, thanks to the correction engine in the digital ASIC.
The correction engine, as described in IEEE standard 1451.2, is a very general algorithm. In the past, Boeing used several different algorithms to convert a sensor’s output to engineering units.
The most common algorithm is the linear single section conversion. In this case, the sensor’s response fits to a single straight-line segment, and that line’s equation converts the data. Another algorithm used is linear multisection, where the sensor’s response models as a series of straight-line segments, each of which applies over some range of the sensor output.
It is easy to see the first conversion algorithm is just a special case of the second algorithm. It is simpler to program because it does not require the logic to determine which segment the sensor output resides in. Still, the second algorithm works for both processes.
To take this one step further, a linear equation is just a special case of a polynomial equation. The limit of the power of the highest term is one, but the same algorithm that will solve a polynomial will also solve the linear equation, given the correct coefficients. This is all well and good, but the people at HP labs (now part of Agilent) went even further. They decided multiple inputs could apply to the same process.
If you use two inputs, they both determine the segment and represent an area on the surface that defines the response of the sensor to two inputs. All of this goes into the correction process defined in IEEE 1451.2. Each input has a series of terms defining that particular input’s segmentation. Each input faces examination in sequence, and the final result is to select a set of coefficients for the multivariable polynomial appropriate for that segment.
The inputs are raised to the appropriate powers, multiplied together to produce the terms in the polynomial. These terms multiply by the coefficients and are summed to produce the final result. This sounds like a complicated process, and it can be if that level of complexity is required. However, if a linear single section conversion is all that is needed, only one extra step is required: to decide the input resides in the only segment defined for that input.
USES FLOATING-POINT MATH
In the pressure belt, we implemented this algorithm in a hardware state machine. Single precision IEEE 759 floating-point math goes on throughout the process, producing a floating-point result.
The correction engine compensates for the variability, nonlinearity, and temperature drift of the sensor and signal conditioner. The correction engine can compensate for most variations, as long as they are repeatable. You cannot remove thermal hysterisis in the sensor or circuit, but you can correct most other errors.
In the pressure belt case, the excitation voltage is a strong function of temperature, and the bridge output is largely a function of pressure. We brought both of these inputs into the correction engine, and a two-input polynomial equation fits to the response of the sensor and signal conditioner. Because both the sensor and the signal conditioner are in the same temperature environment, you can handle the temperature drift of the sensor and signal conditioner at the same time.
The process of calibrating the sensor is now much more complicated. We perform pressure calibrations at several temperatures with the correction engine turned off. The values we obtain determine coefficients for the polynomial. The calibration transducer electronic data sheet is then written into the sensor module. You have to check the calibration at several values of pressure and temperature to verify that the module is set up correctly.
NOT YET PERFECT
How well does the pressure belt perform? At this point in time, not well enough. Some modules work and are within 0.1% uncertainty, but many have errors up to 0.3% over the temperature range of –60° to +50°C. We believe the errors result from stresses that are not repeatable, introduced into the sensor. We are continuing to investigate the problem. The noise performance of the module is excellent. For most applications we have, the accuracy is acceptable as well. However, it is still not good enough for the primary application.
Boeing and Endevco are investigating using these ASICs for other applications, and we are finding some limitations. One application we’re investigating is using the circuit with sensors not in the same temperature environment as the signal conditioner. This would be the case with structural strain gauge bridges in many applications.
Endevco is building some parts to use in this type of application. It wants to add a temperature sensor to the module to measure the signal conditioner temperature. The temperature and the voltage from the remote sensor input to one channel of the correction engine, and a computed output corrects for the temperature drift in the signal conditioner. The output of this calculation then goes into the second channel in the correction engine to linearize the sensor output.
For a lot of applications, this will be sufficient. However, because our current correction engine is just a two-channel device, there is no third input that could be used to temperature compensate the sensor output.
WITH 20/20 HINDSIGHT . . .
If we were doing this design again, we would change several things. For one thing, the correction engine would have at least four input channels and at least four output channels. More thought needs to be given to number of input and output channels, but four would appear to be a minimum.
If we were designing a device for a general market, we would add the ability to produce an n-bit integer output, if required. There are many applications where 16 bits is more than adequate resolution, and the extra bandwidth required to transmit 32 bits is not acceptable.
In Boeing Commercial Airplanes, we expect to use the floating-point output, but many other people in the aerospace industry will not. ADCs with quite good digital filters have come into the market since we started this design. We probably would use those devices instead of putting the digital filter in the ASIC, since this provides a more general solution. IT
Behind the byline
Wayne Catlin, associate technical fellow, Lee Eccles, technical fellow, and Larry Malchodi, associate technical fellow, all work with Boeing Commercial Airplanes–Validation organization, which combines the lab and flight test organizations. They have worked on the pressure belt project since its inception.