原创 Dealing with noisy nuisances

In early 1989, I took an IC design job at a sensor manufacturing company that we'll call SensorCo. It was about three years into a venture-funded start-up staffed by many brilliant PhD scientists and engineers who had pioneered many new processes of making what are now called MEMS (micro-electromechanical systems). They were using break-through (at the time) technology such as finite element analysis for modelling the ideal points to place piezo-resistive elements on a chip to maximise their pressure sensitivity. As one of the few BSEEs there, I felt somewhat under-qualified but figured I could make a contribution if I worked diligently and kept my ears open.

The product line consisted of several very interesting products, including a medical intra-venal pressure sensor to monitor blood pressure in real-time with a small sensor inserted into a vein, a resonant atmospheric pressure sensor intended to be used in harsh environments such as oil drill rigs to measure drill depth, an underwater pressure sensor intended for a dive watch, a ±1g accelerometer, and some absolute and differential air-flow sensors.

My first assignment was to design a voltage-to-frequency converter intended to convert a small differential voltage to a frequency so the pressure value could be easily transmitted through noisy environments. The latest circuit simulation system was available, Analog Workbench from Analog Design Tools, essentially SPICE running behind an X-Windows schematic capture GUI. I had used SPICE before on an Intel 80286 system, so AW was a real treat. Unfortunately there were few models available, and I spent a lot of time learning, "trial and error-ing" and spinning my wheels.

Recall that this was a VC-financed start-up. The company had been in R&D mode for three years with no profit, and the VCs were getting impatient. A couple of months into my time there, a layoff was announced. I don't recall the numbers, but my guess is about 20 per cent of the work force was let go. Some of the really cool projects were scrapped, and a new focus was determined to make the remaining products profitable. I was relieved to learn that I was spared but naturally was very nervous for the future.

The highest volume product was the blood-pressure sensor, a seemingly simple product at first glance: a Wheatstone resistor bridge on a silicon chip. Two of the resistors were positioned to remain constant while the other two changed with pressure. The change of resistance with pressure is highly non-linear, but the desired output was a linear response. The solution is placing serial and parallel resistors around the bridge to linearise the response, similar to the way thermistors act.

If you consider that each sensor has its own specific non-linearity and requires a specific compensation network, it's clear that volume production is a test-and-measurement-intensive process. First the sensor is mounted to its substrate and put in a test fixture. A series of measurements are made at extremes of input excitation, temperature, and pressure. These values are stored on a floppy disc, using bar codes to keep measurements and DUTs (devices under test) in synch. The unit is then put into a laser trimmer and the compensation network is passively trimmed based on the stored values. Next the unit goes to final test, where it again is subjected to extremes of input excitation, temperature, and pressure and tested for linearity.

If that's not complicated enough, consider that the raw full-scale output of the device is about 10mV. The manufacturing floor contained computers, pressure lines, solenoid actuators, heaters, signal sources, X-Y table probers—all the paraphernalia necessary to test the parts. In other words, it was a very noisy environment in which to measure some portion of a 10mV full-scale signal.

Final yield of the blood-pressure sensors was running about 50 per cent, and throwing away half of product starts was not good enough to break even. Because of my prior experience testing and trimming 16bit DACs and ADCs, after the layoff I was approached with a new directive—improve final yield of the blood pressure sensors. This would have a direct impact on the bottom line and was a highly visible project. The pressure was on (pun intended).

The production technician in charge of keeping the line up and running was assigned to acquaint me with the test systems. He showed me schematics of the excitation circuitry, which was custom built for the purpose. It was MS-DOS computer driven, with an interface that drove a DAC08. It was interfaced to a serial port with serial-to-parallel conversion, sending a digital word to an 8 bit MDAC that fed a current to an op amp, which generated voltage to a resistor at the top of the Wheatstone bridge under test.

The op amp was some distance away from the test fixture, maybe 12 or 16 inches  giving lots of opportunity for all the test machinery to inject noise that could easily overwhelm the 10mV signal being measured. When we looked at the bridge outputs with an oscilloscope there was various random noise with peaks of 50mV to 100mV, overwhelming and obscuring the signal being measured. After some study, the tech and I came up with a plan and scheduled a Saturday when production was down to make our changes.

We came in bright and early and went to work. The op amp was remotely sensing, so the summing junction was a long wire from the DUT back to the excitation circuit. We changed the driver from voltage to current, which eliminated the op amp and its noise-sensing summing junction and also the resistor feeding the bridge. Multiple ground connection "antennas" were reduced to a single current return wire. The signal path from the driver to the DUT thus became one wire out and a return wire back, susceptible only to magnetic interference and not electrical noise, due to the signal being a current. A single-point-connected (non-current carrying) ground shield was used to protect the current lines out and back.

I don't recall if changes were made to the software to account for the current output; probably there were. Also I don't recall if the DAC's reference voltage was modified; I suspect it was. The following week, yields went above 80 per cent, and the line became profitable. That allowed the technician and me to survive the next layoff. My job title changed from design engineer to test engineering manager, and ultimately the company became successful and was sold.

I can't claim to have saved the company—certainly it's easier to improve an existing system than to design and build it from nothing. However it was a good feeling to know that the tech and I were able to play a significant role in helping the company succeed.

Brian Lowe submitted this article as part of Frankenstein's Fix, a design contest hosted by EE Times (US).