Looking inside an old Bruker MRI gradient amplifier

[D Straney]

This was being thrown away about 2 years ago at a previous job, and I got the go-ahead to strip it for useful parts before it got hauled out; thought I'd take some photos while I was at it.  Apologies for the heavy flash glare, weird white balance, and generally awful photography: had to do this in a dimly-lit equipment room.

One of the main parts of an MRI scanner are the gradient coils: there are generally 3 of these, one for each physical axis (X, Y, and Z directions), and they're used to create magnetic field gradients across the interior space of the scanner.  These gradients are necessary to separate the magnetic resonance signals from different areas of the patient, and reconstruct an actual 3-dimensional image rather than just a single undifferentiated pixel (like the NMR machines used by chemists, where they don't care about the spatial aspect).  The details and physics of this have been explained much better by other people, so go read up on MRI gradients if you're interested.  What this requires on the electronics side, though, is power amplifiers that can push very high currents through these coils (in the 100s of amps min.) while also being able to put high voltages across the coils to ramp the currents up or down quickly (in the 100s of volts min.).  Modern switch-mode Gradient Power Amplifiers (GPAs) made for high-field MRI systems usually have a 3-phase feed along with water cooling in giant cabinets, and can output multiple kV/kA per channel.  This system here though, from the early 2000's judging by IC date codes, will turn out to use only linear amplifiers, and looks like it's meant for much smaller power levels.


Overall, there's a control & readout box on top, and a rack with 3 large amplifiers and one small 3-channel amplifier (top).  There's also a giant mess of cabling in the back, not shown here.

Control section

This is the brains of the operation for the 3 large amplifiers (likely X, Y, and Z).  A digital board at the back (top-right) with a Siemens CPU accepts commands from a more central MRI computer over an external D-sub connector.  This then feeds digital data to the 3 channels of the DAC board (top-left).  The setpoints from the DAC board then make it (possibly by way of the supervisor board in the middle) to a set of two stacked pre-emphasis boards.  These modify the setpoint signals in a way I'll describe shortly, and then output the final X, Y, and Z current setpoint signals from the back, to be fed to the 3 large amplifiers.  Meanwhile, these amplifiers are returning output current measurements to this control section: the current measurements are both watched by the "supervisor board" in the middle, which seems to compare them against the DAC signals and watch for overcurrent and out-of-regulation events to shut down the power stages if there's a fault, and to the display board at the front of the enclosure which generates the bargraph displays of output on each gradient channel.


Pre-emphasis
Creating a desired magnetic field inside a coil with known geometry is pretty straightforward at DC.  However, when the current in the coil is changing, the changing field induces eddy currents in the metal structures nearby.  With radios, etc. open solenoid coils are normally kept away from large pieces of metal specifically to avoid eddy currents and their effect on inductance and losses: however, in an MRI scanner, having metal near the gradient coils is unavoidable, with the large Dewar that holds the superconducting magnet providing the static field, and even the superconducting coil itself should have some significant coupling to the gradient coils.  This ends up having a low-pass filtering effect on the magnetic field ramp-up and ramp-down slew rate - this diagram illustrates it very nicely, along with how the waveform can be "pre-distorted" to compensate for the effect: https://www.mriquestions.com/what-is-pre-emphasis.html

I'm not sure exactly how the pre-emphasis works in this Bruker GPA, but rather than being done all digitally as it would be handled today, the pre-emphasis board here likely has a multi-stage digitally-programmable filter for each channel, so that its step response can be carefully shaped to be the inverse of the gradient-coil-with-eddy-currents step response.  There's 2 stacked boards to serve 3 channels, each with 9 identical sections, so that adds up to 6 filter sections per channel.

Globally, there's a few OP27 op-amps to handle the inputs & output signals, as well as an AD7514 SPDT analog switch.  An Altera(?) programmable logic device of some kind seems to be responsible for digital interfacing.

Each identical filter section has a 3x OP27 op-amps and an OP15 op-amp, some large 100 nF film caps (presumably for lower-distortion/more precise filter responses than ceramics at the time could provide), an AD7547 dual 12-bit multiplying DAC, and an AD7590 quad SPST analog switch.  I'm guessing that the multiplying DACs are used as a programmable-gain element as part of a programmable filter response (the AD7547 datasheet even shown an example application circuit using 2 as a programmable state variable filter, although this applies to all multiplying DACs: https://www.analog.com/media/en/technical-documentation/data-sheets/AD7547.pdf).

Large amplifier


Inside, these look identical to a high-power linear audio amplifier, which were actually what was used to drive gradient coils in early MRI experiments.  As well as a giant transformer and control-power board at the bottom-left corner here (inc. some logic, relays, and power resistors for pre-charging bulk capacitors), you can see the power stage's bipolar DC supply in the form of the four large metal-can capacitors at the left, and the bridge rectifier attached to the heatsink.

There are two back-to-back heatsinks, each with its own power board that contains many paralleled transistors in TO-3 packages.


Those diode pairs in TO-247 packages seem to be for clamping the output voltage (important with an inductive load) while the two transistors at the left edge, on either side of the thermal-cutout switch, I believe are pre-drivers for the combined base connections of the transistor arrays.


Each power transistor has its own emitter resistor for current sharing, and its own fuse:

You can see some of the biasing components up on posts, likely selected then added as part of the testing/calibration process to match the individual board.


Under that metal shield at the lower-left of the inside overview is the control board, which contains the output-current control loop, as well as some digital signals (shutdown, fault reporting, etc.):


One fun thing about the large amps is these compensation adjustments, made via a set of DIP switches on the front panel.  When running a constant-current control loop that drives a voltage across an inductive load, the load's impedance directly affects the control loop response and stability, and so to be able to make the amplifier work well with a variety of gradient coils (a range of both load inductances and resistances), the control loop values need to be adjusted.  Today the control loop would probably run inside a DSP (which allows for more complicated control schemes, auto-compensation tuning, and all kinds of other nice things) but in this case it was just done by selecting capacitors with switches, and changing a resistance with a multi-turn pot:

[KE5FX]

Very cool piece of hardware.  I'll bet you could sell the coil driver PA on an audiophile forum as an exotic subwoofer amp. 

[TimFox]

Early MRI scanners used commercial audio amplifiers from Crown (now part of the Harman empire owned by Samsung), modified for the inductive load to create these gradient fields.
Basically, the magnetic field gradient makes the nuclear resonant frequency a function of position in the direction of the gradient vector, so that Fourier analysis can reconstruct the distribution of material that resonated.
The waveform and amplitude of each gradient pulse is controlled by the overall system controller, and requires linear amplification, which in more modern units can be done in switch-mode.
The percussive noise you hear from inside the scanner comes from a reaction between the pulsed current in the gradient coils (mounted on a cylindrical form coaxial with the scanner) and the DC magnetic field from the superconducting  electromagnet.

[D Straney]

Small 3-channel amplifier
This seems to be a standalone piece of equipment, which doesn't tie into the control system in the same way as the larger amplifiers, and has much, much lower output power.  I'm guessing it's for driving shim coils: much smaller gradient coils which are given fixed currents to make small corrections to the static magnetic field, and keep it more even throughout the imaging volume to reduce distortions in the image caused by magnetic field imperfections.

It's labeled ACUSTAR both inside and out, which I think is a company bought by Bruker?

Inside, there's a lot of off-the-shelf AC-DC power supplies and EMI filters (inc. a 48VDC supply for each channel), as well as a pre-emphasis board at the back, and 3 power modules.

The pre-emphasis board is all-digital and seems to be just for interfacing & control as opposed to doing any actual filtering or modification of the current setpoints:


Each of the power modules has a heavy heatsink on top, power & setpoint inputs, an amplifier output, connections for control (current measurement, fault reporting, etc.), and a digital data input for pre-emphasis.


Inside, there's a control side and a power side:

The control side connects to the "system I/O" port and so is likely responsible for interfacing, faults, etc.  It also has a DAC connected to the "digital data" port; you can see that at the right side with the digital optoisolators and buffers that feed it.  This is probably how the current setpoint is fed to each power module, digitally from the outside.


You can also see the metal tabs of the power op-amps here, which leads us to the power side.

This contains the actual power components, including those OPA541 power op-amps (10A output capability), likely with two each connected in a bridge configuration for a 48V 20A output off a single power supply.  The actual current control loop seems to be on this board as well.

Looks like swappable compensation components again, to adjust for the load:


I also like those soldered testpoints where you can just drop a scope probe straight in without the "witch hat":


More close-ups:


There's also the analog pre-emphasis board here, which like the pre-emphasis board in the control section for the much larger amplifiers, has a few copies of the AD7547, AD7590, and film capacitors / op-amps.  Presumably it works the same way, with a digitally-programmable multi-stage filter.

[mtwieg]

Thanks for sharing, I love this stuff 

The adjustable filters are really interesting. In a GPA I'd expect that for each axis, you'd have three different adjustable filters. One for feedback (to optimize stability and bandwidth of the overall feedback loop), one for feedforward (to reduce how much error the feedback has to deal with), and one for preemphasis. Each one is probably at least a third order transfer function, and you need at least a few bits per order... overall you need lots of switches and passive components to cover all feasible implementations.

I'm actually surprised by how minimal the implementation of these adjustable filters is on these boards. Perhaps the preemphasis is actually handled by the console side rather than inside the GPA? Or maybe since these are linear amplifiers, the requirement of the feedback network is far simpler than I'm assuming. How did you deduce which ones are used for preemphasis and feedforward?

I'm really interested in how signal integrity is managed with all these cables racing around inside. Personally I would try and use differential signaling whenever possible, but judging from all the coax cables it looks like everything here is single ended. I suppose having all the small signal stuff segregated onto its own boards helps reduce crosstalk with the power electronics.

You're likely correct on the AD7547 acting as an adjustable gain element for the filters. I'm wondering if they're just used to scale the overall output of each filter, or if they're actually used to adjust poles/zeros in the transfer function. Have you tried sketching out the schematic for any of these?

I'm guessing the amplifier outputs are bipolar (as opposed to unipolar differential outputs like you'd get from an H-bridge). I'm wondering how the return path of the output is managed to avoid crosstalk with the small signal systems.

Also, where are the current sensors?

[D Straney]

Ha glad to hear it, I figured you would!

For functions, I'm mostly going off the labeling: Bruker seems to be good at slapping down big silkscreen text about board function everywhere, and the big wide board pair in the control box up top is labeled as "pre-emphasis".  The general layout of the system based on cabling, etc. seems to be that the control box exports current setpoints (with pre-emphasis included) to each of the power amps, and the power amps handle their own feedback & possible feed-forward internally, on that single board next to the power stage.
The labeled "loop" controls for the power amps have both resistors & caps controlled by the DIP switches, so that should be for feedback/control loop compensation; there might be no feed-forward on these or it might be a fixed amplitude, since if you only need to compensate for changing L & R in your load, then the "loop C" setting could change a zero location @ crossover (or a zero to cancel the load pole, depending if you want to do a double-speed descent to crossover, or a simple single-pole descent throughout) while the "load R" setting could set an overall gain based on load resistance that if done correctly would keep both DC feedback gain and DC feed-forward gain correct across load range, I think.

I can definitely imagine noise getting into the signaling being a challenge: keeping most of the small-signal stuff in an entirely separate box does help a lot, and yeah there's probably a lot of care with how the grounding is arranged in the power amps themselves to make sure no output currents flow through the small-signal return paths.  Seems to be a lot of optocouplers involved in isolating digital signals at least from the ground bounce involved.

Adjustable filters: I assumed the DACs were for directly adjusting pole & zero locations, but I guess they could also be mixing in variable amounts to the output from a fixed differentiator circuit, or something like that.  Haven't traced out the schematic but should do that, I still have the pre-emphasis boards and it's a lot easier to do continuity checks than on the conformal-coated avionics stuff I look at in the other posts, which is a huge pain to trace.

Good question about the current sensors.  It could be picking off one of (or mixing signals from a few of) the transistor individual emitter resistors to estimate the total (with calibration for imperfect current sharing), although that seems dangerous in a lot of ways.  Maybe that's what the 4x 0.1Ω (paralleled?) gold-chassis-mount power resistors mounted directly to the heatsink are for (if not part of some kind of RC snubber), and the single coax coming off the top-left corner of the power stage board takes the output from those back to the control loop board?  I also didn't feel like fully disassembling & lifting one of the big heatsink assemblies at the time, so it's plenty possible there's a massive current-sense resistor assembly buried in between the two heatsinks, close to the fan maybe.

[mtwieg]

For functions, I'm mostly going off the labeling: Bruker seems to be good at slapping down big silkscreen text about board function everywhere, and the big wide board pair in the control box up top is labeled as "pre-emphasis".  The general layout of the system based on cabling, etc. seems to be that the control box exports current setpoints (with pre-emphasis included) to each of the power amps, and the power amps handle their own feedback & possible feed-forward internally, on that single board next to the power stage.

Ah, a silkscreen label is self-explanatory. I find it strange that preemphasis is handled inside the GPA, I'd assume that even in the 90s your computer/console would be powerful enough to do all the DSP business. And doing it digitally frees you from the arbitrary limitations of switched filters. 

The labeled "loop" controls for the power amps have both resistors & caps controlled by the DIP switches, so that should be for feedback/control loop compensation; there might be no feed-forward on these or it might be a fixed amplitude, since if you only need to compensate for changing L & R in your load, then the "loop C" setting could change a zero location @ crossover (or a zero to cancel the load pole, depending if you want to do a double-speed descent to crossover, or a simple single-pole descent throughout) while the "load R" setting could set an overall gain based on load resistance that if done correctly would keep both DC feedback gain and DC feed-forward gain correct across load range, I think.

Yeah I follow you. I'm betting for linear amplifiers, the feedforward doesn't need to be super accurate, only need a first-order model (coil L and R). The feedback loop can take care of the residuals (unlike with a digital controller where the feedback is throttled by sampling delays).

Good question about the current sensors.  It could be picking off one of (or mixing signals from a few of) the transistor individual emitter resistors to estimate the total (with calibration for imperfect current sharing), although that seems dangerous in a lot of ways.

Like you pointed out those resistors are to improve current sharing between the power semiconductors. I doubt they measure current for the feedback loop. One of the benefits of a bipolar output amplifier is that the return is effectively grounded, so it's very convenient to put the current sensor there. Makes sense resistors much more viable because you don't need a shunt amplifier with a huge common mode input range.

Maybe that's what the 4x 0.1Ω (paralleled?) gold-chassis-mount power resistors mounted directly to the heatsink are for (if not part of some kind of RC snubber), and the single coax coming off the top-left corner of the power stage board takes the output from those back to the control loop board?

That style or resistor is usually wirewound nichrome, terrible for accurate current sensing.

I also didn't feel like fully disassembling & lifting one of the big heatsink assemblies at the time, so it's plenty possible there's a massive current-sense resistor assembly buried in between the two heatsinks, close to the fan maybe.

Yeah I'd expect there to be a big shunt looking something like this, with kelvin sense connections. Might very well be hidden away from all the power electronics so it's thermally isolated. Or a fluxgate sensor, but less likely in older and medium-current amplifiers like this.