In high frequency switched systems where you measure current with a shunt, you may observe problems such as excessive amplitude on sine wave current ripple, overshoots on square wave ripple or any rapid transition of current, or excessive high-frequency noise. These problems are due to current shunt inductance, which becomes more significant at lower values of shunt resistors, particularly below 1 mΩ.
Troubleshooting requires a tool that anyone designing current sensing systems should have: a top quality clamp-on current probe that has a bandwidth from DC to high frequencies. Tektronix current probes are a popular and ideal choice for this measurement. Highest accuracy is not their forte, you can achieve much better with quality shunts and IC's for accuracy, but to measure and solve problems with dynamic signals the current probe is essential. Not fixing these dynamic problems can compromise your basic current measurement accuracy and corrupt data gathering systems.
This is an equivalent schematic of a shunt inductance problem. A square wave output of a 100 kHz switching regulator is filtered by L1 and C1 such that the current ripple is a sine wave. H1 captures the actual current waveform (probed by ROUT1), and E1 captures the exact voltage across the shunt and its inductance (probed by ROUT), just like a current sense amplifier (the 20 volt supply contributes to convenient offset and scaling to view output waveforms together).
One problem you can run into involves incorrect sine-wave ripple signal amplitude and waveshape. In a real example modeled here, the ripple signal was too large and was casting doubt on the accuracy of the entire measurement. The accompanying schematic supplied to us revealed a mysterious triangle wave drawn on the schematic near the shunt (with no explanation) that at first didn't register to me until I simulated the circuit.
The green trace represents the actual ripple current, while the yellow trace represents the voltage drop across the shunt which is the same signal that would come out of your current sense amplifier with no input filter. Note that the triangle amplitude is much larger than the sine wave (and the sources E and H are scaled such that they will match when everything is right).
This depicts the problem we saw in the application. Since the application had an input filter the waveform out of the amplifier was sinusoidal, but the amplitude was excessive. Simply a matter of too small of a filter capacitor.
This application schematic shows the filter with the incorrect initial values at the RFILT and CFILT locations, yielding the waveforms of Figure 3. A later revision of CFILT to 0.3 μF will provide a correct waveform and amplitude, as shown in Figure 5.
Ripple response with the correct filter values. The waveforms lie on top of each other.
Sure enough, the sine wave ripple does turn into a triangle waveform across a shunt with sufficient inductance. The amplifier initially had a sine wave output because the designer had wisely included a low-pass filter at the amplifier input, but it simply wasn't "tuned" correctly. In this case, the tuning involved adjusting the capacitor value until the ripple matched the correct calculated value. The problem with a real-world shunt is that they defy a disciplined analytical method because of the vagueness of their inductance specifications. You may see something like "0.5 to 5 nH" on the front of the datasheet and no specific value on the spec table, and that is if you are lucky. So, you use a current probe to determine the correct value by iterating your capacitor (obviously if the amplitude is too large you increase the cap and vice-versa).
If in fact, you have a real square-wave of current you may be fortunate enough to have an overshoot that you'd "tune out" the same way. Once you find the correct filter value, it will work in production, and may even hold true if you have to change shunt vendors. There aren't that many ways to build sub-1mΩ shunts. And did I mention this transient response problem, as a result of shunt inductance, gets worse as the shunts get smaller, most often at values less than 1mΩ?
The Importance of Filtering Ahead of the Input
It is important that this filtering be done prior to the current sense IC input. Long-term data gathering of systems without front-end filtering has demonstrated unexplained occasional (but frequent enough to be a problem) large spikes in data plots of current and power values. These spikes were traced to the rising high-frequency response of the shunt causing aliasing in the current sense front end. It doesn't matter if it is a chopper-stabilized amplifier, delta-sigma converter, or SAR with averaging, all of these systems are vulnerable if they are sampling systems. As with any aliasing problem, the correct solution is analog filtering before the input of the current sensing IC. Disregard the vendor that says you don't need a filter. If it is a sampling system and you are collecting data, you need a clean signal into your current sense IC. Also remember that aliasing is not the only potential problem, unfiltered inputs run the risk of these high frequency inputs simply overloading the front end.
Lastly, you certainly can tune to filter at even lower frequencies should you desire even more noise rejection. And filtering before the input to the first amplifier in a chain is always beneficial. Most current-sense ICs limit practical filtering at the input to a single pole, but should always be used, and higher order filtering also implemented, if needed, at the amplifier output.
While this article discusses this problem in the transient domain, any astute observer also realizes it can be viewed as a simple first-order bandwidth problem. Shunt inductance on very low ohm shunts create corner frequencies in the hundreds of kHz, sometimes surprisingly low. No matter how you treat it, as a bandwidth problem, a time constant problem, or transient response problem, the optimum filter will have a time constant equal to the time constant of the shunt resistance and its inductance (or a pole frequency that compensates the shunt zero frequency):
Current sense ICs will always use a differential filter and RFILT will be the sum of both resistors. From a math standpoint, the hard part is getting a real number for LSHUNT.
Lastly, a frequency response plot shows the rising frequency response of a 500 μΩ shunt with 3 nH of inductance in green, and the complementary response of the input filter with a pair of 10Ω resistors and 0.3 μF capacitor. Note that this shunt demonstrates a corner frequency of approximately 30 kHz.