Proper, certificated networks, amplifiers and signal sources for EMC work are expensive! It's very attractive to do these tests yourself, at least to get a vague idea of how close (or beyond) the spec your product is, what sources are the worst offenders, and how much filtering is necessary to treat it. The cost of these products and services may seem intimidating—surely there's something critical they have—but, as with many expensive things, the main cost lies in certification rather than engineering. It's quite reasonable to perform precertification testing, for an outlay of only a few thousand dollars, perhaps less if you have some instruments on hand already.
First and foremost, of course, find out what tests you need to perform! For example, if it's for general U.S. sale, it's going to need FCC Part 15. If it's for the E.U. market as well, then a CE mark is needed. (Usually one test will dominate; for instance, CE is more strict than Part 15 class A, so you don't have to test it twice.) Special regulations may apply (medical, telecomms, industrial, military, etc.).
(10/26/2015) This network is used for telecomm lines that connect to a device. Examples include: telephones, wired internet devices (in this case, Ethernet), industrial (if applicable) networks (Profibus, RS-485 anything, etc.). Presumably, the lines aren't simply equipment wiring, but can potentially run for hundreds of meters and connect to many other devices, carrying conducted RF a long distance. The whole network could potentially radiate over very wide frequency ranges, including very low frequencies. This, obviously, isn't very desirable, so some means is required to test it.
Specifically, CISPR 22 calls for an Impedance Stabilization Network of approximately these values (the link goes to the source PDF):
For present applications (Ethernet), this network was chosen. Two pairs (any general unshielded, common pairs, but Ethernet and telephone would be prime examples) pass through, and have the common mode noise tapped off them, with some attenuation. To improve differential rejection, a centertapped choke is provided. To prevent DC loading, this is AC coupled, however this would also create an LC resonant circuit (at a low frequency, about 5 kHz). This is dampened with a 390 ohm resistor across the choke, which puts it approximately in critical damping. This also loads the line, so the network will have a dB or two of attenuation (and some reflection). Don't expect fantastic signal quality through this network.
Wideband performance is probably also poor. Even using transmission line transformer design methods, the propagation delay through such long windings (necessary to achieve such large inductances) will lead to impedance mismatch, dispersion and reflections. Fortunately, most long-distance signaling methods that would be used by this type of equipment have wide margins (e.g., RS-485 and 10BASE-T have around 20 dB of overhead), or are designed to work with these kinds of things anyway (e.g., DSL's adaptive frequency bands). As for intended network performance, it only needs to work up to 30 MHz for conducted emissions, so it's not insurmountable.
My construction built all the coils on P22/13-3E27 cores (AL ≈ 9 μH t-2), chosen for the very high inductivity and being handy in my junk box. The winding details are:
Superglue was used to fix the windings to the bobbins. The resulting winding lengths were all around 0.5 m. Very little winding space was used, and some savings may be had with a smaller core. The impedance should be pretty stable up to perhaps 75 MHz. The crosstalk might not be so good, though. Perhaps some winding space could be used up with more dielectric, to improve crosstalk.
The completed network. Standard prototyping construction: copper clad PCB stock, soldered together. Screws (use washers!) secure the coils to the board. The connectors are extended on break-out boards with 0.1" pitch headers, so they can be swapped out easily. Grounds are provided on the headers, in case it's necessary to support shielded cables or connectors. The board could've been a little larger (I had to mount L2 on a vertical fin), but this still turned out pretty well for a one-off prototype.
The test setup (click for details). The lines were placed over ground plane (some stacked aluminum sheets—a good enough ground plane, even without foil tape (or better) bonding between them), elevated on foam to give approximately the right characteristic impedance (around 150 Ω common mode). The circuit boards are grounded by contact, held in place with tape. (The coupling network in the middle was later weighted down for better contact; some foil tape would've been handy here.)
The testing coupler is simply a balun transformer, to convert from 50 Ω BNC to 100 Ω differential line. The balanced secondary is isolated, with the centertap connecting to the other BNC via a two resistor impedance match and attenuator (giving the same 9.5 dB attenuation factor the ISN has, for the same reason).
Not shown in the picture, the input signal comes from a wideband noise generator and 5 dB amplifier. The output signal goes to the spectrum analyzer (an old, drifty HP 8590A) through a 20 dB wideband amplifier that has a modest noise floor.
This is the spectrum of the noise source. Setup: [noise generator] → [20 dB attenuator] → [20 dB amplifier] → [spectrum analyzer]. The attenuator here represents intrinsic network loss. Note the frequency scale is different, and that the amplitude and resolution bandwidth are different (oops!). Use this as context, but not an exact input value for the following spectra. The takeaway point is: it's not very flat (about ± 3 dB for 0-60 MHz), but the curve is repeatable and will be recognized below.
This is the system response, differential input coupling (top trace), with the noise floor shown for reference (bottom trace). Setup: [noise generator] → [differential coupler] → [twisted pair] → [ISN] → [20 dB amplifier] → [spectrum analyzer]. For the noise floor, the network was disconnected from the 20 dB amplifier, and its input terminated. (At these gain settings, the analyzer's noise floor is about -96 dB. The amplifier adds quite a lot of noise at the low frequency end, but still has more gain, so the SNR is better overall.)
The first conclusion seems to be: the balun works fantastic at low frequencies! Alas, it's probably to blame for the poor performance at higher frequencies (above 6 MHz or so). Line-to-line coupling in the ISN may also be to blame, but it's hard to conclude based on this. I'll have to retest with a better method (perhaps using a shielded winding, or coupling into the common mode with a current choke approach).
And here's the common mode response (top trace; differential response stored for reference). Same setup, except noise source connected to the common mode input. You can see the isolation ratio is quite poor at high frequencies, even having some gain around 53 MHz. (Since the CM port is attenuated by 9.5 dB, while the differential port is not, this is not inexplicable.)
Overall, the shape (and amplitude; not shown here, but confirmed separately, sorry) is very close to the source's, so the common mode coupling, at least, seems to be very good. As long as the EUT and AE are happy with the signal quality through this network, I think the common mode response should be quite satisfactory for, actually fairly accurate, precompliance testing.