If you are considering this question chances are you are designing a radio receiver for a narrow band communications system. Crystal filters are sometimes used in transmitters as well, in a super heterodyne (superhet) transmit chain, to clean up the modulation spectrum and reduce adjacent channel noise. The same arguments apply here as for the receive case.
Achieving a narrow receive bandwidth
There are several options elaborated here.
At the wider end of the narrowband range SAW filters are usable and offer higher IF frequencies than crystal filters, so making it easier to achieve good image rejection in a superhet receiver. However for the narrower range, the available bandwidths are not small enough. There are two reasons for this. One is the available Q factor of a SAW resonator, abut 2k typical, whereas a crystal achieves typically 100k. The second is temperature stability. A crystal filter follows the AT cut crystal temperature drift characteristic so the shift in centre frequency can be kept within ±20ppm or better over the usual temperature ranges. Narrow SAW filters follow a XT cut characteristic similar to that of a tuning fork crystal. Over -40⁰C to +85⁰C this gives a potential -189ppm shift in the filter centre frequency. This has to be allowed for by making the passband wider. This in turn reduces the adjacent channel rejection achievable.
If your receiver has a wide tuning range over the HF and/or VHF spectrum then an up-conversion superhet architecture should be considered. Narrow SAW filters in the UHF range such as Golledge MA05334 (119.64MHz CF, 30kHz BW), or MP06905 (469.8MHz CF, 710kHz BW) make an excellent 1st choice. For this architecture the front end filter can be a low-pass greatly simplifying the front end design.
For a long time this was the preferred way of achieving tight selectivity and narrow channels. This comes at the expense of higher complexity, as often a double conversion architecture (with a crystal filter at the first IF) is required to get to the 455kHz IF while achieving good image rejection. The performance disadvantage of this approach is that placing the main selectivity further down the receiver chain compromises the blocking performance. The added complexity also brings extra current consumption and cost in the receiver circuits. This design approach is now out of favor, and ceramic filters are getting harder to buy, following the main supplier’s withdrawal from the market.
The low Q of practical inductors makes achieving narrow filters only feasible at low IF. The same problems as the ceramic filters then apply.
This approach has become increasingly popular in recent years with the availability of cheap ADC and DSP devices. High dynamic range and good adjacent channel rejection are possible. The image rejection problem goes away, but is replaced with the need to maintain good quadrature and amplitude control of the two LO and IF paths. Tighter control of filter shape is possible than with analogue techniques, making this suitable for complex signalling techniques that are not tolerant of variations in group delay characteristics. All this comes at the expense of complexity, cost, and the high power consumption of analogue to digital converters and the DSP. Two front end mixers are also required with quadrature LO feed to provide the I and Q signals needed to resolve positive and negative frequencies. There is also the difficulty in removing DC offsets, requiring either a notch in the centre of the receiver characteristic from having a high pass filter in the baseband signal chain, or periodic self calibration to remove the DC offset. As the IF is spread across low frequencies 1/f noise in the baseband circuits gives a poor IF noise figure. This limits the receiver sensitivity, or requires more frontend LNA gain to compensate. Having more LNA gain compromises the receiver dynamic range. Local oscillator leakage from the antenna port is also a problem as the LO is at the wanted frequency so cannot be removed by the frontend filters. Another gotcha is that if you have commutating mixers for the down conversion to achieve good dynamic range, then the front end filter needs to have good rejection at the 3rd harmonic, and above, as the mixers will mix down the odd harmonics very effectively giving the receiver spurious responses.
This is very similar to the Zero IF approach except that the A to D convertor is before the mix to Zero IF, and following lowpass filtering happens using digital processing. The advantage of this is that it eliminates the DC offset and baseband noise issues. The problem is that the very high sample rate and dynamic range required of the ADC makes this a very power hungry and expensive solution. The high data rate from the ADC also requires a FPGA or dedicated digital down converter to do the down conversion, low pass filtering, and decimation to produce a sample rate that a DSP chip can handle.
The superhet receiver architecture has a well established history. Crystal filters have the advantage of excellent signal handling range, low loss, tight selectivity, and being passive, no current consumption. They are also low cost. Advances in packaging mean that, unlike in the past, parts are available in modern surface mount packages that do not take up too much PCB area. Advances in manufacturing techniques mean that higher IF center frequencies are now possible making achieving good image rejection easier. 45MHz is now the new normal for crystal filter based IFs, and 70MHz is becoming increasingly common. Of course there are weaknesses, if a high shape factor is needed than a crystal filter with a large number of poles is required and hence a larger through hole package. The termination impedances are high, but BSL can help with this by suggesting suitable matching networks should working in a 50 ohm system be needed. There is some variation part to part and over temperature, but modern manufacturing process control mean these are minimized. Some of the highest performance narrow band receivers use a hybrid of crystal filters and digital filters to get the best of both worlds. The precision and sharp cut-off of the digital filter but allowing lower sample rate and needing fewer bits resolution in the ADC after the crystal filter, saving considerably on cost and power consumption.
The NanoVNA is tiny one-and-a-half port Vector Network Analyzer. As well as being very small it is also extremely cheap. I picked mine up for £35 on ebay. When it arrived, I was surprised and delighted to find it included a SMA (male) calibration kit, a SMA female back to back (a SMA barrel in RF engineering slang), and a pair of SMA cables. Normally a good calibration kit costs in the order of £500.
Given that much of a VNA’s performance relies on the quality of its calibration, I had a look at this very cheap calibration kit using a very much more expensive VNA, the LA Techniques LA19-13-03.
VNA calibration is performed using a technique called SOTL, which stands for Short Open Through Load. Lets take a look at each of these in turn.
Measurements of the short, open, and load are performed through the supplied SMA barrel. This provides a shift in the reference plane in addition to the variation in the reference plane of the Short or Open.
The return loss was as close to 0dB as can be accurately measured. What is more interesting is the group delay of S11. This tells us what the electrical position of the short circuit relative to the reference plane and how it varies across frequency.
Given that my NanoVNA only operates to 900MHz the wiggles above 6 GHz can be ignored. 92ps equates to an electrical length of 27.6mm. There is however a ~5ps variation over the first 400MHz equating to a 1.5mm shift in reference plane.
Again the return loss was negligible. The S11 group delay looks like this.
Reasonably consistent performance up to 900MHz but different to the short. This causes calibration errors if the calibration software does not alow for different electrical lengths for the Open and Short references. While expensive VNAs alow separate complex models for the calibration references, the NanoVNA does not alow separate reference offset parameters. 98ps equates to 29.4mm. 1.8mm different to the Short, or about 2° at 900MHz.
What maters here is staying as close to an ideal 50 ohms as possible over frequency with zero inductive or capacitive parts to the impedance. A scalar way of quantifying this is the magnitude of S11.
Perfectly adequate for use up to 900MHz.
What matters here is low insertion loss and no impedance mismatch.
Less than 0.05dB loss up to 900MHz counts as very low.
Here we see that the added mismatch due to the SMA barrel is negligible.
Good enough up to 900MHz as I’m not expecting super precision from a £35 VNA. I wouldn’t use this cal kit with an expensive piece of professional kit though.
For those who want to reproduce this or do their own visualization here are the S-parameters.
Not a good combination. Running what had been a 4 minute simulation in a regular directory turned into a 4 and a half hour run when I moved the project to a Dropbox folder.
Just a thought, does the number of programming languages in use now exceed the number of human languages spoken?
Working for Golledge gives me access to some great specialist kit. One example is the phase noise test set.
I bought one of the first batch of redpitayas by joining in the crowd funding of the development. Since I have not seen phase noise data appear elsewhere hopefully people will find this plot of use.
The above was measured at 50MHz.
The frequency setting of the redpitaya was slightly off 50MHz in order to get it as close to 50MHz as possible, and within the tuning range of the OCXO being used as a reference oscillator for this measurement.
It’s good to check simulations against measurement sometimes to ensure all is well. Here is a nice result 🙂
The blue traces are from direct measurement, and the red from measurement of smaller elements combined in simulation. Read on for more detail.
What I was checking today was two approaches to generating characteristic plots of a 4 pole crystal filter. The filter comes split into 2 separate 3-leg UM1 packages. Its a 15kHz wide (-3dB) filter centred on 55MHz. The termination impedance is 500 ohms in parallel with 2pF. The filter requires the user to provide a 10pF shunt capacitor at the junction of the two parts. This arrangement is a right pain in production, but common in the industry, fortunately now being superseded by advanced packaging techniques that can accommodate the two quartz blanks in a single surface mount package.
The method of measuring crystal filters I have been using recently is to measure the S-Parameters of the unmatched filter using the 450 ohm vector network analyser (VNA) I developed at Golledge. The 450 ohm normalised S-Parameter data can then be analysed in a circuit simulator with the termination impedances set as required for the crystal filter. 450 ohms is used in place of the conventional 50 ohms as it is closer to the working impedances of crystal filters, usually in the range 300 ohms to 3k ohms. Measurements in 50 ohms are not accurate enough.
What I was comparing today was approach 2) measuring the two packages of the crystal filter separately then combining them along with an ideal centre coupling capacitor, with approach 1) measuring the two packages connected with a ceramic centre coupling capacitor as a single filter.
Here is the simulation running both approaches
The separate package approach has benefits when measuring the stopband as test fixture stopband leakage is less of a problem.
For those who want to look further into this here are the S-parameters and QucsStudio files.
Our website has been secured. https is now the default, enhancing your privacy a little bit extra.
I have just read a fascinating article entitled “The robots won’t take over because they couldn’t care less” by Professor Margaret Boden. It’s a well researched and thought out piece that also has a fascinating diversion into Japanese culture along the way. Well worth a read.