How Edge Processing Is Enabling Next-Gen Millimeter Wave Scanners


Millimeter wave (mmWave) imaging has become an important part of security scanning systems in airports, public buildings, and stadia. Millimeter wave scanners are superior to traditional metal detectors because they can identify and locate both metallic and nonmetallic threats. This article describes how mmWave imaging hardware works and will present a chipset that uses edge processing to manage massive data loads to enable the development of walkthrough security scanning systems.

How mmWave Imaging Works

Figure 1 illustrates how a mmWave scanner operates. The system consists of an array of transmitters and receivers connected to a spatially dispersed antenna array. The system is analogous to a network analyzer that is measuring return loss or S11. At any one time, one antenna in the array is transmitting a low power signal at a single frequency. This signal reflects off the target and generates backscatter (the illustration shows reflection from a single point on the target but in practice the transmitted signal is omnidirectional, so there will be reflections from multiple points on the target).

The phase and amplitude of the backscatter are measured by all of the receive antennas in the array. Polarization may be used between the transmit and receive antennas to reduce direct transmit-to-receive leakage. Once this measurement is complete, the same signal is transmitted from another transmit antenna (operating at the same frequency) and the measurement process is repeated.

Figure 1. Operation of a mmWave security imaging system.

Because the depth of penetration of RF signals and the nature of the reflection vary with frequency, the scan described earlier is typically repeated at multiple frequencies, across a wide band. The resulting matrix of vectors forms a multidimensional array (vs. frequency and spatial location) that is used to create an image that can identify metallic and nonmetallic objects that are concealed between and under layers of clothing.

The hardware required to complete such a scan must be multichannel and have a wide operating frequency range. The 10 GHz to 40 GHz frequency range is wide enough to differentiate the objects in a typical security scanning scenario (clothing, backpacks, weapons, and explosives). Higher channel count systems tend to have higher resolution, giving them the ability to identify small objects. For example, while detecting a razor blade is critical in airport scanners, securing public buildings and stadia focuses more on the detection of larger items such as weapons or explosives. In these applications, a lower channel count is typically used.

Another critical component in these systems is fast switching time. This enables the realization of scanning systems where the person being scanned only has to pose for a short amount of time (typically one second or less). Next-generation walkthrough systems require faster switching times so that the person does not have to stop and pose.

Figure 2 shows how an Analog Devices mmWave imaging chipset can be used to implement a complete mmWave scanner. An array of transmitters (ADAR2001) is driven from a central agile frequency source. An array of receivers (ADAR2004) detects the reflected signals and downconverts them to a low intermediate frequency at which they are IF sampled by a multichannel continuous-time sigma-delta (CTSD) converter (AD9083).

We will now take a closer look at these components and how their functionality optimizes overall system performance.


As already noted, the transmitter consists of a large channel count of spatially dispersed antennas, each driven by a power amplifier. The ADAR2001 is a 4-channel transmitter that connects directly to the antennas and has an output frequency range of 10 GHz to 40 GHz. Because of the difficulty associated with distributing a 10 GHz to 40 GHz signal in a large array, the ADAR2001 incorporates a 4× multiplier. As a result, all the plumbing and signal distribution in front of the transmitter ICs happens in the 2.5 GHz to 10 GHz frequency range.

The main RF elements of the ADAR2001 transmitter are an RF input buffer, a 4× frequency multiplier with integrated switchable harmonic filters, a 1:4 signal splitter, and four differential-out power amplifiers, which are intended to drive differential antenna structures such as dipole or spiral antennas. A detailed block diagram of the ADAR2001 is shown in Figure 3.

A CW RF input signal between 2.5 GHz to 10 GHz and with a power level of at least –20 dBm is applied to the RFIN port. The broadband frequency multiplier consists of three parallel subcircuits. Each subcircuit (low band, mid band, high band) is optimized to multiply and filter a segment of the total frequency range. Switches at the input and output of the multiplier block are used to select the subcircuit for the desired frequency of operation.

The multiplier output passes through a programmable attenuator before being split into four and applied to four power amplifiers. In addition to the configurable filtering in the multiplier block, each PA contains a low-pass/notch filter that can be enabled or disabled. For output frequencies up to 20 GHz, this filter should be enabled. Above 20 GHz, it should be disabled.

The programmable attenuator is used to help ensure relatively flat output power vs. frequency. This attenuator has approximately 15 dB of digital step attenuation range. As the output frequency sweeps from 10 GHz to 40 GHz, this attenuation should be decreased to maintain the desired output power flatness vs. frequency. This results in a nominal PA output power of +5 dBm on each of the differential PA outputs with harmonic suppression that ranges from –20 dBc to –30 dBc.

To do a complete 10 GHz to 40 GHz frequency sweep, the multiplier/filter block settings must be adjusted seven times to ensure optimum harmonic rejection and output power. In addition, while the system is dwelling at one frequency, each

Figure 2. A complete mmWave imaging system.
Figure 3. The ADAR2001 10 GHz to 40 GHz transmitter.

transmitter channel must be successively turned on and off. To avoid creating a bottleneck of SPI commands, the ADAR2001 includes two state machines that can be preprogrammed with up to 70 states. Once the device’s RAM has been programmed, state advances can be made with a simple pulse to the device’s MADV (advance) pin. These features combine to ensure a 2 ns channel switching time. This switching time is also achievable when switching between ICs (for example, Channel 4 of Device A is switching off as Channel 1 of Device B is switching on). Because a full scan involves a full-channel sweep at multiple frequencies, switching time is critical. For example, if the array has 500 elements and is going to sweep from 10 GHz to 40 GHz in 50 MHz steps, it must perform a total of 300,000 channel switches to complete a full scan.

The RF output power on each channel can be monitored using individual, on-chip RF detectors. Die temperature can also be monitored by an on-chip temperature sensor. These sensors feed into a 5:1 analog multiplexer, which passes the desired signal to an on-chip 8-bit ADC.

The ADF4368 PLL/VCO provides the stimulus to the transmitter network. Its output signal will be split multiple times depending on the number of transmit channels. The ADF4368’s relatively high output power of +9 dBm and the minimum input threshold of the ADAR2001 (–20 dBm) ensure that the ADF4368’s output can be passively split many times before amplifier buffering is required.


The reflections from the transmission are picked up by the receivers, which are an array of multichannel mixers and ADCs. The ADAR2004 is a quad mixer and ADC driver with a digitally programmed gain amplifier (DGA). The LO input, which also has an internal 4× multiplier, is driven by a second PLL whose output frequency is offset from the radio frequency so that the mixer produces a real IF output. The IF outputs of the mixers are then sampled by the AD9083, a 16-channel continuous-time sigma-delta ADC with integrated digital downconversion. An IF sampling architecture was chosen over a zero-IF architecture to avoid DC offsets that result from LO leakage in the receiver and I/Q errors that result from imperfect quadrature balance in the LO’s quadrature splitter. While these imperfections can be mitigated by calibration, calibration would be required at every input frequency because LO leakage and quadrature errors tend to vary with frequency.

Figure 4 shows a block diagram of the ADAR2004 quad mixer. The LO input is driven by a 2.5 GHz to 10 GHz sine wave that produces a 10 GHz to 40 GHz at the output of the multiplier. The multiplier output is fed to the four mixers that have a programmable gain on their IF outputs. Like the ADAR2001 transmitter, the ADAR2004 receiver also has two on-chip state machines that can be preprogrammed.

Multichannel ADC

Figure 5 shows a block diagram of the AD9083, a 16-channel CTSD ADC. The ADC inputs are designed to have the same common-mode voltage as the IF outputs of the ADAR2004. This allows the mixer output and the ADC input to be connected directly. The absence of AC-coupling capacitors ensures that there are no charging/discharging transients when the mixer output switches abruptly (for example, during a frequency step at the input of the mixer).

The use of a first-order CTSD ADC architecture with an integrated single-pole filter saves PCB space by minimizing external filtering. The architecture also enables a fast signal settling time when compared to the settling time of Nyquist rate converters, which require highly selective antialiasing filters to eliminate noise folding. Fast settling time is a key requirement in this application because the ADC settling time must be able to keep up with the fast channel switching on the transmit side.

Figure 4. The ADAR2004 10 GHz to 40 GHz receiver block diagram.

Each ADC has a signal processing tile to filter out-of-band, shaped noise from the sigma-delta ADC and reduce the sample rate. Each tile contains a cascaded integrator comb (CIC) filter, a quadrature digital downconverter (DDC) with multiple finite input response (FIR) decimation filters (decimate by J block), or up to three quadrature DDC channels with averaging decimation filters for data gating applications. The presence of three quadrature DDC channels enables simultaneous demodulation of up to three frequencies. We will see later how this can be used to dramatically speed up the scan time.

System Setup and Operation

The ADAR2001 and ADAR2004 were specifically designed for efficient operation in large arrays. Particular emphasis was placed on reducing wiring overhead. The RFIN and LO input ports of ADAR2001 and ADAR2004 can operate at input levels as low as –20 dBm. Because it is desirable to drive these inputs from a common LO source (the ADF4368 in this case), this low input sensitivity allows lots of passive fanout before amplification is required. For example, if we assume that a Wilkinson power splitter has a net loss of 1 dB, the ADF4368’s output power of 9 dBm can be passively fanned out seven times and can drive 128 devices (512 channels).

The advance and reset pins that drive the ADAR2001 and ADAR2004’s on-chip sequencers are also designed to be driven in parallel to minimize the number of GPIOs the processor or FPGA must provide. By providing enough depth and complexity in the sequencers, it is possible to drive up to 16 ADAR2001 devices with a single set of advance and reset pulses.

Before operation, the ADAR2001 and ADAR2004’s sequencers must be programmed. While it is possible to access all of the functionality of both devices using SPI commands, the associated latency would result in an unacceptably long overall scan time.

Let’s consider how to set up a 64-channel system (64-transmitter, 64-receiver) for a channel-based scan—that is, we cycle through all of the transmit channels at a single frequency before incrementing the frequency and repeating the scan.

Figure 6 shows how the state machines in the 16 ADAR2001 devices are programmed to enable this sweep. A key goal of the architecture is to be able to sequence multiple devices that are doing different things from common control lines.

Notice in Figure 6 that while each IC has 65 states, most of the ICs are programmed to be in sleep mode (SLP) most of the time. For example, IC 1 is only fully active for the first four states as channels 1, 2, 3, and 4 of that IC sequentially transmit. During these four states, all the other ICs are either in SLP or ready (RDY) mode.

Figure 5. The AD9083 block diagram.

Likewise, IC 2 is only fully active during states 5 to 8 as all of the other ICs are either in SLP or RDY mode. By configuring the 16 state machines in this manner with their on cycles offset from one another, it is possible to drive the advance and reset lines of all 16 devices with parallel pulses.

The RDY mode is an intermediate state of what was developed to optimize switching time while saving power. Because most of the transmitters are inactive most of the time, the SLP mode is key to keeping down power consumption. However, the time required to switch from SLP mode to transmit mode (50 ns) is excessive from a system perspective and would result in delays during the scan. The RDY mode is an intermediate state that can be invoked when an IC is preparing to transmit. Notice in Figure 6 that in State 4, Channel 4 of IC 1 is transmitting, and IC 2 is being prepared for transmission by placing it in RDY mode. In the transition from transmitter states 4 to 5, IC 1 transitions from transmit mode to RDY mode, and IC 2 transitions from RDY to transmit mode. This transition takes 10 ns. The subsequent on-chip channel switches (that is, from Channel 1 to Channel 2 to Channel 3 to Channel 4 on IC 2) have a switching time of 2 ns. For a 1024-element array that sweeps from 10 GHz to 40 GHz in 0.1 GHz steps, the complete scan time would be less than 20 ms. This assumes a PLL lock time of 50 μs. If two PLLs operating in ping-pong mode were used to achieve faster frequency settling, the scan time would be well below 5 ms.

Figure 6. Programming 16 ADAR2001 transmitters for a channel sweep that is driven by a single advance pulse.

The operation and sequencing of the ADAR2004 receiver are less complex because it is typical practice to configure all the receiver channels to be receiving all the time. The state machines must still be sequenced so that the correct multiplier path and filter settings are chosen as the receiver sweeps in tandem with the transmitter.

As already noted, each AD9083 ADC channel can access up to three quadrature DDC channels. This means that it can simultaneously demodulate three frequencies, assuming that all three frequencies are within the input frequency range of the ADC’s analog input bandwidth (125 MHz). For example, by positioning three IF tones at 50 MHz, 75 MHz, and 100 MHz, all three can be simultaneously demodulated into I and Q baseband data.

To facilitate this approach on the transmit side, three transmit PLLs must be used instead of one. The three transmit frequencies must always be directed to different physical transmit ICs (the multipliers in the ADAR2001 cannot conduct multitone signals). The three frequencies must always be different but must remain close in frequency to one another as they sweep. For example, if one channel on one of the ADAR2001 devices is transmitting at 10 GHz, two other devices will be transmitting at 10.025 GHz and 10.050 GHz to support IF outputs at 50 MHz, 75 MHz, and 100 MHz. This scheme requires more hardware and switching infrastructure in the transmit path but has the benefit of reducing the overall scan time by a factor of 3.


The chipset consisting of the ADAR2001 quad transmitter, the ADAR2004 quad receiver, the AD9083 16-channel ADC, and the ADF4368 PLL/VCO provides the high level of integration and functionality that is necessary to implement next-generation walkthrough mmWave security scanners. Integrated state machines and on-chip digital downconversion significantly offload traditional centralized processing and move it to the Intelligent Edge. The net result is that the central processor has to worry less about controlling the system during a scan and the data that it receives is already demodulated and decimated. While this chipset was developed specifically for mmWave security imaging applications, the wide frequency range of the ADAR2001 transmitter and ADAR2004 receiver as well as the level of integration of the AD9083 16-channel ADC, make this chipset useful in other applications where high channel density and fast switching are required.


Eamon Nash

Eamon Nash

Eamon Nash是ADI公司RF产品部的应用工程经理。他已在ADI公司工作了20多年,最初他是一名混合信号和DSP产品现场应用工程师,后来成为无线应用领域分立式RF器件应用工程师。