Matt Johnson

Human Compiler


High-performance, low-energy computer systems from the ground up.

Web Links

Blog Work HN

EEWeb Stats

Matt's Projects :

Return to Projects

Underwater Acoustic Pinger Locator

Project Summary

In the field of autonomous underwater robotics, it is often desirable, and indeed necessary, to "ground truth" the position of the robot to a known control point, by detecting and/or navigating to a beacon of some sort. The beacon is necessary because conventional means of correlating on-board position data with a known global coordinate system, such as GPS, do not work well in an underwater environment. One popular type of beacon is an acoustic pinger, which uses a transducer to transmit a relatively high-intensity pressure wave above 20kHz. In this paper, we propose an algorithm and hardware/software implementation for receiving this signal via an ultra-short baseline (\textit{USBL})~array of hydrophones, and calculating a bearing to the pinger. This bearing information can be used to navigate to the pinger, or gain additional knowledge about the vehicle's position.

Project Description

In this paper, we will address all aspects of the conception, design, implementation, and testing of the apparatus, which we will call \textit{APLS}, or the Acoustic Pinger Location System. First, we will elucidate the theory behind our approach to pinger location. Next, we will enumerate the high-level design requirements and constraints which shaped the system as a whole. This discussion will be followed by a description of the iterative process which yielded a finished hardware and software design. Next, we will discuss the process of constructing and debugging the system after the basic design was complete. Finally, we will characterize the finished system, assess the extent to which it met the original design goals, and provide a post-mortem view of opportunities for improvement, not only in the next revision of APLS, but future design projects in general.

Theory Of Operation

It is possible to determine the bearing angle to the beacon by measuring the pairwise differences in the time of signal receipt at the hydrophones in the array. A noncollinear array of three hydrophones is needed to uniquely determine the bearing. A fourth, noncoplanar hydrophone allows for the further calculation of the angle of elevation to the beacon.

The plane wave approximation is a useful and reasonable one to make for the purposes of bearing angle calculation. It is valid if the distance between the array and the beacon is large with respect to the dimensions of the array; in practice, this constraint is nearly always true because the size of the array is constrained for reasons explained later.

It is also assumed in this analysis that the speed of sound in water is constant (but not necessarily known). This simplification is reasonable if the density and temperature of the water in the region of interest are roughly constant.

Refer to Figure {geometry of the hydrophone array} for a diagram of an equilateral hydrophone array. In this figure, a signal from the beacon, approximated as a plane wave, approaches from the trigonometric angle θ. Right-triangle trigonometry can be used to determine relationships between the pairwise time differences and θ. The two triangles of interest are denoted by solid lines. One side length of each is known; it is the distance between two hydrophones in the array. Another side of each is given by the distance traveled by the wave during a pairwise time difference. An unknown angle related to the desired quantity θ also appears in each triangle. Writing trigonometric equations about these unknown angles, it is possible to determine a closed-form expression for θ in terms of the time differences t_3-t_1 and t_3-t_2, the ratio of which we define to be ρ.

Figure:1 geometry of the hydrophone array

Figure 1  geometry of the hydrophone array

(1)\; \; cos(90 \degree – \theta) &= sin(\theta) = \frac{v(t_{3}-t_{2})}{d}

(2)\;\; cos(\theta – 30\degree) &= \frac{v(t_3-t_1)}{d}

(3)\; \; \rho \equiv \frac{t_3-t_1}{t_3-t_2} &= \frac{cos(\theta – 30\degree)}{sin(\theta)}

(3)\; \; \rho &= \frac{cos(\theta) cos(30\degree) – sin(\theta) sin(30\degree)}{sin(\theta)}

(4)\; \; \rho &= \frac{\sqrt{3}}{2} cot(\theta) + \frac{1}{2}

(5)\; \; cot(\theta) &= \frac{2\rho-1}{\sqrt{3}}

(6) \; \; \theta &= tan^{-1} \left( \frac{\sqrt{3}}{2\rho-1} \right)

Strictly speaking, the above expression is only valid for -π/2 < θ < π/2. However, it is easily extended to the domains -π < θ < -π/2 and π/2 < θ < π by noting that (for -π < θ < π) θ must have the same sign as t_3-t_2: if t_3>t_2, then the beacon lies above the x axis and θ is positive; likewise, if t_2>t_3, the beacon lies below the x axis and θ is negative. Thus, if the signs of t_3-t_2 and θ differ, adding (or, equivalently, subtracting) π from θ provides the desired result. In addition, one need address the following corner cases:

  • ρ \Rightarrow 1/2: we consider the arctan function to be defined at \pm \infty as \pm π/2 for the purposes of this analysis, so for ρ=1/2, θ is defined to be π/2 when t_3>t_2 and -π/2 when t_2>t_3.
  • ρ \Rightarrow \pm \infty: in the event that t_3=t_$, the beacon clearly lies on the x axis, so for ρ \rightarrow \pm \infty, we define θ to be 0 if t_3>t_1 and π if t_1>t_3.
  • t_3=t_2=t_1: clearly, this equality is only possible if the source of the ping is the center of the hydrophone array. We arbitrarily define θ to be 0 in this case.

Thus, the trigonometric bearing angle θ is defined fully:

\theta = \left\{\begin{array}{cl}tan^{1} \left( \frac{\sqrt{3}}{2\rho-1}\right) & \mbox{if } t_3-t_1 > \frac{1}{2}(t_3-t_2) \mbox{ and } t_3-t_2 \neq0 \\tan^{-1} \left( \frac{\sqrt{3}}{2\rho-1} \right) + \pi & \mbox{if }t_3-t_1 < \frac{1}{2}(t_3-t_2) \mbox{ and } t_3-t_2 \neq 0 \\\frac{\pi}{2} &\mbox{if } t_3=t_2 \mbox{ and } t_3>t_1 \\\frac{\pi}{2} & \mbox{if } t_3=t_2\mbox{ and } t_1>t_3 \\0 & \mbox{if } t_3-t_1 = \frac{1}{2}(t_3-t_2) \mbox{and } t_3 \geq t_1 \\\pi & \mbox{if } t_3-t_1 = \frac{1}{2}(t_3-t_2) \mbox{and } t_1>t_3 \\\end{array}\right.

Determining the pairwise time differences is a nontrivial task in itself. Ideally, a measurement should be taken every period of the wave to minimize error. To achieve this end, the hydrophones must be arranged such that less than half of a period of the wave may elapse between receipt at any pair of hydrophones. Otherwise, measurements would be ambiguous: unless the beginning of a burst were observed, it would be impossible to determine the period to which a measurement corresponded, so pairs of receipts could not be matched. This solution is also more robust: accurate measurements are still possible if the beginning of a burst is missed or if the signal is not bursted but is instead continuous. Thus, the distance d between two hydrophones is constrained by the acoustic frequency f and speed of sound v as such:

d < \frac{v}{2f}

The resulting hydrophone configuration is known as an ultra-short baseline (USBL) array.

Given this constraint, it is possible to determine the pairwise time differences by observing single periods of the wave. One method for obtaining these time differences is to compare the hydrophones’ output voltage waveforms to a constant threshold value. The times of positive crossings of this threshold are then recorded for each hydrophone. The condition for a successful measurement is signal receipt at all three hydrophones satisfying the following constraint:

max \left\{ |t_2-t_1|, |t_3-t_1|, |t_3-t_2| \right\} < \frac{1}{2f}

As a practical note, the accuracy of this method depends heavily on the resolution of the time measurements. Table 1 provides statistics on the angular error as a function of sample rate, assuming that the hydrophones are in a USBL configuration, separated by 2 cm. It is clear that in order to have reasonably accurate angle measurements, resolution better than 1 µs (i.e. a sample rate of more than 1 MHz) is necessary. Furthermore, this analysis is optimistic since it does not consider other sources of error; as a result, the need for a high sampling rate is clear.

Figure:2 expected error table

Figure 2  expected error table

High-Level System Design

Now that we have described the signals of interest we will be provided by the hydrophones, we must design a means of converting this signal into something useful to a computer, computing a bearing angle, and transmitting the computed angle to a central computer via a standard protocol. With only this constraint, the design space is quite large. We need to further constrain the design to narrow the possibilities to a workable system.

One important consideration for any embedded system is form factor. For a system to be truly useful, its physical dimensions and configuration must mesh well with its operating conditions. This is particularly true in autonomous underwater vehicles, where all electronics are typically housed in an air-tight tube, whose price and vehicle performance cost (in terms of drag) increase dramatically with size. This means that APLS should be small, and should integrate easily with the rest of the electronics in the AUV in a physical sense. One computing form factor that has become extremely popular in the underwater space is PC/104. This specification, itself an extension of IEEE-P996, specifies a bus which connects multiple cards which, due to the common bus specification, can easily be stacked on top of one another (thus coining the term ``PC/104 stack’‘). In addition to the bus, each card in the stack shares common mechanical dimensions for easy stacking, and each card has 4 mounting holes at predefined positions in order to enable common standoffs to be inserted between each pair of cards, to give the stack mechanical rigidity. In order to be physically compatible with a wide range of autonomous underwater vehicle, we decided to house our entire system on a single PC/104-size card. However, since the data communication between APLS and the main computer will be minimal and mostly unidirectional, the large, high-pin-count PC/104 connectors would be overkill, and would occupy a large percentage of the already small (under 15 square inches) available board area. Therefore, we chose to adopt the PC/104 form factor, without using the PC/104 bus. Therefore, APLS can be connected to \textit{any}~existing PC/104 stack without complicating such considerations as stack position and capacitive bus loading, and all data communication can take place over a commodity cable, rather than the PC/104 bus. Our system will be implemented on a custom-designed printed circuit board (PCB).

It became apparent early in the process that our application would require a fair amount of specialized logic to capture the difference in arrival times between all 3 hydrophone signals. After an extensive survey of available embedded processors, none were found that provided the flexibility and customizability we desired. In order to preserve the ability to modify this logic without the need for a PCB redesign, it became clear that a field-programmable gate array (\textit{FPGA}) or complex programmable logic device (\textit{CPLD}) would be an ideal solution. The high dynamic power consumption of these devices was a concern, but a survey of popular FPGA and CPLD families (from industry leaders such as Xilinx, Altera, Lattice Semiconductor, and Atmel) found that although FPGAs consume substantially more power than their ASIC equivalents, when available, even this elevated power consumption is well under 1 Watt, which means that either solution would pale in comparison to some of the other power requirements on an AUV (thrusters, CPU, high-power solenoids and relays, resistive losses in DC/DC converters). At this point, the two alternatives we considered were a completely FPGA-based design, and a CPLD (loaded with timing information capture logic) with a microcontroller handling the processing and communication. At first, we considering designing the FPGA or CPLD/microcontroller combination into our board, but we quickly found that the extremely fine pin/ball pitch and large pin/ball count of these devices would pose major problems in the assembly of our board. Specifically, any FPGA with a gate count sufficient to encompass timing capture, processing, and communication has hundreds of pins and hundreds of pages of PCB design constraints, and attempting to design a board with an integrated FPGA or complex, high-pin-count microcontroller, much less attempting to solder on such a component with the limited toolset available to us, would provide little value to the finished system, while drastically increasing the complexity of our design. We next shifted our focus to finding a commercially-available PCB with an integrated programmable logic solution, which we could somehow connect to our board, which would ostensibly contain all necessary support logic and analog processing hardware.

One potential complication of an FPGA-only design is that designing a relatively general-purpose processing solution using the traditional design entry modes available to FPGA designers (HDL and schematics) is an extremely cumbersome process, particularly given the time constraint present for this project. One solution to this problem is to use a publically-available \textit{soft core}, and embed it within the FPGA. A survey of such cores yielded a wide range of results. Websites dedicated to open hardware design, such as \textit{}, had a variety of cores available, ranging from the 8051 to the 8080, from a simple VLIW superscalar processor to MIPS to JOP, a core optimized to run (stack-based) Java bytecode. These cores were completed and documented to wildly varying extents, and in the end, the total package, including correct, commented RTL, clear documentation, and software development toolsets were lacking for these openly available cores. Given the severe time constraint, we needed to find a more robust, commercially-supported soft core to include in an FPGA, if this was to be a viable option. The two major FPGA vendors, Altera and Xilinx, each offer something along these lines. Their respective solutions are known as NIOS/NIOS II and Picoblaze/MicroBlaze, respectively. The MicroBlaze was the most attractive of these, due to the extensive documentation, a fully ported GCC-based C compiler/assembler/linker toolset, and an IDE integrated with the main synthesis tool (ISE). The NIOS II had many of these same features, but it was found that the majority of commercially-available small-form-factor FPGA boards were Xilinx-based. In the end, we decided to search for a Xilinx Spartan-3-based board. The Spartan-3 FPGA family provided a good compromise of gate count (up to 1 million equivalent ASIC gates available), cost, and power consumption, with the ability to contain a full MicroBlaze processor (including hardware FPU) plus a considerable amount of custom logic.

Small Spartan-3-based boards were available from several vendors, including Xilinx (, Opal Kelly (, and XESS ( The main distinguishing characteristics of these boards were available documentation, on-board logic and interfaces, cost, and ease of interface with a second board. All vendors had adequate documentation, but they varied in terms of how clearly they described the signals entering and leaving the board, and the requirements thereof (voltage, timing, etc.). The boards varied a great deal in complexity; the majority had USB interfaces via an external controller, whereas some had full 100BaseT Ethernet MACs and PHYs. The boards ranged in cost from approximately 100USD to over 400USD. Additionally, some boards appeared to be relatively painless to interface with a second board, with large connectors on the underside of the board, hardwired to FPGA and other pins. Others had no such conveniences, meaning that information would have to be passed between our analog processing board and the FPGA board via a cable of some sort. In the end, the latter solution seemed much less robust, and certainly less elegant. In the end, we decided on a board from Trenz Electronic (\textit{Trenz Electronic}), the TE0140-01. This board satisfied all the design requirements previously specified. In addition to being reasonably priced (99 Euros $\approx$~130 USD at the time of purchase), the board has a UTMI interface to a USB 2.0 (high-speed) interface, two 80-pin connectors to connect to a baseboard, and adequate documentation, including schematics. This board was available in a $400,000$-gate version and a $1,000,000$-gate version. We chose the former, deciding that the increased cost of the larger FPGA did not add significant value to the end system, given our system-level requirements. Additionally, the power consumption of the larger FPGA would be much larger.

The end result of this process was the decision to design a PC/104 form factor PCB containing analog processing and support logic, and board-to-board connectors to connect to a Trenz Electronic TE0140-01, which would house our processing solution, a Xilinx XC3S400FT256-4. With this board, we have the option of communicating via USB via the built-in USB controller and PHY, although another protocol may prove more attractive, due to lower complexity.

The end result of this process was the decision to design a PC/104 form factor PCB containing analog processing and support logic, and board-to-board connectors to connect to a Trenz Electronic TE0140-01, which would house our processing solution, a Xilinx XC3S400FT256-4. With this board, we have the option of communicating via USB via the built-in USB controller and PHY, although another protocol may prove more attractive, due to lower complexity.

Component Selection

Signal Conditioning
High-Level Topology

Once the system topology, form factor, and processing solution had been decided upon, the focus shifted to the analog elements of our design. Specifically, we needed to design a means of converting the signal from the hydrophones, a high-output-impedance analog waveform, into something useful for the FPGA. There were two competing strategies for this part of the design. The first alternative was to buffer and amplify the signals from the hydrophones, and simply pass them into a high-speed, high-precision analog-to-digital converter (ADC). The FPGA could obtain samples from the ADC, and perform digital signal processing (DSP) on them at its leisure. The second strategy was to replace the ADC with a thresholding mechanism, which would essentially transform the hydrophone signal into a one-bit digital signal, representing whether or not the hydrophone signal exceeded some reference voltage. The DSP solution provided slightly more flexibility, in that more information would be captured by the FPGA, and could thus be used in more ways than a single-bit signal. However, the second analog paradigm was much simpler from a software point of view, assuming we could design a properly-functioning amplification, filtering, and digitization solution. We were confident that our algorithm could match the performance of existing DSP-based solutions, most of which make use of the Fast Fourier Transform (FFT)~to calculate the phase difference between received signals, with a much lower hardware cost and software complexity.


The next step in the design process was to decide on a specific sequence of components for amplification, filtering, and digitization. For amplification, a low-noise, low-input-current operation amplifier was desirable. The component that best met these requirements was the AD743, from Analog Devices ( This part offers excellent noise performance (2.9 nV/vHz at 10kHz), a low input offset voltage (maximum 500µV), and a flat frequency response in our region of interest, 20 kHz to 35 kHz (4.5 MHz unity-gain bandwidth). The part was available in an 8-pin DIP package, as well as a 16-pin SOIC package. The surface-mount SOIC part was actually not much smaller than the DIP part, and as we will show, the DIP package was advantageous to us for two reasons. First, this package is directly compatible with solderless breadboard, offering easy prototyping and testing. Also, in a PCB design, DIP components can be useful if signals need to cross layers, since each pin requires a via anyway. For these reasons, we chose the DIP-8 package (Part Number AD743JN). The pinout for this package is shown in Figure (Analog Devices).

Figure:3 Analog Devices

Figure 3  Analog Devices

After selecting the operational amplifier itself, the surrounding passive circuitry was designed, tested, and iterated to an amicable solution. At first, we used the AD743 op-amp in a simple non-inverting configuration. The hydrophone was connected to the non-inverting node of the op-amp. Since the hydrophone is modeled by a capacitor, it acts as an open circuit when it is not receiving a signal; so, to prevent drift, the non-inverting node was pulled down to ground with a resistor large enough not to affect the hydrophone’s response but small compared to the op-amp’s input impedance. A 680 KΩ resistor was chosen for this purpose. Initially, a 100 Ω resistor grounded the inverting node, and a 10 KΩ potentiometer connected the output of the op-amp to its inverting input to provide an adjustable gain of up to 100. During testing, the need for a larger gain arose, so the 100 Ω resistor was reduced to 50 Ω.

It became apparent that because of the rather large gain needed for an output amplitude on the order of volts, non-ideal op-amp characteristics resulted in a large DC output component. This effect was unacceptable because the asymmetry in the output waveform greatly reduced the gain at which clipping began to occur. To address this issue, adjustments were first made to the pull-down resistor to limit the effects of input bias current, but it quickly became evident this was not the problem: even setting the resistor to the parallel combination of the gain resistors, which effectively eliminates input bias current, did not have a significant effect on the DC component of the output voltage. In fact, the input offset voltage caused this artifact: the nominal value of 1 mV becomes nontrivial when amplified by several hundred. To eliminate the input offset voltage, we adjusted the potential at the inverting node until the DC output voltage was zero. To do so without affecting the negative feedback configuration of the op-amp, a potentiometer connected between the 5 V rail and ground was used as an adjustable voltage source, and the wiper was connected to the inverting node via a large (2 MΩ) resistor. This proved to be an effective field-configurable solution to the input offset voltage problem, and it did not have any appreciable effect on the gain.

Finally, we observed that our early choices for gain resistors of 50 Ω and a 10 KΩ potentiometer resulted in a much noisier output than larger resistors with similar ratios. We replaced them with a 20 KΩ resistor and a 5 MΩ potentiometer, which reduced the noise dramatically and provided more headroom in case a larger gain was desired at some point.

The final configuration of the AD743 is shown in Figure (AD743 Final Configuration)

Figure:4 AD743 Final Configuration

Figure 4  AD743 Final Configuration


The next component in the signal processing chain was the filter. The underwater environment is a very noisy one, and without filtering, the acoustic pinger signal could potentially be dwarfed by, or at least mired in, external noise sources, such as ship turbines, animal communication, and pingers at other frequencies. In order to filter out energy at unwanted frequencies, we chose to use a bandpass filter centered around the pinger’s frequency, with a bandwidth of 1 kHz. This bandwidth value was chosen to make sure that the pinger’s frequency would be included in the passband, even if its frequency is only specified to a precision of $\pm 500$~Hz, as many commercial pingers are. One significant challenge posed by the design criteria is that the center frequency of the bandpass filter must be variable across the range 20-35 kHz, in order to be able to track pingers at any frequency. This is where conventional RLC and active filters fail; in order to change the center frequency, many passives would have to be changed, which means a great deal of desoldering and resoldering, which is unacceptable not only from a convenience standpoint, but also for the longevity of the PCB. After some thought, a solution that came to mind was a switched-capacitor filter. Though somewhat rare in many applications, this type of filter satisfies the requirement of variable frequency. Instead of using hardwired passive components to perform the filtering, switched-capacitor filters use an input clock signal to charge and discharge external capacitors, and, in conjunction with external resistors, the input clock frequency determines the filter type (lowpass, bandpass, highpass, notch, allpass), gain, quality factor (Q), and center frequency. There are relatively few switched-capacitor filters available in discrete form. The most suitable one for our application was the LMF100, manufactured by National Semiconductor. Each LMF100 contains two filters, each of which can implement a second-order filter of any of the five types mentioned above. Indeed, one can cascade both filter elements to form a single fourth-order filter. This part is available in a 20-pin DIP package, as well as a 20-pin Wide SOIC package. We chose to purchase three DIP package LMF100s for testing purposes, but decided that the footprint was too large for our area-constrained PCB, so we chose to use the 20-WSOIC (Part Number LMF100CIWM) package in the final implementation. The pinout of this part is shown in Figure (LMF100 20-WSOIC Pinout).

Figure:5 LMF100-20-WSOIC Pinout

Figure 5  LMF100-20-WSOIC Pinout

An additional design constraint was that each hydrophone processing channel should be placed and routed on the PCB as uniformly as possible, to ensure a uniform response. To satisfy this requirement, we chose to use one LMF100 for each of our 3 channels, rather than use 2 LMF100s, one with 2 channels and the other with one. This meant that we could use the aforementioned cascading of second-order filters to implement a fourth-order filter with a much sharper cutoff, yielding a potentially cleaner output signal through stronger rejection of unwanted energy.

The LMF100 offers several modes of operation, set by tying certain control pins in configurations specified in the manufacturer’s datasheet. The simplest mode (fewest external components) which offered the ability to affect the center frequency by modifying resistor values was Mode 3. The basic schematic for Mode 3 filters is shown in Figure (LMF100 Mode 3).

The center frequency’s dependence on external resistor values was important due to the method by which one synthesizes a fourth-order bandpass filter from second-order filters. The basic idea is to cascade two filters with center frequencies slightly above and slightly below the desired fourth-order center frequency, with slightly higher passband gains and higher Q. Multiplying their frequency responses (equivalent to convolution in the time domain, which is equivalent to cascading the filters) yields the desired fourth-order bandpass response. With the ability to shift the center frequency upward and downward with resistors, we could use the same input clock for both second-order filter elements, an important cost-saving feature. The equations used to find the particular f_o, H_OBP (bandpass gain), and Q factors for the individual filters were found in a document distibuted by Texas Instruments, Chapter 16 of a book entitled Op Amps For Everyone ( From this document, the following equations were used to find the three filter attributes listed above. Note that in the equations below, A_mi is the midband gain of each individual filter, Q_i is the quality factor of each filter, and α is the factor by which the desired center frequency should be multiplied and divided to obtain the individual center frequencies.

Figure:6 LMF100-Mode-3

Figure 6  LMF100-Mode-3

(11)\; \; \alpha^{2} + \left[ \frac{\alpha \cdot \Delta \Omega \cdot a_{1} }{b_{1}\left( 1 +\alpha^{2}\right) } \right]^{2} + \frac{1}{\alpha^{2}} – 2 – \frac{\left(\Delta \Omega\right) ^{2}}{b_{1}} = 0

(12)\; \; Q_{i} = Q \cdot \frac{\left( 1 + \alpha^{2} \right) b_{1}}{\alpha \cdot a_{1}}

(13)\; \; A_{mi} = \frac{Q_{i}}{Q} \cdot \sqrt{\frac{A_{m}}{b_{1}}}

Note also that, in the above equations, a_1 and b_1 refer to filter coefficients, specified by the filter type desired. Of the major filter classes (Bessel, Butterworth, and Chebyshev), the Butterworth filter was the most desirable, due to its middle-of-the-road frequency response (it is commonly described as a compromise between Bessel and Chebyshev). The filter coefficients corresponding to this filter type are \sqrt{2} and 1, respectively. The variable ΔΩ corresponds to the inverse of the Q factor of the desired fourth-order filter. Our minimum Q value would be at 20 kHz, which yields 20kHz/1kHz, or 20. This gave ΔΩ = 0.05$. This yielded an alpha value of 1.01752. This, in turn, gave Q and A_mi values of 28.2885 and \sqrt{2}, respectively. This is due to our desire to have the filter by unity-gain at passband. The rationale behind this is that the amplifier is much better suited for low-noise amplification than the filter. With these filter parameters in hand, we proceeded to choose Mode 3 external resistor values for the LMF100.

The LMF100 datasheet specifies that, for Mode 3, the following equations hold:

(14)\; \; f_{o} &= \frac{f_{CLK}}{\{ 50,100 \} } \times \sqrt{\frac{R_{2}}{R_{4}}}

(15)\; \; Q &= \sqrt{\frac{R_{2}}{R_{4}}} \times \frac{R_{3}}{R_{2}}

(16)\; \; H_{OBP} &= -\frac{R_{3}}{R_{1}}

A reasonable value for R_3 is 100kΩ. This yields approximate values for R_1, R_2, and R_4 of 71.5kΩ, $3.48kΩ, and $3.60kΩ, respectively. Note that R_1 and R_3 will have the same value for both second-order elements, whereas R_2 and R_4 will be swapped between the two. This elegant design is a byproduct of selecting the appropriate LMF100 mode. These resistor values are summarized in Table (LMF100 Resistor Values)

Figure:7 LMF100 Resistor Values

Figure 7  LMF100 Resistor Values

Note that the order of the filters does not matter, since convolution is commutative. Note that these resistors are the result of several other design decisions related to the LMF100. First, we decided to power the LMF100 off +5V and -5V, since precisely offsetting the input signal to work on a single-ended power supply would be difficult, and could potentially increase noise by decreasing the amplitude of the signal. Also, the LMF100 can operate in f_CLK/50 or f_CLK/100 mode. The latter mode offers slightly better performance in many situations, since the minimum alisable frequency is doubled with respect to the former. However, the latter mode posed a problem for us in terms of f_CLK adjustability. We intended to provide the LMF100 clock signal from the FPGA, not from a dedicated clocking module or PLL, so it was only possible to provide a limited set of frequencies (relatively even fractions of our input clock frequency, which on the TE0140-01 is 60 MHz). For the center frequency range 20 kHz to 35 kHz, f_CLK/100 mode would require an input clock between 2 MHz and 3.5 MHz, which gives us substantially fewer possible synthesizable frequencies than f_CLK/50 mode. Additionally, we predicted that energy in the 50 * f_CLK to 100 * f_CLK range would be quite low, making the increased aliasing a non-issue. The final LMF100 configuration used is shown in Figure (LMF100 Final Con guration). Note that the resistor numbers are slightly different from those previously discussed, that the analog and digital voltage rails are tied together (since our design is relatively devoid of large digital chips which would dirty an analog power supply), and that this configuration includes several decoupling capacitors, the rationale behind which will be discussed later in this paper.

Digitization And Voltage Translation

After the hydrophone signal has been amplified and filtered, it needs to be digitized and sent to the FPGA. Our proposed digitization scheme consists of an analog comparator with an arbitrary threshold voltage provided by a potentiometer, connected between our power and ground rails. The requirements for this analog comparator are as follows:

1. It must be fast. Every analog comparator has a propagation delay associated with it, on the order of ns or $\mu$s. Any variation in this propagation delay will manifest itself as phase error in the digitized signals, so we must keep the propagation time itself as small as possible, and by extension, limit its variation across the 3 comparators necessary to digitize all 3 hydrophone signals.

2. It must have a dual-supply front end (input) and single-ended output. The signal coming from the LMF100 varies between +5V and -5V, and the digital signal to be produced must vary between 0V and a positive voltage (the FPGA supports various voltages from 1.2V to 3.3V). It turns out that few analog comparators satisfy this requirement.

After an extensive search, a comparator was found that satisfied these requirements, as well as being a quad-comparator part, which means that only one IC would be needed to digitize all 3 hydrophone signals. This not only reduces the bill of materials (BOM) of the PCB, it also drastically reduces propagation time variation. The part is also from Analog Devices, the CMP401. This part comes in 16-pin Narrow SOIC and Thin Shrink Small-Outline Package (TSSOP). We chose the SOIC version for ease of soldering, and greater availability. The full part number is CMP401GS. The pinout of the CMP401GS is shown in Figure (CMP401GS Pinout).

Figure:8 LMF100 Final Configuration

Figure 8  LMF100 Final Configuration

One difficulty encountered in designing the CMP401GS into our system is that, although the front-end and output stages have separate voltage rails, the analog positive supply should not exceed the digital positive supply voltage by more than 200mV. This means that a separate IC would have to translate between the 5V output of the CMP401GS and the FPGA, which only supports I/O standards up to 3.3V. Applying a 5V input would irreparably damage the FPGA. The process of selecting the voltage translator was dramatically simpler than for the rest of the components. All that was required was a quad 3.3V-powered buffer with 5V-tolerant inputs. Such a part was quickly found in Fairchild Semiconductor’s 74LCX125. This part comes in a 14-pin Narrow SOIC package (full part number 74LCX125M). The pinout of the 74LCX125M is shown in Figure (74LCX125M Pinout).

An additional advantage of the 74LCX125M is that it was a quad buffer, which meant that it could translate an additional signal besides the 3 hydrophones, which would turn out to be useful later in the hardware design.

Miscellaneous Components

Serial Communication

The FPGA board selected, the TE0140-01, already contains JTAG and USB interfaces. However, we thought it useful to include a simpler (than UTMI), more standard (than JTAG) debug interface for communicating with our MicroBlaze processor. One such protocol is RS-232, commonly known as “serial”. This protocol is already implemented in a UART module provided by Xilinx as part of the Embedded Development Kit (EDK) software package. Furthermore, RS-232 is supported as one of the debug stdin and stdout ports

Figure:9 CMP401GS Pinout

Figure 9  CMP401GS Pinout

Figure:10 74LCX125M Pinout

Figure 10  74LCX125M Pinout

of the MicroBlaze processor, making RS-232 a valuable system debug tool. However, the FPGA’s input and output voltages were obviously not suitable for the standard, which specifies typical voltages of +12V and -12V. To translate between the 3.3V FPGA I/O’s and a standard RS-232 line, a level shifter was needed. Luckily, these are readily available. We decided on a single transceiver (Rx/Tx signals only) from Texas Instruments (, the MAX3221CDBR. This level shifter required several external capacitors (for decoupling and a charge pump), but they were all of a single, standard value (100nF), so the chip, which comes in a 16-pin SSOP package, was relatively easy to integrate into our design. We decided to power the chip off +5V and GND. The input on the level shifter has a logic threshold suitable for interpreting the FPGA’s 3.3V Tx signal, but some sort of interface was required for the Rx line, since the level shifter was powered off 5V. Luckily, as alluded to earlier, the 74LCX125M’s fourth input served this purpose well. The pinout for the MAX3221CDBR is shown in Figure (MAX3221CDBR Pinout).

Figure:11 MAX3221CDBR Pinout

Figure 11  MAX3221CDBR Pinout

Indicator LEDs

Throughout the course of this design, it became apparent that an easy visual means of quickly verifying the correct operation of the analog processing circuit would be desirable, since oftentimes autonomous underwater vehicles are used in environment where laboratory instrumentation is sparse. A simple means of verifying the entire circuit is to connect a light-emitting diode (LED)~to each of the 3 digitized hydrophone signals. If the circuit is operating correctly, the LED should light up each time the signal goes high (logic `1’). This should only occur during pinger pulses, which typically happen at predefined intervals on the order of seconds.

In order to drive these LEDs, an inverter was needed that could sink approximately 10mA of current per output. The inverter selected for this application was the Texas Instruments SN74LVC3G04DCTR. This triple inverter is housed in an 8-pin SSOP package, which is extremely small, and thus relatively easy to fit in our design (though difficult to solder). The pinout of the inverter is shown in Figure (SN74LVC3G04DCTR Pinout).

For the LEDs themselves, three different colors were desired, in order to be able to easily differentiate which of the 3 channels is triggering, even in an underwater environment. In order to maximize the viewing angle, diffused LEDs were preferred over clear ones. Also, in addition to operating within our voltage constraints, with reasonable power, the LEDs should integrate the current-limiting resistor typically needed in these situations, to reduce the BOM of the PCB and reduce the footprint of the LEDs. A line of LEDs from CML was chosen for the task (part numbers 5602F1-5V (Red), 5602F3-5V (Yellow), and 5602F5-5V (Green)). These LEDs integrate an appropriate current-limiting resistor, and simply have two pins, power and ground.

The final configuration of comparator, voltage translator, inverter, and LEDs is shown in Figure (Digitization And Debug Final Configuration).

Figure:12 SN74LVC3G04DCTR Pinout

Figure 12  SN74LVC3G04DCTR Pinout

Figure:13 Digitization And Debug Final Configuration

Figure 13  Digitization And Debug Final Configuration


Our PCB contains four classes of signals which must be connected to the outside world:

1. Power- Here, we have +5V, -5V, and GND.

2. Hydrophones- There are 3 hydrophone signals, each with their own ground (although they are all tied to system ground on the PCB).

3. Serial- We need at least GND, Rx, and Tx. No flow control will be used. There may be multiple grounds in order to easily interface to a standard DB-9 connector.

4. Program- The FPGA can be reprogrammed by bringing a PROGRAM signal low temporarily. This is a good way to reset the FPGA without power-cycling it (can be done electronically). This signal should be brought out to a connector, along with a GND signal, so that it can be brought low, either by shorting the two signals or by applying 0V to the signal with respect to GND (such as with a digital acquisition card).

A logical partitioning of these signals broke these 4 classes up into three connectors. One connector with large pins was needed for power, one 6-pin connector for hydrophone signals, and a nominally 10-pin connector for serial and program signals. The latter 2 connectors should be relatively standard, so as to facilitate replacement and procurement of new female connectors. The first connector chosen was a Molex 39-30-3045 4-pin right-angle power connector. Two pins were used for GND, one for +5V, and one for -5V. The second connector was a 6-pin right-angle .1’‘ header (Tyco 102617-1). The final connector was a 10-pin right-angle .1’‘ header (Tyco 102617-3). The latter two connectors are two-row connectors, for decreased width (more convenient PCB footprint). Additionally, both header connectors were shrouded for added protection from inadvertent short-circuits.

Decoupling Capacitors

In order to provide a clean voltage rail to each component on the board, a two-level decoupling network was used. The first level consisted of bulk capacitors for the +5V and -5V rails. Two 100$\mu$F radial capacitors (Nichicon UVZ1E101MED) were chosen for this purpose. Their voltage rating (25V) and size (5.5mm diameter) were well-suited to our design. These capacitors will provide moderately-sized surges of current, should they be required by any of our on-board integrated circuits. Also, for each rail of each component, we place a 100nF decoupling capacitor between the power pin and GND. While this increases the component count of the board significantly, it will ensure proper operation of the sensitive analog circuitry by preventing droop on the power rails when current demands change, as they inevitably do. These decoupling caps also ``decouple’‘ the components from the system by absorbing transient spikes in the voltage rail, providing a low-impedance path to ground for high-frequency signals and preventing component damage. The decouplers chosen for the system were Panasonic 25V 0805 form factor capacitors (Panasonic ECJ-2VF1E104Z). The 0805 form factor was chosen as a compromise between solderability and size. Any smaller, and we may not be able to ensure adequate solder joints with the limited equipment available to us. Any bigger, and the capacitors may dominate the board area. As a consequence of choosing the 0805 form factor for the capacitors, we also chose to purchase 0805 resistors, for aesthetic reasons.

In accordance with the manufacturer’s recommendations, each power pin on the FPGA board will be decoupled on our baseboard with a separate 100nF.

Resistor Selection

As mentioned previously, the 0805 form factor was chosen for all onboard resistors. Most resistors only require 1% precision. However, the resistors that determine the center frequency of the LMF100 filter elements (previously referred to as R_2 and R_4) are very important. Any mismatch in these center frequencies will manifest itself as phase error between the processed hydrophone signals, due to the high first derivative of the phase response of a bandpass filter. Therefore, these resistors were purchased with 0.1% precision. Based on availibility, various resistor values from Yageo and Panasonic were purchased.

Potentiometer Selection

The design criteria for the potentiometers is that they should be high-turn, side-adjustment, and through-hole mount (for mechanical stability). The 10kΩ potentiometers selected were Murata PV37P103C01B00s. These are 250mW 12-turn pots. The 5MΩ pots chosen were the Bourns 3296P-1-505. These are 100mW 25-turn pots.

Prototyping Aids

In order to prototype easily with small, surface-mount components on a breadboard, SOIC-to-DIP adapters were needed for the CMP401 and 74LCX125. Such prototyping adapters were readily available in the necessary 14-pin and 16-pin sizes from Aries Electronics (part numbers 14-350000-10 and 16-350000-10, respectively).

PCB Design
High-Level Decisions

There were several decisions to be made before physical PCB layout could begin. First, the manufacturer and process were chosen. A popular manufacturer of prototype PCBs is ExpressPCB ( In addition to a 1-layer or 2-layer process, they also provide free layout software. Unfortunately, it became clear that we required a more advanced process if we were to successfully solder so many fine-pitch components, particularly the 1.9mm-pitch, 80-pin board-to-board connectors. We required a manufacturer who would be able to provide boards with solder mask, which prevents solder bridges between pins of fine-pitch components. Also, since there were so many components placed throughout the board, and many decoupling capacitors which require connections to GND. To make routing all the components possible in such a small area without massive amounts of jumper wires, a 4-layer board was decided upon. The inner two layers would be used for GND and +5V, with the outermost layers used for signal routing. One manfacturer that could provide these boards in a relatively short amount of time, for relatively low cost, was ProtoExpress ( Once they had been decided upon as a manufacturer, we had to find a PCB layout tool that could output Gerber files, the industry-standard format used by many manufacturers to transmit pad, solder mask, drill, board outline, and other data.

One very helpful feature for a PCB layout tool is an integrated schematic capture tool. This allows the designer to design the schematics and physical board within a single environment, and pass information back and forth between the two views. For example, the layout tool can ensure that all connections specified in the schematic have been physically made on the boards. All components instantiated in the schematic can be associated with footprints in the layout tool. There are only a very few tools with these capabilities, including Eagle Layout Package, OrCAD, and ProTel. In the end, due to availablity, a large collection of library schematic components and footprints, ProTel was chosen. Once a ProTel license was obtained, libraries had to be made for all components which weren’t included with the software. This essentially amounted to nearly all components used in the board. This was a very time-consuming process, since all mechanical dimensions of all parts had to be obtained, checked, and meticulously rechecked from manufacturer’s datasheets. However, the process yielded an incredibly useful library of information; once the schematic capture tool and the layout tool were working in concert, the PCB design and design checking process was greatly simplified.


As mentioned previously, the typical 4-layer arrangement of two outer signal layers and internal power and ground planes was used. Most of the literature consulted on the subject suggested that the ground plane should be placed closer to the top, to reduce the inductance of power trace ground return paths.

Component Placement

To provide a reference, the side contaning the FPGA board-to-board connectors was chosen as the “top”. In order to avoid violating height constraints, and maintain proper seating of the FPGA board, only passives (SMT resistors and capacitors) were placed between the two connectors on the top. Most of the integrated circuits were placed on the underside of the board. The 6-pin, 10-pin, and power connectors were placed along the edges of the board, for obvious accessibility reasons. The 6 amplifier adjustment potentiometers were placed according to channel along one side of the board, and the threshold adjustment potentiometer was placed on the opposite side. The decoupling capacitors were placed on either side of the power connector. The AD743 amplifiers were placed in a row on the top of the board, with the leads extending through to the other side. These built-in vias allowed us to minimize the via count on the board while passing the hydrophone signals through to the bottom side of the board, where most of the signal conditioning would take place. The LMF100s are placed in a row next to the AD743s (on the underside), and the other four ICs (analog comparator, voltage translator, RS-232 level shifter, triple inverter) were grouped on the underside as well, on the side opposite the AD743s. The LEDs were placed on the top side, directly on top of the inverter, next to the threshold potentiometer. Also note that the FPGA board was positioned so that the USB connector extended nearly to the edge of the board, for easy connection. Views of the top and bottom of the board are shown in Figures (PCB Top Layer) and (PCB Bottom Layer), and a rough 3D view of the board is shown in Figure (PCB 3D View).

Figure:14 PCB Top Layer

Figure 14  PCB Top Layer


With all the signals and ICs present on this board, the routing was an interesting challenge. ProTel includes an autorouter which can attempt to automatically route signals specified in a schematic based on user-specified design rules. These design rules include routing topology (daisy-chain, shortest, horizontal, vertical, etc.), trace-and-space requirements and preferences, high-priority nets, clearance from the edge of the board, overall routing strategy (net length minimization, via minimization, hybrid), and so forth.

Figure:15 PCB-Bottom Layer

Figure 15  PCB-Bottom Layer

Figure:16 PCB-3D View

Figure 16  PCB-3D View

Though the auto-router proved to be a valuable time-saver on the more mundane nets, and provided a decent starting point, computerized routers still cannot match the results of a skilled human. In order to improve the aesthetic appearance and performance (minimize parasitics and crosstalk) of the board, the majority of the nets were re-routed by hand. An important rule in PCB routing is to avoid right angles and stubs whenever possible. Both of these drastically impact signal integrity (\textit{SI}) by causing discontinuities in the impedance of the trace. Instead, most PCB designs today use 45-degree mitered corners to compromise between right angles and ideal, rounded corners. Also, decoupling capacitors were given their own ground vias whenever possible. It is widely known that the inductance of a capacitor’s route to ground determines how effective it is going to be as a decoupler, and this inductance is greatly decreased when the via is not shared, and when it is placed as close to the capacitor as possible. The trace-and-space requirements of the ProtoExpress process were both 6 mils, but the constraint imposed on ProTel was 8 mils, in order to maintain a safe margin for error, and prevent manufacturability issues.

Additionally, thermal relief was used on all pad vias. This means that a wagon-wheel shape is formed around each via as it passes through the power or ground plane. This is important because, without this, the entire plane will act as a large thermal resistance, making soldering through-hole components extremely difficult. An example of a thermal-relief structure is shown in Figure (Thermal Relief). These structures can be automatically created by ProTel. Another feature offered by ProTel is automatic teardrops. Teardrops spread out the connection between tracks and vias or pads, theoretically creating a more stable, lower-impedance connection. However, both for aesthetic reasons and to maintain maximum spacing between tracks, this feature was not used.

Figure:17 Thermal Relief

Figure 17  Thermal Relief

Board Statistics

The final design had the following characteristics:

  • Pad And Via Holes: 250
  • Pads: 513
  • Tracks: 1352
  • Vias: 165
  • Components: 97 (71 top, 26 bottom)
  • Nets: 75
Release To Manufacturing

After the board design was complete, a design rule check (DRC) was performed within ProTel to check for incomplete connections, short circuits, and clearances below the minimum value. After several iterations, this check passed completely, and the design was ready for manufacturing files to be generated. Luckily, ProTel has a built-in facility for generating a wide variety of Gerber-format files. These were generated, along with a separate drill file, which tells the manufacturer precisely where to drill, and what size drill is needed. These files were submitted to ProtoExpress for manufacturing.

The Second Iteration

We received the first set of 3 boards approximately one week after submitting the Gerber files. Although the boards had been manufactured to our specifications, we quickly found, by test fitting all components, that several mistakes had been made in the first revision of the board. First, several of the vias were too small to fit the pins of their respective components, such as the $5$M$\Omega$~potentiometers and the LEDs. Also, the decoupling capacitor dimensions were coded into ProTel incorrectly, so the component outlines were twice as big as the actual components. Finally, the most serious error was in the LMF100 footprint. The footprint we generated was for a 20-pin SOIC, whereas the LMF100 uses a special Wide SOIC package, which is nearly twice as wide as regular SOIC chips. One potential workaround for this would be to solder down one side of the chips, and solder jumper wires from each pin of the other side to the board. This, however, did not seem like a robust solution, and seemed very fault-prone, both electrically and mechanically. Therefore, a second revision of the board, APLS 2.0, was quickly made, and submitted for manufacturing. All of the above issues were remedied in this revision. The most difficult to correct was the LMF100 footprint, since the much larger footprint required moving a lot of resistors and capacitors around on the board, and rerouting over a hundred traces.

When we received the second board, we found that all components fit correctly, verifying that all the above issues had indeed been fixed. As we would later find, there were several smaller problems which still existed with the design, but these could be fixed without a new board revision (although one will be performed in the near future).


Once the boards and components had been received, a brief conductivity test was performed to ensure that there were no glaring manufacturing defects. The many surface-mount components on the board were attached using reflow soldering, where a thin bead of solder paste is applied to each pad, the component is placed on top, and the entire board is placed in an oven. By carefully following a prescribed temperature profile, all the solder on the board reflows together, effectively soldering an entire side of the board in one pass. This procedure is made considerably more complicated when there are surface-mount components on both sides, for obvious reasons. Luckily, the surface tension of solder paste is such that, even when it reflows, it can support the weight of most surface-mount components, allowing a double-sided board to be assembled in two reflow passes. First, we applied solder paste to one set of pads at a time on the bottom side of the board (where there are fewer components), and placed the components on. For integrated circuits, we placed an individual solder bead on each pad when the pitch was over 50 mils, and a single bead down each row of pads when the pitch was under 50 mils. In the latter case, the solder mask should help attract the solder to one side or the other, ameliorating the serious problem of solder bridges. Once all components on the underside had been placed, we slid the entire board into the oven. After following the reflow profile Figure (Solder Paste Reflow Profile) and allowing the board to cool, we removed the board and inspected it for cold solder joints and solder bridges. Only a few of these were found, and most were easily fixed with a fine-tipped soldering iron. However, one of the resistor pads did not solder at all, and the inverter floated off the pads, so we had to reflow the bottom side again. After this second reflow, all the problems found after the first run were fixed. We could then move onto the top side.

We followed the same procedure on the top side (for surface-mount components only). The board-to-board connectors were particularly worrisome, due to their fine pitch and high pin count. We placed the board in the oven top-side-up, using protoboard and spare FR-4 material as standoffs to prevent the SMT components on the underside from resting on the grill of the oven. After a single reflow run, there were very few solder bridges and no cold solder joints on the top side, and none of the components on the bottom side were dislodged in the process. The lone concern coming out of the reflow process was that some of the joints on the bottom side had a gold tinge to them, indicating that perhaps the oven had gotten too hot on the first reflow run. However, the solder joints appeared to be in order, as evidenced by another conductivity test.

After soldering all the surface mount components, we soldered the through-hole components (all connectors, LEDs, bulk decoupling capacitors, potentiometers, and amplifiers) by hand, using a fine-tipped soldering iron.

Figure:18 Solder Paste Reflow Profile

Figure 18  Solder Paste Reflow Profile

Though the absence of liquid solder flux made the aesthetic appearance of some of the joints somewhat sub-par, they all appeared to be electrically and mechanically sound, and the process went without incident. When all was completed, we were left with a fully-assembled board. A test fit of the FPGA board confirmed that the connectors had been spaced and oriented correctly. Photos of the fully-assembled board are shown in Figures (Top Side Of Baseboard) and (Bottom Side Of Baseboard). Photos of the FPGA board are shown in Figures (Top Side Of Baseboard) and (Bottom Side Of Baseboard). The two boards are shown mounted together in Figure (Both Boards Mounted Together).

Digital System Design
Receipt Time Determination

Given digitized signals indicating that an in-band acoustic signal is being received, the task of determining the order and time of receipts presents itself. A number of plausible solutions were considered, the first of which was analog-to-digital conversion and software thresholding. Given the expense and comparative inelegance of a high sampling rate ADC, we quickly eliminated this option. We also considered using a microcontroller with on-chip analog comparators, such as the Silicon Laboratories ( C8051F series, but the desired sample rate of tens of MHz could not be achieved. Likewise, using extrernal analog comparators and digital input ports on such a microcontroller, we could not achieve the desired time resolution because even the overhead of a polling loop or interrupt service routine virtually guarantees a lower bound of tens of clock cycles between near-coincident arrivals.

A distinct digital system seemed to be a natural solution. The size and wiring complexity of the discrete logic necessary to store several high-resolution timestamps led us to consider a programmable logic device (PLD) implementation. Furthermore, an FPGA implementation eliminated the need for a separate microcontroller because of the availability of soft processor cores.

Initially, the high-level design was as follows: if a comparator output goes high, latch it. Increment a counter every clock cycle, and if, on the rising edge of a clock cycle, a latch is high, store the current count in the corresponding register. When all three registers are populated, provide the two of the pairwise time

Figure:19 Top Side Of Baseboard

Figure 19  Top Side Of Baseboard

Figure:20 Bottom Side Of Baseboard

Figure 20  Bottom Side Of Baseboard

Figure:21 Top Side Of Baseboard

Figure 21  Top Side Of Baseboard

Figure:22 Bottom Side Of Baseboard

Figure 22  Bottom Side Of Baseboard

Figure:23 Both Boards Mounted Together

Figure 23  Both Boards Mounted Together

differences as outputs (The third pairwise time difference can be computed by trivially manipulating the other two.) and indicate that these outputs are valid. A high-level block diagram of this module, dubbed, is provided in Figure (Block diagram of the timecapture module).

Though the concept of this design holds, we saw fit to modify it in several ways. First, timecapture was made synchronous by simply removing the latches. We observed that, since the clock frequency of this module is (ideally) orders of magnitude greater than that of the comparator output, there is little benefit to perceiving a comparator output that is high for a fractional clock cycle to be a valid input; after all, it would have to be high for the setup and hold times, anyway, which are not too much less than the clock period.

An acknowledge signal, which essentially acts as a synchronous reset, was then added. This input signal allows timecapture to hold a sample until the host can acquire it.

A flaw in the design was later identified and addressed. Suppose timecapture was acknowledged or otherwise reset while one or more comparator outputs were high. Then these would immediately be considered valid inputs at the start of the next timecapture cycle, and the recorded time differences would not be meaningful. This problem was handily addressed by the addition of one n-bit shift register per input channel. The most recent input was required to be high and the n-1 preceding inputs were required to be low in order to trigger a receipt. This solution also helped addressed the potentially problematic transient nature of the comparator outputs: the comparators take some time to reach a steady state, so requiring n-1 low inputs suggests that the comparator has been stably low for a sufficient amount of time, and the comparator’s input is just reaching the threshold. This behavior is quite desirable. In this implementation, letting n be 100 proved effective.

Finally, logic was added to reset the device if a valid input has been received on either one or two channels and a certain number of clock cycles has elapsed without valid inputs on all three channels. This so-called reset threshold realizes the constraint of Equation 10. The threshold is not a synthesis-time constant but is instead controlled by the host.

Figure:24 Block diagram of the timecapture module

Figure 24  Block diagram of the timecapture module

HDL Implementation

A benefit of implementing a digital system on a PLD is the ability to take advantage of the useful abstractions that a hardware description language (HDL) offers. One need not, for example, be concerned with the data flow of an up-counter; he need only use VHDL’s addition operator. Furthermore, in a design with extensive parallelism, such as this one—the same operations are being performed on three channels—simplicity in expression can be achieved using constructs like VHDL’s for…GENERATE loop.

Timecapture was initially described in VHDL. It was then translated to Verilog because this design is expressed with greater simplicity and elegance in that HDL. After the numerous improvements discussed above, the design was rewritten again in VHDL, a decision made because the Xilinx MicroBlaze soft processor core lacks a well-documented Verilog interface, and we wanted to avoid the complexities of a mixed-HDL implementation.

The VHDL implementation of timecapture is available in Section 10.1.

Configuring the FPGA

Initial Setup

Using the Spartan-3 FPGA seemed at first to be a straightforward task. The baseboard for the Trenz micromodules allowed for the easy connection of external signals, including a JTAG programmer to load implemented designs into the micromodule’s PROM, which stores the FPGA’s configuration between power cycles. Indeed, Xilinx provides an integrated development environment known as the ISE that makes the synthesis, implementation, and uploading of digital designs fairly streamlined, and a simple test design was run on the FPGA not long after connecting it to a computer with the ISE installed on it.

Unfortunately, the successful implementation of the MicroBlaze soft core was not so simple. Xilinx does provide a robust toolset called the Embedded Development Kit (EDK), a component of which is the Xilinx Platform Studio (XPS). This program implements and synthesizes custom configurations of the MicroBlaze core. The great depth and breadth of this program, unfortunately, gave it a steep learning curve. Our configuration, admittedly, did not ameliorate the situation: while Xilinx provides a detailed XPS tutorial, it assumes the use of a Spartan-3 development kit, which has a very different periphery, memory map, and memory hierarchy. As such, the tutorial could only be followed in parts, and even then, only with caution.

After some effort, a MicroBlaze was instantiated in an ISE project, but the test program (which sends a text string over RS-232) produced no serial output. Simulating the system to determine the cause of this problem, however, turned out to be a difficult task. Because the MicroBlaze VHDL implementation is encrypted, pre-compiled libraries must be given the simulator tool being used. However, the EDK documentation only described the process of exporting these libraries to the commercial simulators ModelSim and NCSim, which we did not have at our disposal; esoterically, the documentation stated only that the libraries could not be exported to Xilinx’s own ISE Simulator. Left only with trial and error to determine the cause of the problem, we tried a variety of MicroBlaze configurations and implementation constraints. After exhaustive testing, the problem was determined to have two causes: first, with the intrusive software debugger enabled, the test program would not begin executing on boot; and second, in a failure to read the Trenz micromodule documentation thoroughly, we had flagged the system clock pin as a local clock rather than as a clock provided on a general I/O pin by the USB interface.

Upon correcting these problems, the test program executed successfully.

Integrating the Digital Subsystems

The MicroBlaze implements the On-Chip Peripheral Bus (OPB), developed by IBM as a means of connecting on-chip devices to the processor core and to each other while abstracting the low-level connection and arbitration details. The OPB, as such, seemed to be the logical choice for connecting timecapture to the MicroBlaze. The XPS ostensibly simplifies implementing an OPB module by generating all of the required interface, allowing the designer to immediately begin implementing the device itself.

Unfortunately, our instance of the MicroBlaze ceased to function when even a stub OPB module was included in the embedded platform. We decided that, given the severe time limitations and the lack of obvious reasons for failure, it would be more productive to pursue a less elegant interface and implement timecapture as an OPB module in a future revision.

The chosen interface uses General Purpose I/O (GPIO) modules to form a fully parallel connection between timecapture and the MicroBlaze. Two 32-bit unidirectional buses provide the two pairwise time differences t_2-t_1 and t_3-t_1, and a (underutilized) 32-bit bidirectional bus provides the Valid signal to the MicroBlaze and provides the Acknowledge signal and the reset threshold to timecapture. Because all of these connections are formed within the FPGA, this parallel interface does not increase the hardware cost but offers much faster and simpler communication between the two devices.


As in most complex embedded systems, the art of debugging was needed to bring the board from its fully-assembled state to its fully-working state. We were pleased to see that upon applying power to the power connector, the current consumption was in the same range as when we had breadboarded the analog circuitry, which reassured us that at least the power pins of the on-board ICs had been soldered correctly. However, through a long debug process, we discovered several minor design errors which prevented the board from functioning fully:

  • Due to insuficient attention to the Trenz Electronic documentation, the pin chosen for the LMF100 filter clock was also connected to a pushbutton on the FPGA board, used for programming. Since we had another signal dedicated to programming, this was not an issue per se. However, the pushbutton had 2 pull-up resistors which would interfere signi cantly with the clock signal. To remedy this problem, these two resistors (0603 and 0402 form factor) were desoldered.
  • Due a simple schematic entry error, the signal ENABLE12, brought out to pin LL2 on the FPGA board, was tied to ground instead of left floating. This disabled the onboard 1.2V voltage regulator, which supplies the power for the Spartan-3’s core logic. In order to ``undo’‘ this connection, tweezers and a utility knife were used to break the board-to-board connector pin and cut the PCB associated with LL2, respectively.
  • Due to another schematic entry error, the signal \overline{FORCEOFF} was tied to ground instead of +5V on the RS-232 level shifter. This caused the transmitter to be constantly powered off. We verified this by observing the input waveform (Tx from the FPGA) on pin 11 of the level shifter, and the lack of any discernable output on pin 13. Also, this signal does not affect the RS-232 receiver, and we verified that receiving still worked by pressing keys in a PC-based terminal program, and observing the waveform on the output side of the level shifter. In order to fix this problem, we disconnected the \overline{FORCEOFF} pin from the PCB, and bridged it to an adjacent pin with a +5V connection with solder.

After diagnosing and remedying these three issues, and modifying a female USB connector to fit onto the FPGA board (to supply power), the system worked flawlessly. These three issues will be addressed in the next revision of the APLS system.

Frequency Response

Once the board had been assembled, we felt it would be useful to measure the frequency response of the two major elements of the analog signal path, the amplifier and the lter. We accomplished this using an Agilent Network Analyzer, with one probe connected to the hydrophone input and the other connected to either the output of the AD743 or the output of the LMF100. The results obtained for the former case are shown in Figure (AD743 Log-Magnitude Frequency Response). The log-magnitude frequency response of the combination of amplifier and filter is shown in Figures (AD743/LMF100 Log-Magnitude Frequency Response) and (AD743/LMF100 Log-Magnitude Frequency Response, Compressed X Axis). Note that these are two views of the same data (different input frequency ranges). The phase response of the combined ampli er and lter is shown in Figure (Figure 26: AD743/LMF100 Phase Response).

Figure:25 AD743 Log Magnitude Frequency Response

Figure 25  AD743 Log Magnitude Frequency Response

From these measurements, several conclusions can be drawn as to the efficacy of the amplifier- filter solution we have chosen. First, the AD743’s frequency response is flat throughout our range of interest (20kHz – 35kHz). Its -3dB frequency is around 38kHz, meaning that our pinger signals will not be significantly attenuated by the amplifier. The combined magnitude response of the amplifier and filter is also encouraging. The peak response is at 20kHz, which is precisely what the FPGA’s clock signal was tuned to, and the lter e ectively attenuates all signals outside the passband. In fact, the response at 10kHz and 40kHz is 55dB lower than the peak value.

One concern that arose from these measurements is the phase response of the fourth-order bandpass filter. While the response observed is predictable, given the nature of bandpass filters in general, the large first derivative of the phase response in the passband is confirmation of our previous belief that proper calibration would be essential for the system to operate correctly. Ostensibly, all three hydrophones should receive the same frequency to within a very small tolerance, but any mismatch in LMF100 device parameters or resistor values can potentially a ect the phase of the output signal, and thus the delay imposed on the digitized square wave, in a significant way. For this reason, it is important to characterize the intrinsic phase shift between all three channels.

Figure:26 AD743-LMF100 Log Magnitude Frequency Response

Figure 26  AD743-LMF100 Log Magnitude Frequency Response

Figure:27 AD743-LMF100 Log Magnitude Frequency Response Compressed X Axis

Figure 27  AD743-LMF100 Log Magnitude Frequency Response Compressed X Axis

Figure:28 AD743-LMF100 Phase Response

Figure 28  AD743-LMF100 Phase Response

Project Files

File NameFile Size
Project report (includes theory, implementation, and test results)1.4 MB
Tags: PCB-Design, DSP, Signal-Processing, FPGA, Embedded Design, Robotics

Comments on this Project:

Joe Wolin
By Joe Wolin (+54) 0Score: 

4 years ago:  Matt,

Thanks for sharing your project - looks like a great multidisciplinary hardware project with sensors and processing together in once nice package.

Good to hear from you; it's been what, 15 years? ;-)

Best Regards,

Login or Register to post comments.
Click Here