Accomplishment
Roshni: Roshni: Retargeted the beamforming RTL to the full 20 to 80 kHz band (384 kS/s, decimation 8, 127-tap filter, 128-sample delay buffers) and regenerated the steering tables for ±52° in 61 steps. All four modules (decimator, filter, fractional delay, scan controller) pass their testbenches against the golden model in both the 384 kS/s and 192 kS/s configurations. Fixed the 192 kS/s filter coefficients and added a clock enable to the filter. The scan controller now outputs a latched 61 × 32-bit energy snapshot with a sweep counter, saturation flag and configurable shift. Also completed a full pass on the Design Report, including verification tests for latency, size and weight, and detection probability.
Wes: The fabricated PCB and all the components to be soldered onto it arrived at the ECE receiving office. I finalized the use-case requirements, microphone array subsystem and design and verification sections of the design report, and sketched a 3D-printed mount that stacks the mic array PCB, Ultra96-V2, Valore powerbank (in a printed slot) and LCD, each on four M2 screws except the powerbank.
Essie: Rewrote the heatmap and capture software to match the final design report (61 angles over ±52°, 105° camera map), running at 30 fps on synthetic data. Added the real data reader and threshold calibration, and drafted a register map and steering ROMs for Roshni to check.
Risks
- All the components have not arrived and most of work depends on the physical components.
Design changes
- Normalization: the heatmap subtracts the sweep’s median noise floor before scaling, where the report says to scale by the peak of the full 61-bin sweep. A flat noise floor otherwise tints the whole image, so update the wording in the report.
- Smoothing Window: The smoothing window is now 4 sweeps (was 3), updated only on new sweeps, and the heatmap subtracts the median noise floor before scaling.
Schedule
On track
Highlights
- The software now matches the design report end to end on mock data: a 61-angle heatmap overlay at 30 fps, with render time cut from 45 ms to about 5 ms. The real-data path (overlay loader, tear-safe reader, threshold calibration) is ready, so integration only waits on Roshni’s bitstream and register map.
Part A Global Factors:
Pressurized-gas leaks occur in plants everywhere, not only in well-funded facilities, and the cost of commercial acoustic cameras is what keeps many smaller or older plants on manual walking inspections. SoundSight targets that gap with 30 low-cost PDM MEMS microphones and a single FPGA board, aiming to reproduce the core function of a commercial acoustic camera at a lower cost. It needs no network connection, cloud service or specialist setup, so it works in plants with poor connectivity. The output is a colored overlay on live video rather than an audio spectrum, so a technician needs no acoustics training to use it.
Part A was written by Wes.
Part B Cultural Factors:
Plant workplaces differ in safety rules and habits, and the device has to fit them. Many technicians are trained to find leaks by ear or with contact tools, so SoundSight is designed to support that judgment, not replace it. It points to a direction and the technician still confirms the leak by visual inspection or an optical gas imager. Because it is non-contact, a technician stays away from hot or pressurized equipment, which fits safety practices that restrict close contact. Technicians often wear hearing protection, and a visual display works when sound cannot be heard. The interface uses a color overlay and no text, so it does not depend on the user’s language.
Part B was written by Essie.
Part C Environmental Factors:
Leaks waste energy and can release harmful substances. The US Department of Energy reports that leaks sometimes waste 20 to 30% of a compressor’s output in poorly maintained systems, and that proactive detection and repair can bring this below 10%. A 1/4-inch leak at 100 psig costs about $8,400 a year in electricity at DOE’s reference rate. Our device aims to help these poorly maintained facilities find leaks earlier and at low cost, which means less wasted electricity and less release of chemicals, steam or process gas. We also considered the device’s own environmental impact. Our worst-case power estimate is 50 W, so one 2-hour inspection uses about 0.1 kWh, which is small compared with the energy a single leak wastes. The device is rechargeable and uses no disposable batteries.
Part C was written by Roshni.
Next week
Roshni:
Roshni: Build the top level and AXI-Lite register wrapper (including an ID register), then deliver the bitstream, .hwh and a written register map to Essie. Settle the mic clock structure with Wes and add the stored-block energy window to the scan controller. Check Essie’s drafted register map against the RTL.
Wes: Solder all components onto the PCB. After PCB assembly, individually test every microphone for a data signal, before handing it off to teammates for the RTL bringup.
Essie: Agree the register map with Roshni and swap the mock source for real sweeps. Test the display, cable and camera on the Ultra96, then calibrate the camera and the false-alarm threshold.
