MSc dissertation · 2025–26 · Supervisor: Dr Deepak Akiwate Data-driven drone noise modelling and prediction Small drone propellers make strong tones at the blade passing frequency and its harmonics. I built a physics-informed model that predicts these tones: a physical source model gives a baseline level, and a random-forest model learns what the physics misses. Each propeller was tested on data the model had never seen. InputsBlade geometry, thrust, RPM, harmonic, observer angle → Physics modelThrust-constrained radial source → baseline SPL → Random forestLearns the measured-minus-physics residual → Predicted tonal SPLH1, H2, H3 at 0°, 40°, 90° Explore the results · leave-one-propeller-out All 207 cases APC 11×7 APC 12×9 APC 13×6 APC 12×14 target 95% Error by direction and harmonic Try it · blade passing frequency fm = m · B · RPM / 60 Blades (B)2 Rotor speed8690 rpm APC 12×14 · 30 N APC 12×14 · 35 N Hear the drone tone H1290 HzH2579 HzH3869 Hz The presets use the RPM my final model predicted for APC 12×14 at 30 N and 35 N, conditions that were never measured. Only the frequencies are calculated here; all markers are drawn the same height because levels are not modelled.
Room Acoustics · coursework Designing a teaching control room A 6.0 × 4.2 × 2.7 m control room in Salford's SEE building, for stereo and 5.1 teaching. I chose the room ratio, predicted the modes and early reflections, then designed bass traps, broadband absorbers and a rear-wall diffuser to create a reflection-free zone at the mix position. Reflections Room modes Treatment Reverberation 3DTopSide Paths DirectFirst reflectionsDiffuser (treated) Listener, across room2.10 m Listener, from front wall3.00 m Untreated Treated ▶ Send a pulse Listen in this room Hand clap Guitar phrase Sound with pulse The left speaker fires a pulse. Each surface acts like a mirror, so its reflection travels from an image source behind it. Six surfaces give six first reflections, including the floor and ceiling. Move the listener with the sliders and every path, time and level is recalculated. Drag the model to turn it. How this is calculated Room 6.0 × 4.2 × 2.7 m. Left speaker at (1.2, 1.5, 1.2) m, listener ear height 1.2 m. Same values as my report. First-order image-source method: the speaker is mirrored in each of the six surfaces. Path length d = distance from image to listener. Arrival time t = d / c, c = 343 m/s Level relative to the direct sound (inverse-square law): ΔL = 20 log10(ddirect / d) Treated surfaces add 10 log10(1 − α) with mid-frequency α from my design: wall absorbers and ceiling cloud 0.8, rug 0.3, rear wall 0.25. Rear diffuser: the reflected energy (1 − α) is shared equally over 7 points across the QRD face, each with its own path speaker → diffuser → listener. This is a simplified picture; real QRD scattering changes with frequency and angle. Not included: second and higher-order reflections, speaker directivity, frequency dependence and edge diffraction. The sound demo adds a decaying tail to stand in for the later reflections. Check: at the default position the direct path is 1.75 m (5.1 ms) and the floor reflection 2.97 m (−4.6 dB), matching the table in my report. Every bar is a room mode below 200 Hz from f = (c/2)√((l/L)² + (m/W)² + (n/H)²). Tall bars are axial modes, shorter ones tangential and oblique. Click a bar to see its standing wave on the plan: blue and gold are opposite pressure, white is a node. Hear this mode Room ratio1 : 1.56 : 2.22Inside Bolt's region Schroeder frequency≈ 130 Hz2000√(0.3 / 68) Problem regions63 Hz · 125 HzTreated with tuned panels Front half absorbs, rear half diffuses (RFZ/LEDE). Hover the plan to see each treatment. Tuned panel absorber f = 60 / √(m · d) Panel mass6 kg/m² Air gap150 mm 63 Hz trap125 Hz trapHear its frequency Tuned to63 HzWavelength5.4 m Rear-wall QRD diffuser · N = 7 Max depth150 mm Wells follow n² mod 7 = 0, 1, 4, 2, 2, 4, 1. Design frequency f0 ≈ c / 2hmax = 1.1 kHz. Predicted T60 of the treated room in octave bands, against the 0.25–0.35 s control-room target. How I got 0.28 s at 1 kHz T = 0.161 × 68 / 38.6 ≈ 0.28 s
Loudspeakers & Microphones · Immersive Sound Reproduction Loudspeaker modelling and design In MATLAB I modelled an eight-driver soundbar (Peerless TC5FB00-04, 40 mm spacing) with a Peerless XLS-P830452 woofer in a 10 litre sealed box, joined by a 4th-order Linkwitz–Riley crossover at 250 Hz. For immersive sound I designed a gramophone-style two-way speaker with a horn tweeter and a transmission line. Soundbar beam Woofer & crossover Gramophone horn Frequency1 kHz 2005001k2k4k8k Listen at0° (sweet spot)30° off-axis The beam is broad at low frequency, narrows through the mid band, then breaks into lobes and nulls at high frequency. This plot is the array factor only: eight ideal point sources at 40 mm spacing with equal weighting, |P(θ)| = |Σ ejkd n sin θ| / 8. It leaves out each driver's own directivity. Max deviation above 1 kHz at 30°28.71 → 24.61 dB Mean deviation above 1 kHz at 30°14.55 → 10.37 dB From my MATLAB model, original design → after a frequency-dependent reduction of the outer drivers. The plot above shows the original, equally weighted array. Sealed box fc59.94 HzQtc0.589Crossover250 Hz Pink noiseWoofer onlySoundbar onlyBoth Crossover250 Hz Gold: woofer in its 10 litre sealed box (2nd-order high-pass, fc 59.94 Hz, Qtc 0.589). Blue: Linkwitz–Riley 4th-order low-pass and high-pass. LR4 outputs are each −6 dB at the crossover and sum flat. Move the slider to shift the crossover. S(x) = S0 emx · fc = cm / 2π Flare (m)15 TL tuning80 Hz Hear the cut-off Horn cut-off820 HzLine length c/4f1.07 m My design: 14 inch (0.356 m) exponential horn with m = 15 for the 1.5 inch tweeter, a folded 80 Hz transmission line for the full-range driver, and a 1.5 kHz crossover. Coverage is about 80 Hz to 20 kHz. Drawing: horn length to scale, mouth about 11 inches across as in my design (10–12 inch). The folds of the line are a sketch; its length is c / 4f.
Principles of acoustics & vibration Machine on springs: finding its natural frequency From a measured impulse response I identified a 100 kg machine on resilient mounts. Press play to see the decay my values predict. fn ≈ 11.2 Hzζ ≈ 0.067k ≈ 4.98 × 10⁵ N/mc ≈ 948 Ns/m I also calibrated a measurement chain (2.87 Pa/V) and found 78.57 dB SPL at the microphone.
Measurement Analysis & Assessment Vibroacoustics: blocked forces Impact-hammer mobility measurements on a coupled source–receiver assembly, then blocked forces to predict its vibration. Adding rotational terms improved the prediction. Validation error (dB). Best coherence 0.9817; repeat impacts agreed to a median ratio of 0.986.
Environmental Noise Measurement & Modelling Aircraft noise next to Manchester Airport An acoustic design statement for flats at Runway Visitor Park, using ProPG and BS 7445-1. Hover a point to compare the average and the loudest event. Façade points were most exposed (up to LAFmax 89.2 dB), so I recommended quiet-side layouts, better glazing and ventilation.
Noise & Vibration Control Heat pump noise and silencer design Checked an air source heat pump against the MCS 020 limit, sized a duct silencer for about 10 dB at the blade passing frequency, and designed a 2 m barrier for at least 8 dB(A) insertion loss. Predicted level at each neighbour before mitigation (with +3 dB façade reflection), against the 37 dB limit.
DSP & Machine Learning Filters, reverb and neural networks Comb and all-pass filters as a simple room reverb, FIR and IIR filter design, and a neural network to map a non-linear system, all in MATLAB. A feedback comb filter, |H| = 1 / |1 − g e−jωM|, drawn with example values g = 0.65 and M = 8 samples: the building block of digital reverb. Hear noise through it