Thermocouples submerged in molten glass drift, degrade and fail, leaving operators blind on the variable that controls fiber diameter, breakage and energy cost. A virtual sensor keeps reading when the physical one stops.
4.37°FRMSE, middle of melt at 100 TPD
Cross-section: crown arch and insulation course, breast walls, the batch line at the charging end, the burner row over the bath, and the three sub-floor thermocouple stations. The wave travels from the charging end toward the throat and each station reads as it passes.
Data
Live process historian, 12 months continuous operation — 2-hour averages plus a 20-minute high-resolution set for transitions
44 raw process tags expanded to 110 engineered features: lags 1–12, rolling means and standard deviations, first and second deltas, oxygen-to-gas stoichiometric ratio, crown thermal gradient, underglass spread
Cleaning carried real weight — missing channels, drifting sensors, unit inconsistencies between pulls, physically impossible readings during faults
Approach
Virtual sensor: six gradient-boosted models predicting underglass temperature at doghouse, middle of melt and throat, from crown thermocouples, gas flow, oxygen flow and cullet percentage
Forecaster: rolling multi-step temperature trajectory forecasts
Prescriber: an optimiser working backward from a target temperature to a recommended gas and oxygen ramp, pre-positioning the furnace ahead of a scheduled production change
Results
Predict the change, not the value — early models scored well by learning the mean rather than the dynamics; reframing the target as ΔT was the decision that made the system work
At higher production rate the crown runs hotter while the glass runs cooler, and the two are nearly uncorrelated within a regime — production rate splits the problem into two physical systems, which is why there are six models rather than three
Temperature travels as a wave down the furnace; the strongest predictors were rates of change at neighbouring thermocouples, not absolute temperatures
LimitsPhysics constrains pre-positioning to roughly 25–30°F of achievable shift in an 8-hour window at safe ramp rates. The rest closes only when pull rate itself changes.
Osteoporosis usually goes undiagnosed until a fracture happens. This work inherited a prior pipeline, audited it end to end, and found that its headline result did not survive honest evaluation.
0.976DiceU-Net bone segmentation
Data
109 hip CT patients — 71 with osteoporosis, 38 without — from University of Toledo Medical Center
15 clinical variables including age, BMI, calcium, haemoglobin and glucose
Patient imaging never leaves the institution; published visuals use segmentation masks or schematics only
Approach
U-Net segmentation to isolate bone regions, then radiomic feature extraction via PyRadiomics — texture, shape and intensity
Radiomic features fused with clinical variables
Six models trained under Bayesian hyperparameter optimisation: Logistic Regression, SVM, Random Forest, AdaBoost, CatBoost and a DNN
Results
Uncovered a data-leakage bug in the inherited pipeline where the model was effectively seeing test data during training
After correction and re-validation a previously reported headline result did not hold, while a simpler and more robust model emerged as the most trustworthy
Replaced a single train/test split with repeated cross-validation, showing the initial results were noise-driven and confirming genuine signal only under stricter testing
LimitsSome images had been exported at different scan settings — a hidden confound that could bias predictions, flagged rather than worked around.
Tumour margins are soft and blurred, which is exactly where hard clustering fails. Fuzzy C-Means segments the ambiguous boundary and a CNN classifies what it finds.
95.8%overall accuracy
Schematic: slices stacking into a volume, with the lesion boundary resolving from a soft membership band into a defined contour — what the fuzzy layer between UNet and UNet++ is for. No patient imaging is used.
Data
MRI brain scans across four classes: glioma, meningioma, pituitary and no tumor
Preprocessing pipeline of noise reduction, intensity normalisation and registration
Approach
Fuzzy C-Means clustering to segment tumour regions, handling soft boundaries better than hard clustering
CNN classification — convolutional and max-pooling stages, dense 1024, transfer learning
Hybrid integration at both feature and decision level, followed by clinical validation
Results
95.8% overall accuracy, a 7.3% improvement over traditional methods
97.87% peak accuracy for the hybrid model — peak, not the headline figure
AUC above 0.97 for all four classes, converging within 25 epochs
Externally validated through the University of Pennsylvania
Studying how cancer cells die normally requires fluorescent labelling, which is invasive and can alter the biology being observed. This asks whether death patterns can be recognised from raw, unlabelled microscopy video alone.
Data
Label-free, unstained microscopy video of live cancer cells
Sequential frame data capturing morphological change over time
Approach
A fuzzy-logic layer to manage noise and ambiguity in unlabelled cell imagery
A deep LSTM modelling the temporal progression of cell death across frames
A hybrid architecture novel to this application
Results
Demonstrated that meaningful cell-death dynamics can be extracted from label-free video, reducing dependence on staining
Published in Neural Computing and Applications, Springer Nature
With only seven years of counts, time-series forecasting is the wrong frame. Normalising the origin-destination matrix by population turns the task into learning a spatial assignment operator that can be re-expanded against any year.
8.92%MAPE across 919 count stations
Schematic: tract nodes on the left scaling with population, trips travelling along sampled assignment edges, and count stations brightening as demand reaches them.
Data
Origin-destination survey across 328 census tracts in four Columbus-area counties
AADT at 919 count stations, 2013–2019
Tract population, income, households and workers, plus TAZ demographic projections to 2050
Approach
Normalised the O-D matrix by origin population to produce per-capita trip rates, then re-expanded against any year’s demographics
Reframed the problem as learning a spatial assignment operator rather than a time series
Dense Keras network, MAE loss, Adam at lr 1e-5, 1000 epochs — trained 2013–2018, tested on held-out 2019
Results
8.92% MAPE across 919 stations
Test MAE 0.0088 scaled, against 0.0033 at convergence
Projection 2020–2050 at 0.91% average annual growth
Limits38,375,647 parameters trained on 6 samples. No baseline comparison against growth-factor projection, and a single-year test set means no error bar. The 2050 curve is near-linear because the model is near-linear in population and the population projections are themselves smooth interpolations — the spatial assignment is real work, the temporal shape is largely inherited.
Built first independently — roadside video capture, detection, tracking and speed derived from tracked movement, with real-time inference on a Raspberry Pi at the roadside. Then integrated into ODOT’s existing statewide detection network.
Speed accuracy was never documented
Schematic: the road as the camera sees it, converging to a vanishing point with tracked boxes and their motion trails, beside the same road rectified to the even ground spacing the speed calculation works in.
Data
Real roadside video captured for the independent build
ODOT’s statewide camera network, which already produced bounding boxes and vehicle class labels
Approach
YOLOv8 detection with DeepSORT multi-object tracking and OpenCV frame processing, deriving speed from tracked movement over time
Real-time inference deployed to a Raspberry Pi camera unit at the roadside
At ODOT, added metric speed estimation via perspective transformation onto the road plane, plus tracking, on top of their existing detection
Results
Working pipeline from video through detection, tracking and movement measurement to speed estimation, with detections, track IDs and motion overlaid on the video
Speed accuracy was never documented and is not estimated here
First of three ODOT assignments. Structured large pavement-performance datasets, built exploratory analytics and KPIs, and delivered interactive Power BI dashboards for engineers to compare conditions and filter datasets.