A machine learning model is trained to detect defective welds on a production line. The training dataset consists of images labelled 'good weld' or 'defective'. What type of machine learning is this?

Unsupervised learning — the model discovers defect patterns without labels Reinforcement learning — the model receives a reward when it identifies a defect Supervised learning — the model learns to predict the correct label from labelled training examples Transfer learning — the model transfers knowledge from weld images to other metals

A predictive maintenance system analyses vibration sensor data from motors and issues warnings 48 hours before a bearing is likely to fail. Compared to scheduled maintenance (replacing parts every 6 months regardless), what is the primary engineering advantage?

Predictive maintenance is cheaper to implement than fixed schedules Parts are replaced when they actually need it — reducing unnecessary replacements and preventing unexpected failures Scheduled maintenance causes more machine damage than predictive systems Predictive systems eliminate the need for any human maintenance technicians

An AI system is used to generate and evaluate 10,000 different bracket designs for an aerospace component, optimising for minimum weight while meeting a stress requirement. What is this technique called?

Simulation — the computer runs physics calculations on each design Computer-aided drafting — CAD software draws the designs automatically Generative design — AI explores the design space far beyond what a human engineer would consider manually Digital twinning — each design variant is mirrored in a physical prototype

A deep learning model for image recognition has many layers. What advantage do many layers provide over a simple single-layer model?

More layers means more training data is generated automatically Each layer learns increasingly abstract features — early layers detect edges, middle layers detect shapes, deep layers detect objects Deeper networks always run faster because each layer is simpler Additional layers reduce the need for labelled training data

An ML model trained on historical sales data from 2018–2022 is deployed in 2024, but its accuracy drops sharply after 6 months. What is the most likely cause?

Overfitting — the model memorised the training data and cannot generalise Hardware degradation — servers slow down over time, reducing model precision Data drift — the real-world patterns the model was trained on have changed, and the model has not been updated The model has used up its maximum number of predictions

FEA (Finite Element Analysis) divides a complex structure into thousands of small elements and solves stress equations for each. What key engineering question does FEA answer that physical testing alone cannot efficiently address?

How quickly a structure can be manufactured in a factory The exact colour appearance of a structure under different lighting conditions Where stress concentrations occur within the structure and whether any point will exceed the material's yield strength — before a physical prototype is built Whether a structure will corrode over 50 years of exposure to seawater

A digital twin of a wind turbine farm receives real-time sensor data from 200 turbines. An engineer notices the twin predicts bearing failure on turbine 47 in 3 days. What actions does the digital twin enable that a simple sensor alarm system cannot?

The twin displays sensor readings on a screen — the same as a simple alarm system The twin simulates future states to predict failures ahead of time and can model the impact of maintenance interventions before they are performed The twin automatically repairs the turbine without human involvement The twin reduces sensor hardware costs by replacing physical sensors

CFD (Computational Fluid Dynamics) simulates how air or fluid flows around a structure. For designing a new car body, what specific advantage does CFD provide over physical wind tunnel testing?

CFD provides more accurate results than wind tunnels in all scenarios CFD can test hundreds of design variants quickly and cheaply, while each wind tunnel session takes days and costs thousands of dollars CFD eliminates the need for any physical aerodynamic testing Wind tunnels cannot measure drag; CFD is the only tool that can

Parametric CAD modelling means dimensions are defined by parameters (variables) rather than fixed numbers. If an engineer changes the 'shaft_diameter' parameter from 25mm to 30mm, what happens to all features that reference that parameter?

Only the shaft itself changes — other features must be updated manually The entire model must be deleted and rebuilt from scratch with the new dimension All dependent features automatically update — holes, fillets, and mating parts that referenced shaft_diameter adjust simultaneously The parameter change is logged but no geometry changes until the engineer approves each feature individually

An unsupervised learning algorithm groups 10,000 customer records into 5 clusters based on purchasing behaviour, without being told in advance what the groups should be. This is most likely which specific technique?

Linear regression — fits a line through the customer data to predict future purchases K-means clustering — the algorithm partitions data into k groups by minimising within-group distance to the centroid Decision tree classification — the model classifies customers using labelled training data Random forest regression — predicts a continuous output for each customer

Training a large neural network requires vast amounts of data and compute. An engineer proposes using a pre-trained model (trained on millions of images) and fine-tuning only the last few layers on a small dataset of 500 domain-specific images. What is this technique called and why is it effective?

Data augmentation — generating more training images from the 500 examples Transfer learning — the pre-trained layers already encode general visual features; only domain-specific adaptation is needed Regularisation — penalising complex models to prevent overfitting on small datasets Ensemble learning — combining many small models trained on subsets of the 500 images

An AI model achieves 99% accuracy on its training dataset but only 72% accuracy on new, unseen data. What problem does this illustrate, and what is the standard remedy?

Underfitting — the model is too simple to capture the training data patterns Data leakage — test data accidentally appeared in the training set Overfitting — the model memorised training examples rather than learning generalisable patterns; remedies include regularisation, dropout, and more training data Hardware error — the GPU calculated training accuracy incorrectly