Use the level shown on each slide to decide how to study it. Exact configuration details support the examples; memorising every constant is not a learning goal.
Core idea — Explain and apply this idea independently.
Worked example — Follow the mechanism, then transfer its reasoning to another case.
Reference detail — Know where to find the exact menu, convention or configuration; memorisation is not required.
Interactive scenes · Actual forecast-quality analysis
| Slide | Level and learning target | Reference scope |
|---|---|---|
| 1. Would you keep this pump running? | Core idea State the operational decision before selecting a model. | |
| 2. Time changes the question | Core idea Explain why order, observation spacing and context change the learning problem. | |
| 3. One physical system, eight channels | Worked example Read a multichannel recording as measurements of one physical system. | |
| 4. Keep the whole abnormal state visible | Core idea Distinguish detecting an entire abnormal state from detecting its onset. | |
| 5. An observation is more than a number | Core idea Identify timestamps, units, channels and data quality before making examples. | |
| 6. What does one row mean? | Core idea Relate the sampling clock to which frequencies can be recovered. | |
| 7. One row is not always one second | Worked example Recognise when row count and elapsed time are different quantities. | |
| 8. What did RMS erase? | Core idea Explain which raw-waveform information an RMS summary cannot preserve. | |
| 9. Same values. A different story. | Worked example Predict which statistics change when the same observations are reordered. | |
| 10. How far does memory reach? | Worked example Connect a lagged pair, its scatterplot and its lag correlation. | |
| 11. Trend, seasonality, and residuals | Worked example Separate trend, repeating structure and residual variation in an example. | |
| 12. A flat mean can hide a change | Worked example Choose a feature that notices a change in spread when the expected level stays fixed. | |
| 13. Stationarity is an assumption to inspect | Core idea Treat stationarity as a modelling assumption that requires evidence. | The exact covariance definition and the null hypotheses of stationarity tests are lookup details. |
| 14. What else can change? | Reference detail Use a diagnostic map to select a property worth inspecting. | |
| 15. Normal depends on the operating regime | Core idea Use operating context to interpret a distribution change. | |
| 16. One signal, three coordinate views | Core idea Choose time, frequency or time–frequency views for the question being asked. | |
| 17. Build a signal from simple waves | Worked example Explain how adding sinusoidal components changes a waveform and its spectrum. | |
| 18. A repeated shape creates harmonics | Worked example Distinguish integer harmonics of a repeated shape from unrelated frequency components. | |
| 19. A Fourier coefficient measures agreement | Worked example Interpret a Fourier coefficient as agreement with a sinusoidal template. | DFT normalisation factors and the special DC/Nyquist conventions are reference details. |
| 20. Keep the phase | Core idea Explain why keeping magnitudes alone loses phase information. | |
| 21. Put units on the frequency axis | Worked example Convert a Fourier-bin index into frequency using the observation clock. | The 256-sample example is a worked conversion; memorise the relation, not the selected bin number. |
| 22. Can these samples tell two waves apart? | Worked example Recognise when different continuous waves fit the same recorded samples. | |
| 23. When did the frequency change? | Worked example Use a local time–frequency view to distinguish different orders of frequency changes. | |
| 24. Choose a window. Choose a resolution. | Core idea Explain the tradeoff between localising transitions and separating nearby frequencies. | |
| 25. A wavelet stretches with the question | Worked example Predict how scale and translation change a local wavelet probe. | The corrected Morlet formula and energy-normalisation convention are reference details. |
| 26. Choose or design the wavelet for the event | Worked example Choose a probe shape for an event and recognise that wavelet design has constraints. | Admissibility and perfect-reconstruction conditions belong to the mathematical reference layer. |
| 27. STFT and CWT ask different local questions | Worked example Compare STFT and CWT localisation without equating independently normalised powers. | |
| 28. A window becomes one training example | Core idea Turn a causal observation window into one feature vector for a declared decision time. | |
| 29. Lags preserve recent history | Worked example Explain which recent history a lag vector preserves. | |
| 30. Change is not the same as state | Worked example Distinguish a sustained abnormal state from its entry and exit transitions. | |
| 31. Smoothing costs response time | Worked example Explain why stronger smoothing can delay a detector response. | |
| 32. Summaries choose what survives | Reference detail Choose a summary by what it preserves and what it can hide. | |
| 33. Spectral features summarize rhythm | Reference detail Look up spectral feature families and specify their physical meaning. | |
| 34. Relationships can fail first | Core idea Notice a broken sensor relationship even when individual channel distributions look normal. | |
| 35. Fit transformations on the past | Core idea Fit data-dependent transformations using only the allowed training information. | |
| 36. Five tasks. Five different outputs. | Core idea Specify the output object for classification, forecasting, detection, segmentation or imputation. | |
| 37. Start with a task-shaped baseline | Core idea Select a simple task-shaped baseline before comparing richer systems. | The named-model catalogue is a lookup aid; choosing an honest task-shaped baseline is essential. |
| 38. What was available at decision time? | Core idea Decide which observations are available at forecast issuance and at state scoring. | |
| 39. Overlapping windows are not independent | Core idea Keep dependent windows together when evaluating generalisation to a new recording. | |
| 40. Roll the forecast origin forward | Worked example Compare expanding and rolling origins with matched forecast horizons. | |
| 41. Hold out the identity you expect to meet | Core idea Match group and time restrictions to the deployment claim. | |
| 42. Guard the information boundary | Core idea Separate unavailable training labels from legitimately available forecast context. | |
| 43. Choose the split for a new person and a later date | Core idea State separately which rule protects identity and which rule protects time. | |
| 44. Choose a feature for a new fault | Core idea Name the property that changed, then explain why an individual-channel threshold can miss it. | |
| 45. Sources, credits, and reproducible controls | Reference detail Locate the sources, assumptions and provenance needed to inspect a claim. |
| Slide | Level and learning target | Reference scope |
|---|---|---|
| 1. What is the signal telling us? | Core idea Match an output contract and its evaluation to the operational question. | |
| 2. Serega has a clever forecaster. Is the pump healthy? | Core idea Explain why an accurate forecast does not by itself establish healthy operation. | |
| 3. Five tasks. Five different outputs | Core idea Identify the distinct outputs produced by the five task types. | |
| 4. Classification: a label for an observed window | Worked example Map an acquired multichannel epoch to one class under a stated split. | |
| 5. Forecasting: predict values that have not arrived | Worked example Separate observed history, a forecast horizon and known-future inputs. | |
| 6. Anomaly detection: unusual relative to what? | Core idea Separate an anomaly score from the thresholded alarm decision. | |
| 7. Scores rank; thresholds decide | Worked example Separate a fixed score ranking from decisions made at one cutoff. | |
| 8. Segmentation: label the timeline, or find its boundaries | Worked example Distinguish assigning states to a timeline from locating boundaries. | |
| 9. Imputation: what information may cross the gap? | Worked example Choose an imputation strategy consistent with online or offline information. | |
| 10. The architecture expresses an assumption | Core idea Connect local filters, recurrent state and attention to different modelling assumptions. | |
| 11. A causal TCN grows its field of view | Reference detail Inspect how dilation and depth determine a temporal receptive field. | Exact receptive-field counts depend on the stated kernel, dilation and depth configuration. |
| 12. Memory and attention answer different questions | Core idea Explain how recurrent summaries and attention access visible history differently. | |
| 13. EEG: time is only one of the axes | Core idea Read electrode and time axes while recognising the change from a pump task to EEG classification. | |
| 14. A classical EEG pipeline already has useful structure | Worked example Connect filter banks, spatial projections and feature summaries to a classical EEG classifier. | |
| 15. A temporal kernel has a frequency response | Worked example Relate a temporal kernel to the frequency structure it emphasises. | |
| 16. One convolution, computed by hand | Worked example Compute a temporal filter output and distinguish it from spatial channel mixing. | |
| 17. Temporal first. Spatial second. | Core idea Identify the axis mixed by a depthwise spatial operation. | |
| 18. Separable convolution: refine, then mix | Worked example Distinguish per-map temporal refinement from pointwise feature-map mixing. | The 512-versus-4096 weight count is a reference calculation for one configuration. |
| 19. Follow one EEG epoch through the network | Worked example Follow which EEGNet operations change maps, electrodes and time. | The 64 × 512 input and 256 → 2 classifier are one reference configuration; preserve the axis logic. |
| 20. An EEG score is only as honest as its split | Core idea Match an EEG claim to held-out trials, sessions or people. | |
| 21. Transfer the filtering idea back to pump sensors | Core idea Which axis collapses, and which axis must stay separate? | |
| 22. What changes when we pretrain once? | Core idea Distinguish sharing pretrained forecasting weights from calibrating a downstream alarm. | |
| 23. Chronos-2: a modern forecasting model we can read | Worked example Trace historical sensor values through patches and attention to forecast quantiles. | Six blocks, eight heads, hidden size 512, REG and the 13-quantile grid are checkpoint details. |
| 24. A forecasting task is data plus a visibility contract | Core idea Specify which values are observed, missing or genuinely known in the future. | |
| 25. Scaling keeps units from dominating the network | Worked example Use historical statistics to transform a series and map quantiles back to its units. | The exact asinh formula and example statistics are reference details; historical-only fitting is essential. |
| 26. A missing value and a measured zero are different | Core idea Explain why a missing value and an observed zero require different masks. | |
| 27. A patch becomes a vector, without value quantization | Worked example Explain how values, time and an observation mask become one patch representation. | The full 16-position table, division by 8192 and 512 embedding coordinates are implementation details. |
| 28. Time attention reads within one variable | Core idea Explain how bidirectional encoding can avoid seeing unknown future targets. | |
| 29. Group attention reads across related variables | Core idea Use group IDs to permit related series to exchange information and isolate unrelated tasks. | |
| 30. The head predicts quantiles for the whole horizon | Core idea Interpret predicted quantiles and distinguish nominal intervals from measured coverage. | |
| 31. Quantile loss makes over- and underprediction cost different | Worked example Explain why a high quantile penalises underprediction more strongly than overprediction. | The exact pinball formula and implementation scale factor are reference details. |
| 32. Pretraining teaches transferable forecasting structure | Core idea Separate the pretraining objective from inference with fixed weights on a new context. | |
| 33. A forecast can become comfortable with a fault | Worked example Explain how a forecast can become accurate inside a sustained abnormal regime. | |
| 34. A detector is a chain of choices | Core idea Trace a complete detection pipeline from observations to a calibrated alarm. | |
| 35. The lab question: detect the entire abnormal state | Core idea State the full-state detection target and the permitted fit, calibration, validation and test roles. | |
| 36. Macro and pooled answer different questions | Worked example Explain which unit receives equal weight in each reported metric. | |
| 37. Rolling features help this model on validation | Worked example Read a validation representation ablation without claiming general transfer. | |
| 38. Compare the complete validation pipelines | Worked example Read the result as a comparison of complete calibrated detection pipelines. | |
| 39. Useful ranking can fail at the chosen cutoff | Worked example Separate useful score ranking from the feasibility of a threshold under an alarm budget. | |
| 40. Averaging helps here; max inherits the saturated tail | Worked example Explain a fusion result using score behaviour and calibration rather than a universal rule. | |
| 41. Compare complete anomaly-detection pipelines | Worked example Compare selected systems and state exactly what their validation results support. | |
| 42. Better forecasts do not guarantee better state detection | Worked example Distinguish average forecast error from discrimination of the entire abnormal state. | |
| 43. Read the horizon and check the interval | Worked example Read empirical coverage together with width and forecast horizon. | |
| 44. Read the score, then read the actual alarm | Worked example Explain a success or failure from raw scores, frozen thresholds and actual alarms. | |
| 45. Serega can read the forecast. You define the question. | Core idea Defend a task definition, an information contract and an honest baseline. |
Select an option, then write a short explanation or discuss it aloud and mark the checkbox. Reveal the reasoning only after this attempt. Compare your explanation with the feedback; the written response is for reflection and is not automatically graded or uploaded.
The PDF has a question frame followed by a reasoning frame for every transfer check.