Malcolm's Talking Points: CD Mapping & Control Presentation

Use this document as a script. Each section matches one slide in mappingmodel.pptx. Read the bullet points out loud, and use the analogies when the audience looks confused.

Slide 1: CD Mapping & Control

• Open with the big picture: we help paper machines make paper that is the same thickness and quality all the way across the sheet.
• CD stands for Cross Direction — meaning the width of the paper, from one edge to the other.
• We use PRBS identification (a smart testing method), adaptive shrinkage modeling, and MPC control to do this automatically.
• Think of it like a self-tuning equalizer for a sound system, but for paper quality.
• This is Stec Corporation's solution for paper mills that want better quality and less waste.

Slide 2: System Architecture

• This slide shows the two main pieces of software that talk to each other.
• automap4.py is the brain. It sits in the office or control room and makes decisions.
• modeldriver.py is the practice machine. It pretends to be the real paper machine so we can test safely in the lab.
• The key idea: automap4 figures everything out by watching what happens. It does not cheat by reading the simulator's hidden settings.
• This matters because in a real mill, the machine does not tell us its secrets. Our software learns them anyway.
• The message system is called ZeroMQ REQ/REP — just think of it as a secure two-way radio between the programs.

Slide 3: Mapping Matrix G[i,j]

• Imagine a spreadsheet with one row for each actuator and one column for each measurement point across the paper.
• Each row shows how much one actuator affects each part of the paper width.
• We say the row sums to 1.0 because we are describing the shape of the effect, not the size.
• The real strength of the effect is figured out separately using the formula: gain = square root of 2π times sigma.
• For dilution actuators, the effect is negative: opening the valve thins the paper, like adding water to soup.
• Before we learn the real shape, we start with a best guess based on an S-curve.

Slide 4: PRBS Identification

• PRBS stands for Pseudo-Random Binary Sequence. It is basically a carefully designed random test pattern.
• Each actuator gets its own random pattern of small moves: down, hold still, or up.
• The values are -1, 0, or +1, with most of them being 0 so the machine is not moving all the time.
• Why do this? Because all 96 zones are tested at the same time instead of one by one. That saves days of testing.
• The moves are small — only 3% of full stroke by default — so the paper quality stays mostly normal during the test.
• Formula on the slide: u(t) = clamp(50 + A × PRBS(t), 0, 100). Translation: take the middle position, add a small random nudge, and keep it between 0 and 100%.

Slide 5: Cross-Correlation & Response Extraction

• After we move the actuators, we wait. The effect does not show up instantly because the paper has to travel to the scanner.
• The transport delay D is the number of scans we wait. It starts at 2 and the system learns the real number.
• We compare the profile change to the PRBS signal using cross-correlation. Think of it as matching a fingerprint.
• The response window is the small area around where we expect to see the effect — about ±35 databoxes in this older slide.
• We use a Gaussian filter to smooth out noise, like blurring a photo to remove grain.
• The window is placed using a hardcoded S-curve, not the learned model. This prevents the software from chasing its own tail and getting confused.

Slide 6: EWMA Update Rules

• EWMA stands for Exponential Weighted Moving Average. It is a fancy way of saying: blend the new measurement with the old estimate.
• Formula: new_estimate = (1 - α) × old_estimate + α × new_measurement.
• During PRBS testing, α is 0.05, so the system learns slowly and carefully.
• During normal control, α is 0.20, so the system adapts faster to real changes.
• After each update, the row is re-normalized so it still sums to 1.
• Quality metric: count how many good observations we have. Progress combines learned quality plus a bonus for unlearned zones that we can predict from nearby learned zones.

Slide 7: Shrinkage Model (S-Curve)

• When paper is made, the edges do not behave the same as the center. The sheet shrinks more near the edges.
• We model this with a third-order polynomial: s(x) = C₃x³ + C₁x.
• The starting guess uses C₃ = -0.4 and C₁ = 0.95.
• x is the normalized actuator position, from -0.5 on the left to +0.5 on the right.
• Expected center position = 250 + s(x) × 460 for a 500-databox sheet.
• As the system learns, it fits a better polynomial from real response centers. But for placing the learning window, it still uses the fallback S-curve to stay stable.

Slide 8: Transport Delay Auto-Estimation

• Transport delay is the time between moving a valve and seeing the result at the scanner.
• It is like turning on a garden hose and waiting for the water to reach the end of a long hose.
• The system estimates this delay automatically by comparing actuator movements to profile changes.
• It uses actual actuator positions, not the PRBS signs, because real positions include clipping and real behavior.
• For each actuator with enough bumps, it extracts the profile time series at the center and finds the lag with the highest correlation.
• The delay is updated during PRBS testing so the learning stays aligned.

Slide 9: Coordinate System & Normalisation

• The paper sheet can wander left and right. If we measured in absolute scanner coordinates, the same actuator would seem to affect different places every scan.
• So we normalize the profile to the sheet itself, using the left and right edges as anchors.
• Think of it like stretching or shrinking a photo so the edges always line up.
• Raw profiles use coordinates 0 to 499 for a 500-databox sheet.
• The normalized index formula converts any raw position into the sheet-relative position.
• This fix is important because without it, edge actuators could be mislearned by 40 or more databoxes.

Slide 10: MPC Supervisory Control

• MPC stands for Model Predictive Control. It is the optimizer that decides how to move the actuators.
• We use an algorithm called SLSQP, which is a standard math solver.
• The objective function has three parts: track the target, do not move too fast, and keep adjacent actuators similar to each other.
• Q, R, and S are tuning weights. In this older slide they are fixed values; in the current software they auto-tune.
• Constraints keep moves small and prevent adjacent actuators from being too different.
• Process gain is recovered using √(2π) × σ, converting the learned shape back into real-world effect size.

Slide 11: Wander Detection & Feed-Forward

• Wander means the paper sheet is moving sideways while it runs through the machine.
• We track the left and right edges with a slow smoothing filter so we see the real movement, not random noise.
• Wander is calculated as the average of the two edges minus the center point.
• We run an FFT — Fast Fourier Transform — on the last 200 scans to find the dominant frequency of the wandering.
• Confidence tells us how sure we are. If the peak is much taller than the background, we trust it.
• Feed-forward means we predict where the sheet will be next scan and compensate the actuators ahead of time.

Slide 12: Stale Scan Protection

• Sometimes the scanner stops scanning for calibration or because of a communication problem.
• The system watches the time between scans. If the gap is bigger than 40 seconds, it pauses.
• While paused, it holds the last setpoints and does not learn or control.
• When fresh scans come back, it resumes automatically.
• This protects the learned model from being corrupted by bad or missing data.
• It can be tested by triggering a scanner standardization cycle in the simulator.

Slide 13: Setpoint Ownership & Bumpless Transfer

• automap4 owns the setpoints. That means it is the only one deciding where the actuators should be.
• When it first connects, it reads the actual actuator positions and uses those as the starting setpoints.
• This is called bumpless transfer — the machine does not jerk when our software takes over.
• Priority order: stale scan holds, PRBS test moves, control moves, idle echoes positions, manual bump overrides selected zones.
• Manual bumps are ignored during PRBS so the test can move all actuators.
• The cached last setpoints are used for mid-scan polls so the machine always has a current command.

Slide 14: CSV Tuning Log

• The Record button starts a detailed log of every scan.
• Each row includes scan number, timestamp, mode, and progress.
• It also records wander data, frequency, confidence, and edge mode.
• Controller tuning values like horizon, R, S, Q, and max move are logged.
• In simulation mode, truth diagnostics show how close the estimates are to the real hidden values.
• This data can be exported to an HTML report for offline analysis and tuning review.

Slide 15: Key Parameter Relationships

• Process gain tells us how strong the actuator effect is. We get it from the learned width: √(2π) × σ.
• Transport delay is found by cross-correlation peak lag, or set manually.
• Control interval is the delay plus zero or one extra scan for stability.
• MPC horizon is at least 3 scans or delay plus one.
• Progress is the average quality across zones, plus credit for predictable unlearned zones.
• Wander confidence comes from how much the FFT peak stands above the background noise.

Slide 16: PaperMachineSimulator Constants

• This slide lists the hidden settings inside the simulator. These are the machine's secrets.
• Defaults are 96 actuators and 500 databoxes.
• The true footprint width is 4.5 databoxes — that is how wide each actuator's effect really is.
• The S-curve uses C₃ = -0.4 and C₁ = 0.95.
• Wander amplitude is 6 databoxes, and machine speed plus path length determine the real transport delay.
• Important: automap4 does not know these numbers. It has to learn them from data, just like it would at a real mill.

Slide 17: Field Report Example

• This slide shows an example of the automatic PDF report the software generates.
• The report is created during or after the PRBS phase, while the system is learning the mapping.
• It includes charts like the mapping heatmap, quality distribution, 2-sigma trend, and wander trend.
• It also shows machine configuration, process estimates, controller tuning, and recommendations.
• This gives the mill engineer a professional document they can save, email, or review in a meeting.
• Bottom line: the customer gets visible proof that the system is learning and working.

General Presentation Tips for Malcolm

• Start every slide by saying what the audience is looking at.
• Use analogies: garden hoses for delay, equalizers for control, fingerprints for correlation.
• Do not read every formula. Say what it means in business terms: faster setup, less waste, better quality.
• When you get a technical question, say you will have the engineer follow up. Do not guess.
• Emphasize the two main customer benefits: the software learns automatically, and it works the same way in simulation and on the real machine.