Day3_​S2_​ok

Stage 1 - Read and align

Read what S1 wrote and put every row on an hour clock

Stage 1

1. Read the S1 hourly file, 2,894 rows
2. 8 hour backward mean of flow = ma8
3. Watch window start as a real date
4. Hours from 1 Jan, negative = before

Stage 2 - Fit the baseline on 2024

Three models fitted on the reference window only, each one scored

Stage 2

1. Split reference (743 h) from watch (2,151 h)
2. temp ~ flow now, R2 0.8015
3. temp ~ ma8, R2 0.9424 - the driver wins
4. Reference mean and SD become flow variables

Stage 3 - Score the watch window

Turn the residuals into three z descriptors and flag 3SD

Stage 3

1. z_temp, z_vib and z_ratio on one axis
2. z_ratio uses no model at all, for comparison
3. Alarm when |z| > 3
4. Roll the hours up to one row a day

Stage 4 - Confusion matrix

Score each descriptor against the work order, day by day

Stage 4

1. Label every day TP / FP / FN / TN
2. Unpivot, then pivot to a 3 x 4 matrix
3. Empty cells become 0 before dividing
4. precision and recall per descriptor