From registration to information, I Anders Ringgaard Kristensen - - PDF document

1

Slide 1

From registration to information, I

Anders Ringgaard Kristensen

Slide 2

Outline

The decision making process The role of models Basic production monitoring as a source of information Key figures and their properties Interpretation of key figures Limitations of traditional production monitoring

2

Slide 3

Focus of this lecture

In last lecture, we discussed:

• Utility concept
• Planning, classical

approach

In this lecture we direct our attention towards the other side of the management cycle:

• Production monitoring

We shall, however, try to look at it from a decision making perspective.

Slide 4

Making decisions

Decision making is based on knowledge:

• General know ledge: What you can read in a textbook on

animal nutrition, animal breeding, agricultural engineering etc.

• Context specific know ledge: What relates directly to the

unique decision problem. Examples:

• The milk yield of dairy cow No. 678 when considered for

culling.

• The estrus status of sow No. 345 when considered for

insemination.

• The current daily gain of the slaughter pigs in House 5

when considering whether or not to adjust protein contents of the diet.

When knowledge is represented in a form that may be used directly as basis for a decision, we call it information.

3

Slide 5

Information sources

General knowledge:

• Look in a textbook
• Ask an expert

Context specific knowledge:

• Obtained through registrations (observations) in the

herd:

• Traditional registrations
• Test day milk yield, cow 567
• Litter size of sow 123
• Sensor based registrations
• Conductivity or temperature of milk from AMS
• Accelerations of a sow from a censor node in

an ear tag

• Computer vision (image analysis)

Slide 6

From registrations to information

We refer to a collection (typically in a database) of registrations of the same kind as data. We don’t use data directly for decision making (huge amounts of data). Before we can use data we need to reduce it through some kind of processing. The resulting information is used for decision making (which again requires processing: optimization).

4

Slide 7

A simple example of the path

Data: Test day milk yields Processing I: Calculating cumulated yields for individual cows over a standardized period and afterwards calculating the herd average. Information: Average milk yield in the herd. Processing II: Linear programming using the Simplex algorithm. Decision: Least cost feed ration for the cows. The path from test day milk yields to feed ration is not unique:

• Both processing steps could be replaced by other methods.
• Choosing a wise processing of data into information is an

important issue in herd management!

Slide 8

Cécile Cornou, LC-2373, IPH, KVL

Advanced example: Hogthrob

Activity Measurements – in Group Housed Pen

• Accelerometer fitted on neck collar
• Acceleration in 2 and 3 dimensions
• Four measurements per second
• Transfer PC via Blue Tooth
• Twenty days (March 1st to March 20th 2005)

Video Recordings

• Four cameras used as web cam
• Twenty days

5

Slide 9

Cécile Cornou, LC-2373, IPH, KVL

Hogthrob: Acceleration data

Eating Walking Rooting Sleeping

Slide 10

Cécile Cornou, LC-2373, IPH, KVL

DLM analysis

Main Results – Sow 5 V-mask

Specificity: 100 % Error Rate: 0 % (FP= 0)

Tabular Cusum

Specificity: 90 % Error Rate: 50 % (FP= 1)

6

Slide 11

An advanced example of the path: Hogthrob

Data: Accelerations of a sow measured 4 times per second in 3 dimensions Processing I: Online time series analysis of the acceleration data using Dynamic Linear Models Information: Sow in heat? (yes/ no). Processing II: Dynamic programming. Decision: Inseminate/ leave open/ cull Notice the reduction in the dimensionality of the information (one binary variable) compared to the data! Next step: Try another kind of processing.

Slide 12

The decision making process: Summary

The purpose of monitoring is to improve the decisions Processing of registrations into information is necessary Choosing the best processing is a key issue Information is a tractable representation of context specific knowledge. Monitoring is the sub-path from registration to information During this course we will follow the path from data to decision:

7

Slide 13

Uncertain information

Classical methods assume certainty. In real life, certain information hardly exists. However, the degree of uncertainty varies. In general, we use distributions for representation

• f knowledge with uncertainty.

For binary information, we may just supply the probability. For continuous information, we supply for instance a normal distribution with a mean and a

• variance. A small variance implies that we are

rather certain about the value. For consistent processing of data, a model is needed.

Slide 14

Uncertainty: Pregnancy status of a cow

For the replacement decision we want to know the pregnancy status of a cow, but it is (most often) unobservable, so we have only indirect observations:

• If a cow has not been inseminated, the probability of

pregnancy is zero.

• If it has been inseminated, the probability is 0.4,

because the conception rate of the herd is 0.4.

• If, after 5 weeks, the cow has not shown heat, the

probability increases. Assuming a heat detection rate of 0.5, the probability increases to 0.7.

• A positive pregnancy diagnosis will further increase the

probability, but only a calving will increase it to 1.

We need a method for consistently combining the indirect observations into an updated probability of pregnancy. We may use a Bayesian Network model for that!

8

Slide 15

Models for monitoring under uncertainty

Assessing the distribution of a key figure from the distribution of data (later this lecture):

• “Black box” approach

Statistical quality control models based on time series analysis (Tuesday 8-10):

• More or less “black box” approach

Dynamic Linear Models based on Bayesian updated time series with Kalman filtering (Tuesday 10- 12):

• Structured model
• Data of same kind

Bayesian networks (Tuesday afternoon)

• Highly structured models
• Data from different sources

Slide 16

Basic production monitoring

Traditional approach to production monitoring:

• Registrations are collected systematically in the herd
• Data is entered into a Management Information

System (MIS).

• At quarterly (or monthly) intervals, the MIS

calculates a bunch of key figures which are presented to the farmer in tabular form in report.

• The farmer looks at the key figure and decides

whether or not to make adjustments to production.

The information provided is the list of key figures. We shall briefly discuss how to interpret such key figures in a sound way.

9

Slide 17

Focus on the monitoring process

Slide 18

Principles of production monitoring

Data recording Database Data processing Report with key figures Analysis

• Statistical
• Utility

Decision

10

Slide 19

Data recording

Events:

• Mating
• Farrowing
• Weaning
• …

Slide 20

A record (or ”registration”)

Event (mating/ farrowing/ weaning … ) Identification (sow # / section # / batch # ) Registration level (animal/ section/ batch/ herd) Time (date or date/ hour/ minute) Property (what is measured) Value (numerical or categorical)

11

Slide 21

A record, example

Event: Farrowing Identification: Sow # 1234 Registration level: Animal Time: February 23rd Property: Live born piglets Value: 12

Slide 22

Principles of production monitoring Data recording Database Data processing Report with key figures Analysis

• Statistical
• Utility

Decision

12

Slide 23

Records: Collect in a table

Basically a table for each event, but several properties

• Example: Farrowing

Easy 1 14 18/ 1 04 1001 … … … … … Diff. 4 9 16/ 1 04 678 Easy 2 12 15/ 1 04 1234 Course Stillborn Liveborn Date Sow #

Slide 24

Database

A collection of tables for different events

Easy 1 14 18/ 1 04 1001 … … … … … Diff. 4 9 16/ 1 04 678 Easy 2 12 15/ 1 04 1234 Course Stillborn Liveborn Date Sow #

13

Slide 25

Principles of production monitoring Data recording Database Data processing Report with key figures Analysis

• Statistical
• Utility

Decision

Slide 26

Data processing

Ask questions to the database (the SQL language):

• Number of events in a period
• Matings per week: # matings/ # weeks
• Farrowings per week: # farrowings/ # weeks
• Averages over a period
• Liveborn piglets per litter

14

Slide 27

Data processing: Complex, I

Total gain, slaughter pigs (this period):

• + Total weight, slaughtered pigs
• + Total weight, dead pigs
• + Valuation weight, end
• – Total weight, inserted pigs
• – Valuation weight, beginning

Daily gain = Total gain / Total days in feed Data sources?

Slide 28

Principles of production monitoring

Data recording Database Data processing Report with key figures Analysis

• Statistical
• Utility

Decision

15

Slide 29

Interdependencies

Key figures are heavily correlated:

• Logically (refer

to figure)

• Biologically

Slide 30

Report with key figures

Key figures have 3 basic properties

• Correctness
• Are all registrations correct (right animal(s),

right event, right value etc.)?

• Validity
• Does the key figure express exactly what we

want to know?

• Precision
• Standard deviation of estimate – Exercise.

16

Slide 31

Correctness

Examples …

Slide 32

Validity

Example: Reproduction in sows Discussion:

• What do you want to know?
• Which figure(s) will provide us with the desired

information?

17

Slide 33

Key figure: Farrowings per week

Utilization of the farrowing department. One of the most important elements from an economical point of view. Presented as an average value. What is of interest is the distribution over weeks.

Slide 34

Principles of production monitoring

Data recording Database Data processing Report with key figures Analysis

• Statistical
• Utility

Decision

18

Slide 35

Precision

Formal definition. The precision ρ is defined as ρ = 1/ σ2 where σ is the standard deviation of the calculated key figure. Let kc be a calculated key figure. We may model kc as follows: kc = kt + es + eo , Where

• kt is the true (unobservable) underlying value
• es is the sample error (the ”biological” variation)
• eo is the observation error (depends on the method)

฀ σ2 = V(es) + V(eo)

We should regard kc as the estimated value of kt, where σ is the standard deviation of the estimate

Slide 36

Precision, statistical evaluation

Farrowing percentage The percentage of matings resulting in a farrowing (sold pregnant sows are excluded from the calculations). Assume that 90 % was expected, but the calculated key figure was 87 % . Should we be worried?

19

Slide 37

Farrowing percentage, I

How many matings are included in the calculation? Assume N. If the expected value is p = 90 % , we want to test the null hypothesis H0: p = 0.90 The number of matings, n, resulting in a farrowing will be binomially distributed with the parameters p = 0.90 and N.

Slide 38

Farrowing percentage, II

Assume that there is no observation error, i.e. V(eo) = 0. The variance of n is equal to the sample variance, which (in a binomial distribution) is: σ2 = Np(1-p) = N × 0.90 × 0.10 = 0.09N

20

Slide 39

Farrowing percentage, III

Evaluation of the result given N

1.7 9 270 261 5.2 27 300 2.0* 12 360 348 6.0 36 400 180 90 Exp. 6 3 Dev. 1.4 174 4.2 18 200 1.0 87 3 9 100 Dev./ σ n σ σ2 N

Slide 40

Principles of production monitoring

Data recording Database Data processing Report with key figures Analysis

• Statistical
• Utility

Decision

21

Slide 41

Utility evaluation

If a deviation is significant from a statistical point

• f view it should also be evaluated from a utility

point of view. Far more difficult than the statistical evaluation! Example: Does it matter that the pregnancy rate is lower than expected? An obvious tool for utility evaluation is a simulation model.

Slide 42

Limitations of traditional monitoring

Dependence on defined targets (expected results). Most often, the precision of the key figures is not calculated. Correlations between key figures are typically ignored. Autocorrelations over time are typically ignored. The processing of data into information is very simple (even though it may be computationally demanding) with loss of information as a consequence. Acknowledgement: Slides 8-10 provided by Cécile Cornou