Support for Diagnosis: Diagnostic window
From DSL
The Diagnostic window is a special feature of GeNIe that facilitates diagnosis.
Consider the Car Network model shown below:
To perform the Car diagnosis, the Battery, Dirty Connections at Starter and Starter have been defined as Target Nodes. The Battery Test, Dash Board Lights On and Starter Noise nodes have been defined as Observation, Ranked Nodes. The Start node has been defined as mandatory and ranked. To begin diagnosis, select Test Diagnosis [ Image:DiagnosisButton.jpg ] button on the toolbar. Once the Diagnostic window has opened, it will appear as shown in the figure below.
The diagnostic window has four panes and they can be characterized as follows:
The top left pane is for the Target nodes, which are non-observable and ranked nodes, typically representing various faulty components.
The top right pane is for the Ranked Observable nodes that have not yet been observed. All ranked observations in this pane are not yet instantiated (i.e., they exclude the Mandatory observations, which have to be instantiated before the diagnostic procedure starts). If the mandatory observations have not been instantiated, it will be listed in the Other Observations window pane as shown below.
To instantiate the mandatory observation, right click the node names listed in this window pane , which is " Start " in our example. The states of this node will open as a pop up menu. Select the relevant state / observation by left clicking the mouse. Once the node is instantiated, it will be listed in bottom right pane as shown in the first Diagnostic Window.
The bottom left pane is for Observable nodes that can been instantiated by the user. The Mandatory nodes will be listed in red (the color red indicates, the node has to be instantiated by the user to begin diagnosis), and the other observable nodes will be listed in black.
The bottom right pane is for the Observable nodes, that have been observed by the user.
The specifications applies to all models and the testing diagnosis window will always be characterized this way.
Entropy/Cost Ratio
At the top of the Diagnosis window, there is a slider labeled Entropy/Cost Ratio. The Entropy/Cost Ratio describes how the measure of informativeness of an observation or a test is combined with its cost. The Entropy/Cost Ratio can range between 0 and 99999. The contents of the box next to the Cost/Entropy Ratio represents the highest value pictured on the slider and can be changed from 0 to 99999. If the Cost ratio is zero, then the Test Ranking Score is based solely on Test Effectiveness with disregard to the costs of testing. Any higher number determines the importance of cost in the ranking.
Here is an example of the how the equation can be applied. This table is generic with some node properties.
| Prob | Test | Cost | Fault |
|---|---|---|---|
| Sigma) | T1 | C1 | F1 |
| Sigma) | T2 | C2 | F2 |
The basis for ranking will be calculated for the table according to the following equation:
E(F1) = P(F1|T1) + alpha * C1(T1).alpha
the Entropy/Cost Ratio as defined above, is the coefficient combining these two disparate factors. The creators of the models can modify alpha. We will now look at an example that will illustrate how alpha affects the ranking. We can see in the two figures below that the ranked observations are seriously affected by Entropy/Cost Ratio, which so low in the first figure and high in the second. The ranking in the first figure, where alpha=0, is based almost solely on the test informativeness. The second figure, however, has a high value of the Entropy/Cost Ratio (alpha = 0.455). We observe a drastic difference between the two figures in terms of the numerical values of the test effectiveness and the resulting test ranking.
By default the target node with the highest probability of fault will be selected for diagnosis. To pursue other faults, you can right click other targets in the top left pane. In this example you can pursue other faults such as 'Starter' or 'Dirty Connections At Starter' apart from the 'Battery node' as shown in the figure below:
Options Menu
Update Immediately: If this option is checked, then the diagnostic window is updated immediately after each change.
Enable Relevance: It is a heuristic for ranking tests and faults that are relevant to the existing set of evidences. The reason behind this heuristic is to exclude those faults and tests which are ranked highly purely because of their priors.
Enable Quick Tests: This option will move the ranked observations with a negative cost ( inexpensive) to the top of the ranked observations list .
Mark Indisputable Targets: If this option is enabled, any target ( fault) nodes with a probability value of 1 or 0 , when right clicked with a mouse on the left window pane will display a message" Can't Pursue Indisputable Fault".
Sort Past Observations by Name: In the bottom right of the window pane, all past observations will be listed in the order of instantiation or name. If this option is enabled, it will list the observations by name.
Nonlinear Scale for E/C ratio: When this option is enabled it switches the scale from linear to logarithmic. This is very useful when the E/C ratio is large.
Multi-fault Algorithm:
GeNIe has algorithms that are based on two approaches :
- . Joint Probability Approach [ first 6 options in the Multi-fault Algorithm sub menu ]
- . Marginal Probability Approach. [ last 2 options in the Multi-fault Algorithm sub menu ]
The Marginal Probability Approach is much faster, however it is not as accurate at the Joint Probability Approach which is more time consuming.
Joint Probability Approach
Basically this approach is separated into two areas, the area of copulas, and the area of differential diagnosis. The first area provides marginal based approximations for the joint probability distribution. The second area basically allows the user to pursue and differentiate between any possible set of scenarios. This area is necessary for presenting the enormous number of scenarios to a user and still providing the user with the ability to pursue any set of scenarios.
Differential Diagnosis
A way in which human diagnosticians cope with the complexity of diagnosis, is by counter opposing competitive hypothesizes and seeking evidence that differentiates between them. This approach, which is referred to as differential diagnosis, is considered standard in medical science, but only applied in a simplified form in current diagnostic expert systems.
Suppose differential diagnosis is performed on the hypothesis set H = {TC,LC,BC} in the DPN with all the possible scenarios where TC, LC and BC are different diseases that may be present in the patient.
Now assume a doctor wants to investigate whether a person has only one disease present. In this case four partitions are formed, three partitions with one disease present and other diseases absent and one partition which contains all other scenarios.
Partition 1 : TC-present, LC-absent, BC-absent
Partition 2 : TC-absent, LC-present, BC-absent
Partition 3 : TC-absent, LC-absent, BC-present
Partition 4 : All other scenarios
Another interesting case is to pursue the scenario that all diseases are present. The partitions, are then, one partition with all diseases present and one partition with all other scenarios. These are but a few of the large no. of partitions possible.
GeNIe takes away the work of defining these partitions from the user, and gives the user three interesting options to make the partitions automatically.
They are :
Partitions with one or more targets: This distribution consists of partitions, where each partition contains a scenario with at least one target present and additional one partition for the rest of the scenarios.
Partitions with only one target: This distribution consists of partitions, where each partition contains a scenario with at most one target present and additional one partition for the rest of the scenarios.
Partitions with all the targets: This distribution consists of partitions, where each partition contains a scenario with all the targets present and ad-ditional one partition for the rest of the scenarios.
These directly map to the following options in the Multi-Fault Algorithm menu, namely :
(In)dependence/At Least One (In)dependence/Only One (In)dependence/All
Marginal Probability Approach
The following two algorithms use the Marginal Probability Approach. They differ in the function they use to select the tests to perform. Both functions are scaled so that they return values between 0 and 1.
Marginal/Strength 1: It uses a function without the support for maximal distance and its minimum is reached when all probabilities of the targets are equal to 0.5
Marginal/Strength 2: It uses a function that has support for maximal distance and is continuous on the area [0,1]
