Diagnosis (artificial intelligence)

The diagnosis is a discipline of the Artificial intelligence which aims at the development of algorithms making it possible to determine if the behavior of a system is in conformity with the hoped behavior. In the contrary case, the algorithm must be able to determine as precisely as possible which parts of the system are faulty and from which types of dysfunctions they suffer. Calculation is based on the observations , which are information on the behavior.

The term diagnosis also refers to calculation like with the result of calculation. This term comes from the medical field where a diagnosis represents the explanation of the symptoms of the patient.

Example

An example of diagnosis (within the meaning of calculation) of a dynamic system is the reasoning of a mechanic opposite a vehicle. The mechanic ensures himself initially that the vehicle is broken down, then uses the observations to discover the origin of the breakdown. In this case, the observations are the state of the vehicle, the observation of the engine, the noises produced by the vehicle, the description of the symptoms by the user, etc an example of diagnosis (with the direction result) is the assertion the battery is discharged .

Diagnosis by expert system

The diagnosis by Expert system is based on the experiment available on the system to build a table of correspondence making it possible to effectively associate the observations with the corresponding diagnoses.

The experiment can be provided:

by a human operator. In this case, human knowledge must be translated into data-processing language;
by a possibly annotated recording of the preceding executions of the system. In this case, a machine algorithm of Learning must be used.

The principal disadvantages of these methods are:

the acquisition of the expertise: the expertise is available only after one certain time of use of the system, which excludes the application for critical systems (nuclear thermal power stations space or robot, for example). In addition, the complétude of the expertise is never assured. Thus, when an unknown behavior takes place on the system, the provided diagnosis will be erroneous.
the training of the expert system: the construction of the expert system is made out-line (i.e. apart from the use) and can be greedy in resources.
size of the expert system: since the expert system captures all the possible observations, it requires sometimes a very important size while a model of the system would be more compact. It happens however that on the contrary, the expert system is more compact than the model since he comprises only relevant information for the diagnosis.
it not robustness: in the event of even light modification of the system, the expert system must be entirely recomputed.

Let us notice that certain expert systems are built not starting from an expertise but directly by a compilation of the model of the system. One can thus give the example of the Diagnostiqueur de Sampath for the diagnosis of the systems to discrete events.

Diagnosis based on the model

The diagnosis is a abductif reasoning basing on the model system and on the observations carried out on the system.

The general outline of the diagnosis based on the model is the following.

One has a model which describes the behavior of the system ( artefact ). This model is an abstraction of the behavior of the system and can be incomplete. In particular, the model of breakdown (i.e. the description of the behavior of the system in the event of breakdown) is generally very partially definite because badly known. The supervisor in charge of the diagnosis also has observations on the system. These observations are provided by sensors placed on the system or can be provided directly by the system (for example when this one transmits observable internal messages or messages towards outside). The supervisor simulates the system thanks to the model and confronts the observations predicted by simulation with the observations provided by the system.

The diagnosis based on the model is a abductif reasoning. Indeed, one can simplify modeling by formulas as the following ones (where $Ab$ is the predicate indicating an abnormal behavior ( $Ab$ for abnormal in English)) :

$\ neg Ab (S) \ Rightarrow Int1 \ wedge Obs1$

$Ab (S) \ Rightarrow Int2 \ wedge Obs2$

The formulas are read in the following way: if the system does not have an abnormal behavior, then it will produce the behavior interns Int1 and the observable behavior Obs1. In the case of an abnormal behavior, it will produce the behavior interns Int2 and the Obs2 observations. Being given the Obs observations, if the behavior of the system is normal or not ( $\ neg Ab (S)$ or $Ab (S) \,$ should be determined) what results well from a abductif reasoning.

Diagnosticability

Diagnosticabilité is an attempt at translation of the English term diagnosability . One also finds the term diagnosability in the French-speaking scientific literature by imitation with English.

The diagnosticability is under field of the diagnosis. The definition varies according to the type of system considered, but one can agree on this definition:

a system is known as diagnosable if whatever the behavior of the system, the supervisor will be able to calculate without ambiguity a diagnosis.

The diagnosticability is generally calculated starting from the model of the system, and cannot thus be checked that within the framework of the diagnosis based on the model. Let us notice that the system can be diagnosable in reality without one being able to prove this property because of the loss of consecutive information to the passage of the system to his model. The question of the diagnosticability thus makes it possible to order models being at various degrees of abstraction of very the système : a more abstract model is more interesting in measurement the diagnosis is often easier to carry out on a small size, but it is less interesting if it does not make it possible any more to provide a precise diagnosis of the system.

The question of the diagnosticability is also very important when designing system since an originator seeks to find a compromise between on the one hand withdrawing sensors as much as possible (to reduce the costs), and on the other hand as much as possible to add sensors to improve the chances to detect and include/understand the dysfunctions.

Random links: