Experimental Adaptive Model-Based 
Diagnostic System 

ABSTRACT 

The purpose of the experimental system is to realize efficient and robust 
diagnosis. To achieve this goal, we have investigated a model-based diagnosis 
technology which utilizes a model of the target device, an adaptive diagnosis 
mechanism with learning capability, and a parallel processing technology. 

FEATURES 

Model-based diagnosis 

It performs model-based diagnosis, utilizing knowledge about the struc-
ture of the target device and the function of each component. This elim-
inates the necessity of interviewing experts in order to build a diagnostic 
knowledge base. 

Adaptive diagnosis mechanism with learning capability 

It learns fault probability distribution, based on its diagnostic experience, 
in order to pinpoint the faulty component with minimum number of tests. 

Parallel processing 

Parallel computation on Multi-PSI machine reduces the computation time 
for the diagnosis and the learning function. 



System Configuration 
- 23 -

RESEARCH GOAL 

Advances in modern technology have produced complex electronic devices, 
such as switching systems, and their maintenance tasks have become seri-
ously difficult. A solution to automate the maintenance task is the expert 
system technology, which represents interviewed expert knowledge in a rule 
form. However, the technology has been applied to only small scale devices, 
because the knowledge base construction is difficult. To eliminate this prob-
lem, model-based diagnosis approaches have been investigated. Since model 
based systems utilize design knowledge of target device structure and be- 
havior, they do not require expert interviews. However, because they lack 
experiential knowledge of human experts, it can not perform efficiently. 

The research goals are, to add a function to utilize experiential knowledge to 
model-based diagnosis systems, to realize an efficient learning function which 
acquires the experiential knowledge from past cases, and to realize parallel 
implementations of the complex functions for fast computation. 







Basic System Flow 

The learning function of the system first estimates fault probability dis-
tribution for each component from the past case record. This learning is 
performed by an inductive learning algorithm, which is based on MDL mea-
surement. This complex task is processed in parallel. 
- 24 -

Then the system calculates the suspect list for given symptom and observa-
tion data. This task is performed by utilizing a model-base knowledge which 
models behavior and interconnection of the target device components. This 
model knowledge is represented in first-order predicate scheme and written in 
the logic programming language KL1. The system computes the suspect list 
by a hypothetical reasoning which calculates a set of explanations of giving 
the faulty output for the given input data on the device. 

The system repeats selecting a test and executing the test, until it reaches a 
termination condition. In order to select the most effective test, it computes 
information gain for each candidate tests by calculating entropy, based on 
fault probability distribution. This function is also complex and implemented 
in parallel. 

EXPERIMENTAL RESULTS 

Several experimental executions have been performed to show the effec-
tiveness of the parallel implementation. The computational time with 16 
processors is four to five times faster for the diagnosis function and eight to 
eleven times faster for the learning function than with one processor. The 
following figure shows the speedup by increasing number of processors. 




Speedup by Parallel Implementation of Learning Algorithm 
- 25 -

OUTLINE OF DEMONSTRATION 

An example model base for a packet switching device with 70 components 
is used for the demonstration. The system first shows the result of the fault 
probability estimation by the learning function. Then, it repeats test selection 
and execution cycle to pinpoint the faulty component, against a simulated 
faulty device model. The process of the diagnosis is shown graphically on 
the display which expresses fault probabilities with different color brightness. 
The parallel execution is performed on a Multi-PSI machine. 




Result of the Learning Function






Diagnosis Session 
- 26 -