Fault Diagnosis Through Responsibility

ABSTRACT

Industry today has a severe problem in the auto~atic testing of analog cards . At the Air Force Institute of Technology, we are developing an Expert System based on the structure and function of an analog circuit card to drive auto~atic test equipment . This system uses the infor~ation contained in t he schematic diagram of the circuit as well as fundamental knowledge of electronics and past experience i n maintaining the card. One of the most important aspects of this system is its ability to reason about possible faults based upon the function of the sub- sections of the circuit. This task is accomplished using the type of "second principles" which an electronic engineer would use .

I INTRODUCTION

Testing and fault diagnosis of printed circuit cards is a ver y important task which is done many times each day . Typically this is performed using pre-determined, static and rigidly structured tests, As a result , the testing tends to be inefficient , missing many faulted components and can often isolate faults t o only a large group of components.

He intr oduce t he "theory of responsibilities" as an approach to automated troubleshooting. Using this approach , the understanding of how a circuit works is recorded by assigning responsibilities for parts of the output waveform to subsections of the circuit. These can be assigned manually ro derived

fror.1 casual simulation .

Our current efforts are directed at automated troubleshoot ing of analog circuit cards . Our overall system, described in (Milne 1984) and (Ramsey 1984) , is designed to automatically test and identify faults in an analog card through the use of automatic test equipment . Our past implementation and paper s have conducted testing based on the structural description of t he circuit. In this paper , our approach to functional testing is

described.

This work differs from others in several significant l'mys. The work of (Cantone 1985) has been entirely within the structural area. His algorithm for deciding which test to perform based on the most i nformation gained and possible cost is very helpful to structural reasoning , but can't help us once we can no longer probe within an sub­ circuit. Cantone only uses structural information to isolate the fault to a single functional area . In this paper , we will show how structural information can be used to further propose faults.

FAULT DIAGNOSIS THROUGH RESPONSIBILITY

Robert Milne

Headquarters Department of the Army Artificial Intelligence Center HQDA DAit1- DO , The Pentagon, Wash DC 20310

(de Kleer 1983) is working to diagnose faults in analoo circuit cards from 'first principles '. That is oiven the low level electronic description of ho: 0 a capacitor works, it should be possible to deduce how a filter would work , and consequently, diagnose faults in it. Although it is agreed that this is the most desirable approach, the author feels that there is too much work still to be done in order to use this approach in testing today. In our work, we start from 'second principles', that is, the type of description that an electronics engineer uses to describes various building blocks of circuits . Several examples are contained below.

(Davis 1983) has done much work in the area of digital troubleshooting based on the function of the components . He relies on computing the function and inverse of each sub- component . In general , it is not possible to compute the inverse of analog functions. Also the digital domain has a very simple output form (1 or 0) , while the analog oomain may have a very rich signal . Hence different techniques are called for. In fact the more complex output signal is one of the key differences between the analog and digital domains .

The work of (Chandrasekaran 1985) is most similar to our own . He describes the object to be diagnosed in a formal language and then compiles this description into a set of production rules to perform the diagnosis . In our approach, we use a different formal description and four simple rules of diagnosis, rather than compiling the system into production rules . Chandrasekaran's work has not defined a clear role for the interaction of functional and structural reasoning. In our work, structural reasoning plays a dual role. It is first used to isolate the possible fault to a single functional area . Secondly , whenever the output is zero, (giving no information) , structural rules are used to propose possible faults .

Qualitative reasoning (Forbus 1981) is important to justify some of our rules and approaches, but since we are working from second principles , many of these results are compiled into the descriptions.

II ST~UCTURE

The traditional role of structure is to isolate the possible fault to one functional module. The system first checks if the output is correct. If it is not , the path of the signal is traced back through the structure of the circuit and a t est is chosen which will split the possible fault path in half. This is done until only one functional module

i s left.

In our work, structure is also used when the output is zero . In this situation, we have no infor~ation on which to base the functional diagnosis, so structural reasoning must be used . I n an analog circuit, we are interested in the output current and voltage drop. Their relationship is controlled by Ohm's Law: E=IR. Qualitatively, we can see that if E is zero , then I will be zero; and that if I is zero , then E will be zero . We can also know that if R is zero , then E and I will be zero. Because of the product of IR, if E is zero we will

need two rules, since I or R could be zero .

Let us look at voltage . R in order to get a voltage shorted to ground , then output voltage . We can diagnosis rule: 'vie know that we need an drop. If an output is R is zero , and hence no translate this into a

Voltage Short Rule:

If one component connects the output and ground ,

and the output is zero,

then that component may be shorted.

\Ve could derive this rule from E=IR , but it can be 'compiled ' to the above form . This compilation is similar to (Chandra. 1985) and only needs to be done once for the fault diagnosis system . This rule is an example of the high level knowledge that an electronics engineer may use. We call this an example of a 'second principle '.

From Krichoff 's Current Law, we know that the current flowing into and out of a node is zero . If there is no current flowing into a node, then there can be no current flowing out of the node. As before , we can compile these facts into a 'second principle ' .

Current Open Rule : If one component connects the input and output, and the output is zero, then that component could be open

We also know that with no current, we get a voltage drop . If the output is zero, be from no R or no I, based upon Ohm's Law. of the dependency between the current and we cannot be sure, based upon a single zero whether 1ve have a short or open, so hypotheses may be produced . ;,ill not it could Because voltage, output , several

\~e will use a simple voltage divider as an example of these structure rules . When the circuit is working properly, the outputs are 01 and 02. Their respective values are determined by the ratio of R1 and R2. If the value of 02 is wrong when 01 is correct , then the ratio R1/R2 is wrong. Qualitatively , we cannot tell which value is wrong .

If ~1 is a shcrt , then 02 will equal 01 and net be zero . This can be predicted by the application of the Voltage Short rule . (Note tha t ground being zero is only a special case). If R1 is open, then 02 will be zero by the Current Open rule. 01 ro1ay or may not be zero depending upon the circuit . If ~2 is shorted , then 02 will be zero by the Voltage Short rule . If R2 is open, then 02 wjl l be zero by the Current Open rul e .

In this section , we have presented our approach to the use of structure for fault diagnosis. In the traditional way, it is used to isolate a fault to a single functional module . 1:/e also use it to propose faults based upon shorts and opens. These two rules are very powerful and alone will propose the correct fault for most of the examples we have encountered .

III ~UNCTION

The Rectifier Circuit:

When t he structural rules have isolated a fault to a single functional module, we turn to functional reasoning. To illustrate the theory of responsibilities presented in this paper, a simple rectifierr circuit will be used. In second principles, the rectifier circuit can be described as follows. To build a rectifier circuit, use a diode to convert each peak in the input waveform to a positive output peak. For a typical sine wave, this means two diodes, one for the positive, and one for the negative pea~ . The output is then filtered. To do this a capacitor is used to store energy and a resistor to drain that energy.

~roo this description r esponsibilities can be assigned: each diode produces a peak in the output , one at the positive peak and one at the negative peak . The capacitor charges, giving us the risins ra~p, and the resistor discharges the rarip . vie can write this formally as:

The diagnosis rules:

If X is Y by Z and not(Y), then Z is bad.

If X is Y by Z and not(X) then use structure rules.

In this brief paper , the imple~entation details have been left out, although the PROLOG user will recognise the role for unification.

These are the only troubleshooting rules we need. To diagnose a fault, the output waveform is compared with tile input waveform. The first rule will t:1en identify which component is faulted .

The above description was tested on an actual circuit . When the resistor was faulted to the open position , the output was a constant level. Couparing the desired output from the real output, we are missing the declining rarip, so the resistor is bad. The DC voltage levels are important in this diagnosis , but have been omitted for this paper.

When one of the diodes was opened , the output was missing one of its peaks. By comparing the desired output with the real output and the input , the rule can identify correctly which diode has been opened . When the capacitor is opened , the rising output is not a raop , so the capacitor is predicted bad . If the resistor is shorted , the output is zero , and the structure rules will correct ly propose the fault . It should be note that we are lucky in this case that each co~ponent ~anifests it to a single part of the output wavefor~ . In this brief paper , the voltage levels have been omitted . These are critical to diagnose sane faults .

IV DS2P FUNCTIONAL REASONING

In the above sections , we !<ave assumed that we understand how to assign responsibilities between the parts of the output waveform and the cooponents of the circuit . In this section , we will see how this can be derived auto~atically .

The basic strategy is to simulate the working of the circuit through our 'second principles '. As each component makes a contribution to the overall output , a responsibility is assigned . We then use the above troubleshooting technique to trace the responsibilities to the faulted component.

\'le will illustration. principles: use the rectifier circuit We start with the following as an second

• Diode: If the input is positive , the output equals the input .
• Capacitor : A capacitor with a rising input will charge
• Capacitor : A charged capacitor with a load , and not a rising input will discharge .
• \lire : \lhen t1-10 11ires oeet , the output is the SUI:J of the two iaputs .

We also assume low level rules that provide for the addition of signal waveforms , the testing for rising and fal ling voltage levels and give the voltage across a capacitor . The input to the clipper is in the foro of a sine wave fron a transformer . When we apply the diode rule, The sine wave is transforoed into a oostive half- wave followed by zero . The second diode is in reverse polarity , so it produces a zero output followed by a half- wave . Responsibility to each half- wave is then assigned . Next , the two wires add , giving us a

half- wave· rectified waveform.

ii011 the signal arrives at the capacitor and the input signal is rising . By the capacitor rule , the capacitor now charoes . The rise in the output voltage is assigne<J the capacitor . \!hen the signal stops r~s~ng , the capactitor has a load , so discharges. This decreasing output voltage is assigned to the resistor .

In this very ~rief explanation, we have seen how the waveforo gets built up from second principles and the responsibilities are assigned . Using this technique , it is possible to derive the resoonsiblities frota c: circuit and then perforr.1

tro~bleshootinz based upon these .

V CONCLUSION

In this paper , the " theory of responsibilities" has been outlined . This states that 1·1e can perfonJ fault diaonosis by assigning responsibilities for parts of the output to various parts of the circuit. These responsibilities are derived froQ a casual simulation of the function of the circuit . It has also demonstrated how this works on a typical buildino block circuit . In our work we have started from ' ~econd principles ', that is , the principles which an electronic engineer would use to describes each sub-circuit . This approach has proved to be very effective in diagnosing faults in analog circuits.

If one has only a limited understand ing of the circuit, then responsibilities can only be assigned in a lioited way , and hence the capability to diavnose faults will be limited . If one has a tho~ough understanding of the circuit, can do a better job of assigning the responsibilities. However , because of limitations within the doMain, it will not always be possible to describe the circuit in s uch a way as to isolate faults to one component . This limitation, however , primarily cones froo the lir:1its of being able to diagnose a circuit as a blac~ box . This technique has been applied to filters , rectifiers , voltage doublers and wave generators . Because of the modularity of :nost circuits, it works even on seer:1ingly complex circujts , by dealing with the sub-circuits .

' Although this technique is very powerful , there is ouch to be done. It should be possible to derive our second principles from first principles . \1e have found that we can generally prove the second principles in this way , but ~e can 't . d~ it autooatically , yet. A r.1ore deta~led descr~pt~on of this work is in (Milne 1985) .

REFERE~lCES

• (l] Cantone , R. ·~odel-Sased Probabilistic Reasoning for Electronic Troubleshooting" , IJCAI 82 , ~crlsruhe , Gernany , Aug 82 .
• (2] Davis , R., "Diagnosing Via Causal ~easoning:
• Paths of interaction and the Locality Pri:1ciple" , AAAI-83 , Yashington D.C ., Aug 83.
• (3] de ~(leer , J . "AI Approaches to Troubleshooting", Joint Services Wor~shop on Art ificial Intelligence
• in ~aintenance , Uct 83, Boulder , Colo.
• (4] Forbus , K. "Qualitative Reasoning about ?hysical
• Processes", IJCAI-7 , 1981. . .
• (5 ] Milne , R. "Using AI in the Test~ng of Pr~nted ~ircuit Boards '', National Aerospace and Slectronics
• Conference , Dayton , Ohio, May 1984.
• (6] ~1ilne , R. "The Theory of Responsibilities"
• SIGART !lEviS , July 1985.
• ( 7] Rarasey , J . "Diagnosis : Using Autooatic Test Equipoent and an Artificia l Intelligence Expert Syster.1 . Dec 84 , Masters Thesis , Air Force Institute
• of Technology .
• [ 8] Ser.1bugamoorthy , V. and Chandrase 1 <aran , B., "Functional Representation of Devices and Compilation of Diagnostic Probler.1 Solving Systems , to appear in Cognitive Science , 1985 .