ACE: An Expert System for Telephone Cable Maintenance

ACE: AN EXPERT SYSTEM FOR TELEPHONE CABLE MAINTENANCE

Gregg T. Vesonder*, Salvatore J. Stolfo**, John E. ZielinsKi*,

FredericK D. Miller*, David H. Copp

• *Bell Laboratories , Whippany , NJ.
• **Columbia University, New YorK, NY

Telephone company engineers use GRAS to pinpoint trouble spots in the local telephone network, to predict future work force . r~qui~ements and to drive plant rehab1l1tat1on decisions. ACE has subsumed some of these tasks and uses its knowledge of cable analysis to retrieve reports from GRAS and to provide "bottom line" information and recorrmendations concerning trouble spots.

Although the problem domain of ACE appears fundamentally different from that of R 1 /X~ON, [ 3] the Digi ta 1 Equipment Corporat1on VAX computer configuration program reported by McDermott, there are enough common characteristics of both domains that suggested the organization of R1 / XCON is suitable for ACE 's app!icatio~ . . Consequently, many of the des1gn dec1s1ons faced in the development of ACE were guided by the methods used in

R1/XCON .

Specifically, the ACE inference engine is a forward-chaining Production system executing the Match problem solving paradigm. The declarative nature of the OPS4 production system notation[4] effected the rapid development of the ACE knowledge base. The team of Knowledge engineers (cognitive psychologists ), who interviewed several human experts, had little difficulty operationalizing the acquired knowledge directly in OPS4. Subsequent cycles of test, debug and modify were carried out with relative ease. The end result was a working, useful prototype ahead of schedule . An advanced version of the system will be present in the telephone companies in the

near future.

The following sections detail the problem domain and organization of ACE and describe how the development and installation of a prototype of this system in a "live" production environment

was completed in a cost effective manner.

ABSTRACT

ACE, a system for Automated Cable Expertise, is a Knowledge-Based Expert System designed to provide trouble­ shooting reports and management analyses for telephone cable maintenance . Design decisions faced during the construction of ACE were guided by recent successes in expert systems technology , most notably R1 / XCON, the Digital Equipment Corporation VAX configuration program. ACE departs from "standard" expert system architectures in its use of a convent ional data base management system as i ts pr imary source of information . Its primary sources of Knowledge are the users of the database system and primers on maintenance analysis strategies.

I. lntroduct ion

In this paper we describe ACE, a Knowledge-based system for Automated Cable Expertise. It was designed to provide support for directing telephone cable maintenance . The development of ACE, begun by Bell Telephone Laboratories two years ago, demonstrates that Artificial Intelligence techniques can be applied to significant and practical "real-world" problems in an efficient manner. The developers of ACE were able to implement, test, modify, and verify the performance of the system ahead of predicted schedules. Furthermore, ACE demonstrates a successful merger of two complementary yet independent technologies : database and Knowledge base systems .

UnliKe, for example, MYCIN[1] and MOLGEN[2] ACE is not a consultant . It is an automatic analysis System digesting hundreds of telephone cable maintenance reports daily. This data is provided by another computer system, CRAS (Cable Repair Administration System), a data management and report generation system .

II . The Problem Domain

In normal operation the telephone company supports a telephone line from a residential or business site to telephone company switching machines in the central office. This line is called a cable pair. A collection of pairs are bundled to form the cables that hang from telephone poles or res ide underground. A collection of cables form a wirecenter . These three levels form the bulk of the local telephone network and the cable maintenance force concentrate their efforts at all three levels. A more detailed description of the local telephone network can be found in a Bell System handbook. [5]

A variety of electrical faults and environmental conditions can cause failure in one or more cables or individual pairs . (Insect infestations in terminal boxes and gnawing rodents are perennial problems.) An important and expensive operation performed by the operating companies is the general maintenance and rehabilitation of these 1 ines.

Customer generated maintenance reports provide information for identifying trouble spots within the local network. In a high density geographic area, the logging and tracking of failure reports has become an expensive and important data processing operation.

In order to identify trouble spots for repair and rehabilitation, many telephone companies use GRAS to monitor the repair activity of the local network on a daily basis. Trained analysts routinely examine impressive volumes of data and attempt to identify spots for maintenance to prevent further disruption of service to customers.

GRAS provides a set of report generating programs each producing a specialized summary of various aspects of customer .generated and employee generated repair tasks . An individual record maintained by GRAS consists of numerous fields detailing a repair task reported by a customer or employee . This is referred to as a trouble in telephone company jargon . Many distinct GRAS records representing troubles may refer to the same pair or cable, indicating a potential chronic problem.

The large volume of detailed information permits the human expert to make an informed selection of candidates for rehabilitative maintenance . However, the limited number of specialists available, and the size of the database inhibits the timely analysis and reporting of persistent problem areas that require rehabilitation. The backlog delays the assessment of future work force needs and the selection of prospective areas for maintenance. The approach of installing an expert system as an adjunct to the GRAS system, assisting management decision making, was proposed as a solution to the problem of timely and accurate selection of areas for rehabilitation.

III. The System Organization

Knowledge-based expert systems have constructed, typically from two loosely coupled modules that form the problem solving engine. (see figure 1)

The knowledge base contains all of the relevant domain-specific information permitting the program to act as a specialized problem solver. Within ACE the knowledge base is represented in rule form. The inference engine controls the ~eductive process. This control strategy 1s also represented in rule form within ACE .

In ACE a third component has been added: a database . Over the past decade database technology has progressed from dull-witted, efficient systems to systems offering abstract data models to facilitate semantically based retrieval functions. However, few, if any, offer the ability to deduce new data from old. At best aggregate data functions can be applied but deductions from raw data are not possible . On the other hand , knowledge based expert systems have been implemented as consultants to specialists in narrow domains . These systems are designed as problem solvers. Expert systems rarely operate with a large amount of data .

ACE attempts to merge the two technologies and this merger may be useful in a variety of applications. Kellog[6) reports on the Knowledge Management I (KM-I) system and its organization is similar in scope to that represented by ACE. However, KM-1 is designed as an intelligent front-end, interfacing the database to the user by an English-like query fac i lity . This makes the database easier to use by making the database access language closer to the language of the application area.

In contrast , in ACE the expert system is the user! Tailored to solve a single class of problems, the knowledge base component of ACE automates the tasks the database supports. Thus the expert system not only answers the questions but also poses them .

A. Details of the Implementation

The knowledge base and inference engine of ACE are written in Franz LISP[7] and the OPS4 Production System language running on a DEC VAX 11/780 . Supporting routines for ACE, including CRAS interfacing and electronic mail facilities (to send ACE output to selected user mailboxes ) are written in resident UNIX* software . The gross system level organization is shown in figure 2.

* UNIX is a registered trademark of Bell Telephone laboratories.

Note that a user is not necessary to initiate ACE on a daily basis . ACE is standalone, awakened nightly by UNIX to study the day 's transactions. The only record an ACE user sees is an electronic mail message on trouble spots it discovered. (see figure 3)

From ace Thu Apr 14 15:42:27 1983

ACE 1.0 output message follows:

···············ACE REPEAT MESSAGE···············

In cable 47 wirecenter 999550 I cable troubles were reported yesterday. There were 15 cable troubles in the cable in the past thirty days. 4 pairs were in the complement range from 600 to 700. 75 percent of the addresses were williams.

The most common disposition code was 0436. This indicates the presence of a problem at one or more crossboxes in this cable. In all likelihood it is one crossbox. This crossbox should be cleaned, rehabilitated and properly closed up, to prevent further troubles, or it should be replaced if it has deteriorated to a point where that is necessary. The detail data for cable 47 is:

End of ACE transmission.

Figure 3: A Sample ACE Message. Produced by Real Data.

B. Details of the Problem Solving Engine

Within ACE, the body of knowledge about wirecenters, CRAS data and commands and analysis strategies is embodied by an OPS4 program.

1. Production Systems In OPS4, data elements in Working Memory have the form of arbitrary LISP list structures. The user can also define other condition and actio~ predicates, the former testing for a var1ety of conditions in Working Memory and _ t~e l~tter adding functions beyond mod1f1cat1on of Working Memory. An English language equivalent of an ACE .production rule is presented in figure 4.
2. IF a range of pairs within a cable have generated a large number of customer reports
3. ANDIF a majority of the work on those pairs was done in the terminal block
4. THEN look for a common address for those repairs .

Figure 4: English Language Equivalent of an ACE Production Rule.

1. Characterizing the Problem Domain The forward-chaining, data-dr iven approach typified by R1 /XCON was chosen for ACE since cable analysis is primarily a bottom-up , data-driven task.

_Using t~e !axonomy of problem doma1ns descr1bed 1n[8] we classify our domain in the following way: ( 1) The data are temporally based . The analyses produced by ACE are dependent on the frequencies of failures over time. Thus the analyses done by ACE are data-driven. (2 ) The data are reliable for the task. The CRAS data base provides most of the necessary information about the cable plant. (3) The knowledge of the domain is reliable. (4) The search of the CRAS database is exhaustive, but the reliability of the knowledge permits sign~ficant pr~ning of the space of poss1ble conclus1ons . (5) The main focus of the problem-solver is the quality of the inf~rences it makes. The ACE design was dr1ven by the necessity of producing the same analyses as its human counterparts .

Because of these characteristics the Match [9] prbblem-solving strategy was chosen for ACE. As in R1/XCON , ACE requires no backtracking search of a large _ prc;>b!em ~pace, rather its primary task 1s d1v1ded 1nto a fi xed sequence of subproblems. These sets of productions are described in the next section.

• 3 . The structure of the Knowledge Base Although no structure is provided (or imposed) by OPS4, the set of about 100 productions and 50 related condition and action functions in ACE 's knowledge base can loosely into sets of related rules that collectively do the analysis.

A set of productions does short term analysis by daily examining the flow of maintenance reports. If troubles are reported for a cable that has no previous history of chronic troubles then information is retained that indicates the cable may soon require attention. When new failures are reported for a cable with a history of persistent problems, ACE requests further detailed reports from CRAS. This information is used to deduce: (1) whether the repair tasK done on that cable suggests that preventive maintenance will reduce future repairs; (2) if preventive maintenance is required then what type is likely to be effective; and (3) if possible, where the rehabilitation should be done. Thus ACE not only identifies trouble spots, but also suggests how to repair them.

Each of these aspects of short term analysis requires substantial deductive power. For ex amp 1 e, 1 oca t i ng the physical spot where rehabilitation should be done is a difficult problem because the employee reporting the repair tasK is permitted to enter the site of the failure in a free textual format. Subsequently, the CRAS record of the tasK may contain an entry that not only may be inaccurate but also subject to typos and capricious abbreviations. As an example, "WASH 5" might refer to the same location as "WASHINGTON AND FIFTH" or "CALENDAR" might refer to the same location as "CALANDER STREET . " Thus ACE contains productions and associated LISP functions encoding heuristics to estimate whether several addresses refer to the same general location.

The main sources of Knowledge for the short term analysis were: ( 1) textbooK Knowledge from primers on telephone cable analysis; (2) expert advise from the developers of CRAS; (3) expert advice from theoreticians of cable analysis both in Bell Telephone Laboratories and in the local operating companies; and (4) local analysts from the operating companies and users of CRAS performing the actual analyses.

Another portion of the ACE PM contains a set of productions that Know how to access CRAS. Based on requests for additional information generated by other productions, these productions assemble CRAS commands and then access the resulting data stream retrieved from CRAS. The actual transmission of requests and data is handled by a UNIX communications program accessible from OPS4.

Finally, a set of rules assemble the appropriate message about the day's events recognized by the system and call on the UNIX mail facilities to deliver them to the appropriate users. ACE Knows the target of each message based on the importance of the message to the user. This information is contained in WM and is user modifiable.

IV. Testing ACE

ACE has been field tested continually since the spring of 1982 by a long distance connection to remote CRAS systems . The local analysts report that they are very satisfied with ACE's performance. Although ACE has not discovered any problems unknown to human users, it did discover them quicker and missed neither the obvious nor the subtle. The praise of the analysts was outweighed only by their enthusiasm for permanent installation of the software.

The execution performance characteristics of the system are encouraging. Each nightly run for a large metropolitan area consisting of about 400,000 lines averaged only about one hour of VAX 11/780 CPU time (which can confidently be improved). On average, 5000 production invocations are sufficient for a complete run . New analysis tasks are under development to expand the scope of ACE's grasp of cable maintenance.