Applicazioni di Plexpert

Articolo presentato al congresso KES 2002
"Knowledge-based Intelligent Information Engineering Systems"
Crema (CR) Italy
16 settembre 2002

e-Maintenance and Business Intelligence

for High Voltage Power Network Components

Diego BISCI

TERNA s.p.a., Via Regina Margherita 137, Roma, Italy

www.enel.it - bisci.diego@enel.it

Massimo GALLANTI

CESI s.p.a., Via Rubattino 54, Milano, Italy

www.cesi.it - gallanti@cesi.it

Alessandro MAZZETTI

Planasia, Via Brin 124 - 16039, Sestri Levante (GE), Italy

www.planasia.it - planasia@planasia.it

Abstract. A new approach for high voltage power network equipment management is described. The underlying technologies are based on Expert Systems and Data Mining. An implementation is presented, featuring an innovative architecture, based on a Web approach. A cost-effective development has been carried out, benefiting of solid Knowledge Representation techniques and effective development tools.

1. Introduction

Management of high voltage transmission networks has been based till now on specialised traditional tools, typically using "time-based" cyclic maintenance scheduling, supported by general purpose information systems, like ERP tools.

The increasing needs of cost-effective operation and maintenance processes, as a consequence of the liberalisation of electricity market, leads to new requirements, towards efficient e-maintenance platform, able to carry out the so called Reliability Centred Methodology (RCM) and Predictive Maintenance techniques. These motivations lead TERNA, a company of ENEL group responsible for operation and maintenance of most of Italian high voltage electricity transmission network, to develop the MBI system (Maintenance and Business Intelligence), together with CESI and PLANASIA.

MBI implements the Condition Based Maintenance, considering the following aspects:

Selecting those components (breakers, transformers, insulators,…) that, because of their criticality and/or their high preventive maintenance cost, are the best candidates for the application of predictive maintenance

Identifying viable condition monitoring and diagnostic solutions that allow to evaluate the component degradation and to timely detect the need of maintenance action

Developing suitable condition monitoring systems able to collect, elaborate and remotely present component data, performance indexes and diagnostic messages. Many solutions have been analysed: continuous monitoring and periodic equipment assessment.

An additional study has been undertaken, addressed to automatic knowledge acquisition, based on Data Mining methodologies.

An Insulator

A Transformer

2. Diagnostic Approach

Our approach to the equipment maintenance problem has been driven by two technologies:

Rule Based Systems for Maintenance and Diagnosis (deductive inference)

Data Mining for automatic acquisition of rules (inductive inference)

The Rule Based System takes into account the following cause-effect relation:

measure à anomaly à action

A typical example could be:

IF oil_dielectrial_rigidity < 24 THEN anomaly = low_insulation

IF anomaly = low_insulation THEN action = perform_gaschromatographic_test

Two kind of anomalies exist:

Measurable anomalies (i.e. originated from the measurement of an equipment value)

Visual anomalies (i.e. detected by the technician during inspections)

The above rule refers to a measurable anomaly, whilst an example of visual anomaly is the following:

IF anomaly = worn_out_gasket THEN action = substitute_pipe

Several actions may be related directly to measures, without any anomaly raising. This is the typical case of degradation situations, as the following:

IF compressor_working_time > 200 hours THEN action = change_oil

This kind of rule is also typical of preventive maintenance handling:

IF last_overhaul > 1 year_ago THEN action = next_overhaul_within_6_months

Actions are related to time: an action may either be urgent or may be deferred. Whenever an action is suggested, its urgency is defined and in case of conflict the most critical one must be considered. As an example, let us consider the following two rules:

IF compressor_working_time > 200 hours THEN action = change_oil (next week)

IF oil_viscosity < 27 THEN action = change_oil (urgently)

If both of the above rules should fire, the action "change_oil" will be considered as "urgent".

Actions are also organised into a hierarchy: in fact some actions are more complex than others. A tree structure is well suited for the representation of actions hierarchical relationship:

breaker_overhaul (higher level action)

substitute_compressor (part level action)

change_oil (subpart level action)

substitute_throttle (subpart level action)

substitute_pipe (part level action)

perform_gaschromatographic_test

clean_contacts (part level action)

overhaul_of_phase_1 (subpart level action)

…

This organisation implies that, besides the cause-effect relation, other tactful issues in diagnosis are to be considered:

Action Merging

Action Predominance

Action Merging is a post-processing step in which suggested actions are analysed and simple-actions are merged into a single more complex action. An example is:

IF action = overhaul_of_phase_1 AND

action = overhaul_of_phase_2

THEN action = overhaul_of_whole_breaker

AND_ALSO delete overhaul_of_phase_1 AND overhaul_of_phase_2

Action Predominance is a further post-processing in which dependencies between actions are considered. If an action is already covered by another one, the former is deleted, provided it is less urgent. As an example, if the compressor includes a throttle, the following rule applies:

IF action = substitute_throttle (within one month) AND

action = substitute_compressor (within one week)

THEN delete substitute_throttle (predominated by substitute_compressor)

Obviously, if the urgency of "substitute_throttle" was less than one week, it could not be deleted.

The overall diagnostic Expert System interface is represented in Fig. 1. The user can set visual anomalies and verify measurable anomalies; the "diagnosis" button launches the reasoning phase including all the post-processing activities. The outcomes are displayed into an "action preview" panel, along with an explanation panel. The "save" button downloads all the session into a Database. Several diagnostic sessions are allowed and actions are more and more cumulated into the Database.

Fig.1 The diagnostic system allows user to set visual anomalies and visual measures. Running the diagnostic task, the measured anomalies are suggested along with actions and explanation.

Fig.1.1 The real mask for anomalies definition.

Fig.1.2 The real mask for actions presentation.

As soon as a considerable amount of actions is stored in the database, the Business Intelligence phase may be launched. This is a "data mining" phase in which the whole set of components is analysed with respect to the related actions suggested; the aim is to discover hidden rules that are not included in Expert System. Fig. 2 shows the decision tree of a mining session in which a relation between type, pressurisation and model has been discovered.

The rules discovered are not automatically included into the diagnostic knowledge base and sounds like this: equipment type is very significant; equipment based on pumps are less affected by fault than those having compressor; model HA-1 is sensible to sea climate; oil equipment are affected by faults.

Fig.2 The Data Mining phase takes into account all emitted actions and discovers correlation with equipment features. The discovered knowledge is represented as a decision tree.

3. System Architecture

The overall system is designed according to a client-server architecture based on Web technology as shown in fig. 3. The Database is centralised into the server and contains all information about equipment to be diagnosed, along with all anomalies and actions suggested.

Two categories of data exists: static data about equipment (location, installation date, brand,…) and dynamic data got from field (number of operations, short-circuits,…)

One of the most interesting feature of this system is the configurability of the Expert System in two modes:

Interactive mode: for on-demand diagnosis, often based on visual anomalies.

Batch mode: for scheduled diagnosis, mainly based on field measurements and degradation.

These modes are two perspective of the same thing: the Expert System is exactly the same, running once as an Applet (on the client browser), and then as a scheduled task.

Fig.3 The System is designed in a Client-Server architecture based on Web Technology.

Another interesting feature of the system architecture is modularity; the Expert System is structured in four independent modules (see fig. 4, right to left):

Knowledge Base: one for each kind of equipment: breaker, transformer, insulator,… up to 10 different equipment types are considered.

General Diagnostic Module: a single module (common to each equipment) dealing with measures, anomalies, actions, merging, predominance and deferability.

Inference Engine: a general module for rules firing. It is the main reasoning module dealing with deductive inference. It is common to each equipment type.

Interface: two separate interfaces exist: a graphical user interface (Applet) and a batch interface. Both of them are connected to the Database through Servlets.

Fig.4 The diagnostic system is structured in four independent module (right to left): knowledge bases (about 10, one for each equipment type), general diagnostic module, inference engine, interface (split in Applet and Batch)

The Data Mining module is an interactive system running on the user browser. It is an Applet that queries the database and presents results to the user. The main technology is Inductive Inference based on Iterative Dichotomizer Tree algorithm (ID3) [8].

4. The Knowledge Technology and Tool

The most suitable tool for supporting the requested inference techniques is the commercial tool Plexpert (http://www.planasia.it/DemoPlex.html).

Plexpert combines forward and backward rules chaining, with evocative (abductive) capabilities (i.e. partially verified rules are considered at low priority) and heuristic search (i.e. data are evaluated only when necessary). Reasoning may be non-monotone (i.e. achieved goals may be re-evaluated when premises change); non-monotonic reasoning is well suited for action merging and predominance.

Reasoning is based on predicate logic, with existential and universal quantifiers (i.e. IF_EXISTS, IF_ALL and FOR_EACH); this is useful for equipment with several instances of the same mechanism.

Data structure is based on hierarchical classes: this is well suited for representing anomalies and actions hierarchy levels.

Particularly interesting is the technique adopted for explanation: each significant rule contains a specific command for generating a portion of an explanation text. The concatenation of all text constitutes a sort of reasoning-log (called "motivation").

One of the most valuable feature of the Inferential Engine is the high integration between inferential and procedural processing: the knowledge base mixes Rules and Java-methods with a bi-directional interaction (a rule can call a method; a method can call a rule, launching a backward session). The integration with Java procedural programming has been very useful for Client-Server communication through URL Stream (HTTP/GET method). Plexpert supports the independence of interface from inference engine; this has been useful for achieving the twofold modes: applet and batch-task.

The Data Mining module has been developed using the product Plaminer, whose demo-system is available at: http://www.planasia.it/DemoMine.html .

5. The Knowledge Acquisition

Knowledge has been acquired in a structured way: few interviews has been carried out, concentrating the attention on a tabular format of the knowledge. Experts were very facilitated by formalising their knowledge in a table like the following:

Part	Measure/Anomaly	Comparison	Action	Deferability
Compressor	Working_time	> 200	Change_oil	One week
Compressor	Worn_out_gasket	True	Substitute_pipe	Urgently
Command	Last_overhaul_more_ than_one_year_ago	True	Overhaul	Six months

Action Merging has been formalised into a table like the following:

Combined Actions	Suggested Action
Overhaul of 2 or 3 phases	Overhaul_of_whole_breaker

Action Predominance has been formalised using a tree structure for representing component hierarchy:

breaker

    compressor

        throttle

        gasket

    pipe

        oil

    contacts

    …

The knowledge testing has been carried out running the system on selected equipment whose defects are well known.

6. Benefits and Field Experience

MBI system has been deployed as a new advanced platform for helping the operation and maintenance activity in TERNA (ENEL). The first results prove that the technology and components models can effectively support the adoption of the new methodology in the electrical transmission environment.

The full application of MBI system will result in the following benefits:

Reducing the frequency of periodic maintenance actions, while preserving reliability in substation operations.

Reducing the number of critical accident failures.

Optimising the frequency and improving the cost-effectiveness of required overhauls.

Creating a sound knowledge base of maintenance related data, to be used for the future tuning of maintenance programs and developing a new strategies.

Improving awareness in equipment acquisition and configuration

Last but not least one of the keys of success is the adoption of the web approach, which facilitates the user by means of familiar tools (browser).

References

[1] Alessandro Mazzetti, Costruire un Sistema Esperto, Muzzio Publ. 1986.

[2] Alessandro Mazzetti, Applicazioni dei Sistemi Esperti, Muzzio Publ 1987.

[3] Hayes-Roth, Waterman, Lenat, Building Expert Systems. Addison Wesley Ed. 1983.

[4] J.F.Glimore, Expert System Tool Evaluation. 6^th International Workshop on Expert System and their applications, Avignon, 1986.

[5] Alessandro Mazzetti, Interpretazione Interattiva dei Dati in ambito Internet/Intranet. Workshop AI*IA on Organizational Tools and Intelligent Access for Heterogeneous Information - Padova, Sept 1998.

[6] Alessandro Mazzetti, Strumenti di sviluppo per Sistemi Basati sulla Conoscenza. Giornata di studio ANIPLA su processi produttivi - Milano, Oct 1996.

[7] Alessandro Mazzetti, Sistemi di ausilio all'evoluzione e manutenzione della conoscenza. 5th Symposium of Italian Association for AI (AI*IA). Napoli, Italy, Sept 1996.

[8] R.Quinlan, Learning efficient classification procedures and their application to chess end games. Machine Learning: an Artificial Intelligence approach. Morgan Kaufmann. 1^st vol 1983.