WO2002069569A1

WO2002069569A1 - Monitoring and diagnosing the operating conditions of a dynamic system

Info

Publication number: WO2002069569A1
Application number: PCT/FR2002/000607
Authority: WO
Inventors: Christophe Dousson; Armen Aghasaryan
Original assignee: France Telecom
Priority date: 2001-02-27
Filing date: 2002-02-15
Publication date: 2002-09-06
Also published as: FR2821508B1; FR2821508A1

Abstract

The invention concerns a process for monitoring a dynamic system, such as a telecommunication network (RT), whereby time-related occurrences (CR) based on events, such as warnings observed in the system are identified by an engine (MRC), at least one time-related occurrence depending on at least two events. It consists in constructing trajectories of a Petri network with transitions consisting of time-related occurrences in each of which an event occurs only once. A diagnostic unit (DI) selects a subset of identified time-related occurrences constituting the transitions of a trajectory, as diagnosis. The invention combines the advantages of time-related occurrences with Petri network trajectories.

Description

Supervision and diagnosis of the functioning of a dynamic system

The present invention relates generally to artificial intelligence, the supervision of a dynamic system in order to produce diagnostics on the operation of the dynamic system, particularly on the history of faults detected in it.

The supervised dynamic system can be a telecommunications network in which monitoring equipment generates alarms.

According to a first known approach, the monitoring of the evolution of a dynamic system is based on a representation of chronicle models using the recognition of relevant events, such as alarms, observed during the evolution of the dynamic system. so as to temporally link their causalities. A chronicle is thus a set of events linked by time constraints.

FIG. 1 illustrates an example of a C chronicle model within the framework of the surveillance of a telecommunications network. The events are, for example, a LOS (Loss Of Signal) signal loss alarm or a LOF (Loss Of Frame) frame loss alarm appearing (Active) or disappearing (Clear), or an LD link loss (LinkDown) or a return to operation of the LU link (LinkUp), etc. The admissible time intervals between two events are indicated in square brackets. This first approach uses a chronicle recognition engine M whose main role is to recognize very efficiently time patterns in a stream of events arriving at the recognition engine compared to chronicles prerecorded in a database of models. Chronic B, as shown in Figure 2.

This first approach is a high-performance technology in the context of the supervision of dynamic systems such as telecommunications networks, for example. The chronicle recognition engine M can process up to approximately 20 events per second.

The search engine produces chronicles recognized as a function of the instants of the events observed, compared to the time constraints of the chronicles recorded, the recognition being incremented as the entry of the events observed. If the recognition of chronicles offers a very effective synthesis of alarms and a rapid detection of malfunctions, in particular using the exploitation of numerical constraints, it does not allow causal linking of events, such as breakdowns and other faults, to each other as they are detected and recognized. In other words, it can always be assumed that any chronicle can follow any other chronicle without it being possible to ascertain whether this sequence is likely in the supervised system. Events can belong to several distinct chronicles without the possibility of knowing whether the belonging of an event to a chronicle is more likely than the fact that the same event belongs to another column.

According to a second known approach, the supervision of the evolution of a dynamic system uses the formalism of Petri nets to model the behavior of the supervised dynamic system. A Petri net is composed of two types of nodes: the places representative of events E which describe the states of the supervised system modeled and the transitions representative of the changes of state of the system. Some of these transitions are associated with observations such as alarms, due to an alarm per transition.

As shown by the causal graph in Figure 3, the crossing of a transition TR corresponding to the signaling of an alarm for example depends on one or more previous events EA, such as breakdowns, placed upstream of the transition and necessary and sufficient to generate the alarm, and influences one or more subsequent events EP, such as breakdowns, placed downstream of the transition. The TR transition is thus a causal link between past events and future events.

A variant of the Viterbi algorithm makes it possible, from the observation of the alarms, to reconstruct the sequence of the different transitions and therefore of the different states through which the system has passed. These sequences are indifferently called Petri network trajectories or stories of the supervised system. A history makes it possible to understand the causes of the alarms observed and therefore the propagation of faults or operating states in the supervised system, and constitutes a diagnosis of the supervised system for a given period.

Viterbi's algorithm builds causal links from a model in the form of a Petri net and therefore describes sequences of possible events. But the Viterbi algorithm does not take into account the numerical time constraints existing in the supervised system. For example, it is impossible with such a diagnosis to differentiate a failure generating two alarms spaced 3 seconds apart from another failure generating the same two alarms spaced 10 seconds apart. Thus the diagnosis based on a Petri net ignores the temporal characteristics of the supervised system.

The objective of the invention is to provide a method for supervising chronicles observed in a dynamic system so as to benefit from the advantages of the two approaches presented above. In other words, the invention aims both to clarify causal relationships between chronicles, and to consider numerical delays in a diagnosis.

To achieve this objective, a method for supervising a dynamic system, comprising recognition of chronicles depending on events observed in the dynamic system among previously recorded chronicles, at least one recorded chronicle depending on at least two events, is characterized in that '' it includes beforehand a construction of trajectories of a Petri net in which transitions are consisting of recorded chronicles and in each of which an event appears only once, and a selection from among said recognized chronicles each comprising at least one event detected in the dynamic system, of a subset of recognized chronicles which constitute transitions of a trajectory in the Petri net.

Compared to a diagnosis according to the second approach, the invention establishes a more precise diagnosis than the simple juxtaposition of the two approaches since two failures can be differentiated by considering the time difference between the different events such as alarms due to these failures.

The Petri net transitions according to the invention are no longer events but chronicles of events.

According to the invention, at least one chronicle in the subset comprises at least two events such as alarms, compared to the transitions of the Petri net labeled by a single event.

According to a characteristic of the invention, the construction of Petri lattice trajectories comprises a recognition of recorded chronicles containing at least one event detected in the dynamic system, and an addition of each of the recognized chronicles to Petri lattice trajectories for each of which all the events of the recognized chronicle are not yet explained in relation to said trajectory.

According to another characteristic of the invention, the selection of a subset of recognized chronicles comprises a calculation of time intervals between the events detected in the dynamic system, a recognition of chronicles recorded in the succession of the detected events in particular by comparison of the calculated intervals with variations in admissible intervals between events in the recorded chronicles, and a selection of a trajectory having transitions made up of recognized chronicles according to their order. The invention also relates to a supervisor for implementing the method of the invention, comprising a chronicle recognition means for recognizing chronicles among chronicles previously recorded as a function of events observed in the dynamic system, and a means of diagnosis having defined trajectories of the Petri network as a function of the chronicles recorded to produce a diagnosis of the dynamic system represented by a trajectory selected as a function of the chronicles recognized by the chronicle recognition means.

Other characteristics and advantages of the present invention will appear more clearly on reading the following description of several preferred embodiments of the invention with reference to the corresponding appended drawings in which:

- Figure 1 is a graph of a chronicle model with its time constraints between events, according to the prior art;

- Figure 2 is a schematic block diagram showing a chronicle recognition engine according to the prior art; FIG. 3 is a causal graph with a Petri network for establishing the diagnosis according to the prior art;

- Figure 4 is a schematic block diagram of a supervisor according to the invention for supervising a telecommunications network;

FIG. 5 is an algorithm of a method for constructing Petri lattice trajectories according to the invention; - Figure 6 is a graph of a Petri network with chronicles according to a first example of the invention; and

- Figure 7 is a graph of a Petri network with chronicles according to a second example of the invention.

According to a preferred embodiment of the invention shown in FIG. 4, the supervision method is based on a management software platform implemented in a network supervisor SU in the form of a workstation in a monitoring system and maintenance for an RT telecommunications network to be supervised.

The telecommunications network RT can be any type of telecommunications network, or set of networks, such as telephone or radiotelephone network, broadband network of ATM type, digital network with ISDN service integration, packet transmission network. , internet network, intranet network, local network, etc. EQ monitoring equipment in the telecommunications network RT in particular detects malfunctions or malfunctions and / or transmission errors for example in network nodes and terminations, such as in switches, multiplexers, cross-connects, routers, recorders, servers, base stations, etc. The operating faults and failures and / or transmission errors are events reported in the form of alarms A by the equipment EQ to the network supervisor SU.

The SU network supervisor contains supervision software essentially consisting of an MRC chronicle recognition engine combined with a DI diagnostician. A Viterbi algorithm is implemented in the DI diagnoser and defines Petri lattice trajectories as a function of chronicles C previously recorded in a database of chronicle models BMC in the search engine MRC. A diagnosis of the supervised network RT produced by the diagnostician DI is displayed as a function of chronicles recognized CR in successions of alarms detected in the supervised network. Each diagnosis displayed is for example in the form of a graph of the Petri grating trajectory in the diagnostic system.

The network supervisor SU builds all the possible trajectories of the telecommunications network RT according to the following main construction steps E0 to E7 shown in FIG. 5.

At an initial stage E0, the diagnostic expertise is described in the form of a combination of a set of chronicles already recorded in the MRC engine and a Petri net in the DI diagnoser. The set of chronicles is defined on the basis of events already observed in the dynamic network RT to be supervised as in the first known approach; the chronicles represent groupings alarms associated with time constraints. The Petri net has transitions labeled no longer by alarms but by chronicles previously defined; the Petri net therefore describes the possible sequences of chronicles.

Taking into account the alarms received, the supervisor builds Petri network trajectories determined by the possible stories of the supervised network in the following way. According to the invention, a story is a chain of recognized chronicles with which a set of "unexplained" alarms ANE is also associated.

When a detected alarm A is transmitted by an equipment EQ to the network supervisor SU in step El, the next step E2 adds the detected alarm to the set of "unexplained" alarms ANE of each HI of the stories in progress in the DI diagnoser. The detected alarm is also applied to the chronicle recognition engine MR which searches for recorded chronicles containing the alarm detected in the base of chronicle models BMC in step E3.

Each recognized chronicle CR is applied by the MRC engine to the DI diagnoser in step E. For each history HI, the diagnostician DI checks that all the alarms of the recognized recognized chronicle CR are contained in the set of "unexplained" alarms ANE of the history HI, in step E5. The diagnostician DI adds the recognized chronicle CR to each of the stories that can accept it, and suppresses the alarms of this recognized chronicle in the set of "unexplained" alarms ANE of the history HI, in step E6.

If steps E3 and E5 are not satisfied, the process returns to step E1. The stories, that is to say the Petri network trajectories for the network supervising RT, are thus completed in response to each alarm A, according to the iterative algorithm El to E7. The mechanism for extending the stories and managing them is that of the Viterbi algorithm already used in known Petri nets without chronicles and by modifying the definition of the state of a trajectory as indicated above. .

Two examples of use of the Petri net constructed according to the invention illustrate below the gain brought to the precision of the diagnosis by the method of supervision of the invention. The second is considered as a unit of time.

In each example, a diagnosis is established according to the main steps ET1 to ETβ shown in FIG. 6.

Following successive alarms received by the supervisor in step ET1, the chronicle recognition engine MRC calculates time intervals between the alarms received, in step ET2. Then, the MRC engine compares each calculated interval with variations in admissible intervals between alarms in chronicles recorded in step ET3.

The chronicles recognized CR as having variations in intervals containing intervals between received alarms and as containing at least one received alarm are applied by the motor MRC to the diagnostician DI, in step ET4.

The DI diagnostician then selects from the Petri lattice trajectories, a trajectory

T whose transitions are made up of recognized chronicles and respect the order of chronicles recognized in step ET5, so as to display for example the graph of this trajectory in step ET6.

Example 1, Figure 7:

In the network monitored RT can occur the following four faults PI to P4 detected and reported by EQ equipment:

- during the PI failure, the network sends an alarm Al constituting a chronicle Cl;

- during the P2 failure, the network sends an A2 alarm followed by an A3 alarm within 10 seconds, which constitutes a C2 chronicle;

- during the P3 failure, the network sends alarms Al and A2 between 5 and 20 seconds later, which constitutes a C3 chronicle;

- during the P4 failure, the network sends the alarm A3 which constitutes a C4 chronicle.

The chronicle recognition engine MRC previously registers the chronicles C1 to C4 constituted according to steps El to E4, as follows: chronicle Cl {event (Al, tl); }, chronicle C2 {event (A2, t2); event (A3, t3);

(t3-t2) in [0, 10]; }, chronicle C3 {event (Al, tl); event (A2, t2);

(t2-tl) in [5, 20]; }, chronicle C4 {event (A3, t3); }, in each of which "event" designates an event A1 to A3 produced at a respective time tl to t3 for the chronicle.

In addition, the DI diagnostician knows that the failure

P2 is always preceded by the PI failure, and the failure

P4 is always preceded by the P3 fault. The DI diagnostician constructs two trajectories T1 and T2 expressing the causal links, that is to say the successions of breakdowns P1-P2 and P3-P4 in a Petri net according to the invention, as shown in FIG. 7, and introduced into the Petri network the chronicles C1 to C4 as transitions in order to describe the behavior of the network monitored RT, according to steps E5 and E6.

If, during supervision, the network supervisor SU receives an alarm Al at t = 20 s, an alarm A2 at t = 50 s and an alarm A3 at t = 55 s (step ET1), the chronicle recognition engine MRC calculates the time intervals between them. successive alarms received such as 50-20 = 30 s and 55-50 = 5 s (step ET2) and compare these intervals to the variations in time intervals stipulated in the chronicles already recorded (step ET3). The engine then recognizes the chronicles C1, C2 and C4, and applies them to the diagnostician DI (step ET4).

The diagnostician DI searches for the trajectories which contain a succession of alarms Al, A2 and A3 and at most the chronicles Cl, C2 and C4 (step ET5). The DI diagnostician displays the trajectory T1 corresponding to the sequence of PI and P2 faults (step ET6) since the alarms received Al, A2 and A3 respect the sequence of PI and P2 faults and the time constraints in the chronicle C2 of the trajectory Tl: 55-50 = 5 se [0, 10] s, and since the time constraints in the chronicle C3 in the other sequence of failures P3 and P4 corresponding to the trajectory T2 are not respected by the alarms received Al and A2 : 50-20 = 30 sg [5, 20] s.

Likewise, if the network supervisor SU receives an alarm Al at t = 10 s, an alarm A2 at t = 20 s and an alarm A3 at t = 35 s, the chronicle recognition engine MRC calculates the time intervals between successive alarms received such as 20-10 = 10 s and 35-20 = 15 s and compares these intervals to variations in time intervals stipulated in the chronicles already recorded. The engine then recognizes chronicles C1, C3 and C4, and applies them to the DI diagnoser.

The diagnostician DI searches for the trajectories which contain a succession of alarms Al, A2 and A3 and at most the chronicles Cl, C3 and C4. The DI diagnostician displays the path T2 corresponding to the sequence of faults P3 and P4 since the alarms received Al, A2 and A3 respect the sequence of faults P3 and P4 and the time constraints in the chronicle C3 of the path T2: 20- 10 = 10 se [5, 20] s, and since the time constraints in the chronicle C2 in the other sequence of faults PI and P2 corresponding to the trajectory Tl are not respected by the alarms received Al and A2: 35-20 = 15 s £ [0, 10] s.

On the other hand, if a Petri network is used according to the prior art without the chronicles and therefore without the time constraints for the two previous examples, a diagnostic expert containing this Petri network successively receives the alarms Al, A2 and A3 and produces a composite diagnosis of the alternative between successive failures PI and P2 and successive failures P3 and P4 without being able to choose one of these two trajectories.

Thanks to the addition of chronicles in the Petri network, the diagnosis is refined according to the invention.

Example 2, Figure 8: In the RT monitored network, the following four faults P5 to P8 can occur, detected and reported by EQ equipment:

- during the P5 failure, the network sends an alarm Al followed by an alarm A2 within 10 seconds which constitutes a chronicle C5;

- during the P6 failure, the network sends an A2 alarm followed by an A3 alarm within 10 seconds which constitutes a C6 chronicle; - during the P7 failure, the network sends the A2 and A3 alarms within 10 seconds, constituting the C6 chronicle;

- during the P8 failure, the network sends the Al and A2 alarms within 10 seconds, constituting the C5 chronicle.

The chronicle recognition engine MRC records the chronicles C5 and C6 constituted according to steps E1 to E4 beforehand, as follows: chronicle C5 {event (Al, tl); event (A2, t2); (t2-tl) in [0, 10]; }, chronicle Cβ {event (A2, t2); event (A2, t3);

(t3-t2) in [0, 10]; }.

In addition, the DI diagnostician knows that the failure

Pβ is always preceded by the P5 failure, and the P8 failure is always preceded by the P7 failure. The DI diagnostician builds two trajectories T3 and

T4 expressing the causal links, that is to say the successions of breakdowns P5-P6 and P7-P8 in a Petri network according to the invention, as shown in FIG. 8, and introduced into the Petri network C5 and Cβ chronicles as transitions in order to describe the behavior of the network monitored RT, according to steps E5 to E6.

If, during supervision, the network supervisor SU receives an alarm Al at t = 20 s, an alarm A2 at t = 25 s, another alarm A2 at t = 42 s and an alarm A3 at t = 45 s (step ET1), the MRC chronicle recognition engine calculates the time intervals between the successive alarms received such as 25- 20 = 5 s and 45-42 = 3 s (step ET2) and compares these intervals to the variations in time intervals stipulated in the chronicles already recorded (step ET3). The engine then recognizes the chronicles C5 and Cβ since 5 sg [0, 10] s and 3 s £ [0, 10] s. The engine applies the recognized chronicles C5 and Cβ according to this order to the DI diagnostician (step ET4).

The diagnostician DI searches for the trajectories which contain a succession of alarms Al, A2 and A3 and at most the recognized chronicles C5 and Cβ (step ET5). The DI diagnostician displays the path T3 corresponding to the sequence of faults P5 and Pβ (step ET6) since the recognized chronicles C5 and Cβ respect the order of the sequence of faults P5 and Pβ in the path T3 and do not respect the order of chronicles Cβ and C5 in the trajectory T4.

On the other hand, if a chronicle recognition engine is used according to the prior art without causal relationships between the chronicles for the preceding example, this chronicle recognition engine successively receives the alarms Al, A2 and A3 and recognizes only the C5 and Cβ chronicles in response to alarms A1-A2 and A2-A3. As the chronicle C5 corresponds to the breakdowns P5 and P8 and that the chronicle Cβ corresponds to the breakdowns Pβ and P7, the final diagnosis would be the breakdown P5 or P8 then the breakdown Pβ or P7 in response to the succession of the chronicles C5 and Cβ without being able choose one of the four possible paths P5-P6, P5-P7, P8-P6 and P8-P7.

Thanks to the introduction of the chronicles as transitions in the Petri net, the diagnosis is refined according to the invention.

Claims

1 - Method for supervising a dynamic system (RT), comprising recognition of chronicles (CR) depending on events (A) observed in the dynamic system among previously recorded chronicles, at least one recorded chronicle (C2; Cβ) dependent on less than two events (A2, A3), characterized in that it previously comprises a construction (E0 - E7) of trajectories (Tl - T4) of a Petri net in which transitions are constituted by recorded chronicles (Cl - Cβ) and in each of which an event appears only once, and a selection (ET1 - ET5) among said recognized chronicles each comprising at least one event (Al - A3) detected in the dynamic system (RT), d '' a subset of recognized chronicles which constitute transitions of a trajectory in the Petri net.

2 - Method according to claim 1, characterized in that the construction of Petri lattice trajectories comprises a recognition (El - E3) of recorded chronicles containing at least one event (A) detected in the dynamic system (RT), and an addition (E4 - E6) of each of the recognized chronicles to Petri network trajectories (HI) for each of which all the events of the recognized chronicle are not yet explained with respect to said trajectory.

3 - Method according to claim 1 to 2, characterized in that the selection of a subset of recognized chronicles comprises a calculation (ET2) of time intervals between the events detected in the dynamic system (RT), a recognition (ET3) of chronicles recorded in the succession of detected events in particular by comparison of the intervals calculated with variations in admissible intervals between events in the chronicles recorded , and a selection (ET5) of a trajectory having transitions constituted by the recognized chronicles according to the order of these.

4 - Network supervisor (SU) for the implementation of the method according to any one of claims 1 to 3, characterized in that it comprises a chronicle recognition means (MRC) for recognizing chronicles (CR) among chronicles previously recorded as a function of events (A) observed in the dynamic system (RT), and a diagnostic means (DI) having defined Petri network trajectories as a function of the recorded chronicles (CR) to produce a diagnosis of the dynamic system (RT) represented by a trajectory selected as a function of the chronicles recognized by the chronicle recognition means (MRC).

5 - Supervisor according to claim 4, in which the dynamic system is a telecommunications network (RT), and the events are alarms (Al - A3) signaling faults (PI - P8) detected in the telecommunications network.